Multi-Scale Simulation of Complex Systems: A Perspective of Integrating Knowledge and Data
Extracted source text · arXiv 2306.10275
\documentclass[acmsmall,nonacm,screen]{acmart}
\AtBeginDocument{%
\providecommand\BibTeX{{%
\normalfont B\kern-0.5em{\scshape i\kern-0.25em b}\kern-0.8em\TeX}}}
\setcopyright{none}
%\copyrightyear{2023}
%\acmYear{2023}
%\acmJournal{CSUR}
\settopmatter{printacmref=false, printccs=false, printfolios=false}
\renewcommand\footnotetextcopyrightpermission[1]{}
\setcopyright{none}
\newcommand{\tabincell}[2]{\begin{tabular}{@{}#1@{}}#2\end{tabular}} %表格自动换行
\usepackage{subfigure}
\usepackage{supertabular}
\usepackage{tabu}
%\usepackage[table,xcdraw]{xcolor}
%\usepackage{booktabs,tabularx}
%\usepackage{multirow}
\usepackage{ragged2e}
\usepackage{booktabs,makecell, multirow, tabularx}
\newcommand{\tabitem}{~~\llap{\textbullet}~~}
\usepackage{makecell}
\newcommand{\der}{\operatorname{d\!}{}}
\newenvironment{packed_itemize}{
\begin{list}{\labelitemi}{\leftmargin=1.5em}
\setlength{\itemsep}{1pt}
\setlength{\parskip}{0pt}
\setlength{\parsep}{0pt}
\setlength{\headsep}{0pt}
\setlength{\topskip}{0pt}
\setlength{\topmargin}{0pt}
\setlength{\topsep}{0pt}
\setlength{\partopsep}{0pt}
}{\end{list}}
\newcommand{\para}[1]{{\vspace{2pt} \bf \noindent #1 \hspace{0.5pt}}}
\begin{document}
\title{Multi-Scale Simulation of Complex Systems: A Perspective of Integrating Knowledge and Data}
%\title{A Survey of Multi-Scale Simulation of Complex Systems: From Perspective of Knowledge and Data}
\author{Huandong~Wang}
\email{wanghuanong@tsinghua.edu.cn}
\author{Huan~Yan}
\email{yanhuanthu@gmail.com}
\author{Can~Rong}
\email{rongcan@pku.edu.cn}
\author{Yuan~Yuan}
\email{y-yuan20@mails.tsinghua.edu.cn}
\author{Fenyu~Jiang}
\email{jiangfenyu@tsinghua.edu.cn}
\author{Zhenyu~Han}
\email{hanzy19@mails.tsinghua.edu.cn}
\author{Hongjie~Sui}
\email{huihj21@mails.tsinghua.edu.cn}
\author{Depeng~Jin}
\email{jindp@tsinghua.edu.cn}
\author{Yong~Li}
\email{liyong07@tsinghua.edu.cn}
\affiliation{%
\institution{Beijing National Research Center for Information Science and Technology (BNRist), Department of Electronic Engineering, Tsinghua University}
\country{China}
}
\begin{abstract}
Complex system simulation has been playing an irreplaceable role in understanding, predicting, and controlling diverse complex systems. In the past few decades, the multi-scale simulation technique has drawn increasing attention for its remarkable ability to overcome the challenges of complex system simulation with unknown mechanisms and expensive computational costs. In this survey, we will systematically review the literature on multi-scale simulation of complex systems from the perspective of knowledge and data. Firstly, we will present background knowledge about simulating complex system simulation and the scales in complex systems. Then, we divide the main objectives of multi-scale modeling and simulation into five categories by considering scenarios with clear scale and scenarios with unclear scale, respectively. After summarizing the general methods for multi-scale simulation based on the clues of knowledge and data, we introduce the adopted methods to achieve different objectives. Finally, we introduce the applications of multi-scale simulation in typical matter systems and social systems.
\end{abstract}
%For example, by exploiting the information obtained from simulations at different scales, we can complement missing components at the target scale to overcome the challenge of unknown mechanisms. By adopting a divide-and-conquer strategy for simulation variables, i.e., simulating variables with different importance with mechanisms of different scales, which have different computational costs and simulation accuracy, we can simulate the complex system with both computational efficiency and simulation accuracy, thus overcoming the challenge of expensive computational costs.
%utilizing less accurate but computationally inexpensive macroscale dynamics to approximate most variables while utilizing computationally expensive microscale mechanisms to simulate only a very small number of important variables,
\keywords{Complex system, simulation, multiple scales, knowledge, data}
\maketitle
\input{section1_introduction}
\input{section2_relatedreviews.tex}
\input{section3_background}
\input{section4_0_objectives}
%\input{section5}
\input{section5_knowledge}
\input{section5_knowledge_data}
\input{section6_clearscale}
\input{section6_unclearscale}
\section{Application of multi-scale simulation in typical complex systems}
\subsection{Matter systems}
\input{section7_molecular}
\input{section7_chemical}
\subsection{Social systems}
\input{section7_transport}
%\input{section7_economics_congnition}
\input{section7_epidemic}
\section{Open Challenges and Future Directions}
A review of the objectives, methods, and applications of multi-scale modeling and simulation for complex systems has been presented in our paper. As we discussed above, unknown mechanisms and expensive computational costs are two key problems in the simulation of complex systems. The former indicates that the mechanism of system evolution is not clear at a specified scale, and the latter refers to the extremely high computational complexity in handling high-dimensional state variables of complex systems. Multi-scale modeling and simulation techniques are regarded as a potential way to solve these problems. To be specific, we can derive the target-scale mechanism by using the knowledge from other scales, which addresses the problem of unclear mechanisms. To deal with the intractable computation, a straightforward solution is to adopt the dimension reduction technique to approximate the massive state variables.
As we surveyed in previous sections, many recent works make efforts to develop different multi-scale methods to address these problems.
However, we still have to face the following challenges:
\begin{packed_itemize}
\item \textbf{How to extract the valuable information at the scale with clear mechanism for deriving target-scale mechanism}.
The state variables at the scale with clear mechanisms may be extremely high-dimensional, which results in unacceptable computation complexity. At the same time, there exist complex correlations between mechanisms at different scales, which makes it difficult to explore. Thus, it is challenging to extract valuable information to derive the target-scale mechanism.
\item \textbf{How to balance the trade-offs between the computation efficiency and simulation accuracy}.
Dimension reduction methods are able to reduce the number of variables that should be handled. They can improve the computation efficiency but may lose some simulation accuracy. A simple way mentioned before is to reduce the dimensions in the unimportant areas. However, it is hard to identify which areas are unimportant in complex systems. Thus, how to balance the trade-offs between computation efficiency and simulation accuracy is still challenging.
\item\textbf{How to jointly utilize the transferable information at different scales in system simulation.} In complex systems, the states at different scales interact with each other. Thus, to simulate the states at a specified scale, we not only need to understand the evolution mechanism of this scale, but also need to explore the information from other scales. Meanwhile, the evolution mechanisms at other scales are also affected by the information from this scale. Thus, it is non-trivial to jointly utilize the transferable information at different scales for simulation.
\end{packed_itemize}
Based on the above challenges, we discuss the future possible research directions.
\begin{packed_itemize}
\item \textbf{Generalized multi-scale simulation framework}. Most existing works focus on specified methods for multi-scale modeling and simulation in their own research areas. These methods are coupled with the knowledge in the specified areas, which makes it hard to generalize to other areas. A potential way is to construct a generalized multi-scale simulation framework. This framework can establish a new paradigm to guide how useful information is transferred across scales.
\item \textbf{Iterative multi-scale modeling and simulation approaches}. The evolution of state variables at a scale would affect the evolution process at other scales. When simulating a specified scale, we not only consider the influence from other scales, but also explore how this scale acts on other scales. Conducting this iterative interaction allows us to better explore the evolution mechanism, which contributes to the accurate simulation. Thus, developing iterative multi-scale modeling and simulation approaches is a promising direction for future work.
\item \textbf{Data and knowledge joint driven multi-scale simulation}. In multi-scale simulation, data-driven methods rely on the observation data at the corresponding scale. Their effectiveness may be greatly affected in the scenario where the observation data is not covered. Knowledge-based methods have a relatively strong generalization ability, but it is hard to model the evolution mechanism of a scale with massive high-dimensional state variables. Thus, we propose to use the advantages of both methods by integrating data and knowledge in multi-scale simulation. Specifically, it integrates the generalization of knowledge-driven methods and the powerful learning ability of data-driven methods, which benefits exploring the complex evolution mechanism at the corresponding scale. Thus, studying data and knowledge joint driven multi-scale simulation is an important tendency for future research.
\end{packed_itemize}
\section{Conclusion and Summary}
Multi-scale modeling and simulation techniques are playing an increasingly important role in understanding, predicting, and controlling diverse complex systems, which is essential for numerous cutting-edge technology applications. In this survey, we systematically review the literature on multi-scale simulation of major matter systems and social systems, and provide a novel in terms of their objectives, methods, and applications. We hope that this survey will help advance progress in the direction of multi-scale simulation of complex systems. We hope that this interdisciplinary survey will help researchers make better use of existing multi-scale simulation techniques, and also encourage the development of better multi-scale simulation methods.
\bibliographystyle{ACM-Reference-Format}
\bibliography{sample-base.bib}
%\input{appendix}
\end{document}
\endinput
\section{Introduction}
Complex systems refer to a category of systems with significant complexity, which is a concept with broad boundaries covering numerous fields of science~\cite{han2017deep,brooks2009charmm,st2021universal,xing2021multi}. For example, the human brain composed of a large scale of neurons, the physical system composed of a large scale of particles, and the transport system composed of a large scale of roads and vehicles are all representative complex systems. Their complexity often arises from the existence of the vast number of elements and components within complex systems, and nonlinearity, uncertainty, feedback loops, and emergent properties caused by the complicated interaction and co-evolution among them~\cite{karsenti2008self,battiston2021physics}. At the same time, in the modern age, numerous cutting-edge technology applications, including drug design, weather forecasting, chemical engineering, and socio-economic governance, are highly relevant to complex systems, and require effective understanding, predicting, and controlling complex systems. However, the complicated properties within complex systems often prevent us from effectively understanding, predicting, and controlling them. For example, the strong uncertainty of the current states of complex systems makes it difficult to predict their future states accurately. In addition, due to the existence of nonlinearities in the system, the aggregate effect of the interventions we impose on the system becomes no longer a simple summation of the effects of individual interventions, and often is accompanied by the effect of feedback loops and emergent properties, making the control of complex systems extremely difficult.
Fortunately, the rapid development of computing technologies has enabled us to obtain unprecedented computing power. The increasing availability of high-performance computing machines with parallel processors has made it possible numerically compute the evolution of the massive variables in complex systems and thus effectively simulate them in this way~\cite{holland1992adaptation}. On the other hand, the rapid development of the artificial intelligence (AI) technology and the data-driven paradigm has made data-driven AI-powered simulation a great success, among which the most representative one is the DeepMind's achievements from protein simulation~\cite{jumper2021highly} to fluid simulation~\cite{sanchez2020learning}. Overall, computer simulation has played an irreplaceable role in understanding, predicting, and controlling complex systems by providing excellent insight and foresight, which includes but is not limited to the following:
\begin{packed_itemize}
\item \textbf{Establish key connections with first-principle simulation:}
In many application scenarios of complex systems, we want to establish the connections between variables, structures, and properties in the systems, which is vague in most cases. Under the circumstances, we can simulate the structure and properties from the variables using first-principles simulation, i.e., ab initio simulation, to establish the connection between them. For example, given the amino acid sequence of a protein, we want to predict its function. In this case, first-principles simulation of its structure of protein folding can help us predict its functional properties to a large extent~\cite{brini2020protein}. Or when we want to understand the phase transition process of interdependent complex networks under random attacks, we can also simulate the network connectivity in the process of cascading failure, based on which we can build connections between whether a cascading collapse occurs and the probability of attack~\cite{buldyrev2010catastrophic,bashan2013extreme}.
\item \textbf{Derive future system states with predictive simulation:}
Simulation technology, on the other hand, is mostly used to help us derive the future state of the system from the current state to provide foresight of the system, whether through kinetic equations~\cite{barzel2013universality, gao2016universal} or other artificial intelligence models such as RNN~\cite{han2021artificial, ghavamian2019accelerating, guo2021electromagnetic} or GCN~\cite{murphy2021deep,lino2021simulating}. For example, by numerical simulation of weather, we can avoid the damage of extreme weather to agriculture. Another example is that by utilizing direct numerical simulation (DNS) to solve Navier–Stokes equations, we can predict fluid dynamics~\cite{kochkov2021machine}, which is helpful in the design of airplanes, vehicles, and engines, etc~\cite{male2019large}.
\item \textbf{Eliminate uncertainty with sampling simulation:}
When dealing with a complex system with strong uncertainty, the high dimensionality and strong mutual dependency of random variables within the complex system makes it hard to analytically understand the system. For example, in a system of atomic spins described by the three-dimensional Ising model~\cite{bohm2022noise}, for the large number of random variables describing the spin directions of atoms in the system, even the simple task of obtaining the probability distribution of the number of atoms with different spin orientations is difficult to be analytically solved. Under this circumstance, sampling simulation, e.g., Monte Carlo Markov simulation~\cite{yang2019high}, can be utilized to generate samples of random variables without requiring analytical solutions to their probability distributions. In this way, we can characterize the system in terms of statistics of these samples, e.g., their mean and variance, thereby eliminating the effects of the strong randomness.
\end{packed_itemize}
Despite the excellent insight and foresight that can be obtained by simulating complex systems, how to effectively simulate complex systems is still an open problem with significant challenges, which can be summarized as follows:
\begin{packed_itemize}
\item \textbf{Unknown mechanisms:} In many cases, the mechanisms of the complex systems which we want to simulate are unclear, incomplete, or difficult to obtain. A direct way to solve this problem is to utilize data-driven methods to directly fit the unknown mechanisms based on observable data. However, this solution has two limitations. First, the data-driven method struggles with generalization. That is, the performance of the fitted dynamics cannot be guaranteed in the scenario not covered by the observable data~\cite{kochkov2021machine}. On the other hand, the observable data are also often incomplete and noisy, which makes fitting the mechanisms very difficult.
\item \textbf{Expensive computational cost:} Even though the dynamics of complex systems are clear and complete, calculating the evolution of extremely high-dimensional state variables in complex systems is usually costly. For example, an accurate direct-numerical simulation of fluid dynamics based on Navier–Stokes equations is impossible at the scale required for solving weather prediction or engine design~\cite{kochkov2021machine}. How to simulate complex systems with acceptable computational costs is another challenge we have to face.
\end{packed_itemize}
The technique of multi-scale simulation, which has been investigated for several decades, can help us solve these two challenges. Specifically, the core idea of multi-scale simulation is to utilize the mechanisms corresponding to multiple scales simultaneously to effectively simulate the dynamics of complex systems. The different scales usually have different spatial resolutions, temporal resolutions, or quantities, and thus sometimes have completely different mechanisms~\cite{brooks2009charmm}. As for the problem of unknown mechanisms, it is trivial that we can complement or derive the macroscale mechanisms based on the simulated microscale dynamics~\cite{weinan2011principles}. Moreover, the microscale dynamics can be better modeled with the boundary conditions or background fields~\cite{wu2020recurrent} obtained based on the simulated macroscale dynamics. Sometimes, the macroscale and microscale mechanisms are both incomplete and tightly coupled to each other, and thus we have to simulate their dynamics together in a concurrent manner~\cite{liu2019exploring}. Overall, through the method of multi-scale simulation, we can complete the incomplete mechanisms of a specific scale, learn or calibrate the parameters of the target scale mechanism, thus overcoming the challenges of unknown mechanisms. As for the expensive computational cost, the dynamics simulated based on macroscale mechanism are more coarse-grained and less accurate, but it is significantly less computationally expensive, since the number of variables that should be handled is greatly reduced. A direct way is to use the macroscale dynamics to directly approximate the microscale dynamics~\cite{vlachas2022multiscale}. Furthermore, the unimportant areas can be simulated and approximated based on macroscale mechanisms, while microscale mechanisms are still used for simulating important areas~\cite{levitt1975computer}. Through appropriate message-passing mechanisms between dynamics of different scales, we can simulate the complex system with both computational efficiency and simulation accuracy, thus overcoming the challenges of expensive computational costs.
In this survey, we will systematically review the recent literature on multi-scale simulation of complex systems, which include both matter systems and social systems. Our goal is to conduct an interdisciplinary literature survey to summarize the common paradigms and methodologies of existing work on multi-scale simulation from perspective of knowledge and data for numerous complex systems in different scientific fields, thereby helping the computer simulation community, machine learning community, transportation society, and others. We hope this survey will help multidisciplinary communities to understand the progress made in existing work and draw useful insights and perspectives from work in other disciplines, as well as potential gaps and opportunities. To provide a comprehensive understanding of interdisciplinary work on multi-scale simulation of complex systems, we provide a novel multi-perspective taxonomy of existing in terms of their objectives, general methods, objective-oriented methods, and applications. Furthermore, the taxonomy of methods in our paper is closely organized around the topic of how existing work leverages knowledge and data. To the best of our knowledge, no existing survey has covered existing studies in terms of both matter complex systems and social complex systems. We hope that this paper will help advance progress in the direction of multi-scale simulation of complex systems.
The structure of this paper is as follows. We begin by introducing related reviews of our paper by comparing their main difference. Then, we introduce background about the simulation of complex systems and scales in complex systems. Next, we summarize the main objectives of multi-scale modeling and simulation techniques. Following the introduction of general methods of multi-scale simulation techniques based on the clues of knowledge and data introduced in Section~\ref{sec:genmethod}, we introduce the mainly adopted methods with different objectives in Section~\ref{sec:objmethod}. Finally, we introduce the application of multi-scale simulation techniques in different systems, including major matter systems and social systems.
\section{Related Reviews}\label{sec:relatedreviews}
\begin{table*}[ht]
\begin{center}
\begin{tabular}{|c|p{12cm}|}
\hline
\textbf{Paper} & \textbf{Description} \\\hline
\cite{gomes2018co} & \tabincell{l}{This paper presented the related techniques, tools, and challenges for co-simulation, \\where a coupled system can be simulated by integrating with multiple simulators.} \\ \hline
\cite{noe2017collective} & \tabincell{l}{This paper summarized the recent developments of finding the collective variables \\of a system, including the related principles, and the state-of-the-art algorithms.} \\ \hline
\cite{klus2018data} & \tabincell{l}{This paper studied different data-driven methods to project dynamics of the \\high-dimensional system into the lower-dimensional space, and further explored the \\similarity and differences between these methods. } \\ \hline
\cite{peng2021multiscale} & \tabincell{l}{This paper summarized the recent multiscale methods that integrate with machine \\learning, and discussed the mutual benefits with each other.}\\ \hline
\cite{albi2019vehicular} & \tabincell{l}{This paper surveyed the multiscale methods in the cases of vehicles and crowds, \\which mainly focused on the bottom-up derivation of models.}\\ \hline
\end{tabular}
\end{center}
\caption{Related reviews of complex system modeling and simulation.}
\label{Tab:existingsurvey}
\end{table*}
There are some comprehensive articles that summarize different techniques, tools, and algorithms in the area of simulation~\cite{noe2017collective,klus2018data,gomes2018co,albi2019vehicular,peng2021multiscale}. Gong et al.~\cite{gomes2018co} gave a comprehensive review about co-simulation, where a coupled system can be simulated by the composition of different simulators. The corresponding theory and techniques were surveyed, which are grouped into two types of paradigms: discrete event co-simulation and continuous time co-simulation. Further, this work studied the hybrid co-simulation that establishes the relationship between these two paradigms. Frank et al.~\cite{noe2017collective} focused on the role of collective variables in studying high-dimension dynamical systems. The collective variables can discover the meaningful collective phenomena over a long period of time, as well as identify different structures at a macroscopic scale. In the system simulation, it is important to find the collective variables of the system to control simulation. The authors reviewed the relevant theory and algorithms that search for the collective variables from simulation data. Similarly, Klus et al.~\cite{klus2018data} studied the recent data-driven methods in discovering the collective variables of a system, including data-driven dimension reduction methods and transfer operator approximation methods. Further, they summarized the similarity and differences between these methods, and discussed how extensions designed for a specified method can be applied to other related ones. Peng et al.~\cite{peng2021multiscale} presented the multiscale modeling methods based on physics-based knowledge and data-driven machine learning. They discussed how multi-scaling modeling integrates with machine learning, and explained the mutual benefits with each other. Albi et al.~\cite{albi2019vehicular} surveyed the multiscale methods in the cases of vehicles and crowds, where the models at the macroscopic scale are derived from individual-based modeling at the microscopic scale. Different from them, our paper gives a comprehensive analysis of multiscale simulation from different aspects, but the above articles point out a small part of them. To be specific, we identify three interaction ways in the scenarios of clear-scale modeling based on the flow of information between scales, including information flow from microscale to macroscale, from macroscale to microscale, and the mutual interaction between microscale and macroscale. However, \cite{albi2019vehicular} only discussed the interaction from microscale to macroscale, and \cite{noe2017collective} and~\cite{klus2018data} focused on the interaction from macroscale to microscale. Furthermore, our paper considers the scenarios of multiscale simulation with unclear scales in complex systems, and further two additional objectives of multiscale simulation, including {\em discovering unknown scales} and {\em learning dynamics at the discovered scale}. Therefore, this paper can provide more insightful and valuable information for researchers to understand the multiscale simulation methods in complex systems.
\section{Background}
To pave the way for reviewing the literature on multiscale simulation of complex systems, we first introduce background about the concepts of simulation and scales of complex systems, respectively.
\subsection{Simulation of complex systems}\label{sec:background_scales}
\begin{figure}[t]
\centering
\includegraphics[width=0.5\textwidth]{figures/define.pdf}
\caption{An illustration of simulation of complex systems, where $X(t)$ is the observable system state variables at time $t$, $C(t)$ is the controllable condition variable of the system. $\boldsymbol{X}$ and $\boldsymbol{C}$ are the set of system state variables and controllable condition variables from $t_0$ to $t$, respectively. $S(t)$ is the target variable that we want to obtain through simulation.}
\label{fig:simulate}
\end{figure}
The goal of the simulation of complex systems is to imitate the operation or behavior of the systems. As shown in Figure~\ref{fig:simulate}, by denoting the observable system state variables at time $t$ as $X(t)$ and the controllable condition variables as $C(t)$, a simulation model $\mathcal{M}(\boldsymbol{X},\boldsymbol{C})$ is built to mimics the target complex system. Specifically, we define $\boldsymbol{X}=\{X(\tau)\}^t_{\tau=t_0}$ and $\boldsymbol{C}=\{C(\tau)\}^t_{\tau=t_0}$, which are the set of all historical observable system state variables and controllable condition variables, respectively.
In some systems, the simulation models only take $X(t)$ and $C(t)$ as inputs. It depends on whether the system is Markovian, i.e., whether the future state depends only on the currently observed system state variables. This is true for a large number of physical systems. For example, in a fluid system of gases, given the current velocity field and external forces, the future state, i.e., the future velocity field, can be accurately simulated. However, this is not true for most social systems, such as collective cognitive systems composed, or transportation systems. The essential reason is that the observable state variables are not enough to completely determine the state of the system, and there exist unobservable hidden state variables in the system. For example, in a transportation system, we can observe the position and speed of each vehicle, but this is not enough to fully determine the future positions and speeds of all vehicles. In addition to the mental states of the vehicle drivers, each vehicle also has a different destination, which is hidden in the system and largely affects the future vehicle behavior. Under these circumstances, the historical sequence of observable state variables should be introduced in the simulation model to estimate the hidden state of the systems.
The simulation target is diverse, and depends on which aspects of the simulation model are required to imitate the operation or behavior of the real complex system. For example, in different simulations, different target variables $S(t)$ are required to be simulated. In addition, sometimes the whole simulated process of system state evolution, i.e., $\{S(t+\Delta t)|\Delta t\in(0,\infty)\}$ is required to imitate the actual complex system, while other times we only focus on the final state of the system $S(\infty)$. At the same time, sometimes we need to simulate the exact value of the system state, i.e., $S=h(X)$, and other times we need to estimate its probability distribution $S=Pr(h(X))$.
\subsection{Scales of complex systems}
There exist various categories of scales in complex systems. As shown in Figure~\ref{fig:scales}(a), typical scales in complex systems can be summarized as the following three categories. The first category of scales is the scale of space, of which the definition is diverse in different scientific fields. For example, in physics, from electrons to atoms, further to molecules, and continuum, the corresponding spatial scales range from a few nanometers to a few meters, which is an obvious change in spatial scales. In biology, it starts from amino acids to polypeptide chains, and finally forms protein molecules. The second category of scales is the scale of time. A representative example is that there exist violent reactions and slow reactions in chemical reaction complex systems, which lead to the formation of fast and slow variables in the system. The last category of scales is the scale of quantity. For example, in society, there exist scales from the individual to the community, and finally to the society. Note that scales can generally be divided into two types. In the first type of scales, the corresponding state is defined over a continuous area, such as the velocity field of a fluid. In the second type of scales, the corresponding state is defined based on a limited number of variables, such as vehicles in a transportation system, and nodes in a complex network. We refer to the former type of scales as the field-based scale and the latter as the particle-based scale.
Despite the typical scales with clear definitions and physical meanings in complex systems, there also exist numerous unclear scales in complex systems. As shown in Figure~\ref{fig:scales}(b), the massive microscale variables in the system form mesoscale variables through complicated combinations and transformations, and further form macroscale variables in the system. These scales often do not correspond to a uniform space size, temporal length, or quantity, but correspond to the emergent structures~\cite{hoel2013quantifying,li1994particle} or laws~\cite{west1999fourth} in the complex system, and play an irreplaceable role in the accurate simulation of the future evolution of the system. In most cases, these scales cannot be effectively extracted based on simple space size, time length, or quantity scale. Under the circumstances, techniques such as manifold learning, and representative learning are required to be utilized to find the hidden scales in complex systems.
\iffalse
Different scales have different characteristics.
Normally, relative microscales are described by particle-based model, while relative macroscales are describe by filed-based model.
\fi
\begin{figure}[t]
\centering
\subfigure[Typical clear scales in complex systems]{\label{fig:rules_tran}\includegraphics[width=.4\textwidth]{figures/clear-scales.pdf}}
\hspace{0.2cm}
\subfigure[Numerous unclear scales in complex systems]{\label{fig:rules_visit} \includegraphics[width=.5\textwidth]{figures/unclear-scales.pdf}}
\caption{An illustration of scales in complex systems.}
\label{fig:scales}
\end{figure}
\section{Objectives of multi-scale modeling and simulation}
We first summarize the main objectives of existing techniques of multi-scale simulation. As discussed in Section~\ref{sec:background_scales}, there exist clear scales corresponding to different space sizes, temporal lengths, and quantities in complex systems, and numerous unclear scales as well. The clarity of the system's useable scales also results in differences in the objectives of the corresponding multi-scale simulation approaches. Thus, in the following, we will introduce the objectives of multi-scale simulation in the scenario with clear scale and unclear scales, respectively. Without loss of generality, we denote the microscale system state variables as $X(t)$, and denote the macroscale system state variables as $\xi(t)$ in this section.
\input{section4_1_clearscale}
\input{section4_2_unclearscale}
\subsection{Scenario with clear scale}
Although the clear-scale modeling and simulation of complex systems are commonly adopted in many fields, such as physics~\cite{lino2022remus,lino2022towards,fortunato2022multiscale,lino2021simulating,karniadakis2021physics,xie2018crystal,faraji2017high}, material science~\cite{liang2018deep,le2015computational,liu2019deep,wang2019coarse,fish2021mesoscopic}, chemistry~\cite{han2017deep,brooks2009charmm} and bioprotein~\cite{bhatia2021machine,ingolfsson2022machine,hambli2011multiscale}. The scale separation can vary in different fields. In the field of fluid/solid state physics, it is often necessary to focus on atomic and above scales modeled using classical mechanics and microscopic electronic scales modeled using quantum mechanics~\cite{brooks2009charmm}. In materials science, researchers usually focus on two scales, continuum strains, and atomic movements, due to the challenges posed by the heterogeneity of materials~\cite{fish2021mesoscopic}. In the study of biological proteins, different scale separations are done in different research scenarios, such as phase space and atomic level in the discovery of new protein structures as well as all-atom proteins and coarse-grained lipids in the study of membrane remodeling~\cite{bhatia2021machine,ingolfsson2022machine}. All the above complex system simulations are explicitly divided into multiple scales, and each scale holds a specific spatial-temporal resolution. Consistent with the existing works, in the clear scale scenario of the existing literature, the scales are ordered according to the coarseness of the spatio-temporal granularity.
Specifically, we divide the main objectives of multi-scale simulation into three categories, which correspond to different ways of interaction between different scales. It is worth noting that the methods do not need to be restricted to sequential modeling or concurrent modeling~\cite{weinan2011principles} and what matters is whether the flow of information between scales is from macroscale to microscale or vice versa. The introduction of three objectives is as follows.
\begin{packed_itemize}
\item \textbf{Resolving local fine-grained dynamics (\textit{Objective A}):} In this objective, the local fine-grained microscale dynamics are what the study is interested in. That is, the simulation target depends only on a local subset of the system state variables, i.e., $S(t)=h(X_{l}(t))$ where $X_{l}(t)\subset X(t)$.
The microscale models are leveraged to resolve the local dynamics $X_{l}(t)$ in detail, while the macroscale models are utilized to simulate the coarse-grained patterns anywhere else, i.e., $X(t)/X_{l}(t)$.
\item \textbf{Completing macroscale mechanisms (\textit{Objective B}):} \textit{Objective B} is the exactly opposite-type of \textit{Objective A}. The macroscale dynamics of systems are what we care about. That is, the simulation target $S(t)=h(\xi(t))$. However, the macroscale mechanism is incomplete. For example, some constitutive information is missing, which is then obtained from microscale models and serves as the key components of macroscale models.
\item \textbf{Co-simulating microscale and macroscale dynamics (\textit{Objective C})}. To achieve this objective, there is a mutual interaction between the microscale models and the macroscale models. The dynamics of a specific scale or its derived statistics will be updated through the models of the corresponding scale and then passed to the other. The simulation of the complex system will be a process of iterative computation of the macroscale models and the microscale models and the information interaction between them.
\end{packed_itemize}
\subsection{Scenario with unclear scale}
Although classic fields, such as physics, chemistry, and biology, have proposed effective clear-scale models, the useable scales in complex systems are not always clear in complex systems, especially for high-dimensional complex systems.
Here, a fundamental problem is that for a particular simulation target which scales are more appropriate for conducting multi-scale simulation.
For example, in a large range of chemical systems, where different reactions compete with each other, it is difficult to study why reactions occur and how to predict and control their behaviors at a clear scale~\cite{crim2008chemical,rossler1976chemical}. In ecosystems, species reside and interact with each other, forming complex scales in communities. Learning the dynamics of communities to reveal how ecosystem functions are essential for ecology~\cite{hu2022emergent}.
Complex systems cannot always be explicitly divided into clear and appropriate scales for different simulation tasks. To deal with these systems, it is necessary to automatically discover appropriate unknown scales in complex systems by reducing the original high-dimensional microscale system state to a low-dimensional macroscale system state, and learn the dynamics at the discovered scale simultaneously.
Thus, the objectives of the corresponding multi-scale simulation approaches can be further classified into two categories, of which an illustration is shown in Figure~\ref{fig:objDE} and the corresponding explanations are as follows:
\begin{packed_itemize}
\item \textbf{Discovering unknown scales (\textit{Objective D}):} Discovering unknown but useful scales is the key to effectively modeling complex systems without clear scales.
This objective aims to extract more tractable scales from complex systems. Further, we must ensure the tractability and usefulness of the extracted scales by adding constraints in terms of specific properties, e.g., linearity or preserved relevant information. More specifically, as shown in Figure~\ref{fig:objDE}, this objective requires obtaining an encoder $\mathcal{E}$, which maps the microscale state variable $X$ to the discovered macroscale state variable $\xi$, and a corresponding decoder $\mathcal{D}$ to solve the inverse problem.
\item \textbf{Learning dynamics at the discovered scale (\textit{Objective E}):} This objective is accompanied by {\em Objective D}, which aims to learn the effective dynamics $\mathcal{M}'(\boldsymbol{\xi},\boldsymbol{C})$ at the discovered target scale. More specifically, it is a variant of {\em Objective B} in the scenario with unclear scale. Unlike {\em Objective B}, in this objective, the macroscale mechanism may be completely unknown rather than incomplete. At the same time, we can also add constraints to the associated problem of {\em objective E} to ensure that the macroscale mechanism to be learned has specific properties, e.g., linearity.
\end{packed_itemize}
\begin{figure}[t]
\centering
\includegraphics[width=0.4\textwidth]{figures/objectiveDE.pdf}
\caption{An illustration of {\em objective D} and {\em objective E}.}
\label{fig:objDE}
\end{figure}
By combining {\em Objective D} and {\em Objective E}, the high-dimensional state variables of complex systems are able to be reduced to a low-dimensional macroscale state variable system with learned dynamics. Thus, the system is able to be simulated at the discovered low-dimensional scale.
\section{General methods of multi-scale modeling and simulation}\label{sec:genmethod}
Before introducing the multi-scale simulation methods used in achieving each objective in detail, we first summarize the general methods of multi-scale simulation techniques based on whether they are knowledge-driven, data-driven, or Data and knowledge jointly driven.
\subsection{Pure knowledge-driven method}
Early approaches mainly adopt knowledge-driven methods to model and simulate complex systems. The commonly used methods mainly include kinetic equation models and statistical probability models.
\textbf{Kinetic equation model.} This model mainly concentrates on modeling the evolution of the system variables in the continuous time domain. Specifically, it uses the differential equation between system state variables and time to model and simulate the system, which is able to display the evolution of the complex system over time intuitively. By denoting the system state variable as $\xi$, which is usually high-dimensional, a kinetic equation model describing the evolution of the system can be represented by:
\begin{equation}
\frac{dX}{dt} = f(X),
\end{equation}
which is an ordinary differential equation with respect to time $t$.
For example, the dynamics of a complex network are modeled by~\cite{barzel2013universality, gao2016universal}:
\begin{equation}
\frac{dx_i}{dt} = F(x_i)+\sum_{(i,j)\in E}G(x_i.x_j),
\end{equation}
where $x_i$ is the state variable for the node $i$ in the complex network, $E$ is the set of edges of the complex network, and $F$ and $G$ are two functions describing the evolution mechanism of the complex network.
The fluid dynamics can be modeled by the Navier–Stokes equations:
\begin{equation}\label{equ:NS-1}
\frac{d\boldsymbol{x}}{dt} = - \bigtriangledown\cdot (x\otimes x) + \frac{1}{Re}\bigtriangledown^2x-\frac{1}{\rho}\bigtriangledown \boldsymbol{p} + f,
\end{equation}
\begin{equation}\label{equ:NS-2}
\bigtriangledown\cdot \boldsymbol{x}=0,
\end{equation}
where $\boldsymbol{x}$ is the flow velocity, which is usually a field defined on three-dimensional space. $\boldsymbol{p}$ is the field of pressure, of which the corresponding term in (\ref{equ:NS-1}) is a Lagrange multiplier corresponding to (\ref{equ:NS-2}).
Kinetic equations can also be utilized to describe complex systems with stochastic dynamics. In this case, stochastic variables are introduced into the system, and derive stochastic differential equations (SDE) as follows:
\begin{equation}\label{equ:SDE}
{dX} = f(X){dt} + \sigma(X)dW,
\end{equation}
where $W$ usually represents a Brownian motion, and thus its derivative represents a Gaussian noise. Specifically, SDE is utilized to model chemical reaction systems~\cite{singer2009detecting}, and biological cell system~\cite{zhou2021dissecting}, financial system~\cite{mcleish1997fitting}, etc.
\textbf{Probability statistical model.} In some complex systems, the state of the system is usually not deterministic, but a probabilistic distribution defined on the state space. Under the circumstances, the system states as well as the system dynamics are modeled based on probabilistic models, in which the most widely used model is the Markov chain (MC) model. Specifically, in quantum mechanics, the system state usually cannot be measured accurately, and thus it is often modeled by a probability density distribution. Furthermore, the SDE in the form of (\ref{equ:SDE}) also falls into the category of Markov models. More generally, the continuous-time Markov chain (CTMC) models the system in the continous time domain, and it describes the system dynamics through the transition density:
\begin{equation}
\mathcal{T}(x, y; \tau) := p(X(t+\tau) = y | X(t) = x),
\end{equation}
where we can observe that the transition density is independent with the time $t$, and $X(t+\tau)$ and $X(t)$ can be either defined on the continuous space or the discrete space.
On the other hand, the discrete-time Markov chain (DTMC) models the system in the discretized time domain. In this case, the future state $X_t$ of the system at the time slot $t$ is only dependent on the previous state $X_{t-1}$, and independent with the system state at earlier time slots, which can be represented by:
\begin{equation}
Pr(X_t|\{X_\tau\}_{\tau< t})=Pr(X_t|X_{t-1}) =: T(X_{t-1},X_t)
\end{equation}
Representative DTMC models include the independent cascade model~\cite{saito2008prediction}, GLobal Epidemic and Mobility (GLEAM) model~\cite{balcan2009multiscale}, etc.
Similar to the kinetic equation model, the Markov chain model also focuses on modeling system evolution over time. However, in some specific applications and scenarios, we do not care about the evolution of complex systems over time. Instead, we focus on the final state of the system $X(\infty)$ or its statistics $S(\infty)$. For example, Hoel et al.~\cite{hoel2013quantifying} only consider one-step state transition of complex systems, where the considered states are referred to as the initial states and the resulting states, respectively. Further, the energy-based model (EBM), e.g., Ising model~\cite{nauenberg1974renormalization}, can be utilized to describe the relation between the system state $S$ and condition variables $C$. Specifically, by denoting the energy as $E(S,C)$, the probability of $S$ can be represented by $Pr(S)=e^{E(S,C)}/\sum_{S'\in\mathcal{S}}e^{E(S',C)}$, where $\mathcal{S}$ is the set of all possible system state.
\begin{table*}[ht]
\begin{center}
\resizebox{1\textwidth}{!}{
\begin{tabular}{c|l|l|l|l}
%\toprule
\hline
\textbf{Categories} & \textbf{Method} & \textbf{Papers} & \textbf{Considering time} &\textbf{Deterministic/stochastic} \\
%\midrule
\hline\hline
\multirow{2}{*}{Kinetic equation model}
& ODE & \cite{barzel2013universality, gao2016universal} & Yes & Deterministic \\ \cline{2-5}
& SDE & \cite{singer2009detecting} & Yes & Stochastic \\ \cline{2-5}
\cline{1-5}
\multirow{3}{*}{Statistical probability model}
& CTMM & \cite{noe2013variational} & Yes & Stochastic\\ \cline{2-5}
& DTMM & \cite{saito2008prediction,balcan2009multiscale,hoel2013quantifying} & Yes \& No & Stochastic \\ \cline{2-5}
& EBM & \cite{mehta2014exact,nauenberg1974renormalization} & No & Stochastic \\
\cline{1-5}
% \multirow{1}{*}{Rule-based model}
% & NFsim & \cite{sneddon2011efficient} & No & Deterministic \\ \cline{2-5}
% & & & & \\ \cline{2-5}
% & & & & \\ \cline{2-5}
\cline{1-5}
\end{tabular}
}
\end{center}
%}
\caption{Knowledge-based multi-scale modeling and simulating methods}\label{Tab:architecture}
\end{table*}
\iffalse
\begin{table*}[ht]
%\resizebox{1\textwidth}{!}{
\begin{center}
\begin{tabular}{c|l|l|l|l}
%\toprule
\hline
\textbf{Categories} & \textbf{Method} & \textbf{Papers} & \textbf{-} &\textbf{Relation between different scales} \\
%\midrule
\hline\hline
\multirow{2}{*}{Single-scale modeling approaches}
& Kinetic equation & \cite{singer2009detecting} & c2-1 & c2-1 \\ \cline{2-5}
& Probability statistic model & \cite{saito2008prediction,balcan2009multiscale,hoel2013quantifying} & c2-1 & c2-1 \\ \cline{2-5}
\cline{1-5}
\multirow{3}{*}{Cross-scales interaction mechanisms}
& Aggregation & \cite{zhang2018deep,yang2019multi,yang2021full,xu2022msidp} & Micro2macro & substitute \\ \cline{2-5}
& Inverse process of the dimensionality reduction & \cite{singer2009detecting} & Macro2micro & substitute \\ \cline{2-5}
& Renormalization & \cite{} & Micro2macro & \\
\cline{1-5}
\end{tabular}
\end{center}
%}
\caption{Knowledge-based multi-scale modeling and simulating method}\label{Tab:architecture}
\end{table*}
\subsubsection{Knowledge-based multi-scale interaction mechanism} \label{Sec:knowledge-inter}
As for the methods to realise transformation between different scales, we divide these methods into the following category:
\begin{packed_itemize}
\item \textbf{Aggregation}: Under the scenario of clear scales, the transformational relations between different scales are usually clear too. Thus, the transformation between different scales is easy to be achieved by directly utilizing the clear transformational relations, which in most cases is aggregation. Specifically, in molecular simulations, the micro-scope atom energy is aggregated into the energy of macro-scope energy of atomic configuration~\cite{zhang2018deep}; in contagion simulation, the micro-scale prediction results of the next infected user is also summarized to obtained the macro-scale total number of infected users~\cite{yang2019multi,yang2021full,xu2022msidp}.
especially from the microscales to marcoscales
In a discrete system~\cite{hoel2013quantifying}, the micro-scale binary elements are mapped into macro-scale variables through a specific coarse-graining function, e.g., Logical mapping~\cite{hoel2013quantifying}, and averaging~\cite{comolatti2022causal}.
\item \textbf{Dimensionality reduction}: e.g., PCA~\cite{xx}, multi-dimensional scaling~\cite{xx}
\item \textbf{Inverse process of the dimensionality reduction}: \cite{lino2021simulating} Under the scenario of unclear scales, existing studies adopt mainly classical dimension reduction method to discover macro scale based on micro-scale observations. Under this circumstance, transformation from micro scale to micro scale can be achieved based on the inverse process of the dimensionality reduction method.
\end{packed_itemize}
\para{Finite element/mesh based models:} This model
On the other hand, the interaction mechanisms between different scales can be summarized as follows:
\begin{itemize}
\item \textbf{Substitute}: In most cases, the micro scale and macro scale are substitutes for each other.
\item \textbf{neighboring}: In some cases, the macro scale is regarded as the environment of the micro scale. That is, they neighbor each other~\cite{bhatia2021machine}
\end{itemize}
Approaches for discovering or constructing new scale:
1)directly grid partition~\cite{lino2021simulating}
\fi
\subsection{Pure data-driven method}
\subsubsection{Data-driven modeling and simulation approaches}
Since the recently rapid development of machine learning technologies, and the availability of massive amount of experimental datasets, data-driven methodologies appear as a surrogate to the conventional numerical methods governed by the physical equations in the field of simulation. Due to the computational efficiency and the acceptable simulation accuracy, the data-driven methods are leveraged in biology, chemistry, material science, and fluid dynamics, with the purpose of single-scale or multi-scale modeling. In this section, we prefer to provide an overview of the various machine learning techniques incorporated into the simulation procedure. These approaches mainly include methods such as the Recurrent Neural Network (RNN), the Convolutional Neural Network (CNN), the Graph Neural Network (GNN), ODENet, and Gaussian Process (GP), etc. Here we briefly introduce each method and discuss how they benefit the simulation work.
\begin{table*}[ht]
%\resizebox{1\textwidth}{!}{
\begin{center}
\begin{tabular}{c|l|l|l}
%\toprule
\hline
\textbf{Categories} & \textbf{Scale} & \textbf{Method} & \textbf{Papers}\\
%\midrule
\hline\hline
\multirow{3}{*}{Single-scale simulation}
& \multirow{2}{*}{Micro}
& RNN & \cite{han2021artificial, ghavamian2019accelerating, guo2021electromagnetic}\\ \cline{3-4}
& & CNN & \cite{guo2021electromagnetic}\\ \cline{2-4}
& Macro & ODENet & \cite{hu2022revealing}\\ \cline{1-4}
\multirow{4}{*}{Multi-scale simulation}
& \multirow{2}{*}{Micro2macro}
& RNN & \cite{zhu2021mixseq}\\ \cline{3-4}
& & GP & \cite{rocha2021fly}\\ \cline{2-4}
& \multirow{2}{*}{Mixed}
& GNN & \cite{lino2021simulating}\\ \cline{3-4}
& & ODENet & \cite{yang2020machine}\\ \cline{1-4}
\end{tabular}
\end{center}
%}
\caption{Data-driven modeling and simulation methods}\label{Tab:data-driven}
\end{table*}
\begin{packed_itemize}
\item \textbf{RNN}: RNN is dedicated to capturing the temporal dependencies, by learning the conditional distributions for current and future temporal status based on the past time series. The MixSeq~\cite{zhu2021mixseq} forecasts the macroscopic time series by aggregating a mixture of microscopic temporal sequences predicted by the seq2seq model. The authors assume that several successive microscopic time series follows a mixture of distributions. Hence cluster them into different components before modeling each component with a seq2seq model configured by specific parameters. It is theoretically shown that with the appropriate estimation for each component, the estimation for the macroscopic time series is enhanced. And the experimental results show that MixSeq performs well on both synthetic and real data. The AI-ATS algorithm proposed in~\cite{han2021artificial} utilizes LSTM and GRU to learn the optimal time step lengths and step jumps for the discrete simulation of platelet dynamics in shear blood flow. By adaptively choosing the simulating time step size and jump, AI-ATS reduces the unnecessary redundant calculations for efficient computation and attains comparable accuracy performance to the standard algorithm. In~\cite{ghavamian2019accelerating}, the authors replace the micro model in a finite element simulation with an RNN framework for acceleration in materials. By sampling the results generated by the finite element method, they train the RNN model to understand the historic status, which resembles a nonlinear finite element analysis. And experiments reveal the qualification of RNN as a surrogate.
\item \textbf{CNN}: For the simulations considering spatial characteristics and correlations, the CNN is regarded as a candidate for its capability of describing spatial dependencies. RCNN is proposed in~\cite{guo2021electromagnetic} for Computational EM (CEM) to play as a replacement of the Finite Difference Time Domain (FDTD) method. It takes both RNN and CNN to imitate the complex computations along time frames and space grids for electromagnetic field simulation in a microscopic view.
\item \textbf{GNN}: The GNN is suitable for representing a set of entities, their attributes, and their relationships with each other. The MultiScaleGNN~\cite{lino2021simulating} treats the concerned physical domain as an unstructured set of nodes, and edges are the distance between each pair of nodes. By establishing several graphs corresponding to different scales of spatial resolution, and realizing the message passing forwardly and inversely between these graphs, MultiScaleGNN is able to perform as a multi-scale simulator for continuum mechanics, which is tested on the advection problems and the incompressible fluid dynamics.
\item \textbf{ODENet}: The Ordinary Differential Equations Network (ODENet)~\cite{chen2018neural} is regarded as a replacement of ResNet~\cite{he2016deep}, which depends on the neural network to represent the derivative of variables. Furthermore, it is capable of modeling the time series with irregular time intervals, showing its superiority over RNN. The authors in~\cite{yang2020machine} utilize the ODENet to take place of the chemical master equations for releasing the burden on both memory and computation costs in the single proliferative compartment model. Moreover, the ODENet executes the model reduction in a gene network with autoregulatory negative feedback through multi-time scales. In~\cite{hu2022revealing} the ODENet is considered for modeling both the classical Lokta-Volterra equations in ecology and the chaotic Lorenz equations in atmospheric turbulence. It embeds the hidden dynamics buried within the massive time sequential data by integrating both machine learning algorithms and ODE modeling.
\item \textbf{GP}: As a typical probabilistic machine learning technique, GP attempts to formulate the observed data with a stochastic process based on Gaussian distribution. When analyzing the stress distribution of a plastic strain band loaded with a transverse tension, the GP method replaces the nested micro-models realized by the finite element approach~\cite{rocha2021fly}. It builds a reduction framework for accelerating the computation on micro-scale before integrating for macroscopic analysis.
\end{packed_itemize}
In order to better understand the relationships of the mentioned methods with the simulation scales, we further provide a categorization as listed in Table~\ref{Tab:data-driven}. Though only several representative works are introduced, it indicates the feasibility and appropriateness of utilizing the data-driven methods in the multi-scale simulation. To sum up, both RNN and ODENet are suitable for temporal simulation, while ODENet is able to handle the irregular time interval cases. CNN can deal with the simulation task from the spatial perspective. GNN is a wise choice when the concerned simulation problem can be treated as a graph. And GP represents the condition when the simulated solution can be fitted by the stochastic distribution. The data-driven methods accelerate the computational process, however, it depends on obtaining the training data as prerequisites.
\subsubsection{Data-driven multi-scale interaction mechanisms}
Besides the utilization of data-driven methods in the simulation of single-scale and multi-scale physical problems, the ways of how the data-driven methods bridge and integrate different scales are discussed here. We list some representative works in Table~\ref{Tab:data-driven interaction} for a clear view.
\begin{table*}[ht]
%\resizebox{1\textwidth}{!}{
\begin{center}
\begin{tabular}{c|l|l|l}
%\toprule
\hline
\textbf{Category} & \textbf{Scale} & \textbf{Method} & \textbf{Papers}\\
%\midrule
\hline\hline
\multirow{6}{*}{Cross-scale interaction mechanisms}
& \multirow{2}{*}{Micro2macro}
& GP & \cite{peirlinck2019using}\\ \cline{3-4}
& & DNN & \cite{wang2020deep}\\ \cline{2-4}
& Macro2micro & VAE & \cite{bhatia2021machine, ingolfsson2022machine}\\ \cline{2-4}
& \multirow{3}{*}{Mixed}
& BN & \cite{hall2021ginns}\\ \cline{3-4}
& & PCA & \cite{louison2021glimps}\\ \cline{3-4}
& & RL & \cite{alshehri2022machine}\\ \cline{1-4}
\end{tabular}
\end{center}
%}
\caption{Data-driven multi-scale interaction mechanisms}\label{Tab:data-driven interaction}
\end{table*}
Peirlinck et al.~\cite{peirlinck2019using} model the heart failure process of pigs from multi-scale perspectives. It propagates the measured uncertainties from the cellular level to the organ level by training a Gaussian process regression. To be specific, it decomposes the heart volume analysis into longitudinal growth monitoring, and correlates such longitudinal growth with the measured cellular level uncertainties via the Gaussian process regression.
%
Wang et al.~\cite{wang2020deep} first design a reduced-order model for generating coarse-grained simulation data in flow dynamics. Then take advantage of the Deep Neural Network (DNN) to learn the correlation between the coarse data and the observed fine-scale data.
%
GINN~\cite{hall2021ginns} is a hybrid approach that combines deep learning with Probabilistic Graphical Models (PGM). It leverages the Bayesian Network (BN), one of the typical PGMs to describe the conditional relationships for key variables across multi-scale in the supercapacitor dynamics. With the data generated by BN, a DNN is developed to establish the correlation between inputs and outputs.
%
~\cite{alshehri2022machine} introduces the machine learning applications in Computational Molecular Design (CMD) from four aspects: property estimation, catalysis, synthesis planning, and design methods. It creatively proposes the usage of Reinforcement Learning (RL) to link the microscopic molecular design with the macroscopic product characteristics. Concretely, the approach selects the candidate design based on a generative model. On passing the decision to a physical simulation system, the current state and reward are produced, which further guides the evolution direction for the RL algorithm.
%
GLIMPS~\cite{louison2021glimps} implements the resolution transformation in molecular models based on the Principle Component Analysis (PCA). It captures the structural characteristics of both low-resolution and high-resolution data, and employs the General Linear Model (GLM) to link them.
%
In~\cite{bhatia2021machine}, a dynamic-importance sampling framework is proposed for bridging the macro and micro scales. It explores the phase space of a macro model, then leverages the Variational Autoencode (VAE) to project the data onto a latent space, and finally matches the simulations on micro-scale. Furthermore, it creates a self-healing mechanism for adjusting the parameters of the macro model inspired by the micro-scale simulation.
%
~\cite{ingolfsson2022machine} follows a similar technical route as that mentioned in~\cite{bhatia2021machine} to reveal the dynamics of RAS signaling proteins. The scale bridging is realized by mapping the lipid composition into a latent space, to instantiate the molecular simulations on micro-scale. There is also a step for refining the macro model parameters based on the feedback.
In summary, the data-driven methods are capable of linking the micro and macro scales. Approaches like GP and BN establish the bridge crossing the scales based on the stochastic distribution or probability theory. PCA and VAE realize the dimensional reduction for matching the feature space from high dimension to low dimension. DNN links the multi-scale simulation via a neural network, while RL observes the current system state and reward on the level of one scale, and provides the action based on the design from another scale. The data-driven methods, mainly supported by machine learning techniques, manage to correlate features across the scales as a result of the ability to construct nonlinear relationships.
\subsection{Data and knowledge jointly driven method}
In the domain of multi-scale simulation, for the purposes of high efficiency and fidelity, the data-driven and knowledge-driven methods are considered together, taking advantages of both sides.
\begin{table*}[ht]
%\resizebox{1\textwidth}{!}{
\begin{center}
\begin{tabular}{l|l|l}
%\toprule
\hline
\textbf{Micro-scale simulation} & \textbf{Macro-scale simulation} & \textbf{Papers}\\
%\midrule
\hline\hline
Data-driven & Knowledge-driven & \cite{peirlinck2019using, wu2021machine}\\ \hline
Knowledge-driven & Data-driven & \cite{alshehri2022machine, liu2019exploring}\\ \hline
Mixed & Mixed & \cite{leung2022nh}\\ \hline
\end{tabular}
\end{center}
%}
\caption{Data and knowledge jointly driven methods}\label{Tab:joint-driven}
\end{table*}
The characterization and analysis of the heart failure procedure are conducted in~\cite{peirlinck2019using}. It models the growth of heart by a stretch-driven growth method, and quantifies the uncertainties via machine learning approaches. The measurements are carried out on the hearts of six pigs. With the help of Bayesian inference, the authors predict the uncertainty probability during the cardiac growth process.
%
NH-PINN, which is proposed in~\cite{leung2022nh}, converts the multi-scale equation into a homogenized one merging the multi-scale properties. And by utilizing the Physics-Informed Neural Network (PINN) for sequentially solving the cell problems and the homogenized equation, NH-PINN manages to improve the accuracy for solving multi-scale problems.
%
In~\cite{alshehri2022machine}, the authors propose a framework based on RL for molecular design. The current decision is determined by the RL algorithm, while the state and reward are generated through a knowledge-based physical simulation system.
%
To tackle the general 3D problems with arbitrary material and geometric non-linearity, ~\cite{liu2019exploring} proposes a two-phase framework based on the Deep Material Network (DMN). It first derives the experimental data for training and testing by Direct Numerical Simulation (DNS) tools in the offline stage. Later it trains the DMN model to learn the relations between the micro-scale nonlinear properties and the overall macro-scale variables, for the purposes of material design and concurrent simulations.
%
In the field of biology, the authors in~\cite{wu2021machine} develop a framework that combines knowledge and data driven modeling in the multi-scale analysis of the reconstruction of the sheep tibia over a period of 12 months. Specifically, the overall model consists of the Finite Element (FE) analyzer at the macroscopic level, along with one neural network and another series of neural networks performing as the replacements at the microscopic level. The results reveal a comparative performance of the proposed model in accuracy, while a superb performance in efficiency.
The import of the data-driven methods into the physical knowledge dominant simulation is a seamless combination. It not only accelerates the simulation efficiency by learning the relations hidden in the data, but also guarantees the simulation fidelity by obeying the knowledge rules. We believe that the data and knowledge jointly driven simulation method will attract more and more attention, with applications in multiple research areas.
\section{Objective-oriented multi-scale simulation method}\label{sec:objmethod}
\subsection{Methods to achieve Objective A --- Resolving local fine-grained dynamics}
\begin{table*}[ht]
\begin{center}
\scalebox{0.93}{
\begin{tabular}{c|l|l|l}
\hline
\textbf{Categories} & \textbf{Method} & \textbf{Papers} & \textbf{Particle/field based} \\ \hline
\hline
\multirow{2}{*}{Domain decomposition} & knowledge-driven & \cite{brooks2009charmm} & Particle \& Field \\ \cline{2-4}
& data-driven & \cite{mosser2017reconstruction,yang2018microstructural,wang2019meta,liu2016self,wang2019coarse} & Particle \& field \\ \hline
\multirow{2}{*}{Multigrid method} & theory-guided & \cite{gravemeier2010algebraic,dai2019excmg} & Field \\ \cline{2-4}
& technique-introduced & \cite{yushu2020image} & Field \\ \hline
\end{tabular}
}
\end{center}
\caption{Multi-scale modeling and simulation methods to achieve {\em Objective A}}\label{Tab:type-a}
\end{table*}
Simulating microscale dynamics with fine-grained information requires huge computational costs leading to obstacles before utilization. One promising solution is that the fine-grained modeling is only applied to key parts of the system while the coarse-grained modeling is used in the remaining part.
The methods to achieve \textit{Objective A}, which resolves the localized region of interests into microscale for the simulation of fine-grained behaviors, are mainly categorized into two subcategories: \textit{domain decomposition-based methods} and \textit{multigrid-based methods}. The domain decomposition based methods segment the regions of numerical calculations, calculate them separately and then integrate them~\cite{weinan2011principles}. The multigrid-based methods, which divide the problem to be solved with different scales of grid splitting, focus on eliminating errors and unnecessary computations by interactions between scales~\cite{weinan2011principles}. Recently, data-driven modeling methods, such as deep learning, have been widely adopted to integrate with domain decomposition-based and multigrid-based methods to enhance the modeling capabilities and computational efficiency, which demonstrates the advantages of synergistic modeling of knowledge and data. These works will be highlighted in the following introduction.
\textbf{Domain decomposition based methods.} As a classical multiscale modeling schema, domain decomposition has played a great role in chemistry~\cite{brooks2009charmm}, biology~\cite{yang2018microstructural}, and physics~\cite{liu2016self} fields. A widely accepted strategy is to use fine-grained accurate models in the local regions of interest and the coarse-grained representation in other regions~\cite{fish2021mesoscopic}. CHARMM~\cite{brooks2009charmm} is a multiscale molecular simulation platform on which multi-field computational simulation experiments can be performed. The quantum mechanics/molecular mechanics (QM/MM) model, which applies the QM model to the region of interest and the MM model to the remaining region, can achieve precision and efficiency at the same time.
Besides, there are many data-driven methods, including variational autoencoder (VAE)~\cite{kingma2013auto}, generative adversarial networks (GAN)~\cite{goodfellow2020generative}, clustering~\cite{xu2005survey} and deep reinforcement learning (DRL)~\cite{mnih2013playing}, utilized to fit the distribution of microscale configuration based on macroscale quantities~\cite{mosser2017reconstruction,yang2018microstructural,wang2019meta,liu2016self,wang2019coarse} in order to handle the heterogeneity of the complex systems. Since recurrent neural networks (RNNs)~\cite{mikolov2010recurrent} have the capability of modeling temporal evolution, Wu et al.~\cite{wu2020recurrent} also model the state changes of microscopic materials affected by history and the macroscale background field using RNNs.
\textbf{Multigrid-based methods.} The multigrid method is an effective computational solution for PDEs (partial differential equations) thanks to its iterative computation and the multiscale property, which leads to the popular utilization in complex system simulation~\cite{miehe2007multiscale}.
In the study of turbulent flow simulation, scale transfer operators implemented through plain aggregation algebraic multigrid method could improve the computational efficiency and accuracy~\cite{gravemeier2010algebraic}.
Yushu et at.~\cite{yushu2020image} use image processing techniques in the multigrid method to improve the computational speed of its modeling multiscale complex systems.
Dai et al.~\cite{dai2019excmg} propose that the multi-scale multigrid method is a high-accuracy solution for PDEs and design a cascadic multigrid method to further enhance the efficiency.
\subsection{Methods to achieve Objective B --- Completing macroscale mechanisms}
\begin{table*}[ht]
\begin{center}
\scalebox{0.93}{
\begin{tabular}{c|l|l|l}
\hline
\textbf{Categories} & \textbf{Method} & \textbf{Papers} & \textbf{Particle/field based} \\ \hline
\hline
\multirow{2}{*}{Knowledge-driven} & FEA & \cite{sridhar2016homogenization,ozdemir2008computational} & Fields \\ \cline{2-4}
& REV & \cite{yuan2009multiple} & Field \\ \hline
\multirow{2}{*}{Data-driven} & NN surrogated DFT & \cite{allam2018application} & Field \\ \cline{2-4}
& NN surrogated FEA & \cite{le2015computational,xie2018crystal} & Field \\ \hline
\multirow{2}{*}{Knowledge and data jointly driven} & NN enhanced FEA & \cite{hambli2011multiscale} & Field \\ \cline{2-4}
& FEA guided NN & \cite{oishi2017computational} & Field \\ \hline
\end{tabular}
}
\end{center}
\caption{Multi-scale modeling and simulation methods to achieve {\em Objective B}}\label{Tab:typeb}
\end{table*}
As mentioned before, the macroscopic simulation on coarse spatial-tempo grain usually has the drawback of low accuracy due to the lack of microscopic fine information. In particular, it is often difficult to obtain high-precision solutions when encountering systems with strong heterogeneity. The \textit{Objective B} proposed in this literature brings clean fine-grained information to macroscale modeling to promote the simulation, which can be classified into three subcategories: knowledge-driven, data-driven as well as knowledge and data jointly driven methods.
\textbf{Knowledge-driven methods.} The most commonly used knowledge-driven methods include finite element analysis (FEA)~\cite{hughes2012finite} and Representative elementary volume (REV)~\cite{hill1963elastic}.
Sridhar et al. \cite{sridhar2016homogenization} use microdynamics differential equations to model the microscopic local resonance phenomena, which is integrated with the macroscale model for continuum materials.
Özdemir et al. \cite{ozdemir2008computational} utilize FEA to model local microstructural heterogeneities and thermal anisotropy with the continuum macroscale.
Yuan et al. \cite{yuan2009multiple} introduce discretized eigendeformation fields at different scales to improve the homogenization method, in which the elasticity solutions of unit cell are precomputed and aggregated to higher scales.
\textbf{Data-driven methods.} In data-driven methods, machine learning models have replaced traditional methods with high computational costs. Density functional theory (DFT), which models the quantum mechanism numerically with heavy computation, is replaced by neural networks in \cite{allam2018application}.
In \cite{le2015computational}, the neural networks take place the FEA to approximate the surface response and to compute the macroscopic stress and tangent tensor components.
Xie et al. \cite{xie2018crystal} utilize graph neural networks, which take the microscale connection of atoms in the crystal and output the material properties, to accelerate the computation.
\textbf{Knowledge and data jointly driven methods.} Recently, physical knowledge and machine learning models are designed jointly to model the complex systems~\cite{karniadakis2021physics}. Deep material network is typical of combining neural networks and REV to model material heterogeneity. The FEA and neural networks are integrated together to model the multi-scale process of bone remodeling \cite{hambli2011multiscale}. Oishi et al. \cite{oishi2017computational} extract the inherent regularity in a computational mechanics application through machine learning models, and the regularity is then leveraged to design a novel FEA method.
\subsection{Methods to achieve Objective C --- Co-simulating microscale and macroscale dynamics}
\begin{table*}[ht]
\begin{center}
\scalebox{1}{
\begin{tabular}{c|l|l|l}
\hline
\textbf{Categories} & \textbf{Method} & \textbf{Papers} & \textbf{Particle/field Based} \\ \hline
\hline
\multirow{2}{*}{Knowledge-driven} & Boundary-constraint & \cite{chakraborty2018hyperdynamics} & Particle \& field \\ \cline{2-4}
& Overlap interaction & \cite{tao2010nonintrusive,guo2016multi} & Field \\ \hline
\multirow{2}{*}{Data-driven} & Sampling-based & \cite{ingolfsson2022machine,bhatia2021machine} & Particle \& field \\ \cline{2-4}
& Fitting-based & \cite{han2017deep,fortunato2022multiscale,lino2022towards,lino2022remus,lino2021simulating} & Particle \& field \\ \hline
\end{tabular}
}
\end{center}
\caption{Multi-scale modeling and simulation methods to achieve {\em Objective C}}\label{Tab:typec}
\end{table*}
In some scenarios, fine-grained simulation and efficient computational efficiency are both important in studying the whole system, such as protein molecular simulation heterogeneous material mechanism. However, fine-grained simulation will inevitably have a computationally intensive limitation due to the high level of information granularity~\cite{di2019massively}. Therefore, works~\cite{ingolfsson2022machine,bhatia2021machine,guo2016multi,chakraborty2018hyperdynamics,han2017deep,tao2010nonintrusive,di2019massively} develop techniques to exploit multi-scale modeling and specially designed interactions between scales to achieve both fine-grained and efficient. Consistent with the overall review, we grouped this literature into two categories, i.e., knowledge-driven and data-driven.
\textbf{Knowledge-driven methods.} Traditionally, molecular dynamics (MD) is applied as a high-precision simulation method at the atomic level and is used as the basis for fine-grained simulation in many applications. Coarse-grained simulations, on the other hand, are usually focused on different spatio-temporal scales in different tasks. Examples include the lipid scale for RAS protein simulations and the continuum scale for heterogeneous material simulations. In order to achieve the numerical integration effect with high accuracy, Tao et al.~\cite{tao2010nonintrusive} propose the two scale flow method combining fast and slow variables. For modeling the crack propagation of heterogeneous materials, Chakraborty et al.~\cite{chakraborty2018hyperdynamics} design a physics-based continuum crack evolution model called concurrent atomistic-continuum computational model, which consists of FEM and MD. And the MD is improved through a novel strain-boost hyperdynamics accelerated time marching scheme to bridge the time gap between different scales. In the electron flow simulation, the Navier-Stokes (NS) equations have been widely used in classic continuum scale while the MD is applied in the microscale for more accurate modeling~\cite{guo2016multi}. In \cite{guo2016multi}, Guo et al. divide the simulation into three parts: the near-wall part, the bulk part, and the overlap part. The near-wall part and the bulk part use the NS equations and MD separately. The overlap part models the coupling between the remaining two parts with a stochastic Eulerian-Lagrangian method.
\textbf{Data-driven methods.}
There are several works~\cite{ingolfsson2022machine,bhatia2021machine,han2017deep} introducing data-driven methods to enhance the coupling of scales in multi-scale modeling, where the main purpose is to more efficiently interact between scales. Bhatia et al.~\cite{ingolfsson2022machine} and Helgi et al.~\cite{bhatia2021machine} adopt variational auto-encoder to embed a latent space for different regions for phase scale in bioprotein Simulation and utilize a novel sampling method to select important regions for fine-grained simulation via MD. Recently, Deep Potential~\cite{han2017deep} leverages a neural network to model the potential surface of atoms in a molecular dynamics system, which accelerates the acquisition of atomic potential energy and achieves the quantum mechanical level of accuracy. With the development of the utilization of GNN in atomic simulation, Lino et al. integrate the idea of multi-scale in simulating the continuum mechanics through specially designed GNN~\cite{lino2021simulating}, and there are works follow this novelty~\cite{fortunato2022multiscale,lino2022towards,lino2022remus}.
\subsection{Methods to achieve Objective D --- Discover unknown scales}
\input{Tabels/sec6.4}
Methods to achieve {\em Objective D} reduce high-dimensional complex systems to scales that are easier to understand and investigate. We introduce existing methods to solve the two subproblems of {\em Objective D}, respectively.
\textbf{Scale with linear properties.} High-dimensional complex systems are highly intractable. To address the problem, a series of methods are proposed to obtain simple scales of complex systems. Among them, a typical method is dynamic mode decomposition (DMD)~\cite{schmid2010dynamic,kutz2016dynamic,proctor2016dynamic,tu2013dynamic}, which is a data-driven dimensionality reduction algorithm proposed by Peter Schmid~\cite{schmid2010dynamic}. DMD processes time series data by computing a set of dynamic models which are separately associated with a fixed oscillation frequency and decay/growth rate. The linear treatment largely reduces the dimension of complex systems without the loss of accuracy, and DMD models are widely applied in spatiotemporal fluid flow analysis~\cite{schmid2011applications,hemati2014dynamic,habibi2020data}.
\textbf{Scale with preserved relevant information.} Researchers seek to identify causal emergence for multi-scale complex systems. For example, Griebenow~\cite{griebenow2019finding} et al. utilize spectral clustering to find the right scale of a complex network with causal emergence. Researchers also discover that macro scales defined by coarse graining the microscale elements show surprising causal emergence~\cite{hoel2013quantifying}. By defining and evaluating the metric of effective information (EI) both at the micro and macro scales, causal interactions are observed to peak at a macro scale. The coarse scales with higher causalities facilitate investigations into complex systems more effectively. Data-driven models~\cite{bar2019learning,lee2020coarse} are proposed to identify coarse-scale partial differential equations (PDEs) that describe the system evolution at the macroscopic level. A series methods to identify slow variables in dynamical complex systems~\cite{perez2013identification,singer2009detecting,noe2013variational,wehmeyer2018time,bramburger2020sparse} are proposed, such as variational approach~\cite{noe2013variational}, autoencoders~\cite{wehmeyer2018time}, diffusion maps~\cite{singer2009detecting}. Meanwhile, governing equations~\cite{champion2019data,brunton2016discovering} or common principles~\cite{huang2018mesoscience,chen2022automated} of the system are extracted. The discovery of governing equations and common principles is essential in data-rich fields that lack well-established theories. Recently, renormalization group (RG) has shown great potential to deal with high-dimensional systems~\cite{koch2018mutual,li2018neural,di2022deep}. A change in scale is called a scale transformation. The renormalization group is intimately related to scale invariance and conformal invariance, symmetries in which a system appears the same at all scales, i.e., so-called self-similarity.
\subsection{Methods to achieve Objective E --- Learning dynamics at the discovered scale}
\input{Tabels/sec6.5}
Methods to achieve {\em Objective E} solve dynamical complex systems without clear scales by learning effective dynamics. To deal with the challenge of high dimensions, there are two types of methods to learn effective dynamics at a lower dimension, which can be further utilized to model dynamical systems.
\textbf{Knowledge-driven methods.}
Knowledge-driven methods leverage mathematical models to learn dynamical complex systems. The widely used approaches include wavelet-based methods~\cite{mallat1989theory,beylkin1998multiresolution,gilbert1998comparison,mehraeen2006wavelet}, multiscale homogenization~\cite{chen2006generalized,yu2002multiscale,li2008generalized}, and stochastic methods~\cite{fish2011nonintrusive,hu2017adaptive}. For example, the wavelet representation~\cite{mallat1989theory}, which is an orthogonal multiresolution representation computed by a pyramidal algorithm, can be applied to various scenarios of scale compression. To deal with large spaces, Jacob et al.~\cite{fish2011nonintrusive} propose a reduced-order homogenization method combined with the Karhunen–Loeve expansion and stochastic collocation method. Molei~\cite{tao2010nonintrusive} et al. also develop a new class of multi-scale integrators for stiff ODEs and SDEs for Hamiltonian systems with hidden dynamics.
\textbf{Data-driven methods.} Data-driven methods leverage machine learning or deep learning models to learn the effective dynamics of complex systems. Vlachas et al.~\cite{vlachas2021accelerated} accelerate simulations of molecular systems by learning dynamics at the coarse scale modeled by autoencoders and applying a probabilistic mapping between coarse and fine scales via a mixture density network. Some other studies also use autoencoders to model dynamical systems~\cite{vlachas2022multiscale,lee2020model,fukami2020convolutional}. Besides, variational autoencoder (VAE) can be utilized to learn latent-space dynamics of high-dimensional systems~\cite{hernandez2018variational}. This is achieved by reducing complex, nonlinear processes to a single embedding with high fidelity. The performance of long short-term memory networks (LSTMs)~\cite{hochreiter1997long} and neural ordinary differential equations (NODEs)~\cite{chen2018neural} is explored in learning latent-space representations of dynamical equations~\cite{maulik2020time,maulik2021reduced}. In addition, sampling strategies~\cite{champion2019discovery,bhatia2021machine,gong2021advanced,dupuis2012importance,manohar2019optimized} are proposed for modeling multi-scale dynamics. For example, Harsh et al.~\cite{bhatia2021machine} present a dynamic-importance sampling approach for adaptive multi-scale simulations. It uses machine learning to dynamically sample the phase space by a macro model using microscale simulations, which enables automatic feedback from the micro to the macro scale.
%Multiscale simulations of complex systems by learning their effective dynamics
%Model reduction of dynamical systems on nonlinear manifolds using deep convolutional autoencoders
%Variational encoding of complex dynamics
%Time-series learning of latent-space dynamics for reduced-order model closure
% \item \textit{Objective D} is to discover unknown but useful scales so as to more effectively model complex systems without clear scales.
% It aims to extract more tractable scales from complex systems. There are two ways to achieve this objective: 1) discovering simple scales, which simplifies multiscale systems to linear or coarse scales; 2) discovering governing scales, which identifies slow variables or governing equations for multiscale systems.
% %For the first type, the primary goal is to obtain scales that are easier to understand and investigate. As for the second type, the goal is to learn governing factors of the system, such as governing equations or slow but robust variables.
% \item \textit{Objective E} is to solve dynamical systems. Learning effective dynamics for multiscale complex systems is quite challenging. This objective seek to address this problem from two perspectives: 1) Knowledge-driven modeling, which learns dynamics of complex systems in a model-based manner; 2) Data-driven modeling, which simplifies or accelerates the learning of effective dynamics via machine learning approaches.
%Equation-free, coarse-grained multiscale computation: Enabling mocroscopic simulators to perform system-level analysis
% \subsubsection{Discovering governing scales.}
% Identification of slow molecular order parameters for Markov model construction.
% Detecting intrinsic slow variables in stochastic dynamical systems by anisotropic diffusion maps.
% A variational approach to modeling slow processes in stochastic dynamical systems
% Time-lagged autoencoders: Deep learning of slow collective variables for molecular kinetics
% Sparse identification of slow timescale dynamics
% Data-driven discovery of coordinates and governing equations
% Discovering governing equations from data by sparse identifcation of nonlinear dynamical systems.
% Mesoscience exploring the common principle at mesoscales
% Automated discovery of fundamental variables hidden in experimental data
% \item \textit{Objective E} is to solve dynamical systems. Learning effective dynamics for multiscale complex systems is quite challenging. This objective seek to address this problem from two perspectives: 1) Knowledge-driven modeling, which learns dynamics of complex systems in a model-based manner; 2) Data-driven modeling, which simplifies or accelerates the learning of effective dynamics via machine learning approaches.
% Ten lectures on wavelets
% Wavelet-based multiscale statistical process monitoring: A literature review
% Equation-free, coarse-grained multiscale computation: Enabling mocroscopic simulators to perform system-level analysis
% The heterogeneous multiscale methods
% Nonintrusive and structure preserving multiscale integration of stiff ODES, SDES, and Hamiltonian systems with hidden slow dynamics via flow averaging
% Multiresolution dynamic mode decomposition
% Dynamic mode decomposition of numerical and experimental data
% Here, we present
% a paradigm of adaptive, multiscale simulations that couple different scales using a dynamic-importance sampling approach.
% Our method uses machine learning to dynamically and exhaustively sample the phase space explored by a macro model using
% microscale simulations and enables an automatic feedback from the micro to the macro scale, leading to a self-healing multiscale simulation. As a result, our approach delivers macro length and time scales, but with the effective precision of the micro
% scale.
%Machine-learning-based dynamic-importance sampling for adaptive multiscale simulations
%Hierarchical deep learning of multiscale differential equation time-steppers
% Equation-free, coarse-grained multiscale computation: enabling microscopic simulators to perform system-level analysis.
% 1) Merging states (PNAS 13)~\cite{hoel2013quantifying,singer2009detecting,perez2013identification}
% causal emergence
% slow variables
% Predicting hidden structure in dynamical systems (Nature Machine Intelligence 2021)
% Nonintrusive and structure preserving multiscale integration of stiff ODEs, SDEs, and Hamiltonian systems with hidden slow dynamics via flow averaging (Multiscale Modeling \& Simulation 2010)
% 2) Classical dimension reduction methods~\cite{adragni2009sufficient,cook2008principal,cook1999dimension,fung2002dimension,li2003dimension}
% 3) Auto-Encoder~\cite{vlachas2022multiscale,maulik2021reduced,wehmeyer2018time,lee2020model}
% discover unknown scales? or accelerate?
% \subsection{Methods to achieve Objective E (Yuanyuan)}
% 1) fit based on ML method
% Learning data-driven discretizations for partial differential equations (PNAS 2019)
% 2) Classical method: e.g., DMD method based on Koopman operator theory~\cite{dylewsky2019dynamic}, which is firstly proposed by Schmid et al.~\cite{schmid2010dynamic}.
\begin{table}[]
\centering
\setlength{\leftmargini}{0.4cm}
\begin{tabular}{m{4cm} | m{5cm} | m{4cm} }
\hline
\textbf{Category} & \textbf{Method} & \textbf{Examples} \\ \hline\hline
Scale with linear properties & Dynamic mode decomposition~\cite{schmid2010dynamic,kutz2016dynamic,proctor2016dynamic,tu2013dynamic} & ~\cite{schmid2011applications,hemati2014dynamic,habibi2020data} \\ \hline
\multirow{4}{*}{\makecell[c]{Scale with preserved \\ relevant information}}
& Causal emergence & \cite{griebenow2019finding,hoel2013quantifying} \\ \cline{2-3}
& Coarse-scale partial differential equations & \cite{bar2019learning,lee2020coarse} \\ \cline{2-3}
& Slow variables & \cite{perez2013identification,singer2009detecting,noe2013variational,wehmeyer2018time,bramburger2020sparse}\\ \cline{2-3}
& Governing equations & \cite{champion2019data, brunton2016discovering, huang2018mesoscience, chen2022automated} \\ \cline{2-3}
& Renormalization group & \cite{koch2018mutual,li2018neural,di2022deep} \\ \hline
\end{tabular}%
\caption{Multi-scale modeling and simulation methods to achieve {\em Objective D}}
\end{table}
\begin{table}[]
\centering
\setlength{\leftmargini}{0.4cm}
\begin{tabular}{m{4cm} | m{5cm} | m{4cm} }
\hline
\textbf{Category} & \textbf{Method} & \textbf{Examples} \\ \hline\hline
\multirow{3}{*}{Knowledge-driven} & Wavelet-based methods & \cite{mallat1989theory,beylkin1998multiresolution,gilbert1998comparison,mehraeen2006wavelet}\\ \cline{2-3}
& Multiscale homogenization & \cite{chen2006generalized,yu2002multiscale,li2008generalized}\\ \cline{2-3}
& Stochastic methods & \cite{chen2006generalized,yu2002multiscale,li2008generalized}\\ \hline
\multirow{3}{*}{\makecell[c]{Data-driven}}
& Autoencoders & \cite{vlachas2021accelerated,vlachas2022multiscale,lee2020model,fukami2020convolutional,hernandez2018variational} \\ \cline{2-3}
& Neural ODEs & \cite{chen2018neural,maulik2020time,maulik2021reduced}\\ \cline{2-3}
& Sampling strategies & \cite{champion2019discovery,bhatia2021machine,gong2021advanced,dupuis2012importance,manohar2019optimized}\\ \hline
\end{tabular}%
\caption{Multi-scale modeling and simulation methods to achieve {\em Objective E}}
\end{table}
\subsubsection{Atomic and molecular systems}
Traditionally, Car-Parrinello molecular dynamics (CPMD) is the widely adopted multiscale simulation method for calculating the configuration of the atomic and molecular system, where the density functional theory (DFT)~\cite{car1985unified} deals with the electronic level quantum mechanics and the classic mechanics are modeling with Newton's laws.
Even though CPMD will bring very high simulation accuracy, the computation will be very large and the utilization is limited. To achieve both effectiveness and efficiency, CHARMM proposed that the local region of interest could be different from the remaining region in simulating methods, which is called quantum mechanics - molecular mechanics (QM-MM)~\cite{brooks2009charmm}.
Recently, Weinan et al. propose deep potential~\cite{han2017deep}, which introduced deep learning techniques into the complex system simulation, and the effectiveness and efficiency are both obtained at the system level. Following this, GNNs are also introduced to solve the complex problem of turbulence modeling and continuum modeling~\cite{fortunato2022multiscale,lino2022towards,lino2022remus,lino2021simulating}.
\subsubsection{Biomolecular system}
Multiscale modeling provides better solutions in many complicated tasks, such as novel protein structure discovery, remodeling of membrane, and biological stress modeling.
For example, in \cite{bhatia2021machine}, which is aimed at achieving simulation with macroscale space size and temporal length, and microscale precision, the different scales interact with each other via the consistency at the boundary between different scales.
Stansfeld et al. design a multiscale method to simulate the assembly and the interactions of membrane protein/lipid, in which the macroscale is at molecular dynamics while the microscale reaches atomistic resolution~\cite{stansfeld2011coarse}. In the collagen modeling application, Masic et al. reveal how the physics of the microscale at the molecular level contributes to the mechanics at the macro behavior level~\cite{masic2011observations}.
%蛋白质（生物）Machine-learning-based dynamic-importance sampling for adaptive multiscale simulations[J]. Nature Machine Intelligence放进来
%鄂院士的分子势能模拟等相关工作都放进来。我们在workshop中讲的几个典型案例都放进来。
\iffalse
Atomic and molecular system simulation is necessary in many fields to solve large scale complex problems that cannot be solved directly by analytical means, such as turbulence modeling and bioproteins discovering. Because of the differences in molecular size, simulations of biomolecular systems often involve larger scales than traditional molecular dynamics. So that in traditional matter systems, the electronic level is microscopic and the atomic and higher levels are macroscopic, while in biomolecular systems, the atomic level is microscopic and the higher scales are macroscopic.
\subsubsection{Molecular dynamics system (Rongcan)}
\fi
\subsubsection{Chemical systems}
\begin{figure}[t]
\centering
\subfigure[]{\label{fig:rules_tran}\includegraphics[width=.18\textwidth]{figures/Sec7.1.3/chemical_1.pdf}}
\hspace{-0.1cm}
\subfigure[]{\label{fig:rules_visit} \includegraphics[width=.18\textwidth]{figures/Sec7.1.3/chemical_2.pdf}}
\hspace{-0.1cm}
\subfigure[]{\label{fig:rules_visit} \includegraphics[width=.18\textwidth]{figures/Sec7.1.3/chemical_3.pdf}}
\vspace{-0.2cm}
\caption{Three different gas–solid two-phase flow with the same amount of solids and the same gas flow rate $U_g$. The different mesoscale structures lead to different properties of the system~\cite{li1994particle}.}\label{fig:chonlyglobal}
\end{figure}
Chemical systems are also a category of systems with multi-scale structures.
For example, a widely existing complex system in nature and engineering with multi-scale structures is the Gas–solid two-phase flow, which is a non-equilibrium system with significant heterogeneity.
Specifically, there are two mechanisms in this system, including mechanisms of particles and gas.
Then, in a concurrent-up gas–solid two-phase flow as shown in Figure~\ref{fig:chonlyglobal}, the gas tends to find an upward path with minimum resistance. If the mechanism of gas dominates the systems, which occurs when the flow velocity is particularly large, particles will be separated as much as possible to minimize the resistance to gas, thereby minimizing energy consumption for transporting and suspending particles per unit volume of gas $W_{st}$.
On the other hand, the particles tend to minimize their potential energy.
If the mechanism of particles dominates the systems, which occurs when the flow velocity is particularly small, particles will be as concentrated as possible to minimize their potential energy, thereby the voidage $\epsilon$ of the gas–solid flow.
However, in most scenarios in nature and engineering, the flow velocity is neither too large nor too small, and neither mechanism is able to dominate the system.
Thus, two mechanisms co-exist and compete with each other in the system, which causes a compromise in reaching the optimal states corresponding to the two mechanisms, i.e., minimal $W_{st}$ and minimal $\epsilon$, respectively.
During the competition of the two principles of minimum potential energy for particles and minimum resistance to gas, the mesoscale structure is formed, i.e., dense particle clusters surrounded by the dilute broth, which are denoted by dense phase and dilute phase, respectively.
From the micro scale, by only observing the interaction between each particle and its surrounding circular area, we can only observe two different mechanisms corresponding to the dense phase and the dilute phase.
From the macro scale, only the global state variables cannot fully characterize the system. For example, as we can observe from Figure~\ref{fig:chonlyglobal}, the three systems have the same global state variables including the same gas flow rate $U_g$, and the amount of solids.
The different meso-scale particle cluster sizes lead to the difference between Figure~\ref{fig:chonlyglobal}(a) and (b); The existence of a core-annulus flow further leads to the difference between Figure~\ref{fig:chonlyglobal}(b) and (c).
Thus, the mesoscale structure must be observed and modeled to fully characterize the complex system.
How to model the mesoscale structure of the Gas–solid two-phase flow has been studied in numerous existing studies.
\citeauthor{li1994particle} et al.~\cite{li1994particle} propose the energy-minimization multiscale model (EMMS) model.
In this model, the Gas–solid two-phase flow is described by eight variables, of which the most important variables include the fraction of dense phase $f$, the voidage of the dense phase $\epsilon_c$, the voidage of the dilute phase $\epsilon_d$, and the diameter of particle clusters $d_{cl}$, etc.
They find that instead of minimizing the energy consumption for transporting and suspending particles per unit volume of gas $W_{st}$ or the voidage $\epsilon$ in the system, the energy consumption for the suspension and transportation of particles with respect to unit mass of particles $N_{st}$ is minimized.
Further, \citeauthor{ge2002physical} et al.~\cite{ge2002physical} propose a numerical algorithm to solve this problem.
Specifically, this algorithm defines a two-dimensional solution space of $(\epsilon_c,\epsilon_d)$. For each pair of $(\epsilon_c,\epsilon_d)$, this algorithm searches for the value of $f$ such that $U_{si}$ calculated based on its definition equals $U_{si}$ calculated based the balance conditions including mass balance, momentum balance, and pressure balance, etc.
Specifically, $U_{si}$ represents the slip velocity between two phases.
Then, the problem can be solved by traversing all feasible $(\epsilon_c,\epsilon_d)$ within the solution space and finding the $(\epsilon^*_c,\epsilon^*_d)$ with minimum $W_{st}$.
The EMMS model is further extended to a more general multi-objective variational methodology~\cite{mo2020analysis, li2018mesoscience}. Based on this methodology, the target system can be mathematically expressed by the multi-object optimization (MOP) problem.
\citeauthor{mo2020analysis} et al.~\cite{mo2020analysis} validate the consistency of the solutions obtained based on the EMMS model compared with the solutions obtained based on the MOP model through numerical analysis. What's more, the MOP model can further explore those solutions not found by the EMMS model.
\iffalse
Specifically, in the dense phase, the mechanism of the solid, i.e., particles dominates the whole systems, driving the local voidage $\epsilon_c$ to be minimized.
In the dilute phase, the mechanism of the gas dominates the whole systems, driving the energy consumption for transporting and suspending particles $W_{st}$ to be minimized.
In actual Gas–solid two-phase flow, two mechanisms co-exist and compete with each other, resulting in the formation of mesoscales, i.e., the dense-phase particle clusters surrounded by the dilute-phase broth.
, there exists significant multiscale hierarchical structures.
Specifically, there are two phases in this system, including dense phase and dilute phase.
In the dense phase, the mechanism of the solid, i.e., particles dominates the whole systems, driving the voidage in the system to be minimized.
In the dilute phase, the mechanism of the gas dominates the whole systems, driving the energy consumption for transporting and suspending particles $W_{st}$ to be minimized.
In actual Gas–solid two-phase flow, two mechanisms co-exist and compete with each other, resulting in the formation of mesoscales, i.e., the dense-phase particle clusters surrounded by the dilute-phase broth.
\fi
\subsubsection{Transportation system}
\begin{table*}
\resizebox{1\textwidth}{!}{
\begin{tabular}{c|p{3cm}|p{2.5cm}|p{2.5cm}|p{2.5cm}|p{2.5cm}}
\hline
\textbf{Categories} & \textbf{Research Problem} & \textbf{Method} & \textbf{Macroscopic} & \textbf{Microscopic} & \textbf{Papers} \\ \hline
\multirow{4}{*}{Spatial} & How traffic performance can be improved is unclear (Objective C) & Multi-scale model predictive control approach & Network level perimeter control & Local level signal control & \cite{yang2017multi} \\ \cline{2-6}
& Unclear the connectivity relationship within a multi-modal urban transportation network (Objective B) & Multi-modal and multi-scale GIS-T data model & Bus, metro and street transit networks & Bus route segment, metro line, street segments, walking links & \cite{chen2011multi} \\ \cline{2-6}
% & Simulation of traffic flow at road intersections & Incremental node model & Traffic network & Intersections & \cite{flotterod2011operational,adacher2018macroscopic} \\ \cline{2-6}
& Unclear changes of future traffic (Objective B) & Hierarchical graph convolution network & Regions & Road segments & \cite{guo2021hierarchical} \\ \cline{2-6}
& Unclear mechanism of the heterogeneous traffic streams (Objective B) & Mean field games theory & Network-level traffic streams & Individual vehicle behavior & \cite{kachroo2017multiscale} \\ \hline
\multirow{3}{*}{Temporal}
& Unclear mechanism of driver behavior (Objective A) & A combined CNN-RNN based multi-task learning model & Long time-scale mental cognitive process & Short time-scale driver’s physical behaviors & \cite{xing2021multi} \\ \cline{2-6}
& Unclear changes of future traffic (Objective B) & Multi-Step dependency relation model & Short time-scale traffic & Long time-scale traffic & \cite{liu2022msdr} \\ \cline{2-6}
& Unclear mechanism of urban freight operations (Objective A) & Multi-scale agent-based simulation model & Long-term strategic decision & Short-term operational decision & \cite{alho2017multi} \\ \hline
\end{tabular}
}
\caption{Multi-scale modeling and simulating methods in transportation.}\label{Tab:tranportation_multiscale}
\end{table*}
In the field of transportation, the traffic flow theory is to explore the spatial and temporal patterns of traffic, which has developed several related models. These models can be classified into two groups: microscopic models and macroscopic models. The microscopic models aim to characterize the interactions between the vehicles, while the macroscopic ones focus on the aggregated behaviors of a group of vehicles. Their basis is to investigate the relationship between the density, speed, and flow volume. Specifically, the density $\rho$ is the average number of vehicles passing a unit length of the road segment. The speed $v$ is the average speed of all the vehicles in a road segment. The flow volume $q$ is the average number of vehicles in a road segment in a time slot. Their relationship can be formulated as follow:
\begin{equation}
q=\rho \cdot v.
\end{equation}
\begin{packed_itemize}
\item \textbf{Macroscopic traffic models}. Such models focus on the macroscopic characteristics of traffic flow, which regards the traffic flow as the continuous fluid composed of a large amount of vehicles. Based on the theory of fluid mechanics, the researchers derive the partial differential equations for the speed and density to characterize the traffic dynamics. Formally, according to the conservation principle of traffic flow, it satisfies the following continuity equation:
\begin{equation}
\frac{\partial (\rho v)}{\partial x} + \frac{\partial v}{\partial t}=0,
\label{equ:lwr}
\end{equation}
where $\rho(x,t)$ and $v(x,t)$ are the flow density and average speed at location $x$ at time $t$, respectively.
This equation is built upon the traffic fluid mechanics model called Lighthill-Whitham-Richards (LWR) model~\cite{lighthill1955kinematic}.
Particularly, for the speed $v(x,t)$, the LWR model assumes that the traffic speed and density satisfy the relationship of $v(x,t)=v_e(\rho(x,t))$. Thus, the Equation~\ref{equ:lwr} can be rewritten as
\begin{equation}
(v_e+\rho \frac{\partial v_e}{\partial \rho})\frac{\partial (\rho v)}{\partial x} + \frac{\partial v}{\partial t}=0.
\label{equ:lwr2}
\end{equation}
The above equation captures the nonlinear traffic features such as the formation and propagation of shock waves in traffic flow, which facilitates to explore various complex traffic phenomenon, e.g., the stop-and-go trend in traffic congestion.
\item \textbf{Microscopic traffic models}. These models are to learn the moving states of each vehicle in a certain traffic environment, including car following model and cellular automaton model.
The car following models study the dynamic process of each vehicle following the leading vehicle based on the assumption that there is no overtaking phenomenon. Traditionally, their general mathematical expression can be shown as follow,
\begin{equation}
\frac{dv_n(t)}{dt}=f(v_n(t),\Delta v_n(t), \Delta x_n(t)),
\end{equation}
where $v(t)$ denotes the speed of vehicle $n$, and $\Delta v_n(t), \Delta x_n(t)$ are the speed difference and displacement difference between vehicle $n$ and its front vehicle. $f(\cdot)$ represents the car following model.
To implement $f(\cdot)$, many researchers propose various methods by modeling drivers' behavior of acceleration or deceleration~\cite{newell1961nonlinear,bando1995dynamical,helbing1998generalized,treiber2000congested,nakayama2001effect,zhao2005new,ge2006effect,xiaomei2007stability,xie2008stabilization,peng2010dynamical}. For example, Reusche and Pipes~\cite{pipes1953operational} proposed the first car following model
\begin{equation}
\frac{dv_n(t+\tau)}{dt}=\lambda \Delta v_n(t),
\end{equation}
which considers the strength of the driver's response to the stimulus he receives from his driving environment. In this equation, $\tau$ is the delayed response time of the vehicle, and $\lambda$ is the strength coefficient of a driver in response to the stimulus.
The cellular automaton models are discrete in both space and time, so it is suitable for parallel computation to accelerate the simulation of microscopic traffic flow. The simplest cellular automaton model is one-dimensional where a single road segment is divided into discrete regular grids. Each grid represents a cell with the finite states. At any time, the state of a cell is either empty or occupying a vehicle. Further, the state of a cell at the next time is determined by itself and the state of its neighboring cells at the current time. Since a cell cannot have more than one vehicle, we define the following rule: if the cell ahead is empty, the vehicle can move one step forward; otherwise, it has to keep waiting. According to this rule, all vehicles could update their states at the next time. Based on it, several extended models~\cite{helbing1999cellular,chowdhury2000statistical,kerner2002cellular,gao2007cellular,chakroborty2008microscopic,tao2008synchronized} are proposed to capture more realistic phenomena that occur in traffic flows, which helps to better understand the traffic dynamics and implement the effective traffic control.
\end{packed_itemize}
Although we can adopt the microscopic traffic models to infer the dynamics of massive microscopic states in transportation systems, it is not easy to derive the macroscopic structure emerged from these states. One important problem is that we cannot accurately realize the scale at which the macroscopic structure will emerge. This is because the accumulated noise from the microscopic scale will prevent us from identifying the macroscopic structure. To solve it, multi-scale modeling is proposed to fully utilize the advantages of microscopic and macroscopic scales for various traffic applications. Through the reviews of recent works, we classify them into two groups in terms of multi-scale properties, i.e., space and time, as listed in Table~\ref{Tab:tranportation_multiscale}.
\begin{packed_itemize}
\item \textbf{Spatial multi-scale modeling}. The works belonging to this type adopt different scales of space to design the corresponding models. Authors in~\cite{guo2021hierarchical} first used the spectral clustering method to construct the macro graph of regions from the road network, and then proposed the hierarchical graph convolutional network to forecast the traffic by operating both the micro and macro traffic graphs. Yang et al.~\cite{yang2017multi} proposed a model predictive control based approach to model the interaction between the network level control and the local level control to optimize the performance at the local and the network level as a whole. Chen et al.~\cite{chen2011multi} developed a multi-scale and multi-modal Geographical Information Systems for Transportation (GIS-T) data model to implement the integration and multiple representation of the urban transportation networks that include bus route segment, metro line, street segments, walking links. Authors in~\cite{kachroo2017multiscale} focused on analyzing the heterogeneous traffic streams from connected vehicles and normal vehicles. They proposed a mean field game framework to establish the connection between the individual vehicle behavior and overall macroscopic traffic streams. In this framework, the connected vehicles controlled microscopically can influence and control overall traffic streams.
\item \textbf{Temporal multi-scale modeling}. This type of works focuses on exploring the advantages of multi-scale modeling in the time dimension. To be specific, Yang et al.~\cite{xing2021multi} proposed to jointly model the driver behaviors in different time scales for behavior reasoning. In this work, the mental cognitive process is recognized to have longer time-scale patterns, while drivers' physical behavior has fast dynamics. A multi-task framework with the CNN-RNN network is developed to process the task of multi-scale driver behavior recognition. Authors in~\cite{liu2022msdr} designed a multi-step dependency relation network to learn the relations between the short time-scale traffic features and long time-scale ones. Alho et al.~\cite{alho2017multi} focused on modeling urban freight operations. They proposed a novel agent-based framework integrating the models with different time resolutions, including long-term strategic decision model and short-term operational decision model. The former simulates the agent decisions made for one year, and the latter simulates the fast decision making process. The authors also implemented it on an open-source mobility simulation platform and verified the effectiveness of their proposed approach.
\end{packed_itemize}
\subsubsection{Epidemic contagion system}
Accurately modeling the transmission pattern of the epidemic is a long-standing research topic for decades, which is attracting more and more public attentions with the outbreak of COVID-19. The epidemic dynamics is tightly entangled with human mobility \cite{jia2020population,kraemer2020effect,jewell2021s,wu2020nowcasting}, economic development \cite{haw2022optimizing,guan2020global,spelta2020after}, and vaccine distribution \cite{chen2022strategic,rotesi2021national,wagner2021vaccine,yang2022equitable,mcadams2020incentivising}, forming a complex system. In this survey, we focus on the simulation paradigm of epidemic contagion systems of the recent advances, which can be divided into 3 categories as shown in Table~\ref{Tab:epidemic}.
\begin{table*}[ht]
\begin{center}
\resizebox{1\textwidth}{!}{
\begin{tabular}{c|l|l|l}
%\toprule
\hline
\textbf{Categories} & \textbf{Scales} & \textbf{Method} & \textbf{Papers} \\
%\midrule
\hline\hline
\multirow{3}{*}{Single-scale simulation} &
\multirow{2}{*}{Macroscopic model}
& Regression model & \cite{jia2020population,kraemer2020effect,jewell2021s,iacus2020human} \\ \cline{3-4}
& & Compartmental model & \cite{maier2020effective,lai2020effect,ross1916application} \\ \cline{2-4}
& Microscopic model & Contact Network & \cite{murphy2021deep,qian2021scaling,aleta2020modelling,abueg2021modeling,nishi2020network} \\ \cline{1-4}
\multirow{2}{*}{Multi-scale simulation} &
\multirow{2}{*}{Multi-scale model}
& Metapopulation & \cite{balcan2010modeling,balcan2009multiscale,zhang2017spread,chinazzi2020effect,wu2020nowcasting,zhou2020effects,chang2021mobility,zhang2020changes,brockmann2013hidden} \\ \cline{3-4}
& & High-order network & \cite{iacopini2019simplicial,st2021universal} \\ \cline{1-4}
\end{tabular}
}
\end{center}
%}
\caption{Simulation methods of epidemic contagion systems}
\label{Tab:epidemic}
\end{table*}
\textbf{Regression model.} For macroscopic simulation, regression models are widely used. During the first strike of COVID-19 pandemic, researchers use generalized linear model (GLM) \cite{kraemer2020effect}, multiplicative exponential model \cite{jia2020population}, or multilevel Bayesian regression model \cite{jewell2021s} to investigate how human mobility and control measures affect the early transmission of the epidemic. Specifically, the risk source model proposed in \cite{jia2020population} is shown in \eqref{eq:risk_source}:
\begin{equation}
\label{eq:risk_source}
y_{i}=c \prod_{j=1}^{m} \mathrm{e}^{\beta_{j} x_{j i}} \mathrm{e}^{\sum_{k=1}^{n} \lambda_{k} I_{i k}},
\end{equation}
where $y_i$ is the number of the cumulative (or daily) confirmed cases in prefecture $i$; $x_{1i}$ is the cumulative population outflow from the source region of the epidemic to prefecture $i$ in different time periods. $x_{2i}$ and $x_{3i}$ are the GDP and the population size of prefecture $i$, respectively. $\lambda_k$ is the fixed effect for province $k$, $n$ is the number of prefectures considered in the analysis, $I_{ik}$ is the dummy for prefecture $i$, where $I_{ik}=1$ if $i\in k$, otherwise $I_{ik}=0$. Finally, $c$ and $\beta_j$ are the free parameters to be calibrated. From this equation, the author accurately models the macroscopic epidemic dynamics. Similarly, the regression model in \cite{jewell2021s} and \cite{kraemer2020effect} are as shown in \eqref{eq:multilevel_bayesian_regression} and \eqref{eq:glm} accordingly. All these methods use a simple regression model to depict the aggregated epidemic outcome ($y_{i, t}$ in \eqref{eq:multilevel_bayesian_regression} and $Y(t)$ in \eqref{eq:glm}) with other variants such as mobility ($M_{i,t}$ in \eqref{eq:multilevel_bayesian_regression} and $M(t-5)$ in \eqref{eq:glm}), population ($X_{i}$ in \eqref{eq:multilevel_bayesian_regression}), the polymerase chain reaction (PCR) test ($IT(t)$ in \eqref{eq:glm}) and temperature ($T_{i,t}$ in \eqref{eq:multilevel_bayesian_regression}).
\begin{equation}
\label{eq:multilevel_bayesian_regression}
y_{i, t}=\alpha_{c_{t}}+X_{i} \beta+T_{i, t} \theta+C_{s_{i}, t} \phi+M_{i, t} \gamma_{c_{i}, t}+\epsilon_{i, t}
\end{equation}
\begin{equation}
\label{eq:glm}
Y(t) = Y(t-4) + IT(t) + M(t-5) + IM(t)
\end{equation}
{\textbf{Compartmental model.}} Another type of macroscopic model for epidemic modeling is the famous compartmental model, which can be traced to the 20th century \cite{ross1916application}. Classic compartmental models, such as the SEIR model, divide the whole population in a region into different compartments as Susceptible~(S), Exposed~(E), Infected~(I) and Recovered~(R) and use the following ordinary differential equations (ODE) to depict the aggregated transmission process:
\begin{equation}
\begin{aligned}
{\der S}/{\der t} &= - {\beta SI}/{N} \\
{\der E}/{\der t} &= {\beta SI}/{N} - \sigma E \\
{\der I}/{\der t} &= \sigma E - \gamma I\\
{\der R}/{\der t} &= \gamma I,
\end{aligned}
\end{equation}
where $S,E,I,R$ are the susceptible, asymptomatic, infected, and recovered people, and $N=S+E+I+R$ is the total population of the region. For the parameters, $\beta$ is the infection rate, $\gamma$ is the recovery rate, and $\sigma$ is the incubation period of the disease. Specifically, the basic reproductive number of the disease is $R_0 = \beta / \gamma$, which denotes the expected number of cases directly generated by one case in a population where all individuals are susceptible. When $R_0>1$, the disease can start to transmit in an exponential pattern. When $R_0 < 1$, the disease will not become an epidemic since the infected people will recover before it can transmit to more people. These compartmental models are macroscopic since the variables are the total number of people in different states, where the heterogeneous contact patterns are replaced by a homogeneous mixing assumption. It limits their expressive power since they can only capture the exponential growth pattern, which only exists during the early beginning of the epidemic and quickly changed due to the introduction of non-pharmaceutical interventions \cite{flaxman2020estimating,lai2020effect,snoeijer2021measuring}.
To extend the expressive power of classic compartmental models, Maier et al. \cite{maier2020effective} introduce a ``containment and quarantine'' mechanism to the classic SIR model as follows:
\begin{equation}
\begin{aligned}
\partial_{t} S &= -\alpha SI - \kappa_0 S \\
\partial_{t} I &= \alpha SI - \beta I - \kappa_0 I - \kappa I \\
\partial_{t} R &= \beta I + \kappa_0 S \\
\partial_{t} X &= (\kappa + \kappa_0) I,
\end{aligned}
\end{equation}
where $\alpha$ and $\beta$ are the infection rate and the recovery rate accordingly, and the self-containment rate that affects both susceptible and infected people is $\kappa_0$, which assumes the self-containment process will prevent any transmission for these parts of people and directly turn the susceptible and infected people to the new $X$ state (which do not act with any states). Besides, there is also a quarantine rate $\kappa$ that only influences the infected people. By the new mechanism that protects part of the susceptible people, the proposed model successfully extends the expressive power of the SIR model to capture sub-exponential increase patterns.
\textbf{Contact network model.}
For microscopic modeling of the contagion system, most of the research adopts the network approach to represent individual-level contact patterns \cite{murphy2021deep,qian2021scaling,aleta2020modelling,abueg2021modeling,nishi2020network}. Qian et al. \cite{qian2021scaling} use smart card data in urban transit systems to construct individual-level contact networks. Specifically, a classic susceptible-infectious-susceptible (SIS) model is used to model the disease transmission on the network as follows:
\begin{equation}
p_{i, t}=1+p_{i, t-1}\left(q_{i, t}-r\right)-q_{i, t}, \forall i \in V
\end{equation}
In the above equation, $p_{i,t}$ represents the probability for node $i$ infected at time $t$, and the recovery rate is noted as $r$. The variable $q_{i,t}$ represents the probability that node $i$ is in the susceptible state at time $t$. It depends on all the neighborhood $j\in \mathbb{N}(i)$ are either in the susceptible state or in the infected state but fails the transmission:
\begin{equation}
q_{i, t}=\prod_{j \in \mathcal{N}(i)}\left(1-p_{j, t}+\left(1-\beta_{i, j}\right) p_{j, t}\right).
\end{equation}
The parameter $\beta_{i, j} = \beta t_{i,j}$ represents the transmission rate between node $i$ and $j$.
Similarly, Alberto et al. \cite{aleta2020modelling} construct a multi-layer network with a household layer, workplace, and community layer, and a school layer to investigate how testing, contact tracing, and household quarantine affect the disease transmission. Abueg et al. \cite{abueg2021modeling} simulate the contact network in Washington state by three types of sub-networks: fully connected networks for households, small world networks \cite{watts1998collective} for workplaces, schools or social circles, and random networks for random interactions. Nishi et al. \cite{nishi2020network} further investigate how interventions on the structure of contact networks affect the disease transmission. Murphy et al. \cite{murphy2021deep} propose a novel graph convolution network (GCN) model to capture the contagion dynamics in complex contact networks, which is the recent advantage that combines classic epidemiological model with deep learning technology.
From the above discussion, macroscopic models \cite{jia2020population,kraemer2020effect,lai2020effect} only require the empirical epidemic records at the aggregated level, which can be easily obtained in most parts of the world. However, it also forbids them to explain the underlying mechanism due to the over-aggregation of the model parameters, such as the superspreading phenomenon \cite{Lloyd-Smith2005-ao,adam2020clustering}. Microscopic models greatly improve the expressive power of macroscopic models by introducing individual mobility \cite{murphy2021deep,qian2021scaling,aleta2020modelling}, where the heterogeneous contact patterns explain the emergence of such phenomena. However, it also poses huge demand for mobility data \cite{anwari2021exploring}, which are often difficult to collect in low- and middle-income countries and raises significant privacy concerns~\cite{ahmad2020state}. This dilemma impels the emergence of multi-scale models that achieves a great balance between the expressive power and data demand of the contagion system.
\textbf{Metapopulation model.} A classic solution is the metapopulation model, where the whole simulation region is divided into several sub-regions, where each sub-region maintains a compartmental model (such as SEIR), where an aggregated mobility network connects these sub-regions. The famous global epidemic and mobility (GLEaM) model \cite{balcan2009multiscale,balcan2010modeling} considers the multi-scale mobility of short-scale commuting flows and long-range airline traffic on an SEIR model, which is still inspiring recent studies \cite{zhang2020changes,chinazzi2020effect}. Wu et al. \cite{wu2020nowcasting} propose a metapopulation model using aggregated human mobility data across more than 300 prefecture-level cities in mainland China to investigate how domestic and international movement affect the outbreak of the disease,
where the number of international outbound and inbound air passengers and the daily number of all domestic outbound and inbound travelers are all considered.
\iffalse
Specifically, we have:
\begin{equation}
\begin{aligned}
\frac{\mathrm{d} S(t)}{\mathrm{d} t} &=-\frac{S(t)}{N}\left(\frac{R_{0}}{D_{I}} I(t)+z(t)\right) +L_{I, W}+L_{C, W}(t) -\left(\frac{L_{W, I}}{N}+\frac{L_{W, C}(t)}{N}\right) S(t) \\
\frac{\mathrm{d} E}{\mathrm{~d} t} &=\frac{S(t)}{N}\left(\frac{R_{0}}{D_{I}} I(t)+z(t)\right) -\frac{E(t)}{D_{E}} -\left(\frac{L_{W, I}}{N}+\frac{L_{W, C}(t)}{N}\right) E(t) \\
\frac{\mathrm{d} I(t)}{\mathrm{d} t} &=\frac{E(t)}{D_{E}}-\frac{I(t)}{D_{I}}-\left(\frac{L_{W, I}}{N}+\frac{L_{W, C}(t)}{N}\right) I(t),
\end{aligned}
\end{equation}
where $L_{W,I}$ and $L_{I,W}$ is the daily average number of international outbound and inbound air passengers accordingly, and $L_{W,C}$ and $L_{C,W}$ are the daily number of all domestic outbound and inbound travelers.
\fi
Then, according to the mobility network, the sub-regions are connected. The regional mobility data is easier to obtain, which greatly reduces the difficulty of applying the model in real-world scenarios.
Zhou et al. \cite{zhou2020effects} use anonymous mobile phone data to construct an individual-level SEIR model in a metapopulation manner, where the city is divided into multiple regions and runs different SEIR models with different regions, where the metapopulation matrix is incorporated.
\iffalse
index $i$ as follows:
\begin{equation}
\begin{aligned}
\frac{\mathrm{d} S_i(t)}{\mathrm{d} t} &= -\lambda_i S_i(t) \\
\frac{\mathrm{d} E_i(t)}{\mathrm{d} t} &= \lambda_i S_i(t) - \sigma_i E_i(t)\\
\frac{\mathrm{d} I_i(t)}{\mathrm{d} t} &= \sigma_i E_i(t) - \gamma_i I_i(t)\\
\frac{\mathrm{d} R_i(t)}{\mathrm{d} t} &= \gamma_i I_i(t),
\end{aligned}
\end{equation}
and the infection rate $\lambda$ is coupled with the mobility pattern by the following equations:
\begin{equation}
\lambda_{i}=\sum_{j=1}^{k+1} m_{i j} \beta_{j} \frac{\sum_{l=1}^{k} m_{l j}\left(E_{l}+P_{m} I_{l}\right)}{\sum_{l=1}^{k} m_{l j} N_{l}},
\end{equation}
where $N_l$ is the total population in region $l$, and $P_m$ is the proportion of mobile infectious cases. $\beta_j$ represents the transmission rate from infected people. Specifically, the mixing matrix $\bm{M} = {m_{ab}}$ is constructed as follows:
\begin{equation}
\begin{aligned}
m_{ij} &= \frac{L_{ij}}{N_i} \times \frac{8}{24}, 0\leq i,j \leq K \\
m_{i,isolate} &=1-\sum_{j=1}^{K} m_{ij},0\leq i,j \leq K, isolate=K+1.
\end{aligned}
\end{equation}
$L_{ij}$ is the number of people moving from region $i$ to $j$ according to the mobile phone data.
\fi
Furthermore, Chang et al. \cite{chang2021mobility} propose a bipartite network with time-varying edges between census block groups (CBGs) and points of interest (POIs), which represents the residential areas and venues where the transmission may occur. Based on this highly flexible framework, the authors successfully reproduce the superspreading phenomenon and evaluate the disproportionate impact of the epidemic towards disadvantaged racial and socioeconomic groups.
\textbf{High-order network model.}
Another problem that lies in the single-scale simulation methods is the ignorance of high-order effects during the disease transmission. For example, the macroscopic compartmental models assume a heterogeneous first-order contact between different compartments, and the microscopic models only consider direct connections between individuals. However, such assumptions disobey real-world observations: the contact between family members are closer than the contact between acquaintance, which leads to a higher probability of infection. To tackle this challenge, researchers propose high-order networks, where simplicial complexes \cite{iacopini2019simplicial} or hypergraphs \cite{st2021universal} is introduced to model the transmission patterns in a small group of people. Lacopini et al. \cite{iacopini2019simplicial} assign different infection rates to different motifs to depict the high-order pattern in a social network, where the transition of infection density becomes discontinued and bistable. Similarly, St-Onge et al. \cite{st2021universal} also observe such discontinuous transitions with a different method, where they introduce hyperedges (an edge that connects multiple nodes) to model the high-order feature.
To summarize, going beyond single-scale simulation methods that stuck in the dilemma of expressive power and data demand, multi-scale simulation uses the metapopulation technique to improve the simulation scale with aggregated level mobility data. It also reveals the underlying high-order effect of disease transmission by considering the contact pattern of a group of people, which brings the simulation of epidemic contagion systems to a new height.
%%% -*-BibTeX-*-
%%% Do NOT edit. File created by BibTeX with style
%%% ACM-Reference-Format-Journals [18-Jan-2012].
\begin{thebibliography}{203}
%%% ====================================================================
%%% NOTE TO THE USER: you can override these defaults by providing
%%% customized versions of any of these macros before the \bibliography
%%% command. Each of them MUST provide its own final punctuation,
%%% except for \shownote{}, \showDOI{}, and \showURL{}. The latter two
%%% do not use final punctuation, in order to avoid confusing it with
%%% the Web address.
%%%
%%% To suppress output of a particular field, define its macro to expand
%%% to an empty string, or better, \unskip, like this:
%%%
%%% \newcommand{\showDOI}[1]{\unskip} % LaTeX syntax
%%%
%%% \def \showDOI #1{\unskip} % plain TeX syntax
%%%
%%% ====================================================================
\ifx \showCODEN \undefined \def \showCODEN #1{\unskip} \fi
\ifx \showDOI \undefined \def \showDOI #1{#1}\fi
\ifx \showISBNx \undefined \def \showISBNx #1{\unskip} \fi
\ifx \showISBNxiii \undefined \def \showISBNxiii #1{\unskip} \fi
\ifx \showISSN \undefined \def \showISSN #1{\unskip} \fi
\ifx \showLCCN \undefined \def \showLCCN #1{\unskip} \fi
\ifx \shownote \undefined \def \shownote #1{#1} \fi
\ifx \showarticletitle \undefined \def \showarticletitle #1{#1} \fi
\ifx \showURL \undefined \def \showURL {\relax} \fi
% The following commands are used for tagged output and should be
% invisible to TeX
\providecommand\bibfield[2]{#2}
\providecommand\bibinfo[2]{#2}
\providecommand\natexlab[1]{#1}
\providecommand\showeprint[2][]{arXiv:#2}
\bibitem[Abueg et~al\mbox{.}(2021)]%
{abueg2021modeling}
\bibfield{author}{\bibinfo{person}{Matthew Abueg}, \bibinfo{person}{Robert
Hinch}, \bibinfo{person}{Neo Wu}, \bibinfo{person}{Luyang Liu},
\bibinfo{person}{William Probert}, \bibinfo{person}{Austin Wu},
\bibinfo{person}{Paul Eastham}, \bibinfo{person}{Yusef Shafi},
\bibinfo{person}{Matt Rosencrantz}, \bibinfo{person}{Michael Dikovsky},
{et~al\mbox{.}}} \bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{Modeling the effect of exposure notification and
non-pharmaceutical interventions on COVID-19 transmission in Washington
state}.
\newblock \bibinfo{journal}{\emph{NPJ digital medicine}} \bibinfo{volume}{4},
\bibinfo{number}{1} (\bibinfo{year}{2021}), \bibinfo{pages}{1--10}.
\newblock
\bibitem[Adam et~al\mbox{.}(2020)]%
{adam2020clustering}
\bibfield{author}{\bibinfo{person}{Dillon~C Adam}, \bibinfo{person}{Peng Wu},
\bibinfo{person}{Jessica~Y Wong}, \bibinfo{person}{Eric~HY Lau},
\bibinfo{person}{Tim~K Tsang}, \bibinfo{person}{Simon Cauchemez},
\bibinfo{person}{Gabriel~M Leung}, {and} \bibinfo{person}{Benjamin~J
Cowling}.} \bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{Clustering and superspreading potential of
{SARS-CoV-2} infections in {Hong Kong}}.
\newblock \bibinfo{journal}{\emph{Nature Medicine}} \bibinfo{volume}{26},
\bibinfo{number}{11} (\bibinfo{year}{2020}), \bibinfo{pages}{1714--1719}.
\newblock
\bibitem[Ahmad and Chauhan(2020)]%
{ahmad2020state}
\bibfield{author}{\bibinfo{person}{Norita Ahmad} {and} \bibinfo{person}{Preeti
Chauhan}.} \bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{State of data privacy during {COVID-19}}.
\newblock \bibinfo{journal}{\emph{Computer}} \bibinfo{volume}{53},
\bibinfo{number}{10} (\bibinfo{year}{2020}), \bibinfo{pages}{119--122}.
\newblock
\bibitem[Albi et~al\mbox{.}(2019)]%
{albi2019vehicular}
\bibfield{author}{\bibinfo{person}{Giacomo Albi}, \bibinfo{person}{Nicola
Bellomo}, \bibinfo{person}{Luisa Fermo}, \bibinfo{person}{S-Y Ha},
\bibinfo{person}{J Kim}, \bibinfo{person}{Lorenzo Pareschi},
\bibinfo{person}{David Poyato}, {and} \bibinfo{person}{Juan Soler}.}
\bibinfo{year}{2019}\natexlab{}.
\newblock \showarticletitle{Vehicular traffic, crowds, and swarms: From kinetic
theory and multiscale methods to applications and research perspectives}.
\newblock \bibinfo{journal}{\emph{Mathematical Models and Methods in Applied
Sciences}} \bibinfo{volume}{29}, \bibinfo{number}{10} (\bibinfo{year}{2019}),
\bibinfo{pages}{1901--2005}.
\newblock
\bibitem[Aleta et~al\mbox{.}(2020)]%
{aleta2020modelling}
\bibfield{author}{\bibinfo{person}{Alberto Aleta}, \bibinfo{person}{David
Martin-Corral}, \bibinfo{person}{Ana Pastore~y Piontti},
\bibinfo{person}{Marco Ajelli}, \bibinfo{person}{Maria Litvinova},
\bibinfo{person}{Matteo Chinazzi}, \bibinfo{person}{Natalie~E Dean},
\bibinfo{person}{M~Elizabeth Halloran}, \bibinfo{person}{Ira~M Longini~Jr},
\bibinfo{person}{Stefano Merler}, {et~al\mbox{.}}}
\bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{Modelling the impact of testing, contact tracing
and household quarantine on second waves of COVID-19}.
\newblock \bibinfo{journal}{\emph{Nature Human Behaviour}} \bibinfo{volume}{4},
\bibinfo{number}{9} (\bibinfo{year}{2020}), \bibinfo{pages}{964--971}.
\newblock
\bibitem[Alho et~al\mbox{.}(2017)]%
{alho2017multi}
\bibfield{author}{\bibinfo{person}{Andr{\'e} Alho}, \bibinfo{person}{BK
Bhavathrathan}, \bibinfo{person}{Monique Stinson}, \bibinfo{person}{Raja
Gopalakrishnan}, \bibinfo{person}{Diem-Trinh Le}, {and}
\bibinfo{person}{Moshe Ben-Akiva}.} \bibinfo{year}{2017}\natexlab{}.
\newblock \showarticletitle{A multi-scale agent-based modelling framework for
urban freight distribution}.
\newblock \bibinfo{journal}{\emph{Transportation Research Procedia}}
\bibinfo{volume}{27} (\bibinfo{year}{2017}), \bibinfo{pages}{188--196}.
\newblock
\bibitem[Allam et~al\mbox{.}(2018)]%
{allam2018application}
\bibfield{author}{\bibinfo{person}{Omar Allam}, \bibinfo{person}{Byung~Woo
Cho}, \bibinfo{person}{Ki~Chul Kim}, {and} \bibinfo{person}{Seung~Soon
Jang}.} \bibinfo{year}{2018}\natexlab{}.
\newblock \showarticletitle{Application of DFT-based machine learning for
developing molecular electrode materials in Li-ion batteries}.
\newblock \bibinfo{journal}{\emph{RSC advances}} \bibinfo{volume}{8},
\bibinfo{number}{69} (\bibinfo{year}{2018}), \bibinfo{pages}{39414--39420}.
\newblock
\bibitem[Alshehri and You(2022)]%
{alshehri2022machine}
\bibfield{author}{\bibinfo{person}{Abdulelah~S Alshehri} {and}
\bibinfo{person}{Fengqi You}.} \bibinfo{year}{2022}\natexlab{}.
\newblock \showarticletitle{Machine learning for multiscale modeling in
computational molecular design}.
\newblock \bibinfo{journal}{\emph{Current Opinion in Chemical Engineering}}
\bibinfo{volume}{36} (\bibinfo{year}{2022}), \bibinfo{pages}{100752}.
\newblock
\bibitem[Anwari et~al\mbox{.}(2021)]%
{anwari2021exploring}
\bibfield{author}{\bibinfo{person}{Nafis Anwari}, \bibinfo{person}{Md~Tawkir
Ahmed}, \bibinfo{person}{Md~Rakibul Islam}, \bibinfo{person}{Md Hadiuzzaman},
{and} \bibinfo{person}{Shohel Amin}.} \bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{Exploring the travel behavior changes caused by the
COVID-19 crisis: A case study for a developing country}.
\newblock \bibinfo{journal}{\emph{Transportation Research Interdisciplinary
Perspectives}} \bibinfo{volume}{9} (\bibinfo{year}{2021}),
\bibinfo{pages}{100334}.
\newblock
\bibitem[Balcan et~al\mbox{.}(2009)]%
{balcan2009multiscale}
\bibfield{author}{\bibinfo{person}{Duygu Balcan}, \bibinfo{person}{Vittoria
Colizza}, \bibinfo{person}{Bruno Gon{\c{c}}alves}, \bibinfo{person}{Hao Hu},
\bibinfo{person}{Jos{\'e}~J Ramasco}, {and} \bibinfo{person}{Alessandro
Vespignani}.} \bibinfo{year}{2009}\natexlab{}.
\newblock \showarticletitle{Multiscale mobility networks and the spatial
spreading of infectious diseases}.
\newblock \bibinfo{journal}{\emph{Proceedings of the National Academy of
Sciences}} \bibinfo{volume}{106}, \bibinfo{number}{51}
(\bibinfo{year}{2009}), \bibinfo{pages}{21484--21489}.
\newblock
\bibitem[Balcan et~al\mbox{.}(2010)]%
{balcan2010modeling}
\bibfield{author}{\bibinfo{person}{Duygu Balcan}, \bibinfo{person}{Bruno
Gon{\c{c}}alves}, \bibinfo{person}{Hao Hu}, \bibinfo{person}{Jos{\'e}~J
Ramasco}, \bibinfo{person}{Vittoria Colizza}, {and}
\bibinfo{person}{Alessandro Vespignani}.} \bibinfo{year}{2010}\natexlab{}.
\newblock \showarticletitle{Modeling the spatial spread of infectious diseases:
The GLobal Epidemic and Mobility computational model}.
\newblock \bibinfo{journal}{\emph{Journal of computational science}}
\bibinfo{volume}{1}, \bibinfo{number}{3} (\bibinfo{year}{2010}),
\bibinfo{pages}{132--145}.
\newblock
\bibitem[Bando et~al\mbox{.}(1995)]%
{bando1995dynamical}
\bibfield{author}{\bibinfo{person}{Masako Bando}, \bibinfo{person}{Katsuya
Hasebe}, \bibinfo{person}{Akihiro Nakayama}, \bibinfo{person}{Akihiro
Shibata}, {and} \bibinfo{person}{Yuki Sugiyama}.}
\bibinfo{year}{1995}\natexlab{}.
\newblock \showarticletitle{Dynamical model of traffic congestion and numerical
simulation}.
\newblock \bibinfo{journal}{\emph{Physical review E}} \bibinfo{volume}{51},
\bibinfo{number}{2} (\bibinfo{year}{1995}), \bibinfo{pages}{1035}.
\newblock
\bibitem[Bar-Sinai et~al\mbox{.}(2019)]%
{bar2019learning}
\bibfield{author}{\bibinfo{person}{Yohai Bar-Sinai}, \bibinfo{person}{Stephan
Hoyer}, \bibinfo{person}{Jason Hickey}, {and} \bibinfo{person}{Michael~P
Brenner}.} \bibinfo{year}{2019}\natexlab{}.
\newblock \showarticletitle{Learning data-driven discretizations for partial
differential equations}.
\newblock \bibinfo{journal}{\emph{Proceedings of the National Academy of
Sciences}} \bibinfo{volume}{116}, \bibinfo{number}{31}
(\bibinfo{year}{2019}), \bibinfo{pages}{15344--15349}.
\newblock
\bibitem[Barzel and Barab{\'a}si(2013)]%
{barzel2013universality}
\bibfield{author}{\bibinfo{person}{Baruch Barzel} {and}
\bibinfo{person}{Albert-L{\'a}szl{\'o} Barab{\'a}si}.}
\bibinfo{year}{2013}\natexlab{}.
\newblock \showarticletitle{Universality in network dynamics}.
\newblock \bibinfo{journal}{\emph{Nature physics}} \bibinfo{volume}{9},
\bibinfo{number}{10} (\bibinfo{year}{2013}), \bibinfo{pages}{673--681}.
\newblock
\bibitem[Bashan et~al\mbox{.}(2013)]%
{bashan2013extreme}
\bibfield{author}{\bibinfo{person}{Amir Bashan}, \bibinfo{person}{Yehiel
Berezin}, \bibinfo{person}{Sergey~V Buldyrev}, {and} \bibinfo{person}{Shlomo
Havlin}.} \bibinfo{year}{2013}\natexlab{}.
\newblock \showarticletitle{The extreme vulnerability of interdependent
spatially embedded networks}.
\newblock \bibinfo{journal}{\emph{Nature Physics}} \bibinfo{volume}{9},
\bibinfo{number}{10} (\bibinfo{year}{2013}), \bibinfo{pages}{667--672}.
\newblock
\bibitem[Battiston et~al\mbox{.}(2021)]%
{battiston2021physics}
\bibfield{author}{\bibinfo{person}{Federico Battiston}, \bibinfo{person}{Enrico
Amico}, \bibinfo{person}{Alain Barrat}, \bibinfo{person}{Ginestra Bianconi},
\bibinfo{person}{Guilherme Ferraz~de Arruda}, \bibinfo{person}{Benedetta
Franceschiello}, \bibinfo{person}{Iacopo Iacopini}, \bibinfo{person}{Sonia
K{\'e}fi}, \bibinfo{person}{Vito Latora}, \bibinfo{person}{Yamir Moreno},
{et~al\mbox{.}}} \bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{The physics of higher-order interactions in complex
systems}.
\newblock \bibinfo{journal}{\emph{Nature Physics}} \bibinfo{volume}{17},
\bibinfo{number}{10} (\bibinfo{year}{2021}), \bibinfo{pages}{1093--1098}.
\newblock
\bibitem[Beylkin and Coult(1998)]%
{beylkin1998multiresolution}
\bibfield{author}{\bibinfo{person}{Gregory Beylkin} {and}
\bibinfo{person}{Nicholas Coult}.} \bibinfo{year}{1998}\natexlab{}.
\newblock \showarticletitle{A multiresolution strategy for reduction of
elliptic PDEs and eigenvalue problems}.
\newblock \bibinfo{journal}{\emph{Applied and Computational Harmonic Analysis}}
\bibinfo{volume}{5}, \bibinfo{number}{2} (\bibinfo{year}{1998}),
\bibinfo{pages}{129--155}.
\newblock
\bibitem[Bhatia et~al\mbox{.}(2021)]%
{bhatia2021machine}
\bibfield{author}{\bibinfo{person}{Harsh Bhatia}, \bibinfo{person}{Timothy~S
Carpenter}, \bibinfo{person}{Helgi~I Ing{\'o}lfsson},
\bibinfo{person}{Gautham Dharuman}, \bibinfo{person}{Piyush Karande},
\bibinfo{person}{Shusen Liu}, \bibinfo{person}{Tomas Oppelstrup},
\bibinfo{person}{Chris Neale}, \bibinfo{person}{Felice~C Lightstone},
\bibinfo{person}{Brian Van~Essen}, {et~al\mbox{.}}}
\bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{Machine-learning-based dynamic-importance sampling
for adaptive multiscale simulations}.
\newblock \bibinfo{journal}{\emph{Nature Machine Intelligence}}
\bibinfo{volume}{3}, \bibinfo{number}{5} (\bibinfo{year}{2021}),
\bibinfo{pages}{401--409}.
\newblock
\bibitem[B{\"o}hm et~al\mbox{.}(2022)]%
{bohm2022noise}
\bibfield{author}{\bibinfo{person}{Fabian B{\"o}hm}, \bibinfo{person}{Diego
Alonso-Urquijo}, \bibinfo{person}{Guy Verschaffelt}, {and}
\bibinfo{person}{Guy Van~der Sande}.} \bibinfo{year}{2022}\natexlab{}.
\newblock \showarticletitle{Noise-injected analog Ising machines enable
ultrafast statistical sampling and machine learning}.
\newblock \bibinfo{journal}{\emph{Nature Communications}} \bibinfo{volume}{13},
\bibinfo{number}{1} (\bibinfo{year}{2022}), \bibinfo{pages}{1--13}.
\newblock
\bibitem[Bramburger et~al\mbox{.}(2020)]%
{bramburger2020sparse}
\bibfield{author}{\bibinfo{person}{Jason~J Bramburger}, \bibinfo{person}{Daniel
Dylewsky}, {and} \bibinfo{person}{J~Nathan Kutz}.}
\bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{Sparse identification of slow timescale dynamics}.
\newblock \bibinfo{journal}{\emph{Physical Review E}} \bibinfo{volume}{102},
\bibinfo{number}{2} (\bibinfo{year}{2020}), \bibinfo{pages}{022204}.
\newblock
\bibitem[Brini et~al\mbox{.}(2020)]%
{brini2020protein}
\bibfield{author}{\bibinfo{person}{Emiliano Brini}, \bibinfo{person}{Carlos
Simmerling}, {and} \bibinfo{person}{Ken Dill}.}
\bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{Protein storytelling through physics}.
\newblock \bibinfo{journal}{\emph{Science}} \bibinfo{volume}{370},
\bibinfo{number}{6520} (\bibinfo{year}{2020}), \bibinfo{pages}{eaaz3041}.
\newblock
\bibitem[Brockmann and Helbing(2013)]%
{brockmann2013hidden}
\bibfield{author}{\bibinfo{person}{Dirk Brockmann} {and} \bibinfo{person}{Dirk
Helbing}.} \bibinfo{year}{2013}\natexlab{}.
\newblock \showarticletitle{The hidden geometry of complex, network-driven
contagion phenomena}.
\newblock \bibinfo{journal}{\emph{science}} \bibinfo{volume}{342},
\bibinfo{number}{6164} (\bibinfo{year}{2013}), \bibinfo{pages}{1337--1342}.
\newblock
\bibitem[Brooks et~al\mbox{.}(2009)]%
{brooks2009charmm}
\bibfield{author}{\bibinfo{person}{Bernard~R Brooks},
\bibinfo{person}{Charles~L Brooks~III}, \bibinfo{person}{Alexander~D
Mackerell~Jr}, \bibinfo{person}{Lennart Nilsson}, \bibinfo{person}{Robert~J
Petrella}, \bibinfo{person}{Beno{\^\i}t Roux}, \bibinfo{person}{Youngdo Won},
\bibinfo{person}{Georgios Archontis}, \bibinfo{person}{Christian Bartels},
\bibinfo{person}{Stefan Boresch}, {et~al\mbox{.}}}
\bibinfo{year}{2009}\natexlab{}.
\newblock \showarticletitle{CHARMM: the biomolecular simulation program}.
\newblock \bibinfo{journal}{\emph{Journal of computational chemistry}}
\bibinfo{volume}{30}, \bibinfo{number}{10} (\bibinfo{year}{2009}),
\bibinfo{pages}{1545--1614}.
\newblock
\bibitem[Brunton et~al\mbox{.}(2016)]%
{brunton2016discovering}
\bibfield{author}{\bibinfo{person}{Steven~L Brunton}, \bibinfo{person}{Joshua~L
Proctor}, {and} \bibinfo{person}{J~Nathan Kutz}.}
\bibinfo{year}{2016}\natexlab{}.
\newblock \showarticletitle{Discovering governing equations from data by sparse
identification of nonlinear dynamical systems}.
\newblock \bibinfo{journal}{\emph{Proceedings of the national academy of
sciences}} \bibinfo{volume}{113}, \bibinfo{number}{15}
(\bibinfo{year}{2016}), \bibinfo{pages}{3932--3937}.
\newblock
\bibitem[Buldyrev et~al\mbox{.}(2010)]%
{buldyrev2010catastrophic}
\bibfield{author}{\bibinfo{person}{Sergey~V Buldyrev}, \bibinfo{person}{Roni
Parshani}, \bibinfo{person}{Gerald Paul}, \bibinfo{person}{H~Eugene Stanley},
{and} \bibinfo{person}{Shlomo Havlin}.} \bibinfo{year}{2010}\natexlab{}.
\newblock \showarticletitle{Catastrophic cascade of failures in interdependent
networks}.
\newblock \bibinfo{journal}{\emph{Nature}} \bibinfo{volume}{464},
\bibinfo{number}{7291} (\bibinfo{year}{2010}), \bibinfo{pages}{1025--1028}.
\newblock
\bibitem[Car and Parrinello(1985)]%
{car1985unified}
\bibfield{author}{\bibinfo{person}{Richard Car} {and} \bibinfo{person}{Mark
Parrinello}.} \bibinfo{year}{1985}\natexlab{}.
\newblock \showarticletitle{Unified approach for molecular dynamics and
density-functional theory}.
\newblock \bibinfo{journal}{\emph{Physical review letters}}
\bibinfo{volume}{55}, \bibinfo{number}{22} (\bibinfo{year}{1985}),
\bibinfo{pages}{2471}.
\newblock
\bibitem[Chakraborty and Ghosh(2018)]%
{chakraborty2018hyperdynamics}
\bibfield{author}{\bibinfo{person}{Subhendu Chakraborty} {and}
\bibinfo{person}{Somnath Ghosh}.} \bibinfo{year}{2018}\natexlab{}.
\newblock \showarticletitle{Hyperdynamics accelerated concurrent
atomistic-continuum model for developing crack propagation models in elastic
crystalline materials}.
\newblock \bibinfo{journal}{\emph{Computational Materials Science}}
\bibinfo{volume}{154} (\bibinfo{year}{2018}), \bibinfo{pages}{212--224}.
\newblock
\bibitem[Chakroborty and Maurya(2008)]%
{chakroborty2008microscopic}
\bibfield{author}{\bibinfo{person}{Partha Chakroborty} {and}
\bibinfo{person}{Akhilesh~Kumar Maurya}.} \bibinfo{year}{2008}\natexlab{}.
\newblock \showarticletitle{Microscopic analysis of cellular automata based
traffic flow models and an improved model}.
\newblock \bibinfo{journal}{\emph{Transport reviews}} \bibinfo{volume}{28},
\bibinfo{number}{6} (\bibinfo{year}{2008}), \bibinfo{pages}{717--734}.
\newblock
\bibitem[Champion et~al\mbox{.}(2019b)]%
{champion2019data}
\bibfield{author}{\bibinfo{person}{Kathleen Champion}, \bibinfo{person}{Bethany
Lusch}, \bibinfo{person}{J~Nathan Kutz}, {and} \bibinfo{person}{Steven~L
Brunton}.} \bibinfo{year}{2019}\natexlab{b}.
\newblock \showarticletitle{Data-driven discovery of coordinates and governing
equations}.
\newblock \bibinfo{journal}{\emph{Proceedings of the National Academy of
Sciences}} \bibinfo{volume}{116}, \bibinfo{number}{45}
(\bibinfo{year}{2019}), \bibinfo{pages}{22445--22451}.
\newblock
\bibitem[Champion et~al\mbox{.}(2019a)]%
{champion2019discovery}
\bibfield{author}{\bibinfo{person}{Kathleen~P Champion},
\bibinfo{person}{Steven~L Brunton}, {and} \bibinfo{person}{J~Nathan Kutz}.}
\bibinfo{year}{2019}\natexlab{a}.
\newblock \showarticletitle{Discovery of nonlinear multiscale systems: Sampling
strategies and embeddings}.
\newblock \bibinfo{journal}{\emph{SIAM Journal on Applied Dynamical Systems}}
\bibinfo{volume}{18}, \bibinfo{number}{1} (\bibinfo{year}{2019}),
\bibinfo{pages}{312--333}.
\newblock
\bibitem[Chang et~al\mbox{.}(2021)]%
{chang2021mobility}
\bibfield{author}{\bibinfo{person}{Serina Chang}, \bibinfo{person}{Emma
Pierson}, \bibinfo{person}{Pang~Wei Koh}, \bibinfo{person}{Jaline Gerardin},
\bibinfo{person}{Beth Redbird}, \bibinfo{person}{David Grusky}, {and}
\bibinfo{person}{Jure Leskovec}.} \bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{Mobility network models of COVID-19 explain
inequities and inform reopening}.
\newblock \bibinfo{journal}{\emph{Nature}} \bibinfo{volume}{589},
\bibinfo{number}{7840} (\bibinfo{year}{2021}), \bibinfo{pages}{82--87}.
\newblock
\bibitem[Chen et~al\mbox{.}(2022a)]%
{chen2022automated}
\bibfield{author}{\bibinfo{person}{Boyuan Chen}, \bibinfo{person}{Kuang Huang},
\bibinfo{person}{Sunand Raghupathi}, \bibinfo{person}{Ishaan Chandratreya},
\bibinfo{person}{Qiang Du}, {and} \bibinfo{person}{Hod Lipson}.}
\bibinfo{year}{2022}\natexlab{a}.
\newblock \showarticletitle{Automated discovery of fundamental variables hidden
in experimental data}.
\newblock \bibinfo{journal}{\emph{Nature Computational Science}}
\bibinfo{volume}{2}, \bibinfo{number}{7} (\bibinfo{year}{2022}),
\bibinfo{pages}{433--442}.
\newblock
\bibitem[Chen et~al\mbox{.}(2022b)]%
{chen2022strategic}
\bibfield{author}{\bibinfo{person}{Lin Chen}, \bibinfo{person}{Fengli Xu},
\bibinfo{person}{Zhenyu Han}, \bibinfo{person}{Kun Tang},
\bibinfo{person}{Pan Hui}, \bibinfo{person}{James Evans}, {and}
\bibinfo{person}{Yong Li}.} \bibinfo{year}{2022}\natexlab{b}.
\newblock \showarticletitle{Strategic COVID-19 vaccine distribution can
simultaneously elevate social utility and equity}.
\newblock \bibinfo{journal}{\emph{Nature Human Behaviour}}
(\bibinfo{year}{2022}), \bibinfo{pages}{1--12}.
\newblock
\bibitem[Chen et~al\mbox{.}(2018)]%
{chen2018neural}
\bibfield{author}{\bibinfo{person}{Ricky~TQ Chen}, \bibinfo{person}{Yulia
Rubanova}, \bibinfo{person}{Jesse Bettencourt}, {and}
\bibinfo{person}{David~K Duvenaud}.} \bibinfo{year}{2018}\natexlab{}.
\newblock \showarticletitle{Neural ordinary differential equations}.
\newblock \bibinfo{journal}{\emph{Advances in neural information processing
systems}} \bibinfo{volume}{31} (\bibinfo{year}{2018}).
\newblock
\bibitem[Chen et~al\mbox{.}(2011)]%
{chen2011multi}
\bibfield{author}{\bibinfo{person}{Shaopei Chen}, \bibinfo{person}{Jianjun
Tan}, \bibinfo{person}{Christophe Claramunt}, {and} \bibinfo{person}{Cyril
Ray}.} \bibinfo{year}{2011}\natexlab{}.
\newblock \showarticletitle{Multi-scale and multi-modal GIS-T data model}.
\newblock \bibinfo{journal}{\emph{Journal of Transport Geography}}
\bibinfo{volume}{19}, \bibinfo{number}{1} (\bibinfo{year}{2011}),
\bibinfo{pages}{147--161}.
\newblock
\bibitem[Chen and Fish(2006)]%
{chen2006generalized}
\bibfield{author}{\bibinfo{person}{Wen Chen} {and} \bibinfo{person}{Jacob
Fish}.} \bibinfo{year}{2006}\natexlab{}.
\newblock \showarticletitle{A generalized space--time mathematical
homogenization theory for bridging atomistic and continuum scales}.
\newblock \bibinfo{journal}{\emph{Internat. J. Numer. Methods Engrg.}}
\bibinfo{volume}{67}, \bibinfo{number}{2} (\bibinfo{year}{2006}),
\bibinfo{pages}{253--271}.
\newblock
\bibitem[Chinazzi et~al\mbox{.}(2020)]%
{chinazzi2020effect}
\bibfield{author}{\bibinfo{person}{Matteo Chinazzi}, \bibinfo{person}{Jessica~T
Davis}, \bibinfo{person}{Marco Ajelli}, \bibinfo{person}{Corrado Gioannini},
\bibinfo{person}{Maria Litvinova}, \bibinfo{person}{Stefano Merler},
\bibinfo{person}{Ana Pastore~y Piontti}, \bibinfo{person}{Kunpeng Mu},
\bibinfo{person}{Luca Rossi}, \bibinfo{person}{Kaiyuan Sun}, {et~al\mbox{.}}}
\bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{The effect of travel restrictions on the spread of
the 2019 novel coronavirus (COVID-19) outbreak}.
\newblock \bibinfo{journal}{\emph{Science}} \bibinfo{volume}{368},
\bibinfo{number}{6489} (\bibinfo{year}{2020}), \bibinfo{pages}{395--400}.
\newblock
\bibitem[Chowdhury et~al\mbox{.}(2000)]%
{chowdhury2000statistical}
\bibfield{author}{\bibinfo{person}{Debashish Chowdhury},
\bibinfo{person}{Ludger Santen}, {and} \bibinfo{person}{Andreas
Schadschneider}.} \bibinfo{year}{2000}\natexlab{}.
\newblock \showarticletitle{Statistical physics of vehicular traffic and some
related systems}.
\newblock \bibinfo{journal}{\emph{Physics Reports}} \bibinfo{volume}{329},
\bibinfo{number}{4-6} (\bibinfo{year}{2000}), \bibinfo{pages}{199--329}.
\newblock
\bibitem[Crim(2008)]%
{crim2008chemical}
\bibfield{author}{\bibinfo{person}{F~Fleming Crim}.}
\bibinfo{year}{2008}\natexlab{}.
\newblock \showarticletitle{Chemical reaction dynamics}.
\newblock \bibinfo{journal}{\emph{Proceedings of the National Academy of
Sciences}} \bibinfo{volume}{105}, \bibinfo{number}{35}
(\bibinfo{year}{2008}), \bibinfo{pages}{12647--12648}.
\newblock
\bibitem[Dai et~al\mbox{.}(2019)]%
{dai2019excmg}
\bibfield{author}{\bibinfo{person}{Ruxin Dai}, \bibinfo{person}{Pengpeng Lin},
{and} \bibinfo{person}{Jun Zhang}.} \bibinfo{year}{2019}\natexlab{}.
\newblock \showarticletitle{An EXCMG accelerated multiscale multigrid
computation for 3D Poisson equation}.
\newblock \bibinfo{journal}{\emph{Computers \& Mathematics with Applications}}
\bibinfo{volume}{77}, \bibinfo{number}{8} (\bibinfo{year}{2019}),
\bibinfo{pages}{2051--2060}.
\newblock
\bibitem[Di~Natale et~al\mbox{.}(2019)]%
{di2019massively}
\bibfield{author}{\bibinfo{person}{Francesco Di~Natale}, \bibinfo{person}{Harsh
Bhatia}, \bibinfo{person}{Timothy~S Carpenter}, \bibinfo{person}{Chris
Neale}, \bibinfo{person}{Sara Kokkila-Schumacher}, \bibinfo{person}{Tomas
Oppelstrup}, \bibinfo{person}{Liam Stanton}, \bibinfo{person}{Xiaohua Zhang},
\bibinfo{person}{Shiv Sundram}, \bibinfo{person}{Thomas~RW Scogland},
{et~al\mbox{.}}} \bibinfo{year}{2019}\natexlab{}.
\newblock \showarticletitle{A massively parallel infrastructure for adaptive
multiscale simulations: modeling RAS initiation pathway for cancer}. In
\bibinfo{booktitle}{\emph{Proceedings of the International Conference for
High Performance Computing, Networking, Storage and Analysis}}.
\bibinfo{pages}{1--16}.
\newblock
\bibitem[Di~Sante et~al\mbox{.}(2022)]%
{di2022deep}
\bibfield{author}{\bibinfo{person}{Domenico Di~Sante}, \bibinfo{person}{Matija
Medvidovi{\'c}}, \bibinfo{person}{Alessandro Toschi},
\bibinfo{person}{Giorgio Sangiovanni}, \bibinfo{person}{Cesare Franchini},
\bibinfo{person}{Anirvan~M Sengupta}, {and} \bibinfo{person}{Andrew~J
Millis}.} \bibinfo{year}{2022}\natexlab{}.
\newblock \showarticletitle{Deep Learning the Functional Renormalization
Group}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2202.13268}}
(\bibinfo{year}{2022}).
\newblock
\bibitem[Dupuis et~al\mbox{.}(2012)]%
{dupuis2012importance}
\bibfield{author}{\bibinfo{person}{Paul Dupuis}, \bibinfo{person}{Konstantinos
Spiliopoulos}, {and} \bibinfo{person}{Hui Wang}.}
\bibinfo{year}{2012}\natexlab{}.
\newblock \showarticletitle{Importance sampling for multiscale diffusions}.
\newblock \bibinfo{journal}{\emph{Multiscale Modeling \& Simulation}}
\bibinfo{volume}{10}, \bibinfo{number}{1} (\bibinfo{year}{2012}),
\bibinfo{pages}{1--27}.
\newblock
\bibitem[Faraji et~al\mbox{.}(2017)]%
{faraji2017high}
\bibfield{author}{\bibinfo{person}{Somayeh Faraji}, \bibinfo{person}{S~Alireza
Ghasemi}, \bibinfo{person}{Samare Rostami}, \bibinfo{person}{Robabe
Rasoulkhani}, \bibinfo{person}{Bastian Schaefer}, \bibinfo{person}{Stefan
Goedecker}, {and} \bibinfo{person}{Maximilian Amsler}.}
\bibinfo{year}{2017}\natexlab{}.
\newblock \showarticletitle{High accuracy and transferability of a neural
network potential through charge equilibration for calcium fluoride}.
\newblock \bibinfo{journal}{\emph{Physical Review B}} \bibinfo{volume}{95},
\bibinfo{number}{10} (\bibinfo{year}{2017}), \bibinfo{pages}{104105}.
\newblock
\bibitem[Fish et~al\mbox{.}(2021)]%
{fish2021mesoscopic}
\bibfield{author}{\bibinfo{person}{Jacob Fish}, \bibinfo{person}{Gregory~J
Wagner}, {and} \bibinfo{person}{Sinan Keten}.}
\bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{Mesoscopic and multiscale modelling in materials}.
\newblock \bibinfo{journal}{\emph{Nature materials}} \bibinfo{volume}{20},
\bibinfo{number}{6} (\bibinfo{year}{2021}), \bibinfo{pages}{774--786}.
\newblock
\bibitem[Fish and Wu(2011)]%
{fish2011nonintrusive}
\bibfield{author}{\bibinfo{person}{Jacob Fish} {and} \bibinfo{person}{Wei Wu}.}
\bibinfo{year}{2011}\natexlab{}.
\newblock \showarticletitle{A nonintrusive stochastic multiscale solver}.
\newblock \bibinfo{journal}{\emph{Internat. J. Numer. Methods Engrg.}}
\bibinfo{volume}{88}, \bibinfo{number}{9} (\bibinfo{year}{2011}),
\bibinfo{pages}{862--879}.
\newblock
\bibitem[Flaxman et~al\mbox{.}(2020)]%
{flaxman2020estimating}
\bibfield{author}{\bibinfo{person}{Seth Flaxman}, \bibinfo{person}{Swapnil
Mishra}, \bibinfo{person}{Axel Gandy}, \bibinfo{person}{H~Juliette~T Unwin},
\bibinfo{person}{Thomas~A Mellan}, \bibinfo{person}{Helen Coupland},
\bibinfo{person}{Charles Whittaker}, \bibinfo{person}{Harrison Zhu},
\bibinfo{person}{Tresnia Berah}, \bibinfo{person}{Jeffrey~W Eaton},
{et~al\mbox{.}}} \bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{Estimating the effects of non-pharmaceutical
interventions on COVID-19 in Europe}.
\newblock \bibinfo{journal}{\emph{Nature}} \bibinfo{volume}{584},
\bibinfo{number}{7820} (\bibinfo{year}{2020}), \bibinfo{pages}{257--261}.
\newblock
\bibitem[Fortunato et~al\mbox{.}(2022)]%
{fortunato2022multiscale}
\bibfield{author}{\bibinfo{person}{Meire Fortunato}, \bibinfo{person}{Tobias
Pfaff}, \bibinfo{person}{Peter Wirnsberger}, \bibinfo{person}{Alexander
Pritzel}, {and} \bibinfo{person}{Peter Battaglia}.}
\bibinfo{year}{2022}\natexlab{}.
\newblock \showarticletitle{Multiscale meshgraphnets}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2210.00612}}
(\bibinfo{year}{2022}).
\newblock
\bibitem[Fukami et~al\mbox{.}(2020)]%
{fukami2020convolutional}
\bibfield{author}{\bibinfo{person}{Kai Fukami}, \bibinfo{person}{Taichi
Nakamura}, {and} \bibinfo{person}{Koji Fukagata}.}
\bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{Convolutional neural network based hierarchical
autoencoder for nonlinear mode decomposition of fluid field data}.
\newblock \bibinfo{journal}{\emph{Physics of Fluids}} \bibinfo{volume}{32},
\bibinfo{number}{9} (\bibinfo{year}{2020}), \bibinfo{pages}{095110}.
\newblock
\bibitem[Gao et~al\mbox{.}(2016)]%
{gao2016universal}
\bibfield{author}{\bibinfo{person}{Jianxi Gao}, \bibinfo{person}{Baruch
Barzel}, {and} \bibinfo{person}{Albert-L{\'a}szl{\'o} Barab{\'a}si}.}
\bibinfo{year}{2016}\natexlab{}.
\newblock \showarticletitle{Universal resilience patterns in complex networks}.
\newblock \bibinfo{journal}{\emph{Nature}} \bibinfo{volume}{530},
\bibinfo{number}{7590} (\bibinfo{year}{2016}), \bibinfo{pages}{307--312}.
\newblock
\bibitem[Gao et~al\mbox{.}(2007)]%
{gao2007cellular}
\bibfield{author}{\bibinfo{person}{Kun Gao}, \bibinfo{person}{Rui Jiang},
\bibinfo{person}{Shou-Xin Hu}, \bibinfo{person}{Bing-Hong Wang}, {and}
\bibinfo{person}{Qing-Song Wu}.} \bibinfo{year}{2007}\natexlab{}.
\newblock \showarticletitle{Cellular-automaton model with velocity adaptation
in the framework of Kerner’s three-phase traffic theory}.
\newblock \bibinfo{journal}{\emph{Physical Review E}} \bibinfo{volume}{76},
\bibinfo{number}{2} (\bibinfo{year}{2007}), \bibinfo{pages}{026105}.
\newblock
\bibitem[Ge et~al\mbox{.}(2006)]%
{ge2006effect}
\bibfield{author}{\bibinfo{person}{HX Ge}, \bibinfo{person}{HB Zhu}, {and}
\bibinfo{person}{SQ Dai}.} \bibinfo{year}{2006}\natexlab{}.
\newblock \showarticletitle{Effect of looking backward on traffic flow in a
cooperative driving car following model}.
\newblock \bibinfo{journal}{\emph{The European Physical Journal B-Condensed
Matter and Complex Systems}} \bibinfo{volume}{54}, \bibinfo{number}{4}
(\bibinfo{year}{2006}), \bibinfo{pages}{503--507}.
\newblock
\bibitem[Ge and Li(2002)]%
{ge2002physical}
\bibfield{author}{\bibinfo{person}{Wei Ge} {and} \bibinfo{person}{Jinghai Li}.}
\bibinfo{year}{2002}\natexlab{}.
\newblock \showarticletitle{Physical mapping of fluidization regimes—the EMMS
approach}.
\newblock \bibinfo{journal}{\emph{Chemical Engineering Science}}
\bibinfo{volume}{57}, \bibinfo{number}{18} (\bibinfo{year}{2002}),
\bibinfo{pages}{3993--4004}.
\newblock
\bibitem[Ghavamian and Simone(2019)]%
{ghavamian2019accelerating}
\bibfield{author}{\bibinfo{person}{F Ghavamian} {and} \bibinfo{person}{A
Simone}.} \bibinfo{year}{2019}\natexlab{}.
\newblock \showarticletitle{Accelerating multiscale finite element simulations
of history-dependent materials using a recurrent neural network}.
\newblock \bibinfo{journal}{\emph{Computer Methods in Applied Mechanics and
Engineering}} \bibinfo{volume}{357} (\bibinfo{year}{2019}),
\bibinfo{pages}{112594}.
\newblock
\bibitem[Gilbert(1998)]%
{gilbert1998comparison}
\bibfield{author}{\bibinfo{person}{AC14927910896 Gilbert}.}
\bibinfo{year}{1998}\natexlab{}.
\newblock \showarticletitle{A comparison of multiresolution and classical
one-dimensional homogenization schemes}.
\newblock \bibinfo{journal}{\emph{Applied and computational harmonic analysis}}
\bibinfo{volume}{5}, \bibinfo{number}{1} (\bibinfo{year}{1998}),
\bibinfo{pages}{1--35}.
\newblock
\bibitem[Gomes et~al\mbox{.}(2018)]%
{gomes2018co}
\bibfield{author}{\bibinfo{person}{Cl{\'a}udio Gomes}, \bibinfo{person}{Casper
Thule}, \bibinfo{person}{David Broman}, \bibinfo{person}{Peter~Gorm Larsen},
{and} \bibinfo{person}{Hans Vangheluwe}.} \bibinfo{year}{2018}\natexlab{}.
\newblock \showarticletitle{Co-simulation: a survey}.
\newblock \bibinfo{journal}{\emph{ACM Computing Surveys (CSUR)}}
\bibinfo{volume}{51}, \bibinfo{number}{3} (\bibinfo{year}{2018}),
\bibinfo{pages}{1--33}.
\newblock
\bibitem[Gong et~al\mbox{.}(2021)]%
{gong2021advanced}
\bibfield{author}{\bibinfo{person}{Xiping Gong}, \bibinfo{person}{Yumeng
Zhang}, {and} \bibinfo{person}{Jianhan Chen}.}
\bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{Advanced Sampling Methods for Multiscale Simulation
of Disordered Proteins and Dynamic Interactions}.
\newblock \bibinfo{journal}{\emph{Biomolecules}} \bibinfo{volume}{11},
\bibinfo{number}{10} (\bibinfo{year}{2021}), \bibinfo{pages}{1416}.
\newblock
\bibitem[Goodfellow et~al\mbox{.}(2020)]%
{goodfellow2020generative}
\bibfield{author}{\bibinfo{person}{Ian Goodfellow}, \bibinfo{person}{Jean
Pouget-Abadie}, \bibinfo{person}{Mehdi Mirza}, \bibinfo{person}{Bing Xu},
\bibinfo{person}{David Warde-Farley}, \bibinfo{person}{Sherjil Ozair},
\bibinfo{person}{Aaron Courville}, {and} \bibinfo{person}{Yoshua Bengio}.}
\bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{Generative adversarial networks}.
\newblock \bibinfo{journal}{\emph{Commun. ACM}} \bibinfo{volume}{63},
\bibinfo{number}{11} (\bibinfo{year}{2020}), \bibinfo{pages}{139--144}.
\newblock
\bibitem[Gravemeier et~al\mbox{.}(2010)]%
{gravemeier2010algebraic}
\bibfield{author}{\bibinfo{person}{Volker Gravemeier},
\bibinfo{person}{Michael~W Gee}, \bibinfo{person}{Martin Kronbichler}, {and}
\bibinfo{person}{Wolfgang~A Wall}.} \bibinfo{year}{2010}\natexlab{}.
\newblock \showarticletitle{An algebraic variational multiscale--multigrid
method for large eddy simulation of turbulent flow}.
\newblock \bibinfo{journal}{\emph{Computer Methods in Applied Mechanics and
Engineering}} \bibinfo{volume}{199}, \bibinfo{number}{13-16}
(\bibinfo{year}{2010}), \bibinfo{pages}{853--864}.
\newblock
\bibitem[Griebenow et~al\mbox{.}(2019)]%
{griebenow2019finding}
\bibfield{author}{\bibinfo{person}{Ross Griebenow}, \bibinfo{person}{Brennan
Klein}, {and} \bibinfo{person}{Erik Hoel}.} \bibinfo{year}{2019}\natexlab{}.
\newblock \showarticletitle{Finding the right scale of a network: efficient
identification of causal emergence through spectral clustering}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:1908.07565}}
(\bibinfo{year}{2019}).
\newblock
\bibitem[Guan et~al\mbox{.}(2020)]%
{guan2020global}
\bibfield{author}{\bibinfo{person}{Dabo Guan}, \bibinfo{person}{Daoping Wang},
\bibinfo{person}{Stephane Hallegatte}, \bibinfo{person}{Steven~J Davis},
\bibinfo{person}{Jingwen Huo}, \bibinfo{person}{Shuping Li},
\bibinfo{person}{Yangchun Bai}, \bibinfo{person}{Tianyang Lei},
\bibinfo{person}{Qianyu Xue}, \bibinfo{person}{D’Maris Coffman},
{et~al\mbox{.}}} \bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{Global supply-chain effects of COVID-19 control
measures}.
\newblock \bibinfo{journal}{\emph{Nature human behaviour}} \bibinfo{volume}{4},
\bibinfo{number}{6} (\bibinfo{year}{2020}), \bibinfo{pages}{577--587}.
\newblock
\bibitem[Guo et~al\mbox{.}(2021a)]%
{guo2021hierarchical}
\bibfield{author}{\bibinfo{person}{Kan Guo}, \bibinfo{person}{Yongli Hu},
\bibinfo{person}{Yanfeng Sun}, \bibinfo{person}{Sean Qian},
\bibinfo{person}{Junbin Gao}, {and} \bibinfo{person}{Baocai Yin}.}
\bibinfo{year}{2021}\natexlab{a}.
\newblock \showarticletitle{Hierarchical Graph Convolution Network for Traffic
Forecasting}. In \bibinfo{booktitle}{\emph{Proceedings of the AAAI Conference
on Artificial Intelligence}}, Vol.~\bibinfo{volume}{35}.
\bibinfo{pages}{151--159}.
\newblock
\bibitem[Guo et~al\mbox{.}(2016)]%
{guo2016multi}
\bibfield{author}{\bibinfo{person}{Lin Guo}, \bibinfo{person}{Shiyi Chen},
{and} \bibinfo{person}{Mark~O Robbins}.} \bibinfo{year}{2016}\natexlab{}.
\newblock \showarticletitle{Multi-scale simulation method for electroosmotic
flows}.
\newblock \bibinfo{journal}{\emph{The European Physical Journal Special
Topics}} \bibinfo{volume}{225}, \bibinfo{number}{8} (\bibinfo{year}{2016}),
\bibinfo{pages}{1551--1582}.
\newblock
\bibitem[Guo et~al\mbox{.}(2021b)]%
{guo2021electromagnetic}
\bibfield{author}{\bibinfo{person}{Liangshuai Guo}, \bibinfo{person}{Maokun
Li}, \bibinfo{person}{Shenheng Xu}, \bibinfo{person}{Fan Yang}, {and}
\bibinfo{person}{Li Liu}.} \bibinfo{year}{2021}\natexlab{b}.
\newblock \showarticletitle{Electromagnetic Modeling Using an FDTD-Equivalent
Recurrent Convolution Neural Network: Accurate Computing on a Deep Learning
Framework.}
\newblock \bibinfo{journal}{\emph{IEEE Antennas and Propagation Magazine}}
(\bibinfo{year}{2021}).
\newblock
\bibitem[Habibi et~al\mbox{.}(2020)]%
{habibi2020data}
\bibfield{author}{\bibinfo{person}{Milad Habibi}, \bibinfo{person}{Scott~TM
Dawson}, {and} \bibinfo{person}{Amirhossein Arzani}.}
\bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{Data-driven pulsatile blood flow physics with
dynamic mode decomposition}.
\newblock \bibinfo{journal}{\emph{Fluids}} \bibinfo{volume}{5},
\bibinfo{number}{3} (\bibinfo{year}{2020}), \bibinfo{pages}{111}.
\newblock
\bibitem[Hall et~al\mbox{.}(2021)]%
{hall2021ginns}
\bibfield{author}{\bibinfo{person}{Eric~J Hall}, \bibinfo{person}{S{\o}ren
Taverniers}, \bibinfo{person}{Markos~A Katsoulakis}, {and}
\bibinfo{person}{Daniel~M Tartakovsky}.} \bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{Ginns: Graph-informed neural networks for
multiscale physics}.
\newblock \bibinfo{journal}{\emph{J. Comput. Phys.}} \bibinfo{volume}{433}
(\bibinfo{year}{2021}), \bibinfo{pages}{110192}.
\newblock
\bibitem[Hambli et~al\mbox{.}(2011)]%
{hambli2011multiscale}
\bibfield{author}{\bibinfo{person}{Ridha Hambli}, \bibinfo{person}{Houda
Katerchi}, {and} \bibinfo{person}{Claude-Laurent Benhamou}.}
\bibinfo{year}{2011}\natexlab{}.
\newblock \showarticletitle{Multiscale methodology for bone remodelling
simulation using coupled finite element and neural network computation}.
\newblock \bibinfo{journal}{\emph{Biomechanics and modeling in mechanobiology}}
\bibinfo{volume}{10}, \bibinfo{number}{1} (\bibinfo{year}{2011}),
\bibinfo{pages}{133--145}.
\newblock
\bibitem[Han et~al\mbox{.}(2021)]%
{han2021artificial}
\bibfield{author}{\bibinfo{person}{Changnian Han}, \bibinfo{person}{Peng
Zhang}, \bibinfo{person}{Danny Bluestein}, \bibinfo{person}{Guojing Cong},
{and} \bibinfo{person}{Yuefan Deng}.} \bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{Artificial intelligence for accelerating time
integrations in multiscale modeling}.
\newblock \bibinfo{journal}{\emph{J. Comput. Phys.}} \bibinfo{volume}{427}
(\bibinfo{year}{2021}), \bibinfo{pages}{110053}.
\newblock
\bibitem[Han et~al\mbox{.}(2017)]%
{han2017deep}
\bibfield{author}{\bibinfo{person}{Jiequn Han}, \bibinfo{person}{Linfeng
Zhang}, \bibinfo{person}{Roberto Car}, {et~al\mbox{.}}}
\bibinfo{year}{2017}\natexlab{}.
\newblock \showarticletitle{Deep potential: A general representation of a
many-body potential energy surface}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:1707.01478}}
(\bibinfo{year}{2017}).
\newblock
\bibitem[Haw et~al\mbox{.}(2022)]%
{haw2022optimizing}
\bibfield{author}{\bibinfo{person}{David~J Haw}, \bibinfo{person}{Giovanni
Forchini}, \bibinfo{person}{Patrick Doohan}, \bibinfo{person}{Paula
Christen}, \bibinfo{person}{Matteo Pianella}, \bibinfo{person}{Robert
Johnson}, \bibinfo{person}{Sumali Bajaj}, \bibinfo{person}{Alexandra~B
Hogan}, \bibinfo{person}{Peter Winskill}, \bibinfo{person}{Marisa Miraldo},
{et~al\mbox{.}}} \bibinfo{year}{2022}\natexlab{}.
\newblock \showarticletitle{Optimizing social and economic activity while
containing SARS-CoV-2 transmission using DAEDALUS}.
\newblock \bibinfo{journal}{\emph{Nature Computational Science}}
\bibinfo{volume}{2}, \bibinfo{number}{4} (\bibinfo{year}{2022}),
\bibinfo{pages}{223--233}.
\newblock
\bibitem[He et~al\mbox{.}(2016)]%
{he2016deep}
\bibfield{author}{\bibinfo{person}{Kaiming He}, \bibinfo{person}{Xiangyu
Zhang}, \bibinfo{person}{Shaoqing Ren}, {and} \bibinfo{person}{Jian Sun}.}
\bibinfo{year}{2016}\natexlab{}.
\newblock \showarticletitle{Deep residual learning for image recognition}. In
\bibinfo{booktitle}{\emph{Proceedings of the IEEE conference on computer
vision and pattern recognition}}. \bibinfo{pages}{770--778}.
\newblock
\bibitem[Helbing and Schreckenberg(1999)]%
{helbing1999cellular}
\bibfield{author}{\bibinfo{person}{Dirk Helbing} {and} \bibinfo{person}{Michael
Schreckenberg}.} \bibinfo{year}{1999}\natexlab{}.
\newblock \showarticletitle{Cellular automata simulating experimental
properties of traffic flow}.
\newblock \bibinfo{journal}{\emph{Physical review E}} \bibinfo{volume}{59},
\bibinfo{number}{3} (\bibinfo{year}{1999}), \bibinfo{pages}{R2505}.
\newblock
\bibitem[Helbing and Tilch(1998)]%
{helbing1998generalized}
\bibfield{author}{\bibinfo{person}{Dirk Helbing} {and} \bibinfo{person}{Benno
Tilch}.} \bibinfo{year}{1998}\natexlab{}.
\newblock \showarticletitle{Generalized force model of traffic dynamics}.
\newblock \bibinfo{journal}{\emph{Physical review E}} \bibinfo{volume}{58},
\bibinfo{number}{1} (\bibinfo{year}{1998}), \bibinfo{pages}{133}.
\newblock
\bibitem[Hemati et~al\mbox{.}(2014)]%
{hemati2014dynamic}
\bibfield{author}{\bibinfo{person}{Maziar~S Hemati}, \bibinfo{person}{Matthew~O
Williams}, {and} \bibinfo{person}{Clarence~W Rowley}.}
\bibinfo{year}{2014}\natexlab{}.
\newblock \showarticletitle{Dynamic mode decomposition for large and streaming
datasets}.
\newblock \bibinfo{journal}{\emph{Physics of Fluids}} \bibinfo{volume}{26},
\bibinfo{number}{11} (\bibinfo{year}{2014}), \bibinfo{pages}{111701}.
\newblock
\bibitem[Hern{\'a}ndez et~al\mbox{.}(2018)]%
{hernandez2018variational}
\bibfield{author}{\bibinfo{person}{Carlos~X Hern{\'a}ndez},
\bibinfo{person}{Hannah~K Wayment-Steele}, \bibinfo{person}{Mohammad~M
Sultan}, \bibinfo{person}{Brooke~E Husic}, {and} \bibinfo{person}{Vijay~S
Pande}.} \bibinfo{year}{2018}\natexlab{}.
\newblock \showarticletitle{Variational encoding of complex dynamics}.
\newblock \bibinfo{journal}{\emph{Physical Review E}} \bibinfo{volume}{97},
\bibinfo{number}{6} (\bibinfo{year}{2018}), \bibinfo{pages}{062412}.
\newblock
\bibitem[Hill(1963)]%
{hill1963elastic}
\bibfield{author}{\bibinfo{person}{Rodney Hill}.}
\bibinfo{year}{1963}\natexlab{}.
\newblock \showarticletitle{Elastic properties of reinforced solids: some
theoretical principles}.
\newblock \bibinfo{journal}{\emph{Journal of the Mechanics and Physics of
Solids}} \bibinfo{volume}{11}, \bibinfo{number}{5} (\bibinfo{year}{1963}),
\bibinfo{pages}{357--372}.
\newblock
\bibitem[Hochreiter and Schmidhuber(1997)]%
{hochreiter1997long}
\bibfield{author}{\bibinfo{person}{Sepp Hochreiter} {and}
\bibinfo{person}{J{\"u}rgen Schmidhuber}.} \bibinfo{year}{1997}\natexlab{}.
\newblock \showarticletitle{Long short-term memory}.
\newblock \bibinfo{journal}{\emph{Neural computation}} \bibinfo{volume}{9},
\bibinfo{number}{8} (\bibinfo{year}{1997}), \bibinfo{pages}{1735--1780}.
\newblock
\bibitem[Hoel et~al\mbox{.}(2013)]%
{hoel2013quantifying}
\bibfield{author}{\bibinfo{person}{Erik~P Hoel}, \bibinfo{person}{Larissa
Albantakis}, {and} \bibinfo{person}{Giulio Tononi}.}
\bibinfo{year}{2013}\natexlab{}.
\newblock \showarticletitle{Quantifying causal emergence shows that macro can
beat micro}.
\newblock \bibinfo{journal}{\emph{Proceedings of the National Academy of
Sciences (PNAS)}} \bibinfo{volume}{110}, \bibinfo{number}{49}
(\bibinfo{year}{2013}), \bibinfo{pages}{19790--19795}.
\newblock
\bibitem[Holland(1992)]%
{holland1992adaptation}
\bibfield{author}{\bibinfo{person}{John~H Holland}.}
\bibinfo{year}{1992}\natexlab{}.
\newblock \bibinfo{booktitle}{\emph{Adaptation in natural and artificial
systems: an introductory analysis with applications to biology, control, and
artificial intelligence}}.
\newblock
\bibitem[Hu et~al\mbox{.}(2022a)]%
{hu2022emergent}
\bibfield{author}{\bibinfo{person}{Jiliang Hu}, \bibinfo{person}{Daniel~R
Amor}, \bibinfo{person}{Matthieu Barbier}, \bibinfo{person}{Guy Bunin}, {and}
\bibinfo{person}{Jeff Gore}.} \bibinfo{year}{2022}\natexlab{a}.
\newblock \showarticletitle{Emergent phases of ecological diversity and
dynamics mapped in microcosms}.
\newblock \bibinfo{journal}{\emph{Science}} \bibinfo{volume}{378},
\bibinfo{number}{6615} (\bibinfo{year}{2022}), \bibinfo{pages}{85--89}.
\newblock
\bibitem[Hu et~al\mbox{.}(2017)]%
{hu2017adaptive}
\bibfield{author}{\bibinfo{person}{Nan Hu}, \bibinfo{person}{Jacob Fish}, {and}
\bibinfo{person}{Colin McAuliffe}.} \bibinfo{year}{2017}\natexlab{}.
\newblock \showarticletitle{An adaptive stochastic inverse solver for
multiscale characterization of composite materials}.
\newblock \bibinfo{journal}{\emph{Internat. J. Numer. Methods Engrg.}}
\bibinfo{volume}{109}, \bibinfo{number}{12} (\bibinfo{year}{2017}),
\bibinfo{pages}{1679--1700}.
\newblock
\bibitem[Hu et~al\mbox{.}(2022b)]%
{hu2022revealing}
\bibfield{author}{\bibinfo{person}{Pipi Hu}, \bibinfo{person}{Wuyue Yang},
\bibinfo{person}{Yi Zhu}, {and} \bibinfo{person}{Liu Hong}.}
\bibinfo{year}{2022}\natexlab{b}.
\newblock \showarticletitle{Revealing hidden dynamics from time-series data by
ODENet}.
\newblock \bibinfo{journal}{\emph{J. Comput. Phys.}} \bibinfo{volume}{461}
(\bibinfo{year}{2022}), \bibinfo{pages}{111203}.
\newblock
\bibitem[Huang et~al\mbox{.}(2018)]%
{huang2018mesoscience}
\bibfield{author}{\bibinfo{person}{Wenlai Huang}, \bibinfo{person}{Jinghai Li},
{and} \bibinfo{person}{Peter~P Edwards}.} \bibinfo{year}{2018}\natexlab{}.
\newblock \showarticletitle{Mesoscience: exploring the common principle at
mesoscales}.
\newblock \bibinfo{journal}{\emph{National science review}}
(\bibinfo{year}{2018}).
\newblock
\bibitem[Hughes(2012)]%
{hughes2012finite}
\bibfield{author}{\bibinfo{person}{Thomas~JR Hughes}.}
\bibinfo{year}{2012}\natexlab{}.
\newblock \bibinfo{booktitle}{\emph{The finite element method: linear static
and dynamic finite element analysis}}.
\newblock \bibinfo{publisher}{Courier Corporation}.
\newblock
\bibitem[Iacopini et~al\mbox{.}(2019)]%
{iacopini2019simplicial}
\bibfield{author}{\bibinfo{person}{Iacopo Iacopini}, \bibinfo{person}{Giovanni
Petri}, \bibinfo{person}{Alain Barrat}, {and} \bibinfo{person}{Vito Latora}.}
\bibinfo{year}{2019}\natexlab{}.
\newblock \showarticletitle{Simplicial models of social contagion}.
\newblock \bibinfo{journal}{\emph{Nature communications}} \bibinfo{volume}{10},
\bibinfo{number}{1} (\bibinfo{year}{2019}), \bibinfo{pages}{1--9}.
\newblock
\bibitem[Iacus et~al\mbox{.}(2020)]%
{iacus2020human}
\bibfield{author}{\bibinfo{person}{Stefano~Maria Iacus},
\bibinfo{person}{Carlos Santamaria}, \bibinfo{person}{Francesco Sermi},
\bibinfo{person}{Spyros Spyratos}, \bibinfo{person}{Dario Tarchi}, {and}
\bibinfo{person}{Michele Vespe}.} \bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{Human mobility and COVID-19 initial dynamics}.
\newblock \bibinfo{journal}{\emph{Nonlinear Dynamics}} \bibinfo{volume}{101},
\bibinfo{number}{3} (\bibinfo{year}{2020}), \bibinfo{pages}{1901--1919}.
\newblock
\bibitem[Ing{\'o}lfsson et~al\mbox{.}(2022)]%
{ingolfsson2022machine}
\bibfield{author}{\bibinfo{person}{Helgi~I Ing{\'o}lfsson},
\bibinfo{person}{Chris Neale}, \bibinfo{person}{Timothy~S Carpenter},
\bibinfo{person}{Rebika Shrestha}, \bibinfo{person}{Cesar~A L{\'o}pez},
\bibinfo{person}{Timothy~H Tran}, \bibinfo{person}{Tomas Oppelstrup},
\bibinfo{person}{Harsh Bhatia}, \bibinfo{person}{Liam~G Stanton},
\bibinfo{person}{Xiaohua Zhang}, {et~al\mbox{.}}}
\bibinfo{year}{2022}\natexlab{}.
\newblock \showarticletitle{Machine learning--driven multiscale modeling
reveals lipid-dependent dynamics of RAS signaling proteins}.
\newblock \bibinfo{journal}{\emph{Proceedings of the National Academy of
Sciences}} \bibinfo{volume}{119}, \bibinfo{number}{1} (\bibinfo{year}{2022}),
\bibinfo{pages}{e2113297119}.
\newblock
\bibitem[Jewell et~al\mbox{.}(2021)]%
{jewell2021s}
\bibfield{author}{\bibinfo{person}{Sean Jewell}, \bibinfo{person}{Joseph
Futoma}, \bibinfo{person}{Lauren Hannah}, \bibinfo{person}{Andrew~C Miller},
\bibinfo{person}{Nicholas~J Foti}, {and} \bibinfo{person}{Emily~B Fox}.}
\bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{It’s complicated: Characterizing the time-varying
relationship between cell phone mobility and COVID-19 spread in the US}.
\newblock \bibinfo{journal}{\emph{NPJ digital medicine}} \bibinfo{volume}{4},
\bibinfo{number}{1} (\bibinfo{year}{2021}), \bibinfo{pages}{1--11}.
\newblock
\bibitem[Jia et~al\mbox{.}(2020)]%
{jia2020population}
\bibfield{author}{\bibinfo{person}{Jayson~S Jia}, \bibinfo{person}{Xin Lu},
\bibinfo{person}{Yun Yuan}, \bibinfo{person}{Ge Xu}, \bibinfo{person}{Jianmin
Jia}, {and} \bibinfo{person}{Nicholas~A Christakis}.}
\bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{Population flow drives spatio-temporal distribution
of COVID-19 in China}.
\newblock \bibinfo{journal}{\emph{Nature}} \bibinfo{volume}{582},
\bibinfo{number}{7812} (\bibinfo{year}{2020}), \bibinfo{pages}{389--394}.
\newblock
\bibitem[Jumper et~al\mbox{.}(2021)]%
{jumper2021highly}
\bibfield{author}{\bibinfo{person}{John Jumper}, \bibinfo{person}{Richard
Evans}, \bibinfo{person}{Alexander Pritzel}, \bibinfo{person}{Tim Green},
\bibinfo{person}{Michael Figurnov}, \bibinfo{person}{Olaf Ronneberger},
\bibinfo{person}{Kathryn Tunyasuvunakool}, \bibinfo{person}{Russ Bates},
\bibinfo{person}{Augustin {\v{Z}}{\'\i}dek}, \bibinfo{person}{Anna
Potapenko}, {et~al\mbox{.}}} \bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{Highly accurate protein structure prediction with
AlphaFold}.
\newblock \bibinfo{journal}{\emph{Nature}} \bibinfo{volume}{596},
\bibinfo{number}{7873} (\bibinfo{year}{2021}), \bibinfo{pages}{583--589}.
\newblock
\bibitem[Kachroo et~al\mbox{.}(2017)]%
{kachroo2017multiscale}
\bibfield{author}{\bibinfo{person}{Pushkin Kachroo}, \bibinfo{person}{Shaurya
Agarwal}, \bibinfo{person}{Benedetto Piccoli}, {and} \bibinfo{person}{Kaan
{\"O}zbay}.} \bibinfo{year}{2017}\natexlab{}.
\newblock \showarticletitle{Multiscale modeling and control architecture for
V2X enabled traffic streams}.
\newblock \bibinfo{journal}{\emph{IEEE Transactions on Vehicular Technology}}
\bibinfo{volume}{66}, \bibinfo{number}{6} (\bibinfo{year}{2017}),
\bibinfo{pages}{4616--4626}.
\newblock
\bibitem[Karniadakis et~al\mbox{.}(2021)]%
{karniadakis2021physics}
\bibfield{author}{\bibinfo{person}{George~Em Karniadakis},
\bibinfo{person}{Ioannis~G Kevrekidis}, \bibinfo{person}{Lu Lu},
\bibinfo{person}{Paris Perdikaris}, \bibinfo{person}{Sifan Wang}, {and}
\bibinfo{person}{Liu Yang}.} \bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{Physics-informed machine learning}.
\newblock \bibinfo{journal}{\emph{Nature Reviews Physics}} \bibinfo{volume}{3},
\bibinfo{number}{6} (\bibinfo{year}{2021}), \bibinfo{pages}{422--440}.
\newblock
\bibitem[Karsenti(2008)]%
{karsenti2008self}
\bibfield{author}{\bibinfo{person}{Eric Karsenti}.}
\bibinfo{year}{2008}\natexlab{}.
\newblock \showarticletitle{Self-organization in cell biology: a brief
history}.
\newblock \bibinfo{journal}{\emph{Nature reviews Molecular cell biology}}
\bibinfo{volume}{9}, \bibinfo{number}{3} (\bibinfo{year}{2008}),
\bibinfo{pages}{255--262}.
\newblock
\bibitem[Kerner et~al\mbox{.}(2002)]%
{kerner2002cellular}
\bibfield{author}{\bibinfo{person}{Boris~S Kerner}, \bibinfo{person}{Sergey~L
Klenov}, {and} \bibinfo{person}{Dietrich~E Wolf}.}
\bibinfo{year}{2002}\natexlab{}.
\newblock \showarticletitle{Cellular automata approach to three-phase traffic
theory}.
\newblock \bibinfo{journal}{\emph{Journal of Physics A: Mathematical and
General}} \bibinfo{volume}{35}, \bibinfo{number}{47} (\bibinfo{year}{2002}),
\bibinfo{pages}{9971}.
\newblock
\bibitem[Kingma and Welling(2013)]%
{kingma2013auto}
\bibfield{author}{\bibinfo{person}{Diederik~P Kingma} {and}
\bibinfo{person}{Max Welling}.} \bibinfo{year}{2013}\natexlab{}.
\newblock \showarticletitle{Auto-encoding variational bayes}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:1312.6114}}
(\bibinfo{year}{2013}).
\newblock
\bibitem[Klus et~al\mbox{.}(2018)]%
{klus2018data}
\bibfield{author}{\bibinfo{person}{Stefan Klus}, \bibinfo{person}{Feliks
N{\"u}ske}, \bibinfo{person}{P{\'e}ter Koltai}, \bibinfo{person}{Hao Wu},
\bibinfo{person}{Ioannis Kevrekidis}, \bibinfo{person}{Christof Sch{\"u}tte},
{and} \bibinfo{person}{Frank No{\'e}}.} \bibinfo{year}{2018}\natexlab{}.
\newblock \showarticletitle{Data-driven model reduction and transfer operator
approximation}.
\newblock \bibinfo{journal}{\emph{Journal of Nonlinear Science}}
\bibinfo{volume}{28}, \bibinfo{number}{3} (\bibinfo{year}{2018}),
\bibinfo{pages}{985--1010}.
\newblock
\bibitem[Koch-Janusz and Ringel(2018)]%
{koch2018mutual}
\bibfield{author}{\bibinfo{person}{Maciej Koch-Janusz} {and}
\bibinfo{person}{Zohar Ringel}.} \bibinfo{year}{2018}\natexlab{}.
\newblock \showarticletitle{Mutual information, neural networks and the
renormalization group}.
\newblock \bibinfo{journal}{\emph{Nature Physics}} \bibinfo{volume}{14},
\bibinfo{number}{6} (\bibinfo{year}{2018}), \bibinfo{pages}{578--582}.
\newblock
\bibitem[Kochkov et~al\mbox{.}(2021)]%
{kochkov2021machine}
\bibfield{author}{\bibinfo{person}{Dmitrii Kochkov}, \bibinfo{person}{Jamie~A
Smith}, \bibinfo{person}{Ayya Alieva}, \bibinfo{person}{Qing Wang},
\bibinfo{person}{Michael~P Brenner}, {and} \bibinfo{person}{Stephan Hoyer}.}
\bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{Machine learning--accelerated computational fluid
dynamics}.
\newblock \bibinfo{journal}{\emph{Proceedings of the National Academy of
Sciences}} \bibinfo{volume}{118}, \bibinfo{number}{21}
(\bibinfo{year}{2021}), \bibinfo{pages}{e2101784118}.
\newblock
\bibitem[Kraemer et~al\mbox{.}(2020)]%
{kraemer2020effect}
\bibfield{author}{\bibinfo{person}{Moritz~UG Kraemer},
\bibinfo{person}{Chia-Hung Yang}, \bibinfo{person}{Bernardo Gutierrez},
\bibinfo{person}{Chieh-Hsi Wu}, \bibinfo{person}{Brennan Klein},
\bibinfo{person}{David~M Pigott}, \bibinfo{person}{Open COVID-19 Data~Working
Group†}, \bibinfo{person}{Louis Du~Plessis}, \bibinfo{person}{Nuno~R
Faria}, \bibinfo{person}{Ruoran Li}, {et~al\mbox{.}}}
\bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{The effect of human mobility and control measures
on the COVID-19 epidemic in China}.
\newblock \bibinfo{journal}{\emph{Science}} \bibinfo{volume}{368},
\bibinfo{number}{6490} (\bibinfo{year}{2020}), \bibinfo{pages}{493--497}.
\newblock
\bibitem[Kutz et~al\mbox{.}(2016)]%
{kutz2016dynamic}
\bibfield{author}{\bibinfo{person}{J~Nathan Kutz}, \bibinfo{person}{Steven~L
Brunton}, \bibinfo{person}{Bingni~W Brunton}, {and} \bibinfo{person}{Joshua~L
Proctor}.} \bibinfo{year}{2016}\natexlab{}.
\newblock \bibinfo{booktitle}{\emph{Dynamic mode decomposition: data-driven
modeling of complex systems}}.
\newblock \bibinfo{publisher}{SIAM}.
\newblock
\bibitem[Lai et~al\mbox{.}(2020)]%
{lai2020effect}
\bibfield{author}{\bibinfo{person}{Shengjie Lai}, \bibinfo{person}{Nick~W
Ruktanonchai}, \bibinfo{person}{Liangcai Zhou}, \bibinfo{person}{Olivia
Prosper}, \bibinfo{person}{Wei Luo}, \bibinfo{person}{Jessica~R Floyd},
\bibinfo{person}{Amy Wesolowski}, \bibinfo{person}{Mauricio Santillana},
\bibinfo{person}{Chi Zhang}, \bibinfo{person}{Xiangjun Du}, {et~al\mbox{.}}}
\bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{Effect of non-pharmaceutical interventions to
contain COVID-19 in China}.
\newblock \bibinfo{journal}{\emph{nature}} \bibinfo{volume}{585},
\bibinfo{number}{7825} (\bibinfo{year}{2020}), \bibinfo{pages}{410--413}.
\newblock
\bibitem[Le et~al\mbox{.}(2015)]%
{le2015computational}
\bibfield{author}{\bibinfo{person}{BA Le}, \bibinfo{person}{Julien Yvonnet},
{and} \bibinfo{person}{Q-C He}.} \bibinfo{year}{2015}\natexlab{}.
\newblock \showarticletitle{Computational homogenization of nonlinear elastic
materials using neural networks}.
\newblock \bibinfo{journal}{\emph{Internat. J. Numer. Methods Engrg.}}
\bibinfo{volume}{104}, \bibinfo{number}{12} (\bibinfo{year}{2015}),
\bibinfo{pages}{1061--1084}.
\newblock
\bibitem[Lee and Carlberg(2020)]%
{lee2020model}
\bibfield{author}{\bibinfo{person}{Kookjin Lee} {and} \bibinfo{person}{Kevin~T
Carlberg}.} \bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{Model reduction of dynamical systems on nonlinear
manifolds using deep convolutional autoencoders}.
\newblock \bibinfo{journal}{\emph{J. Comput. Phys.}} \bibinfo{volume}{404}
(\bibinfo{year}{2020}), \bibinfo{pages}{108973}.
\newblock
\bibitem[Lee et~al\mbox{.}(2020)]%
{lee2020coarse}
\bibfield{author}{\bibinfo{person}{Seungjoon Lee}, \bibinfo{person}{Mahdi
Kooshkbaghi}, \bibinfo{person}{Konstantinos Spiliotis},
\bibinfo{person}{Constantinos~I Siettos}, {and} \bibinfo{person}{Ioannis~G
Kevrekidis}.} \bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{Coarse-scale PDEs from fine-scale observations via
machine learning}.
\newblock \bibinfo{journal}{\emph{Chaos: An Interdisciplinary Journal of
Nonlinear Science}} \bibinfo{volume}{30}, \bibinfo{number}{1}
(\bibinfo{year}{2020}), \bibinfo{pages}{013141}.
\newblock
\bibitem[Leung et~al\mbox{.}(2022)]%
{leung2022nh}
\bibfield{author}{\bibinfo{person}{Wing~Tat Leung}, \bibinfo{person}{Guang
Lin}, {and} \bibinfo{person}{Zecheng Zhang}.}
\bibinfo{year}{2022}\natexlab{}.
\newblock \showarticletitle{NH-PINN: Neural homogenization-based
physics-informed neural network for multiscale problems}.
\newblock \bibinfo{journal}{\emph{J. Comput. Phys.}} \bibinfo{volume}{470}
(\bibinfo{year}{2022}), \bibinfo{pages}{111539}.
\newblock
\bibitem[Levitt and Warshel(1975)]%
{levitt1975computer}
\bibfield{author}{\bibinfo{person}{Michael Levitt} {and} \bibinfo{person}{Arieh
Warshel}.} \bibinfo{year}{1975}\natexlab{}.
\newblock \showarticletitle{Computer simulation of protein folding}.
\newblock \bibinfo{journal}{\emph{Nature}} \bibinfo{volume}{253},
\bibinfo{number}{5494} (\bibinfo{year}{1975}), \bibinfo{pages}{694--698}.
\newblock
\bibitem[Li et~al\mbox{.}(2008)]%
{li2008generalized}
\bibfield{author}{\bibinfo{person}{Aiqin Li}, \bibinfo{person}{Renge Li}, {and}
\bibinfo{person}{Jacob Fish}.} \bibinfo{year}{2008}\natexlab{}.
\newblock \showarticletitle{Generalized mathematical homogenization: from
theory to practice}.
\newblock \bibinfo{journal}{\emph{Computer methods in applied mechanics and
engineering}} \bibinfo{volume}{197}, \bibinfo{number}{41-42}
(\bibinfo{year}{2008}), \bibinfo{pages}{3225--3248}.
\newblock
\bibitem[Li et~al\mbox{.}(2018)]%
{li2018mesoscience}
\bibfield{author}{\bibinfo{person}{Jinghai Li}, \bibinfo{person}{Wenlai Huang},
\bibinfo{person}{Jianhua Chen}, \bibinfo{person}{Wei Ge}, {and}
\bibinfo{person}{Chaofeng Hou}.} \bibinfo{year}{2018}\natexlab{}.
\newblock \showarticletitle{Mesoscience based on the EMMS principle of
compromise in competition}.
\newblock \bibinfo{journal}{\emph{Chemical Engineering Journal}}
\bibinfo{volume}{333} (\bibinfo{year}{2018}), \bibinfo{pages}{327--335}.
\newblock
\bibitem[Li and Kwauk(1994)]%
{li1994particle}
\bibfield{author}{\bibinfo{person}{J Li} {and} \bibinfo{person}{M Kwauk}.}
\bibinfo{year}{1994}\natexlab{}.
\newblock \bibinfo{title}{Particle-Fluid Two-Phase Flow: the
Energy-Minimization Multi-Scale Method; Minimization Multi-Scale Method;
Metallurg}.
\newblock
\newblock
\bibitem[Li and Wang(2018)]%
{li2018neural}
\bibfield{author}{\bibinfo{person}{Shuo-Hui Li} {and} \bibinfo{person}{Lei
Wang}.} \bibinfo{year}{2018}\natexlab{}.
\newblock \showarticletitle{Neural network renormalization group}.
\newblock \bibinfo{journal}{\emph{Physical review letters}}
\bibinfo{volume}{121}, \bibinfo{number}{26} (\bibinfo{year}{2018}),
\bibinfo{pages}{260601}.
\newblock
\bibitem[Liang et~al\mbox{.}(2018)]%
{liang2018deep}
\bibfield{author}{\bibinfo{person}{Liang Liang}, \bibinfo{person}{Minliang
Liu}, \bibinfo{person}{Caitlin Martin}, {and} \bibinfo{person}{Wei Sun}.}
\bibinfo{year}{2018}\natexlab{}.
\newblock \showarticletitle{A deep learning approach to estimate stress
distribution: a fast and accurate surrogate of finite-element analysis}.
\newblock \bibinfo{journal}{\emph{Journal of The Royal Society Interface}}
\bibinfo{volume}{15}, \bibinfo{number}{138} (\bibinfo{year}{2018}),
\bibinfo{pages}{20170844}.
\newblock
\bibitem[Lighthill and Whitham(1955)]%
{lighthill1955kinematic}
\bibfield{author}{\bibinfo{person}{Michael~James Lighthill} {and}
\bibinfo{person}{Gerald~Beresford Whitham}.} \bibinfo{year}{1955}\natexlab{}.
\newblock \showarticletitle{On kinematic waves II. A theory of traffic flow on
long crowded roads}.
\newblock \bibinfo{journal}{\emph{Proceedings of the Royal Society of London.
Series A. Mathematical and Physical Sciences}} \bibinfo{volume}{229},
\bibinfo{number}{1178} (\bibinfo{year}{1955}), \bibinfo{pages}{317--345}.
\newblock
\bibitem[Lino et~al\mbox{.}(2021)]%
{lino2021simulating}
\bibfield{author}{\bibinfo{person}{Mario Lino}, \bibinfo{person}{Chris
Cantwell}, \bibinfo{person}{Anil~A Bharath}, {and} \bibinfo{person}{Stathi
Fotiadis}.} \bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{Simulating continuum mechanics with multi-scale
graph neural networks}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2106.04900}}
(\bibinfo{year}{2021}).
\newblock
\bibitem[Lino et~al\mbox{.}(2022a)]%
{lino2022remus}
\bibfield{author}{\bibinfo{person}{Mario Lino}, \bibinfo{person}{Stati
Fotiadis}, \bibinfo{person}{Anil~A Bharath}, {and} \bibinfo{person}{Chris
Cantwell}.} \bibinfo{year}{2022}\natexlab{a}.
\newblock \showarticletitle{REMuS-GNN: A Rotation-Equivariant Model for
Simulating Continuum Dynamics}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2205.07852}}
(\bibinfo{year}{2022}).
\newblock
\bibitem[Lino et~al\mbox{.}(2022b)]%
{lino2022towards}
\bibfield{author}{\bibinfo{person}{Mario Lino}, \bibinfo{person}{Stathi
Fotiadis}, \bibinfo{person}{Anil~A Bharath}, {and} \bibinfo{person}{Chris
Cantwell}.} \bibinfo{year}{2022}\natexlab{b}.
\newblock \showarticletitle{Towards Fast Simulation of Environmental Fluid
Mechanics with Multi-Scale Graph Neural Networks}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2205.02637}}
(\bibinfo{year}{2022}).
\newblock
\bibitem[Liu et~al\mbox{.}(2022)]%
{liu2022msdr}
\bibfield{author}{\bibinfo{person}{Dachuan Liu}, \bibinfo{person}{Jin Wang},
\bibinfo{person}{Shuo Shang}, {and} \bibinfo{person}{Peng Han}.}
\bibinfo{year}{2022}\natexlab{}.
\newblock \showarticletitle{Msdr: Multi-step dependency relation networks for
spatial temporal forecasting}. In \bibinfo{booktitle}{\emph{Proceedings of
the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining}}.
\bibinfo{pages}{1042--1050}.
\newblock
\bibitem[Liu et~al\mbox{.}(2016)]%
{liu2016self}
\bibfield{author}{\bibinfo{person}{Zeliang Liu}, \bibinfo{person}{MA Bessa},
{and} \bibinfo{person}{Wing~Kam Liu}.} \bibinfo{year}{2016}\natexlab{}.
\newblock \showarticletitle{Self-consistent clustering analysis: an efficient
multi-scale scheme for inelastic heterogeneous materials}.
\newblock \bibinfo{journal}{\emph{Computer Methods in Applied Mechanics and
Engineering}} \bibinfo{volume}{306} (\bibinfo{year}{2016}),
\bibinfo{pages}{319--341}.
\newblock
\bibitem[Liu and Wu(2019)]%
{liu2019exploring}
\bibfield{author}{\bibinfo{person}{Zeliang Liu} {and} \bibinfo{person}{CT Wu}.}
\bibinfo{year}{2019}\natexlab{}.
\newblock \showarticletitle{Exploring the 3D architectures of deep material
network in data-driven multiscale mechanics}.
\newblock \bibinfo{journal}{\emph{Journal of the Mechanics and Physics of
Solids}} \bibinfo{volume}{127} (\bibinfo{year}{2019}),
\bibinfo{pages}{20--46}.
\newblock
\bibitem[Liu et~al\mbox{.}(2019)]%
{liu2019deep}
\bibfield{author}{\bibinfo{person}{Zeliang Liu}, \bibinfo{person}{CT Wu}, {and}
\bibinfo{person}{M391298807188754 Koishi}.} \bibinfo{year}{2019}\natexlab{}.
\newblock \showarticletitle{A deep material network for multiscale topology
learning and accelerated nonlinear modeling of heterogeneous materials}.
\newblock \bibinfo{journal}{\emph{Computer Methods in Applied Mechanics and
Engineering}} \bibinfo{volume}{345} (\bibinfo{year}{2019}),
\bibinfo{pages}{1138--1168}.
\newblock
\bibitem[Lloyd-Smith et~al\mbox{.}(2005)]%
{Lloyd-Smith2005-ao}
\bibfield{author}{\bibinfo{person}{J~O Lloyd-Smith}, \bibinfo{person}{S~J
Schreiber}, \bibinfo{person}{P~E Kopp}, {and} \bibinfo{person}{W~M Getz}.}
\bibinfo{year}{2005}\natexlab{}.
\newblock \showarticletitle{Superspreading and the effect of individual
variation on disease emergence}.
\newblock \bibinfo{journal}{\emph{Nature}} \bibinfo{volume}{438},
\bibinfo{number}{7066} (\bibinfo{date}{Nov.} \bibinfo{year}{2005}),
\bibinfo{pages}{355--359}.
\newblock
\bibitem[Louison et~al\mbox{.}(2021)]%
{louison2021glimps}
\bibfield{author}{\bibinfo{person}{Keverne~A Louison}, \bibinfo{person}{Ian~L
Dryden}, {and} \bibinfo{person}{Charles~A Laughton}.}
\bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{GLIMPS: A Machine Learning Approach to Resolution
Transformation for Multiscale Modeling}.
\newblock \bibinfo{journal}{\emph{Journal of Chemical Theory and Computation}}
\bibinfo{volume}{17}, \bibinfo{number}{12} (\bibinfo{year}{2021}),
\bibinfo{pages}{7930--7937}.
\newblock
\bibitem[Maier and Brockmann(2020)]%
{maier2020effective}
\bibfield{author}{\bibinfo{person}{Benjamin~F Maier} {and}
\bibinfo{person}{Dirk Brockmann}.} \bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{Effective containment explains subexponential
growth in recent confirmed COVID-19 cases in China}.
\newblock \bibinfo{journal}{\emph{Science}} \bibinfo{volume}{368},
\bibinfo{number}{6492} (\bibinfo{year}{2020}), \bibinfo{pages}{742--746}.
\newblock
\bibitem[Mal{\'e} et~al\mbox{.}(2019)]%
{male2019large}
\bibfield{author}{\bibinfo{person}{Quentin Mal{\'e}}, \bibinfo{person}{Gabriel
Staffelbach}, \bibinfo{person}{Olivier Vermorel}, \bibinfo{person}{Antony
Misdariis}, \bibinfo{person}{Fr{\'e}d{\'e}ric Ravet}, {and}
\bibinfo{person}{Thierry Poinsot}.} \bibinfo{year}{2019}\natexlab{}.
\newblock \showarticletitle{Large eddy simulation of pre-chamber ignition in an
internal combustion engine}.
\newblock \bibinfo{journal}{\emph{Flow, Turbulence and Combustion}}
\bibinfo{volume}{103}, \bibinfo{number}{2} (\bibinfo{year}{2019}),
\bibinfo{pages}{465--483}.
\newblock
\bibitem[Mallat(1989)]%
{mallat1989theory}
\bibfield{author}{\bibinfo{person}{Stephane~G Mallat}.}
\bibinfo{year}{1989}\natexlab{}.
\newblock \showarticletitle{A theory for multiresolution signal decomposition:
the wavelet representation}.
\newblock \bibinfo{journal}{\emph{IEEE transactions on pattern analysis and
machine intelligence}} \bibinfo{volume}{11}, \bibinfo{number}{7}
(\bibinfo{year}{1989}), \bibinfo{pages}{674--693}.
\newblock
\bibitem[Manohar et~al\mbox{.}(2019)]%
{manohar2019optimized}
\bibfield{author}{\bibinfo{person}{Krithika Manohar}, \bibinfo{person}{Eurika
Kaiser}, \bibinfo{person}{Steven~L Brunton}, {and} \bibinfo{person}{J~Nathan
Kutz}.} \bibinfo{year}{2019}\natexlab{}.
\newblock \showarticletitle{Optimized sampling for multiscale dynamics}.
\newblock \bibinfo{journal}{\emph{Multiscale Modeling \& Simulation}}
\bibinfo{volume}{17}, \bibinfo{number}{1} (\bibinfo{year}{2019}),
\bibinfo{pages}{117--136}.
\newblock
\bibitem[Masic et~al\mbox{.}(2011)]%
{masic2011observations}
\bibfield{author}{\bibinfo{person}{Admir Masic}, \bibinfo{person}{Luca
Bertinetti}, \bibinfo{person}{Roman Schuetz}, \bibinfo{person}{Leonardo
Galvis}, \bibinfo{person}{Nadya Timofeeva}, \bibinfo{person}{John~WC Dunlop},
\bibinfo{person}{Jong Seto}, \bibinfo{person}{Markus~A Hartmann}, {and}
\bibinfo{person}{Peter Fratzl}.} \bibinfo{year}{2011}\natexlab{}.
\newblock \showarticletitle{Observations of multiscale, stress-induced changes
of collagen orientation in tendon by polarized Raman spectroscopy}.
\newblock \bibinfo{journal}{\emph{Biomacromolecules}} \bibinfo{volume}{12},
\bibinfo{number}{11} (\bibinfo{year}{2011}), \bibinfo{pages}{3989--3996}.
\newblock
\bibitem[Maulik et~al\mbox{.}(2021)]%
{maulik2021reduced}
\bibfield{author}{\bibinfo{person}{Romit Maulik}, \bibinfo{person}{Bethany
Lusch}, {and} \bibinfo{person}{Prasanna Balaprakash}.}
\bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{Reduced-order modeling of advection-dominated
systems with recurrent neural networks and convolutional autoencoders}.
\newblock \bibinfo{journal}{\emph{Physics of Fluids}} \bibinfo{volume}{33},
\bibinfo{number}{3} (\bibinfo{year}{2021}), \bibinfo{pages}{037106}.
\newblock
\bibitem[Maulik et~al\mbox{.}(2020)]%
{maulik2020time}
\bibfield{author}{\bibinfo{person}{Romit Maulik}, \bibinfo{person}{Arvind
Mohan}, \bibinfo{person}{Bethany Lusch}, \bibinfo{person}{Sandeep Madireddy},
\bibinfo{person}{Prasanna Balaprakash}, {and} \bibinfo{person}{Daniel
Livescu}.} \bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{Time-series learning of latent-space dynamics for
reduced-order model closure}.
\newblock \bibinfo{journal}{\emph{Physica D: Nonlinear Phenomena}}
\bibinfo{volume}{405} (\bibinfo{year}{2020}), \bibinfo{pages}{132368}.
\newblock
\bibitem[McAdams et~al\mbox{.}(2020)]%
{mcadams2020incentivising}
\bibfield{author}{\bibinfo{person}{David McAdams},
\bibinfo{person}{Kaci~Kennedy McDade}, \bibinfo{person}{Osondu Ogbuoji},
\bibinfo{person}{Matthew Johnson}, \bibinfo{person}{Siddharth Dixit}, {and}
\bibinfo{person}{Gavin Yamey}.} \bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{Incentivising wealthy nations to participate in the
{COVID-19} {V}accine {G}lobal {A}ccess {F}acility {(COVAX)}: a game theory
perspective}.
\newblock \bibinfo{journal}{\emph{{BMJ} global health}} \bibinfo{volume}{5},
\bibinfo{number}{11} (\bibinfo{year}{2020}), \bibinfo{pages}{e003627}.
\newblock
\bibitem[McLeish and Kolkiewicz(1997)]%
{mcleish1997fitting}
\bibfield{author}{\bibinfo{person}{Don~L McLeish} {and} \bibinfo{person}{Adam~W
Kolkiewicz}.} \bibinfo{year}{1997}\natexlab{}.
\newblock \showarticletitle{Fitting diffusion models in finance}.
\newblock \bibinfo{journal}{\emph{Lecture Notes-Monograph Series}}
(\bibinfo{year}{1997}), \bibinfo{pages}{327--350}.
\newblock
\bibitem[Mehraeen and Chen(2006)]%
{mehraeen2006wavelet}
\bibfield{author}{\bibinfo{person}{Shafigh Mehraeen} {and}
\bibinfo{person}{Jiun-Shyan Chen}.} \bibinfo{year}{2006}\natexlab{}.
\newblock \showarticletitle{Wavelet Galerkin method in multi-scale
homogenization of heterogeneous media}.
\newblock \bibinfo{journal}{\emph{International journal for numerical methods
in engineering}} \bibinfo{volume}{66}, \bibinfo{number}{3}
(\bibinfo{year}{2006}), \bibinfo{pages}{381--403}.
\newblock
\bibitem[Mehta and Schwab(2014)]%
{mehta2014exact}
\bibfield{author}{\bibinfo{person}{Pankaj Mehta} {and} \bibinfo{person}{David~J
Schwab}.} \bibinfo{year}{2014}\natexlab{}.
\newblock \showarticletitle{An exact mapping between the variational
renormalization group and deep learning}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:1410.3831}}
(\bibinfo{year}{2014}).
\newblock
\bibitem[Miehe and Bayreuther(2007)]%
{miehe2007multiscale}
\bibfield{author}{\bibinfo{person}{Christian Miehe} {and}
\bibinfo{person}{CG2344771 Bayreuther}.} \bibinfo{year}{2007}\natexlab{}.
\newblock \showarticletitle{On multiscale FE analyses of heterogeneous
structures: from homogenization to multigrid solvers}.
\newblock \bibinfo{journal}{\emph{Internat. J. Numer. Methods Engrg.}}
\bibinfo{volume}{71}, \bibinfo{number}{10} (\bibinfo{year}{2007}),
\bibinfo{pages}{1135--1180}.
\newblock
\bibitem[Mikolov et~al\mbox{.}(2010)]%
{mikolov2010recurrent}
\bibfield{author}{\bibinfo{person}{Tomas Mikolov}, \bibinfo{person}{Martin
Karafi{\'a}t}, \bibinfo{person}{Lukas Burget}, \bibinfo{person}{Jan
Cernock{\`y}}, {and} \bibinfo{person}{Sanjeev Khudanpur}.}
\bibinfo{year}{2010}\natexlab{}.
\newblock \showarticletitle{Recurrent neural network based language model.}. In
\bibinfo{booktitle}{\emph{Interspeech}}, Vol.~\bibinfo{volume}{2}. Makuhari,
\bibinfo{pages}{1045--1048}.
\newblock
\bibitem[Mnih et~al\mbox{.}(2013)]%
{mnih2013playing}
\bibfield{author}{\bibinfo{person}{Volodymyr Mnih}, \bibinfo{person}{Koray
Kavukcuoglu}, \bibinfo{person}{David Silver}, \bibinfo{person}{Alex Graves},
\bibinfo{person}{Ioannis Antonoglou}, \bibinfo{person}{Daan Wierstra}, {and}
\bibinfo{person}{Martin Riedmiller}.} \bibinfo{year}{2013}\natexlab{}.
\newblock \showarticletitle{Playing atari with deep reinforcement learning}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:1312.5602}}
(\bibinfo{year}{2013}).
\newblock
\bibitem[Mo et~al\mbox{.}(2020)]%
{mo2020analysis}
\bibfield{author}{\bibinfo{person}{Yi Mo}, \bibinfo{person}{Mengjie Du},
\bibinfo{person}{Wei Ge}, {and} \bibinfo{person}{Pingwen Zhang}.}
\bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{Analysis of the energy-minimization multiscale
model with multiobjective optimization}.
\newblock \bibinfo{journal}{\emph{Particuology}} \bibinfo{volume}{48}
(\bibinfo{year}{2020}), \bibinfo{pages}{109--115}.
\newblock
\bibitem[Mosser et~al\mbox{.}(2017)]%
{mosser2017reconstruction}
\bibfield{author}{\bibinfo{person}{Lukas Mosser}, \bibinfo{person}{Olivier
Dubrule}, {and} \bibinfo{person}{Martin~J Blunt}.}
\bibinfo{year}{2017}\natexlab{}.
\newblock \showarticletitle{Reconstruction of three-dimensional porous media
using generative adversarial neural networks}.
\newblock \bibinfo{journal}{\emph{Physical Review E}} \bibinfo{volume}{96},
\bibinfo{number}{4} (\bibinfo{year}{2017}), \bibinfo{pages}{043309}.
\newblock
\bibitem[Murphy et~al\mbox{.}(2021)]%
{murphy2021deep}
\bibfield{author}{\bibinfo{person}{Charles Murphy}, \bibinfo{person}{Edward
Laurence}, {and} \bibinfo{person}{Antoine Allard}.}
\bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{Deep learning of contagion dynamics on complex
networks}.
\newblock \bibinfo{journal}{\emph{Nature Communications}} \bibinfo{volume}{12},
\bibinfo{number}{1} (\bibinfo{year}{2021}), \bibinfo{pages}{1--11}.
\newblock
\bibitem[Nakayama et~al\mbox{.}(2001)]%
{nakayama2001effect}
\bibfield{author}{\bibinfo{person}{Akihiro Nakayama}, \bibinfo{person}{Y{\=u}ki
Sugiyama}, {and} \bibinfo{person}{Katsuya Hasebe}.}
\bibinfo{year}{2001}\natexlab{}.
\newblock \showarticletitle{Effect of looking at the car that follows in an
optimal velocity model of traffic flow}.
\newblock \bibinfo{journal}{\emph{Physical Review E}} \bibinfo{volume}{65},
\bibinfo{number}{1} (\bibinfo{year}{2001}), \bibinfo{pages}{016112}.
\newblock
\bibitem[Nauenberg and Nienhuis(1974)]%
{nauenberg1974renormalization}
\bibfield{author}{\bibinfo{person}{M Nauenberg} {and} \bibinfo{person}{B
Nienhuis}.} \bibinfo{year}{1974}\natexlab{}.
\newblock \showarticletitle{Renormalization-group approach to the solution of
general Ising models}.
\newblock \bibinfo{journal}{\emph{Physical Review Letters}}
\bibinfo{volume}{33}, \bibinfo{number}{27} (\bibinfo{year}{1974}),
\bibinfo{pages}{1598}.
\newblock
\bibitem[Newell(1961)]%
{newell1961nonlinear}
\bibfield{author}{\bibinfo{person}{Gordon~Frank Newell}.}
\bibinfo{year}{1961}\natexlab{}.
\newblock \showarticletitle{Nonlinear effects in the dynamics of car
following}.
\newblock \bibinfo{journal}{\emph{Operations research}} \bibinfo{volume}{9},
\bibinfo{number}{2} (\bibinfo{year}{1961}), \bibinfo{pages}{209--229}.
\newblock
\bibitem[Nishi et~al\mbox{.}(2020)]%
{nishi2020network}
\bibfield{author}{\bibinfo{person}{Akihiro Nishi}, \bibinfo{person}{George
Dewey}, \bibinfo{person}{Akira Endo}, \bibinfo{person}{Sophia Neman},
\bibinfo{person}{Sage~K Iwamoto}, \bibinfo{person}{Michael~Y Ni},
\bibinfo{person}{Yusuke Tsugawa}, \bibinfo{person}{Georgios Iosifidis},
\bibinfo{person}{Justin~D Smith}, {and} \bibinfo{person}{Sean~D Young}.}
\bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{Network interventions for managing the COVID-19
pandemic and sustaining economy}.
\newblock \bibinfo{journal}{\emph{Proceedings of the National Academy of
Sciences}} \bibinfo{volume}{117}, \bibinfo{number}{48}
(\bibinfo{year}{2020}), \bibinfo{pages}{30285--30294}.
\newblock
\bibitem[No{\'e} and Clementi(2017)]%
{noe2017collective}
\bibfield{author}{\bibinfo{person}{Frank No{\'e}} {and}
\bibinfo{person}{Cecilia Clementi}.} \bibinfo{year}{2017}\natexlab{}.
\newblock \showarticletitle{Collective variables for the study of long-time
kinetics from molecular trajectories: theory and methods}.
\newblock \bibinfo{journal}{\emph{Current opinion in structural biology}}
\bibinfo{volume}{43} (\bibinfo{year}{2017}), \bibinfo{pages}{141--147}.
\newblock
\bibitem[No{\'e} and Nuske(2013)]%
{noe2013variational}
\bibfield{author}{\bibinfo{person}{Frank No{\'e}} {and} \bibinfo{person}{Feliks
Nuske}.} \bibinfo{year}{2013}\natexlab{}.
\newblock \showarticletitle{A variational approach to modeling slow processes
in stochastic dynamical systems}.
\newblock \bibinfo{journal}{\emph{Multiscale Modeling \& Simulation}}
\bibinfo{volume}{11}, \bibinfo{number}{2} (\bibinfo{year}{2013}),
\bibinfo{pages}{635--655}.
\newblock
\bibitem[Oishi and Yagawa(2017)]%
{oishi2017computational}
\bibfield{author}{\bibinfo{person}{Atsuya Oishi} {and} \bibinfo{person}{Genki
Yagawa}.} \bibinfo{year}{2017}\natexlab{}.
\newblock \showarticletitle{Computational mechanics enhanced by deep learning}.
\newblock \bibinfo{journal}{\emph{Computer Methods in Applied Mechanics and
Engineering}} \bibinfo{volume}{327} (\bibinfo{year}{2017}),
\bibinfo{pages}{327--351}.
\newblock
\bibitem[{\"O}zdemir et~al\mbox{.}(2008)]%
{ozdemir2008computational}
\bibfield{author}{\bibinfo{person}{I {\"O}zdemir}, \bibinfo{person}{WAM
Brekelmans}, {and} \bibinfo{person}{MGD2376223 Geers}.}
\bibinfo{year}{2008}\natexlab{}.
\newblock \showarticletitle{Computational homogenization for heat conduction in
heterogeneous solids}.
\newblock \bibinfo{journal}{\emph{International journal for numerical methods
in engineering}} \bibinfo{volume}{73}, \bibinfo{number}{2}
(\bibinfo{year}{2008}), \bibinfo{pages}{185--204}.
\newblock
\bibitem[Peirlinck et~al\mbox{.}(2019)]%
{peirlinck2019using}
\bibfield{author}{\bibinfo{person}{Mathias Peirlinck}, \bibinfo{person}{F
Sahli~Costabal}, \bibinfo{person}{KL Sack}, \bibinfo{person}{JS Choy},
\bibinfo{person}{GS Kassab}, \bibinfo{person}{JM Guccione},
\bibinfo{person}{M De~Beule}, \bibinfo{person}{Patrick Segers}, {and}
\bibinfo{person}{E Kuhl}.} \bibinfo{year}{2019}\natexlab{}.
\newblock \showarticletitle{Using machine learning to characterize heart
failure across the scales}.
\newblock \bibinfo{journal}{\emph{Biomechanics and modeling in mechanobiology}}
\bibinfo{volume}{18}, \bibinfo{number}{6} (\bibinfo{year}{2019}),
\bibinfo{pages}{1987--2001}.
\newblock
\bibitem[Peng and Sun(2010)]%
{peng2010dynamical}
\bibfield{author}{\bibinfo{person}{GH Peng} {and} \bibinfo{person}{DH Sun}.}
\bibinfo{year}{2010}\natexlab{}.
\newblock \showarticletitle{A dynamical model of car-following with the
consideration of the multiple information of preceding cars}.
\newblock \bibinfo{journal}{\emph{Physics Letters A}} \bibinfo{volume}{374},
\bibinfo{number}{15-16} (\bibinfo{year}{2010}), \bibinfo{pages}{1694--1698}.
\newblock
\bibitem[Peng et~al\mbox{.}(2021)]%
{peng2021multiscale}
\bibfield{author}{\bibinfo{person}{Grace~CY Peng}, \bibinfo{person}{Mark
Alber}, \bibinfo{person}{Adrian Buganza~Tepole}, \bibinfo{person}{William~R
Cannon}, \bibinfo{person}{Suvranu De}, \bibinfo{person}{Savador Dura-Bernal},
\bibinfo{person}{Krishna Garikipati}, \bibinfo{person}{George Karniadakis},
\bibinfo{person}{William~W Lytton}, \bibinfo{person}{Paris Perdikaris},
{et~al\mbox{.}}} \bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{Multiscale modeling meets machine learning: What
can we learn?}
\newblock \bibinfo{journal}{\emph{Archives of Computational Methods in
Engineering}} \bibinfo{volume}{28}, \bibinfo{number}{3}
(\bibinfo{year}{2021}), \bibinfo{pages}{1017--1037}.
\newblock
\bibitem[P{\'e}rez-Hern{\'a}ndez et~al\mbox{.}(2013)]%
{perez2013identification}
\bibfield{author}{\bibinfo{person}{Guillermo P{\'e}rez-Hern{\'a}ndez},
\bibinfo{person}{Fabian Paul}, \bibinfo{person}{Toni Giorgino},
\bibinfo{person}{Gianni De~Fabritiis}, {and} \bibinfo{person}{Frank
No{\'e}}.} \bibinfo{year}{2013}\natexlab{}.
\newblock \showarticletitle{Identification of slow molecular order parameters
for Markov model construction}.
\newblock \bibinfo{journal}{\emph{The Journal of chemical physics}}
\bibinfo{volume}{139}, \bibinfo{number}{1} (\bibinfo{year}{2013}),
\bibinfo{pages}{07B604\_1}.
\newblock
\bibitem[Pipes(1953)]%
{pipes1953operational}
\bibfield{author}{\bibinfo{person}{Louis~A Pipes}.}
\bibinfo{year}{1953}\natexlab{}.
\newblock \showarticletitle{An operational analysis of traffic dynamics}.
\newblock \bibinfo{journal}{\emph{Journal of applied physics}}
\bibinfo{volume}{24}, \bibinfo{number}{3} (\bibinfo{year}{1953}),
\bibinfo{pages}{274--281}.
\newblock
\bibitem[Proctor et~al\mbox{.}(2016)]%
{proctor2016dynamic}
\bibfield{author}{\bibinfo{person}{Joshua~L Proctor}, \bibinfo{person}{Steven~L
Brunton}, {and} \bibinfo{person}{J~Nathan Kutz}.}
\bibinfo{year}{2016}\natexlab{}.
\newblock \showarticletitle{Dynamic mode decomposition with control}.
\newblock \bibinfo{journal}{\emph{SIAM Journal on Applied Dynamical Systems}}
\bibinfo{volume}{15}, \bibinfo{number}{1} (\bibinfo{year}{2016}),
\bibinfo{pages}{142--161}.
\newblock
\bibitem[Qian et~al\mbox{.}(2021)]%
{qian2021scaling}
\bibfield{author}{\bibinfo{person}{Xinwu Qian}, \bibinfo{person}{Lijun Sun},
{and} \bibinfo{person}{Satish~V Ukkusuri}.} \bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{Scaling of contact networks for epidemic spreading
in urban transit systems}.
\newblock \bibinfo{journal}{\emph{Scientific reports}} \bibinfo{volume}{11},
\bibinfo{number}{1} (\bibinfo{year}{2021}), \bibinfo{pages}{1--12}.
\newblock
\bibitem[Rocha et~al\mbox{.}(2021)]%
{rocha2021fly}
\bibfield{author}{\bibinfo{person}{IBCM Rocha}, \bibinfo{person}{Pierre
Kerfriden}, {and} \bibinfo{person}{FP van~der Meer}.}
\bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{On-the-fly construction of surrogate constitutive
models for concurrent multiscale mechanical analysis through probabilistic
machine learning}.
\newblock \bibinfo{journal}{\emph{Journal of Computational Physics: X}}
\bibinfo{volume}{9} (\bibinfo{year}{2021}), \bibinfo{pages}{100083}.
\newblock
\bibitem[Ross(1916)]%
{ross1916application}
\bibfield{author}{\bibinfo{person}{Ronald Ross}.}
\bibinfo{year}{1916}\natexlab{}.
\newblock \showarticletitle{An application of the theory of probabilities to
the study of a priori pathometry.—Part I}.
\newblock \bibinfo{journal}{\emph{Proceedings of the Royal Society of London.
Series A, Containing papers of a mathematical and physical character}}
\bibinfo{volume}{92}, \bibinfo{number}{638} (\bibinfo{year}{1916}),
\bibinfo{pages}{204--230}.
\newblock
\bibitem[R{\"o}ssler(1976)]%
{rossler1976chemical}
\bibfield{author}{\bibinfo{person}{Otto~E R{\"o}ssler}.}
\bibinfo{year}{1976}\natexlab{}.
\newblock \showarticletitle{Chemical turbulence: chaos in a simple
reaction-diffusion system}.
\newblock \bibinfo{journal}{\emph{Zeitschrift f{\"u}r Naturforschung A}}
\bibinfo{volume}{31}, \bibinfo{number}{10} (\bibinfo{year}{1976}),
\bibinfo{pages}{1168--1172}.
\newblock
\bibitem[Rotesi et~al\mbox{.}(2021)]%
{rotesi2021national}
\bibfield{author}{\bibinfo{person}{Tiziano Rotesi}, \bibinfo{person}{Paolo
Pin}, \bibinfo{person}{Maria Cucciniello}, \bibinfo{person}{Amyn~A Malik},
\bibinfo{person}{Elliott~E Paintsil}, \bibinfo{person}{Scott~E Bokemper},
\bibinfo{person}{Kathryn Willebrand}, \bibinfo{person}{Gregory~A Huber},
\bibinfo{person}{Alessia Melegaro}, {and} \bibinfo{person}{Saad~B Omer}.}
\bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{National interest may require distributing
{COVID-19} vaccines to other countries}.
\newblock \bibinfo{journal}{\emph{Scientific reports}} \bibinfo{volume}{11},
\bibinfo{number}{1} (\bibinfo{year}{2021}), \bibinfo{pages}{1--8}.
\newblock
\bibitem[Saito et~al\mbox{.}(2008)]%
{saito2008prediction}
\bibfield{author}{\bibinfo{person}{Kazumi Saito}, \bibinfo{person}{Ryohei
Nakano}, {and} \bibinfo{person}{Masahiro Kimura}.}
\bibinfo{year}{2008}\natexlab{}.
\newblock \showarticletitle{Prediction of information diffusion probabilities
for independent cascade model}. In \bibinfo{booktitle}{\emph{International
conference on knowledge-based and intelligent information and engineering
systems}}. \bibinfo{pages}{67--75}.
\newblock
\bibitem[Sanchez-Gonzalez et~al\mbox{.}(2020)]%
{sanchez2020learning}
\bibfield{author}{\bibinfo{person}{Alvaro Sanchez-Gonzalez},
\bibinfo{person}{Jonathan Godwin}, \bibinfo{person}{Tobias Pfaff},
\bibinfo{person}{Rex Ying}, \bibinfo{person}{Jure Leskovec}, {and}
\bibinfo{person}{Peter Battaglia}.} \bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{Learning to simulate complex physics with graph
networks}. In \bibinfo{booktitle}{\emph{International Conference on Machine
Learning}}. \bibinfo{pages}{8459--8468}.
\newblock
\bibitem[Schmid(2010)]%
{schmid2010dynamic}
\bibfield{author}{\bibinfo{person}{Peter~J Schmid}.}
\bibinfo{year}{2010}\natexlab{}.
\newblock \showarticletitle{Dynamic mode decomposition of numerical and
experimental data}.
\newblock \bibinfo{journal}{\emph{Journal of fluid mechanics}}
\bibinfo{volume}{656} (\bibinfo{year}{2010}), \bibinfo{pages}{5--28}.
\newblock
\bibitem[Schmid et~al\mbox{.}(2011)]%
{schmid2011applications}
\bibfield{author}{\bibinfo{person}{Peter~J Schmid}, \bibinfo{person}{Larry Li},
\bibinfo{person}{Matthew~P Juniper}, {and} \bibinfo{person}{O Pust}.}
\bibinfo{year}{2011}\natexlab{}.
\newblock \showarticletitle{Applications of the dynamic mode decomposition}.
\newblock \bibinfo{journal}{\emph{Theoretical and Computational Fluid
Dynamics}} \bibinfo{volume}{25}, \bibinfo{number}{1} (\bibinfo{year}{2011}),
\bibinfo{pages}{249--259}.
\newblock
\bibitem[Singer et~al\mbox{.}(2009)]%
{singer2009detecting}
\bibfield{author}{\bibinfo{person}{Amit Singer}, \bibinfo{person}{Radek Erban},
\bibinfo{person}{Ioannis~G Kevrekidis}, {and} \bibinfo{person}{Ronald~R
Coifman}.} \bibinfo{year}{2009}\natexlab{}.
\newblock \showarticletitle{Detecting intrinsic slow variables in stochastic
dynamical systems by anisotropic diffusion maps}.
\newblock \bibinfo{journal}{\emph{Proceedings of the National Academy of
Sciences (PNAS)}} \bibinfo{volume}{106}, \bibinfo{number}{38}
(\bibinfo{year}{2009}), \bibinfo{pages}{16090--16095}.
\newblock
\bibitem[Snoeijer et~al\mbox{.}(2021)]%
{snoeijer2021measuring}
\bibfield{author}{\bibinfo{person}{Berber~T Snoeijer}, \bibinfo{person}{Mariska
Burger}, \bibinfo{person}{Shaoxiong Sun}, \bibinfo{person}{Richard~JB
Dobson}, {and} \bibinfo{person}{Amos~A Folarin}.}
\bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{Measuring the effect of Non-Pharmaceutical
Interventions (NPIs) on mobility during the COVID-19 pandemic using global
mobility data}.
\newblock \bibinfo{journal}{\emph{NPJ digital medicine}} \bibinfo{volume}{4},
\bibinfo{number}{1} (\bibinfo{year}{2021}), \bibinfo{pages}{1--12}.
\newblock
\bibitem[Spelta et~al\mbox{.}(2020)]%
{spelta2020after}
\bibfield{author}{\bibinfo{person}{Alessandro Spelta}, \bibinfo{person}{Andrea
Flori}, \bibinfo{person}{Francesco Pierri}, \bibinfo{person}{Giovanni
Bonaccorsi}, {and} \bibinfo{person}{Fabio Pammolli}.}
\bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{After the lockdown: simulating mobility, public
health and economic recovery scenarios}.
\newblock \bibinfo{journal}{\emph{Scientific reports}} \bibinfo{volume}{10},
\bibinfo{number}{1} (\bibinfo{year}{2020}), \bibinfo{pages}{1--13}.
\newblock
\bibitem[Sridhar et~al\mbox{.}(2016)]%
{sridhar2016homogenization}
\bibfield{author}{\bibinfo{person}{Ashwin Sridhar}, \bibinfo{person}{Varvara~G
Kouznetsova}, {and} \bibinfo{person}{Marc~GD Geers}.}
\bibinfo{year}{2016}\natexlab{}.
\newblock \showarticletitle{Homogenization of locally resonant acoustic
metamaterials towards an emergent enriched continuum}.
\newblock \bibinfo{journal}{\emph{Computational mechanics}}
\bibinfo{volume}{57}, \bibinfo{number}{3} (\bibinfo{year}{2016}),
\bibinfo{pages}{423--435}.
\newblock
\bibitem[St-Onge et~al\mbox{.}(2021)]%
{st2021universal}
\bibfield{author}{\bibinfo{person}{Guillaume St-Onge}, \bibinfo{person}{Hanlin
Sun}, \bibinfo{person}{Antoine Allard}, \bibinfo{person}{Laurent
H{\'e}bert-Dufresne}, {and} \bibinfo{person}{Ginestra Bianconi}.}
\bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{Universal nonlinear infection kernel from
heterogeneous exposure on higher-order networks}.
\newblock \bibinfo{journal}{\emph{Physical Review Letters}}
\bibinfo{volume}{127}, \bibinfo{number}{15} (\bibinfo{year}{2021}),
\bibinfo{pages}{158301}.
\newblock
\bibitem[Stansfeld and Sansom(2011)]%
{stansfeld2011coarse}
\bibfield{author}{\bibinfo{person}{Phillip~J Stansfeld} {and}
\bibinfo{person}{Mark~SP Sansom}.} \bibinfo{year}{2011}\natexlab{}.
\newblock \showarticletitle{From coarse grained to atomistic: a serial
multiscale approach to membrane protein simulations}.
\newblock \bibinfo{journal}{\emph{Journal of chemical theory and computation}}
\bibinfo{volume}{7}, \bibinfo{number}{4} (\bibinfo{year}{2011}),
\bibinfo{pages}{1157--1166}.
\newblock
\bibitem[Tao et~al\mbox{.}(2008)]%
{tao2008synchronized}
\bibfield{author}{\bibinfo{person}{Chen Tao}, \bibinfo{person}{Jia Bin},
\bibinfo{person}{Li Xin-Gang}, \bibinfo{person}{Jiang Rui}, {and}
\bibinfo{person}{Gao Zi-You}.} \bibinfo{year}{2008}\natexlab{}.
\newblock \showarticletitle{Synchronized flow in a cellular automaton model
with time headway dependent randomization}.
\newblock \bibinfo{journal}{\emph{Chinese Physics Letters}}
\bibinfo{volume}{25}, \bibinfo{number}{8} (\bibinfo{year}{2008}),
\bibinfo{pages}{2795}.
\newblock
\bibitem[Tao et~al\mbox{.}(2010)]%
{tao2010nonintrusive}
\bibfield{author}{\bibinfo{person}{Molei Tao}, \bibinfo{person}{Houman Owhadi},
{and} \bibinfo{person}{Jerrold~E Marsden}.} \bibinfo{year}{2010}\natexlab{}.
\newblock \showarticletitle{Nonintrusive and structure preserving multiscale
integration of stiff ODEs, SDEs, and Hamiltonian systems with hidden slow
dynamics via flow averaging}.
\newblock \bibinfo{journal}{\emph{Multiscale Modeling \& Simulation}}
\bibinfo{volume}{8}, \bibinfo{number}{4} (\bibinfo{year}{2010}),
\bibinfo{pages}{1269--1324}.
\newblock
\bibitem[Treiber et~al\mbox{.}(2000)]%
{treiber2000congested}
\bibfield{author}{\bibinfo{person}{Martin Treiber}, \bibinfo{person}{Ansgar
Hennecke}, {and} \bibinfo{person}{Dirk Helbing}.}
\bibinfo{year}{2000}\natexlab{}.
\newblock \showarticletitle{Congested traffic states in empirical observations
and microscopic simulations}.
\newblock \bibinfo{journal}{\emph{Physical review E}} \bibinfo{volume}{62},
\bibinfo{number}{2} (\bibinfo{year}{2000}), \bibinfo{pages}{1805}.
\newblock
\bibitem[Tu(2013)]%
{tu2013dynamic}
\bibfield{author}{\bibinfo{person}{Jonathan~H Tu}.}
\bibinfo{year}{2013}\natexlab{}.
\newblock \emph{\bibinfo{title}{Dynamic mode decomposition: Theory and
applications}}.
\newblock \bibinfo{thesistype}{Ph.\,D. Dissertation}.
\bibinfo{school}{Princeton University}.
\newblock
\bibitem[Vlachas et~al\mbox{.}(2022)]%
{vlachas2022multiscale}
\bibfield{author}{\bibinfo{person}{Pantelis~R Vlachas},
\bibinfo{person}{Georgios Arampatzis}, \bibinfo{person}{Caroline Uhler},
{and} \bibinfo{person}{Petros Koumoutsakos}.}
\bibinfo{year}{2022}\natexlab{}.
\newblock \showarticletitle{Multiscale simulations of complex systems by
learning their effective dynamics}.
\newblock \bibinfo{journal}{\emph{Nature Machine Intelligence}}
\bibinfo{volume}{4}, \bibinfo{number}{4} (\bibinfo{year}{2022}),
\bibinfo{pages}{359--366}.
\newblock
\bibitem[Vlachas et~al\mbox{.}(2021)]%
{vlachas2021accelerated}
\bibfield{author}{\bibinfo{person}{Pantelis~R Vlachas}, \bibinfo{person}{Julija
Zavadlav}, \bibinfo{person}{Matej Praprotnik}, {and} \bibinfo{person}{Petros
Koumoutsakos}.} \bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{Accelerated Simulations of Molecular Systems
through Learning of Effective Dynamics}.
\newblock \bibinfo{journal}{\emph{Journal of Chemical Theory and Computation}}
\bibinfo{volume}{18}, \bibinfo{number}{1} (\bibinfo{year}{2021}),
\bibinfo{pages}{538--549}.
\newblock
\bibitem[Wagner et~al\mbox{.}(2021)]%
{wagner2021vaccine}
\bibfield{author}{\bibinfo{person}{Caroline~E Wagner}, \bibinfo{person}{Chadi~M
Saad-Roy}, \bibinfo{person}{Sinead~E Morris}, \bibinfo{person}{Rachel~E
Baker}, \bibinfo{person}{Michael~J Mina}, \bibinfo{person}{Jeremy Farrar},
\bibinfo{person}{Edward~C Holmes}, \bibinfo{person}{Oliver~G Pybus},
\bibinfo{person}{Andrea~L Graham}, \bibinfo{person}{Ezekiel~J Emanuel},
{et~al\mbox{.}}} \bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{Vaccine nationalism and the dynamics and control of
{SARS-CoV-2}}.
\newblock \bibinfo{journal}{\emph{Science}} \bibinfo{volume}{373},
\bibinfo{number}{6562} (\bibinfo{year}{2021}), \bibinfo{pages}{1--8}.
\newblock
\bibitem[Wang and Sun(2019)]%
{wang2019meta}
\bibfield{author}{\bibinfo{person}{Kun Wang} {and} \bibinfo{person}{WaiChing
Sun}.} \bibinfo{year}{2019}\natexlab{}.
\newblock \showarticletitle{Meta-modeling game for deriving theory-consistent,
microstructure-based traction--separation laws via deep reinforcement
learning}.
\newblock \bibinfo{journal}{\emph{Computer Methods in Applied Mechanics and
Engineering}} \bibinfo{volume}{346} (\bibinfo{year}{2019}),
\bibinfo{pages}{216--241}.
\newblock
\bibitem[Wang and G{\'o}mez-Bombarelli(2019)]%
{wang2019coarse}
\bibfield{author}{\bibinfo{person}{Wujie Wang} {and} \bibinfo{person}{Rafael
G{\'o}mez-Bombarelli}.} \bibinfo{year}{2019}\natexlab{}.
\newblock \showarticletitle{Coarse-graining auto-encoders for molecular
dynamics}.
\newblock \bibinfo{journal}{\emph{npj Computational Materials}}
\bibinfo{volume}{5}, \bibinfo{number}{1} (\bibinfo{year}{2019}),
\bibinfo{pages}{1--9}.
\newblock
\bibitem[Wang et~al\mbox{.}(2020)]%
{wang2020deep}
\bibfield{author}{\bibinfo{person}{Yating Wang}, \bibinfo{person}{Siu~Wun
Cheung}, \bibinfo{person}{Eric~T Chung}, \bibinfo{person}{Yalchin Efendiev},
{and} \bibinfo{person}{Min Wang}.} \bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{Deep multiscale model learning}.
\newblock \bibinfo{journal}{\emph{J. Comput. Phys.}} \bibinfo{volume}{406}
(\bibinfo{year}{2020}), \bibinfo{pages}{109071}.
\newblock
\bibitem[Watts and Strogatz(1998)]%
{watts1998collective}
\bibfield{author}{\bibinfo{person}{Duncan~J Watts} {and}
\bibinfo{person}{Steven~H Strogatz}.} \bibinfo{year}{1998}\natexlab{}.
\newblock \showarticletitle{Collective dynamics of ‘small-world’networks}.
\newblock \bibinfo{journal}{\emph{nature}} \bibinfo{volume}{393},
\bibinfo{number}{6684} (\bibinfo{year}{1998}), \bibinfo{pages}{440--442}.
\newblock
\bibitem[Wehmeyer and No{\'e}(2018)]%
{wehmeyer2018time}
\bibfield{author}{\bibinfo{person}{Christoph Wehmeyer} {and}
\bibinfo{person}{Frank No{\'e}}.} \bibinfo{year}{2018}\natexlab{}.
\newblock \showarticletitle{Time-lagged autoencoders: Deep learning of slow
collective variables for molecular kinetics}.
\newblock \bibinfo{journal}{\emph{The Journal of chemical physics}}
\bibinfo{volume}{148}, \bibinfo{number}{24} (\bibinfo{year}{2018}),
\bibinfo{pages}{241703}.
\newblock
\bibitem[Weinan(2011)]%
{weinan2011principles}
\bibfield{author}{\bibinfo{person}{E Weinan}.} \bibinfo{year}{2011}\natexlab{}.
\newblock \bibinfo{booktitle}{\emph{Principles of multiscale modeling}}.
\newblock \bibinfo{publisher}{Cambridge University Press}.
\newblock
\bibitem[West et~al\mbox{.}(1999)]%
{west1999fourth}
\bibfield{author}{\bibinfo{person}{Geoffrey~B West}, \bibinfo{person}{James~H
Brown}, {and} \bibinfo{person}{Brian~J Enquist}.}
\bibinfo{year}{1999}\natexlab{}.
\newblock \showarticletitle{The fourth dimension of life: fractal geometry and
allometric scaling of organisms}.
\newblock \bibinfo{journal}{\emph{science}} \bibinfo{volume}{284},
\bibinfo{number}{5420} (\bibinfo{year}{1999}), \bibinfo{pages}{1677--1679}.
\newblock
\bibitem[Wu et~al\mbox{.}(2021)]%
{wu2021machine}
\bibfield{author}{\bibinfo{person}{Chi Wu}, \bibinfo{person}{Ali Entezari},
\bibinfo{person}{Keke Zheng}, \bibinfo{person}{Jianguang Fang},
\bibinfo{person}{Hala Zreiqat}, \bibinfo{person}{Grant~P Steven},
\bibinfo{person}{Michael~V Swain}, {and} \bibinfo{person}{Qing Li}.}
\bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{A machine learning-based multiscale model to
predict bone formation in scaffolds}.
\newblock \bibinfo{journal}{\emph{Nature Computational Science}}
\bibinfo{volume}{1}, \bibinfo{number}{8} (\bibinfo{year}{2021}),
\bibinfo{pages}{532--541}.
\newblock
\bibitem[Wu et~al\mbox{.}(2020b)]%
{wu2020nowcasting}
\bibfield{author}{\bibinfo{person}{Joseph~T Wu}, \bibinfo{person}{Kathy Leung},
{and} \bibinfo{person}{Gabriel~M Leung}.} \bibinfo{year}{2020}\natexlab{b}.
\newblock \showarticletitle{Nowcasting and forecasting the potential domestic
and international spread of the 2019-nCoV outbreak originating in Wuhan,
China: a modelling study}.
\newblock \bibinfo{journal}{\emph{The Lancet}} \bibinfo{volume}{395},
\bibinfo{number}{10225} (\bibinfo{year}{2020}), \bibinfo{pages}{689--697}.
\newblock
\bibitem[Wu et~al\mbox{.}(2020a)]%
{wu2020recurrent}
\bibfield{author}{\bibinfo{person}{Ling Wu}, \bibinfo{person}{Nanda~Gopala
Kilingar}, \bibinfo{person}{Ludovic Noels}, {et~al\mbox{.}}}
\bibinfo{year}{2020}\natexlab{a}.
\newblock \showarticletitle{A recurrent neural network-accelerated multi-scale
model for elasto-plastic heterogeneous materials subjected to random cyclic
and non-proportional loading paths}.
\newblock \bibinfo{journal}{\emph{Computer Methods in Applied Mechanics and
Engineering}} \bibinfo{volume}{369} (\bibinfo{year}{2020}),
\bibinfo{pages}{113234}.
\newblock
\bibitem[Xiaomei and Ziyou(2007)]%
{xiaomei2007stability}
\bibfield{author}{\bibinfo{person}{Zhao Xiaomei} {and} \bibinfo{person}{Gao
Ziyou}.} \bibinfo{year}{2007}\natexlab{}.
\newblock \showarticletitle{The stability analysis of the full velocity and
acceleration velocity model}.
\newblock \bibinfo{journal}{\emph{Physica A: Statistical Mechanics and its
Applications}} \bibinfo{volume}{375}, \bibinfo{number}{2}
(\bibinfo{year}{2007}), \bibinfo{pages}{679--686}.
\newblock
\bibitem[Xie et~al\mbox{.}(2008)]%
{xie2008stabilization}
\bibfield{author}{\bibinfo{person}{DF Xie}, \bibinfo{person}{ZY Gao}, {and}
\bibinfo{person}{XM Zhao}.} \bibinfo{year}{2008}\natexlab{}.
\newblock \showarticletitle{Stabilization of traffic flow based on the multiple
information of preceding cars}.
\newblock \bibinfo{journal}{\emph{Communications in Computational Physics}}
\bibinfo{volume}{3}, \bibinfo{number}{4} (\bibinfo{year}{2008}),
\bibinfo{pages}{899--912}.
\newblock
\bibitem[Xie and Grossman(2018)]%
{xie2018crystal}
\bibfield{author}{\bibinfo{person}{Tian Xie} {and} \bibinfo{person}{Jeffrey~C
Grossman}.} \bibinfo{year}{2018}\natexlab{}.
\newblock \showarticletitle{Crystal graph convolutional neural networks for an
accurate and interpretable prediction of material properties}.
\newblock \bibinfo{journal}{\emph{Physical review letters}}
\bibinfo{volume}{120}, \bibinfo{number}{14} (\bibinfo{year}{2018}),
\bibinfo{pages}{145301}.
\newblock
\bibitem[Xing et~al\mbox{.}(2021)]%
{xing2021multi}
\bibfield{author}{\bibinfo{person}{Yang Xing}, \bibinfo{person}{Chen Lv},
\bibinfo{person}{Dongpu Cao}, {and} \bibinfo{person}{Efstathios Velenis}.}
\bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{Multi-scale driver behavior modeling based on deep
spatial-temporal representation for intelligent vehicles}.
\newblock \bibinfo{journal}{\emph{Transportation research part C: emerging
technologies}} \bibinfo{volume}{130} (\bibinfo{year}{2021}),
\bibinfo{pages}{103288}.
\newblock
\bibitem[Xu and Wunsch(2005)]%
{xu2005survey}
\bibfield{author}{\bibinfo{person}{Rui Xu} {and} \bibinfo{person}{Donald
Wunsch}.} \bibinfo{year}{2005}\natexlab{}.
\newblock \showarticletitle{Survey of clustering algorithms}.
\newblock \bibinfo{journal}{\emph{IEEE Transactions on neural networks}}
\bibinfo{volume}{16}, \bibinfo{number}{3} (\bibinfo{year}{2005}),
\bibinfo{pages}{645--678}.
\newblock
\bibitem[Yang et~al\mbox{.}(2019)]%
{yang2019high}
\bibfield{author}{\bibinfo{person}{Kun Yang}, \bibinfo{person}{Yi-Fan Chen},
\bibinfo{person}{Georgios Roumpos}, \bibinfo{person}{Chris Colby}, {and}
\bibinfo{person}{John Anderson}.} \bibinfo{year}{2019}\natexlab{}.
\newblock \showarticletitle{High performance Monte Carlo simulation of Ising
model on TPU clusters}. In \bibinfo{booktitle}{\emph{Proceedings of the
International Conference for High Performance Computing, Networking, Storage
and Analysis}}. \bibinfo{pages}{1--15}.
\newblock
\bibitem[Yang et~al\mbox{.}(2017)]%
{yang2017multi}
\bibfield{author}{\bibinfo{person}{Kaidi Yang}, \bibinfo{person}{Nan Zheng},
{and} \bibinfo{person}{Monica Menendez}.} \bibinfo{year}{2017}\natexlab{}.
\newblock \showarticletitle{Multi-scale perimeter control approach in a
connected-vehicle environment}.
\newblock \bibinfo{journal}{\emph{Transportation research procedia}}
\bibinfo{volume}{23} (\bibinfo{year}{2017}), \bibinfo{pages}{101--120}.
\newblock
\bibitem[Yang et~al\mbox{.}(2020)]%
{yang2020machine}
\bibfield{author}{\bibinfo{person}{Wuyue Yang}, \bibinfo{person}{Liangrong
Peng}, \bibinfo{person}{Yi Zhu}, {and} \bibinfo{person}{Liu Hong}.}
\bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{When machine learning meets multiscale modeling in
chemical reactions}.
\newblock \bibinfo{journal}{\emph{The Journal of Chemical Physics}}
\bibinfo{volume}{153}, \bibinfo{number}{9} (\bibinfo{year}{2020}),
\bibinfo{pages}{094117}.
\newblock
\bibitem[Yang et~al\mbox{.}(2018)]%
{yang2018microstructural}
\bibfield{author}{\bibinfo{person}{Zijiang Yang}, \bibinfo{person}{Xiaolin Li},
\bibinfo{person}{L Catherine~Brinson}, \bibinfo{person}{Alok~N Choudhary},
\bibinfo{person}{Wei Chen}, {and} \bibinfo{person}{Ankit Agrawal}.}
\bibinfo{year}{2018}\natexlab{}.
\newblock \showarticletitle{Microstructural materials design via deep
adversarial learning methodology}.
\newblock \bibinfo{journal}{\emph{Journal of Mechanical Design}}
\bibinfo{volume}{140}, \bibinfo{number}{11} (\bibinfo{year}{2018}).
\newblock
\bibitem[Ye et~al\mbox{.}(2022)]%
{yang2022equitable}
\bibfield{author}{\bibinfo{person}{Yang Ye}, \bibinfo{person}{Qingpeng Zhang},
\bibinfo{person}{Xuan Wei}, \bibinfo{person}{Zhidong Cao},
\bibinfo{person}{Hsiang-Yu Yuan}, {and} \bibinfo{person}{Daniel~Dajun Zeng}.}
\bibinfo{year}{2022}\natexlab{}.
\newblock \showarticletitle{Equitable access to {COVID-19} vaccines makes a
life-saving difference to all countries}.
\newblock \bibinfo{journal}{\emph{Nature Human Behaviour}} \bibinfo{volume}{6},
\bibinfo{number}{2} (\bibinfo{year}{2022}), \bibinfo{pages}{207–216}.
\newblock
\bibitem[Yu and Fish(2002)]%
{yu2002multiscale}
\bibfield{author}{\bibinfo{person}{Qing Yu} {and} \bibinfo{person}{Jacob
Fish}.} \bibinfo{year}{2002}\natexlab{}.
\newblock \showarticletitle{Multiscale asymptotic homogenization for
multiphysics problems with multiple spatial and temporal scales: a coupled
thermo-viscoelastic example problem}.
\newblock \bibinfo{journal}{\emph{International journal of solids and
structures}} \bibinfo{volume}{39}, \bibinfo{number}{26}
(\bibinfo{year}{2002}), \bibinfo{pages}{6429--6452}.
\newblock
\bibitem[Yuan and Fish(2009)]%
{yuan2009multiple}
\bibfield{author}{\bibinfo{person}{Zheng Yuan} {and} \bibinfo{person}{Jacob
Fish}.} \bibinfo{year}{2009}\natexlab{}.
\newblock \showarticletitle{Multiple scale eigendeformation-based reduced order
homogenization}.
\newblock \bibinfo{journal}{\emph{Computer Methods in Applied Mechanics and
Engineering}} \bibinfo{volume}{198}, \bibinfo{number}{21-26}
(\bibinfo{year}{2009}), \bibinfo{pages}{2016--2038}.
\newblock
\bibitem[Yushu and Matou{\v{s}}(2020)]%
{yushu2020image}
\bibfield{author}{\bibinfo{person}{Dewen Yushu} {and} \bibinfo{person}{Karel
Matou{\v{s}}}.} \bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{The image-based multiscale multigrid solver,
preconditioner, and reduced order model}.
\newblock \bibinfo{journal}{\emph{J. Comput. Phys.}} \bibinfo{volume}{406}
(\bibinfo{year}{2020}), \bibinfo{pages}{109165}.
\newblock
\bibitem[Zhang et~al\mbox{.}(2020)]%
{zhang2020changes}
\bibfield{author}{\bibinfo{person}{Juanjuan Zhang}, \bibinfo{person}{Maria
Litvinova}, \bibinfo{person}{Yuxia Liang}, \bibinfo{person}{Yan Wang},
\bibinfo{person}{Wei Wang}, \bibinfo{person}{Shanlu Zhao},
\bibinfo{person}{Qianhui Wu}, \bibinfo{person}{Stefano Merler},
\bibinfo{person}{C{\'e}cile Viboud}, \bibinfo{person}{Alessandro Vespignani},
{et~al\mbox{.}}} \bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{Changes in contact patterns shape the dynamics of
the COVID-19 outbreak in China}.
\newblock \bibinfo{journal}{\emph{Science}} \bibinfo{volume}{368},
\bibinfo{number}{6498} (\bibinfo{year}{2020}), \bibinfo{pages}{1481--1486}.
\newblock
\bibitem[Zhang et~al\mbox{.}(2017)]%
{zhang2017spread}
\bibfield{author}{\bibinfo{person}{Qian Zhang}, \bibinfo{person}{Kaiyuan Sun},
\bibinfo{person}{Matteo Chinazzi}, \bibinfo{person}{Ana Pastore~y Piontti},
\bibinfo{person}{Natalie~E Dean}, \bibinfo{person}{Diana~Patricia Rojas},
\bibinfo{person}{Stefano Merler}, \bibinfo{person}{Dina Mistry},
\bibinfo{person}{Piero Poletti}, \bibinfo{person}{Luca Rossi},
{et~al\mbox{.}}} \bibinfo{year}{2017}\natexlab{}.
\newblock \showarticletitle{Spread of Zika virus in the Americas}.
\newblock \bibinfo{journal}{\emph{Proceedings of the national academy of
sciences}} \bibinfo{volume}{114}, \bibinfo{number}{22}
(\bibinfo{year}{2017}), \bibinfo{pages}{E4334--E4343}.
\newblock
\bibitem[Zhao and Gao(2005)]%
{zhao2005new}
\bibfield{author}{\bibinfo{person}{Xiao-Mei Zhao} {and} \bibinfo{person}{Zi-You
Gao}.} \bibinfo{year}{2005}\natexlab{}.
\newblock \showarticletitle{A new car-following model: full velocity and
acceleration difference model}.
\newblock \bibinfo{journal}{\emph{The European Physical Journal B-Condensed
Matter and Complex Systems}} \bibinfo{volume}{47}, \bibinfo{number}{1}
(\bibinfo{year}{2005}), \bibinfo{pages}{145--150}.
\newblock
\bibitem[Zhou et~al\mbox{.}(2021)]%
{zhou2021dissecting}
\bibfield{author}{\bibinfo{person}{Peijie Zhou}, \bibinfo{person}{Shuxiong
Wang}, \bibinfo{person}{Tiejun Li}, {and} \bibinfo{person}{Qing Nie}.}
\bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{Dissecting transition cells from single-cell
transcriptome data through multiscale stochastic dynamics}.
\newblock \bibinfo{journal}{\emph{Nature communications}} \bibinfo{volume}{12},
\bibinfo{number}{1} (\bibinfo{year}{2021}), \bibinfo{pages}{1--15}.
\newblock
\bibitem[Zhou et~al\mbox{.}(2020)]%
{zhou2020effects}
\bibfield{author}{\bibinfo{person}{Ying Zhou}, \bibinfo{person}{Renzhe Xu},
\bibinfo{person}{Dongsheng Hu}, \bibinfo{person}{Yang Yue},
\bibinfo{person}{Qingquan Li}, {and} \bibinfo{person}{Jizhe Xia}.}
\bibinfo{year}{2020}\natexlab{}.
\newblock \showarticletitle{Effects of human mobility restrictions on the
spread of COVID-19 in Shenzhen, China: a modelling study using mobile phone
data}.
\newblock \bibinfo{journal}{\emph{The Lancet Digital Health}}
\bibinfo{volume}{2}, \bibinfo{number}{8} (\bibinfo{year}{2020}),
\bibinfo{pages}{e417--e424}.
\newblock
\bibitem[Zhu et~al\mbox{.}(2021)]%
{zhu2021mixseq}
\bibfield{author}{\bibinfo{person}{Zhibo Zhu}, \bibinfo{person}{Ziqi Liu},
\bibinfo{person}{Ge Jin}, \bibinfo{person}{Zhiqiang Zhang},
\bibinfo{person}{Lei Chen}, \bibinfo{person}{Jun Zhou}, {and}
\bibinfo{person}{Jianyong Zhou}.} \bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{MixSeq: Connecting Macroscopic Time Series
Forecasting with Microscopic Time Series Data}.
\newblock \bibinfo{journal}{\emph{Advances in Neural Information Processing
Systems}} \bibinfo{volume}{34} (\bibinfo{year}{2021}),
\bibinfo{pages}{12904--12916}.
\newblock
\end{thebibliography}