Learning Climate Variability from Scarce Data with Diffusion Models: A Test Case for ENSO
Extracted source text · arXiv 2606.27094
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% AGUJournalTemplate.tex: this template file is for articles formatted with LaTeX
%
% This file includes commands and instructions
% given in the order necessary to produce a final output that will
% satisfy AGU requirements, including customized APA reference formatting.
%
% You may copy this file and give it your
% article name, and enter your text.
%
% guidelines and troubleshooting are here:
%% To submit your paper:
\documentclass[draft]{agujournal2019}
\usepackage{url} %this package should fix any errors with URLs in refs.
\usepackage{lineno}
\usepackage[inline]{trackchanges} %for better track changes. finalnew option will compile document with changes incorporated.
\usepackage{soul}
\usepackage{amsmath,amssymb}
\usepackage[utf8]{inputenc}
%\linenumbers % disabled for arXiv version
%%%%%%%
% As of 2018 we recommend use of the TrackChanges package to mark revisions.
% The trackchanges package adds five new LaTeX commands:
%
% \note[editor]{The note}
% \annote[editor]{Text to annotate}{The note}
% \add[editor]{Text to add}
% \remove[editor]{Text to remove}
% \change[editor]{Text to remove}{Text to add}
%
% complete documentation is here: http://trackchanges.sourceforge.net/
%%%%%%%
\draftfalse
%% Enter journal name below.
%% Choose from this list of Journals:
%
% JGR: Atmospheres
% JGR: Biogeosciences
% JGR: Earth Surface
% JGR: Oceans
% JGR: Planets
% JGR: Solid Earth
% JGR: Space Physics
% Global Biogeochemical Cycles
% Geophysical Research Letters
% Paleoceanography and Paleoclimatology
% Radio Science
% Reviews of Geophysics
% Tectonics
% Space Weather
% Water Resources Research
% Geochemistry, Geophysics, Geosystems
% Journal of Advances in Modeling Earth Systems (JAMES)
% Earth's Future
% Earth and Space Science
% Geohealth
%
% ie, \journalname{Water Resources Research}
\journalname{Geophysical Research Letters}
\begin{document}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% TITLE
%
% (A title should be specific, informative, and brief. Use
% abbreviations only if they are defined in the abstract. Titles that
% start with general keywords then specific terms are optimized in
% searches)
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Example: \title{This is a test title}
\title{Learning Climate Variability from Scarce Data with Diffusion Models: A Test Case for ENSO}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% AUTHORS AND AFFILIATIONS
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Authors are individuals who have significantly contributed to the
% research and preparation of the article. Group authors are allowed, if
% each author in the group is separately identified in an appendix.)
% List authors by first name or initial followed by last name and
% separated by commas. Use \affil{} to number affiliations, and
% \thanks{} for author notes.
% Additional author notes should be indicated with \thanks{} (for
% example, for current addresses).
% Example: \authors{A. B. Author\affil{1}\thanks{Current address, Antartica}, B. C. Author\affil{2,3}, and D. E.
% Author\affil{3,4}\thanks{Also funded by Monsanto.}}
\authors{Lluís Palma\affil{1,2}, Vincent Verjans\affil{1}, Amanda Duarte\affil{1}, Albert Soret\affil{1} and Markus Donat\affil{1,3}}
% \affiliation{1}{First Affiliation}
% \affiliation{2}{Second Affiliation}
% \affiliation{3}{Third Affiliation}
% \affiliation{4}{Fourth Affiliation}
\affiliation{1}{Barcelona Supercomputing Center, Earth Sciences Department, Barcelona, Spain. 2}
\affiliation{2}{Facultat de Física, Universitat de Barcelona, Barcelona, Spain. 3}
\affiliation{3}{Institució Catalana de Recerca i Estudis Avançats (ICREA), Barcelona, Spain}
%(repeat as many times as is necessary)
% Corresponding author mailing address and e-mail address:
% (include name and email addresses of the corresponding author. More
% than one corresponding author is allowed in this LaTeX file and for
% publication; but only one corresponding author is allowed in our
% editorial system.)
% Example: \correspondingauthor{First and Last Name}{email@address.edu}
\correspondingauthor{Lluís Palma}{lluis.palma@bsc.es}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% KEY POINTS
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% List up to three key points (at least one is required)
% Key Points summarize the main points and conclusions of the article
% Each must be 140 characters or fewer with no special characters or punctuation and must be complete sentences
% Example:
% \begin{keypoints}
% \item List up to three key points (at least one is required)
% \item Key Points summarize the main points and conclusions of the article
% \item Each must be 140 characters or fewer with no special characters or punctuation and must be complete sentences
% \end{keypoints}
\begin{keypoints}
\item We train diffusion models on synthetic (LIM-based) tropical Pacific SST fields to assess whether they recover the underlying low-dimensional structure of the data.
\item Diffusion models recover the low-order structure given sufficient data, but the $\sim700$ monthly observations in ERSSTv5 fall an order of magnitude below the $\sim7,000$ needed for convergence.
\item Pre-training on CMIP6 with a learned model embedding, followed by fine-tuning on scarce observations, overcomes this data limitation.
\end{keypoints}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% ABSTRACT and PLAIN LANGUAGE SUMMARY
%
% A good Abstract will begin with a short description of the problem
% being addressed, briefly describe the new data or analyses, then
% briefly states the main conclusion(s) and how they are supported and
% uncertainties.
% The Plain Language Summary should be written for a broad audience,
% including journalists and the science-interested public, that will not have
% a background in your field.
%
% A Plain Language Summary is required in GRL, JGR: Planets, JGR: Biogeosciences,
% JGR: Oceans, G-Cubed, Reviews of Geophysics, and JAMES.
% see http://sharingscience.agu.org/creating-plain-language-summary/)
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% \begin{abstract} starts the second page
\begin{abstract}
Diffusion models are increasingly applied to climate emulation, but whether they capture the correct modes of variability remains unclear, a concern amplified by data scarcity at longer timescales. We investigate this using synthetic tropical Pacific SST fields from Linear Inverse Models (LIMs), whose known low-order structure bypasses the overlapping and confounding modes of real observations. With sufficient training data, our model recovers the correct structure of both Gaussian and non-Gaussian LIMs, including ENSO's Eastern-Pacific/Central-Pacific El Niño/La Niña asymmetry. Yet an ablation on the number of monthly training samples reveals that the ~700 observations in ERSSTv5 fall an order of magnitude below the ~7,000 samples needed for convergence, and not all diffusion parameterisations recover the correct low-order structure. Pre-training on CMIP6 with a learned model embedding, followed by fine-tuning on scarce observations, closes this gap — reproducing observed statistics more faithfully than both Gaussian and non-Gaussian LIMs.
\end{abstract}
\section*{Plain Language Summary}
El Niño and La Niña are major climate phenomena that affect weather patterns worldwide, yet we only have about 70 years of detailed sea surface temperature observations — too few for
modern AI methods that typically need thousands of examples to learn reliably. We tested whether diffusion models, a type of AI originally developed for image generation, can learn the essential patterns of tropical Pacific variability using synthetic datasets with known characteristics. These models succeed when given enough data but struggle with the limited observational record, often memorising training examples rather than learning general patterns. To overcome data scarcity, we developed a strategy that first trains the model on output from many different climate simulations, embedding each simulation's identity, and then fine-tunes on real observations. This approach produces realistic variability patterns without memorisation.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% BODY TEXT
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%% Suggested section heads:
% \section{Introduction}
%
% The main text should start with an introduction. Except for short
% manuscripts (such as comments and replies), the text should be divided
% into sections, each with its own heading.
% Headings should be sentence fragments and do not begin with a
% lowercase letter or number. Examples of good headings are:
% \section{Materials and Methods}
% Here is text on Materials and Methods.
%
% \subsection{A descriptive heading about methods}
% More about Methods.
%
% \section{Data} (Or section title might be a descriptive heading about data)
%
% \section{Results} (Or section title might be a descriptive heading about the
% results)
%
% \section{Conclusions}
\section{Introduction}
Deep learning has driven recent advances in data-driven weather prediction, rivalling traditional Numerical Weather Prediction (NWP) models \cite{pathakFourCastNetGlobalDatadriven2022,lamLearningSkillfulMediumrange2023,kochkovNeuralGeneralCirculation2024,priceGenCastDiffusionbasedEnsemble2024}. These advances have motivated extending such approaches to longer timescales \cite{panImprovingSeasonalForecast2022,raderOptimizingSeasonalToDecadalAnalog2023, watt-meyerACE2AccuratelyLearning2025,kentSkilfulGlobalSeasonal2025,palmaDatadrivenSeasonalClimate2026}, where NWPs and climate models are computationally expensive and limited in skill. Yet progress is hampered by the short span of available reanalyses (40--70 years). Unlike weather forecasting, where high-frequency variability yields many independent samples over short periods, longer timescales yield far fewer \cite{gibsonTrainingMachineLearning2021,miloshevichProbabilisticForecastsExtreme2023, materiaArtificialIntelligenceClimate2024}. One example is the El Ni\~no--Southern Oscillation (ENSO), the dominant mode of interannual variability in the tropical Pacific \cite{wangNinoSouthernOscillation2017}. Its warm phase (El Ni\~no), characterised by anomalous warming of the central and eastern equatorial Pacific, recurs every 2--7 years and alternates with the cold La Ni\~na phase. ENSO-driven anomalies alter tropical convection and trigger atmospheric teleconnections that drive temperature and precipitation impacts in large parts of the world with major societal consequences \cite{mcphadenENSOIntegratingConcept2006,liuNonlinearNinoImpacts2023}. Yet a single reanalysis may contain only a few dozen ENSO cycles, limiting the application of deep learning algorithms with a large number of trainable parameters.
Some Machine learning (ML) approaches address observational data scarcity by training on climate model output \cite{anderssonSeasonalArcticSea2021,gibsonTrainingMachineLearning2021,panImprovingSeasonalForecast2022,palmaDatadrivenSeasonalClimate2026}, benefiting from thousands of simulated years that span regime shifts and unobserved trends, though at the cost of inheriting model biases, such as misrepresented teleconnections and unresolved small-scale processes. An emerging alternative trains AI weather models on reanalysis data and runs them autoregressively to longer lead times \cite{watt-meyerACE2AccuratelyLearning2025,kentSkilfulGlobalSeasonal2025}, avoiding model errors but reintroducing small test sets with no guarantee of extrapolation beyond the training distribution. Across all approaches, complex architectures make it difficult to verify that learned relationships are physically consistent, with skill scores and summary statistics typically serving as proxy validation \cite{raspWeatherBench2Benchmark2024}.
From all the variety of ML architectures, diffusion models \cite{hoDenoisingDiffusionProbabilistic2020, songScoreBasedGenerativeModeling2021,karrasElucidatingDesignSpace2022,lipmanFlowMatchingGenerative2023} have demonstrated strong performance in weather and climate emulation tasks owing to their ability to model the nonlinear, non-Gaussian behaviour intrinsic to the climate system. Their progressive denoising objective provides implicit regularisation: the model must produce coherent fields from heavily corrupted to nearly clean inputs, helping prevent overfitting and avoiding the oversmoothed fields typical of direct RMSE objectives. Diffusion models have been used for applications ranging from climate emulation to downscaling and bias adjustment \cite{bassettiDiffESMConditionalEmulation2024,priceGenCastDiffusionbasedEnsemble2024,aichConditionalDiffusionModels2026}. Yet they necessitate large data amounts and are not exempt from the interpretability challenges shared among deep learning methods \cite{regtUnderstandingValuesAims2020,krennScientificUnderstandingArtificial2022,channingAIScientificDiscovery2026}.
Assessing whether machine-learned relationships are physically meaningful can be approached from different perspectives. Here, we adopt a distribution-matching perspective which assumes that a small number of modes capture the dominant fraction of climate variability \cite{devironGlobalModesClimate2013}. Verifying whether an ML model has learned such structure from real data is challenging, as overlapping timescales and confounding modes can obscure the signal, and true low-order structure is a priori unknown. To address this, we leverage a synthetic multi-millennial SST dataset derived from a low-order ENSO model \cite{martinez-villalobosLowOrderDataDrivenModel2025}. This dataset is long enough to serve as ground truth for evaluating the model's learned representation, in a controlled, idealised setup free from trends, distributional shifts, or other higher-order modes affecting ENSO --- something not possible with real-world reanalysis. By evaluating if diffusion models can recover an idealised low-dimensional representation of ENSO, this study assesses a prerequisite for their application to the full climate system.
Thus, we formulate the following questions: Can diffusion models recover the low-dimensional structure of a synthetic tropical Pacific SST dataset? Are there differences between diffusion parametrisations? Is the sample size of current reanalyses sufficient to reach optimal performance in terms of low-order recovery? If not, how can we mitigate this through different training strategies?
\section{Data and Methods}\label{sec:data}
\subsection{Data}
% To learn a low-order ENSO structure, we use 120k monthly tropical Pacific SST fields generated from Linear Inverse Models (Lims; Penland, 1989; Martinez-Villalobos et al., 2025). The LIMs are calibrated to 1948-2022 ERSSTv5 reanalysis (Huang et al., 2015).
To learn ENSO variability, we use 120k monthly tropical Pacific SST fields generated from two sets of Linear Inverse Models (LIMs; \citeA{penlandRandomForcingForecasting1989,martinez-villalobosLowOrderDataDrivenModel2025,martinez-villalobosTropicalPacificSST2026}). Specifically, we use both the standard LIM and non-Gaussian (NG-LIM) versions \cite{martinez-villalobosLowOrderDataDrivenModel2025}. In a standard stationary LIM, the evolution of the tropical Pacific state vector is parametrised as follows:
\begin{equation}
\frac{d\mathbf{x}}{dt} = \mathbf{M}\mathbf{x} + \mathbf{B}\boldsymbol{\eta}
\end{equation}
where $\mathbf{x}$ is a 10-dimensional state vector whose first two components are the East and Central Pacific (EP, CP) indices, defined as linear combinations of the leading two Principal Components (PCs) of tropical Pacific SST variability, and whose remaining eight components are PCs 3--10. The matrix $\mathbf{M}$ represents the linearized deterministic dynamics derived from the empirical 1 month lag covariance matrix \cite{penlandOptimalGrowthTropical1995}, $\mathbf{B}$ is the noise covariance matrix derived from the fluctuation-dissipation relationship \cite{penlandBalanceConditionStochastic1994}, and $\boldsymbol{\eta}$ is a Gaussian white noise vector. The NG-LIM applies an additional Yeo-Johnson transformation \cite{yeoNewFamilyPower2000} to map the state variables $\mathbf{x}$, to near-Gaussianity, better capturing ENSO nonlinearities and asymmetries \cite{choiENSOTransitionDuration2013, dinezioNonlinearControlsPersistence2014,martinez-villalobosObservedNinoLaNina2019}. Thus, both LIM and NG-LIM synthetic datasets have a known low-dimensional structure (the first 10 PCs explaining 90\% of tropical Pacific SST variability).
All LIMs are calibrated to the 1948-2022 ERSSTv5 observations \cite{huangExtendedReconstructedSea2017}. Additionally, we use the same raw ERSSTv5 SST fields and NG-LIM emulations of 30 CMIP6 \cite{eyringOverviewCoupledModel2016} models also obtained from \cite{martinez-villalobosTropicalPacificSST2026}, and with the same calibration period.
\subsection{Diffusion Model}
Generative models --- and diffusion models (DMs; \citeA{hoDenoisingDiffusionProbabilistic2020,songScoreBasedGenerativeModeling2021}) in particular --- learn to draw samples from complex data distributions for which we have empirical observations but whose true form is unknown. Inspired by non-equilibrium thermodynamics \cite{sohl-dicksteinDeepUnsupervisedLearning2015}, DMs follow a two-stage process: a forward process that gradually corrupts data with Gaussian noise, and a reverse process that generates new samples by denoising. More formally, the DM forward process progressively adds noise to data samples according to:
\begin{equation}
\mathbf{z}_t = \alpha_t \mathbf{x} + \sigma_t \boldsymbol{\epsilon} \label{eq:forward}
\end{equation}
where $\alpha_t$ is the signal scaling coefficient at time $t$, controlling how much of the original data $\mathbf{x}$ is preserved, and $\sigma_t$ controls the noise level ($\boldsymbol{\epsilon} \sim \mathcal{N}(0, I)$). Note that diffusion time $t$ indexes the noising process, not the physical time over which weather evolves. Thus, as the noising process progresses (in our convention, from $t=1$ at the original data to $t=0$ at pure noise), the ratio $\alpha_t / \sigma_t$ decreases monotonically, making the data indistinguishable from noise. These marginals form a time-continuous, increasingly noisy sequence — the probability path:
\begin{equation}
p(\mathbf{z}_t) = \int p(\mathbf{z}_t \mid \mathbf{x})\, p(\mathbf{x})\, d\mathbf{x} = \int \mathcal{N}(\mathbf{z}_t\mid\alpha_t \mathbf{x}, \sigma_t^2 I)\, p(\mathbf{x})\, d\mathbf{x} \label{eq:probpath}
\end{equation}
In this work, we adopt the ordinary differential equation (ODE)-based \cite{hoDenoisingDiffusionProbabilistic2020} formulation of flow matching \cite{lipmanFlowMatchingGenerative2023}, though it is closely intertwined with other diffusion formalisms (we refer the reader to \citeA{laiPrinciplesDiffusionModels2025} for a comprehensive overview). We parameterise the time evolution of $p(\mathbf{z}_t)$ via the time derivative of $\mathbf{z}$, a velocity field $\mathbf{v}$, which prescribes how samples must evolve so that their distribution matches $p(\mathbf{z}_t)$ at every $t$ --- a constraint formalised by the continuity equation. For the Gaussian path defined by Eq.~\ref{eq:probpath}, the corresponding probability flow ODE can be expressed as:
\begin{equation}
\frac{d\mathbf{z}_t}{dt} = \mathbf{v}_t(\mathbf{z}_t) \label{eq:velocity}
\end{equation}
We adopt a simple linear schedule \cite{lipmanFlowMatchingGenerative2023}, for which $\alpha_t = t$ and $\sigma_t = 1 - t$. The objective is then to learn a neural network $\mathbf{v}_\theta(\mathbf{z}_t, t) \approx \mathbf{v}_t(\mathbf{z}_t)$ that approximates this velocity field. As $\mathbf{v}_t(\mathbf{z}_t)$ is generally intractable, we seek to minimise the network output against the conditional velocity $\mathbf{v}_t(\mathbf{z}_t \mid \mathbf{x})$, which can be derived from the time derivative of Equation~\ref{eq:forward} as:
\begin{equation}
\mathbf{v}_t(\mathbf{z}_t \mid \mathbf{x}) = \alpha'_t \mathbf{x} + \sigma'_t \boldsymbol{\epsilon} = \mathbf{x} - \boldsymbol{\epsilon} = \frac{\mathbf{x} - \mathbf{z}_t}{1-t} = \frac{\mathbf{z}_t - \boldsymbol{\epsilon}}{t}
\label{eq:cond_vel}
\end{equation}
yielding the conditional (on $\mathbf{x}$) flow matching loss:
\begin{equation}
\mathcal{L} = \mathbb{E}_{t \sim \mathcal{U}(0,1),\, \mathbf{x} \sim p(\mathbf{x}),\, \boldsymbol{\epsilon} \sim \mathcal{N}(0,I)} \left[\left\| \mathbf{v}_\theta(\mathbf{z}_t, t) - \mathbf{v}_t(\mathbf{z}_t \mid \mathbf{x}) \right\|^2 \right] \label{eq:loss}
\end{equation}
\citeA{lipmanFlowMatchingGenerative2023} shows that this loss shares the same gradient with respect to $\theta$ as the marginal flow matching loss (Mean Squared Error against $\mathbf{v}_t(\mathbf{z}_t)$), so minimising one also minimises the other. Intuitively, sampling multiple $(\mathbf{x}, \boldsymbol{\epsilon})$ pairs drives the conditional expectation of $\mathbf{v}_t(\mathbf{z}_t\mid\mathbf{x})$ over $\mathbf{x}$ towards $\mathbf{v}_t(\mathbf{z}_t)$. Once $\mathbf{v}_\theta$ is learned, the generative process recovers samples from the data distribution: we integrate Equation~\ref{eq:velocity} from $t=0$ to $t=1$, replacing $\mathbf{v}_t(\mathbf{z}_t)$ with $\mathbf{v}_\theta(\mathbf{z}_t, t)$.
It is important to note that, given the relationship established in Equation~\ref{eq:cond_vel}, our network $\mathbf{d}_\theta$ can predict different quantities at each time step $t$: the clean data $\mathbf{x}$ ($\mathbf{v}_t = \frac{\mathbf{d}_\theta - \mathbf{z}_t}{1-t}$) --- as in the original Denoising Diffusion Probabilistic Models (DDPM) formulation \cite{sohl-dicksteinDeepUnsupervisedLearning2015,songGenerativeModelingEstimating2020,hoDenoisingDiffusionProbabilistic2020} ---, the noise $\boldsymbol{\epsilon}$ ($\mathbf{v}_t = \frac{\mathbf{z}_t - \mathbf{d}_\theta}{t}$) \cite{hoDenoisingDiffusionProbabilistic2020}, or $\mathbf{v}_t$ directly \cite{salimansProgressiveDistillationFast2022, lipmanFlowMatchingGenerative2023}. These are related by simple identities and differ in training only by time-dependent reweighting factors \cite{laiPrinciplesDiffusionModels2025,liBackBasicsLet2025}.
In practice, however, these parameterisations behave very differently. \citeA{liBackBasicsLet2025} argue that natural data occupy a low-dimensional manifold within the high-dimensional observation space, while noise spans the full space. A network predicting clean data ($\mathbf{x}$-prediction) therefore only needs to capture this low-dimensional structure, whereas predicting noise ($\boldsymbol{\epsilon}$-prediction), or an intermediate quantity ($\mathbf{v}$-prediction), requires preserving high-dimensional information. We hypothesise that this reasoning extends to climate fields, where a low-dimensional structure can explain a large fraction of climate variability \cite{devironGlobalModesClimate2013}.
Thus, our model ingests monthly SST anomaly fields from the tropical Pacific ($\mathbf{x} \in \mathbb{R}^{n_{lat} \times n_{lon}}$) and learns the velocity field ($\mathbf{v}_\theta \in \mathbb{R}^{n_{lat} \times n_{lon}}$) that drives the distribution underlying SST variability in the region. For details on the network architecture, training, and sampling strategy, we refer the reader to the Supplementary Information.
\section{Results}\label{sec:results}
To evaluate whether diffusion models recover the correct low-dimensional structure, we start by training our model ($\mathbf{x}$-prediction) on both the LIM and NG-LIM synthetic datasets and compare the generated and true fields.
In line with \citeA{martinez-villalobosLowOrderDataDrivenModel2025}, Figure 1 shows the joint PC1–PC2 scatter plots (a, f) and spatial skewness maps, testing whether the model preserves the non-Gaussian warm–cold asymmetry
characteristic of ENSO, for true (b, g) and generated (c, h) data. Overall, the diffusion model successfully reproduces both the Gaussian (LIM) and non-Gaussian (NG-LIM) distributions. When trained on the NG-LIM dataset, generated samples reproduce the characteristic U shape (highlighted by the quadratic fit curves), reflecting the tendency for EP El Niños (positive PC1, positive PC2) and CP La Niñas (negative PC1, positive PC2) to reach larger anomalies \cite{choiENSOTransitionDuration2013,martinez-villalobosObservedNinoLaNina2019}. The skewness maps (b, c, g, h) further validate that the model preserves the non-Gaussian warm–cold asymmetry, with a horseshoe shape with positive skewness in the eastern Pacific and negative skewness in the western Pacific. Beyond these, we compare the spatial EOF patterns of true and generated data via a cross-correlation matrix (d, i). Again, the diffusion model consistently recovers the first 10 EOFs from both the LIM and NG-LIM datasets, as indicated by strong (above 0.9) diagonal and weak (below 0.35) off-diagonal correlations. To assess whether EOF relationships are preserved, we project the generated data onto the true EOFs. The projected variance per EOF (e, j) closely matches the true variance structure (max. $\sim 7\%$ difference), confirming that the model faithfully reproduces both the spatial patterns and their relative importance on both datasets.
\begin{figure}
\noindent\includegraphics[width=\textwidth]{fig1_lim_vs_nglim.png}
\caption{Top row (a,b,c,d,e): model trained on Linear Inverse Model (LIM) data. Bottom row (f,g,h,i,j): trained on nonlinear-Gaussian LIM (NG-LIM). Note that ``true'' refers to LIM- and NG-LIM-generated data in the top and bottom row, respectively. (a,f) Scatter plots of the two leading principal components (PC1 vs PC2) for true data (green) and generated samples (grey/black), with quadratic fit curves highlighting the nonlinear PC1--PC2 relationship. (b,c,g,h) Spatial skewness maps for true (b,g) and generated (c,h) data. (d,i) EOF spatial pattern cross-correlation heatmaps between true and generated samples (note the nonlinear color scaling). (e,j) Percentage variance explained by the leading projected (on True EOFs) PCs for true (dashed) and generated (solid) data.}
\label{fig:lim_vs_nglim}
\end{figure}
Motivated by \citeA{liBackBasicsLet2025}, we test on the NG-LIM dataset whether results differ across diffusion parameterisations: $\mathbf{x}$-prediction, $\mathbf{v}$-prediction, and $\boldsymbol{\epsilon}$-prediction \cite{laiPrinciplesDiffusionModels2025}. Figure S1 shows PC1–PC2 scatter plots, EOF spatial cross-correlation matrices, and projected variance per EOF for each diffusion variant. $\boldsymbol{\epsilon}$-prediction fails all three diagnostics. $\mathbf{v}$-prediction reproduces the PC1–PC2 distribution and EOF correlations well (diagonal mean of 0.92), but projects substantially less variance onto the true EOF patterns. Projecting $\mathbf{v}$-prediction samples onto their own leading 10 EOFs recovers only 19.6\% of their variance (not shown), indicating that $\mathbf{v}$-prediction dramatically overestimates the effective dimensionality of the data. Thus, only $\mathbf{x}$-prediction fully recovers the true low-dimensional structure of the data.
We further test the sensitivity to the input dimension $D$ (Fig. S2) by reducing it from its full dimensionality to the manifold dimension $d$ using PCA. The grey shaded region marks the under-capacity regime ($D$ \textgreater 256, the hidden layer width), where the input must be compressed through a narrower internal representation, forcing the network to learn a lower-dimensional encoding. $\boldsymbol{\epsilon}$-prediction recovers the data only near the manifold dimension and degrades as $D$ grows, consistent with its failure at full $D$ in Figure~S1. Overall $\mathbf{x}$-prediction is the only parameterisation that holds across dimensionalities and diagnostics --- we hypothesise that the degradation at $D \sim 110$ reflects a double-descent-like transition: as $D$ decreases, the network shifts from a regime where the bottleneck forces generalisation through compression to one where excess capacity lets it interpolate the training data, improving generalisation. Thus, consistent with \citeA{liBackBasicsLet2025}, these results confirm that directly predicting clean data enables under-capacity networks to operate effectively in high-dimensional spaces.
So far, the results indicate that diffusion models can recover the low-dimensional structure of both Gaussian and non-Gaussian ENSO synthetic datasets. We now ask whether this methodology could be extended to real observational data, which is far scarcer (~720 vs. 120k monthly samples). We perform an ablation study by randomly subsampling the training set from 120k down to 50 samples.
Figure 2 shows a reduced version of the previously shown diagnostics. The cross-correlation matrix is summarised as a single variance-weighted diagonal correlation score. We also compute the mean projected variance ratio per mode — the mean variance of generated samples projected onto each true EOF, divided by the corresponding true variance — and the spatial correlation of the skewness pattern. In addition, as diffusion models are prone to memorising training data, we track a nearest-neighbour memorisation ratio: for each generated and held-out test sample, we find the nearest training sample by L2 distance and compare median distances. A ratio near 1.0 indicates healthy generalisation; values well below 1.0 signal memorisation. The vertical dotted line marks N = 720, the number of monthly samples in the observational record (ERSSTv5) available for training (80\% of the 1948-2022 set). Results show a clear convergence at around 7000 samples, well beyond the number of available samples in ERSSTv5. Some metrics, such as the EOF diagonal correlation or the skewness correlation, show high values even though the memorisation ratio shows high degrees of memorisation. This highlights limitations of standard statistical verification when applied to generative models with strong memorisation capabilities. Additionally, we remove extreme values (defined by the El Niño 3.4 index) from the training set and recompute skewness maps (Supplementary Figure S3) and El Niño composites (Supplementary Figure S4). Without extremes in the training data, the model fails to reproduce them — as reflected in both the skewness maps and composite differences — underscoring the sensitivity of diffusion models to training set composition.
\begin{figure}
\noindent\includegraphics[width=\textwidth]{fig2_ablation_vs_N.png}
\caption{Performance of $\mathbf{x}$-prediction diffusion as a function of training set size ($N$), evaluated on the NG-LIM dataset. (a) Variance-weighted PC pattern correlation. (b) Projected
variance ratio. (c) Training (solid) and validation (dashed) loss. (d) Spatial skewness correlation. (e) Memorization ratio (nearest-neighbour distance of generated vs held-out to training set).
Shaded bands show $\pm 1$ standard deviation across 10 random seeds. Vertical dotted line marks $N = 720$, the number of monthly samples in the observational record (ERSSTv5). Subsampling is performed at random across the training set.}
\label{fig:ablation_vs_N}
\end{figure}
Deep learning algorithms are known to underperform in data-scarce regimes. Several strategies have been proposed to address this, ranging from fine-tuning and transfer learning to dimensionality reduction \cite{anderssonSeasonalArcticSea2021,gibsonTrainingMachineLearning2021,panImprovingSeasonalForecast2022,materiaArtificialIntelligenceClimate2024}. We test several of these to bridge the gap shown in Figure 2 between model performance with abundant and scarce training data.
Figure 3 shows the same metrics as Figure 2, but restricted to 720 NG-LIM samples. Black bars show the performance achieved when training directly on 720 samples (as in Figure 2 for N = 720). Hollow white bars represent "optimal" performance achieved by training on the full 120k samples dataset. We test several approaches to close this gap. First (blue), we test whether dimensionality reduction (to 100D) via Principal Component Analysis (PCA) helps overcome the data scarcity gap. Next (purple), inspired by several works \cite{anderssonSeasonalArcticSea2021, gibsonTrainingMachineLearning2021,panImprovingSeasonalForecast2022, miloshevichProbabilisticForecastsExtreme2023,palmaDatadrivenSeasonalClimate2026}, we train a diffusion model on the NG-LIM CMIP6 dataset, pooling all models into a single training set. Once trained, we evaluate this model against the NG-LIM target samples. Building on this, we extend the approach in two ways. The first (orange) involves fine-tuning the CMIP6 pre-trained model on 720 samples of the ERSSTv5 NG-LIM dataset. The second (red) goes a step further by adding a model embedding during CMIP6 pre-training \cite{guoEntityEmbeddingsCategorical2016,raspNeuralNetworksPostprocessing2018,panImprovingSeasonalForecast2022}: a one-hot encoding of the CMIP6 model identifier, linearly projected into an 8-dimensional vector fed to the network. In a subsequent stage, we fine-tune only this 8-dimensional vector using the 720 ERSSTv5 NG-LIM samples, keeping the rest of the network frozen. This serves two purposes. First, during training, it makes the network aware of inter-model differences, preventing it from confounding different model dynamics or collapsing them into an averaged-out signal. Second, it allows the network to learn directions of inter-model variability in the 8-dimensional embedding space. During fine-tuning, the network can navigate along these directions — and even extrapolate beyond them — to better approximate the observational target, even when no single CMIP6 model fully captures the observed behaviour.
\begin{figure}
\noindent\includegraphics[width=\textwidth]{fig3_strategies.png}
\caption{Comparison of training strategies for learning the NG-LIM distribution from only $N = 720$ samples. (a) Variance-weighted PC pattern correlation, (b) spatial skewness
correlation, (c) projected variance ratio, and (d) memorisation ratio. Strategies shown: direct training in full space (black), PCA reduction to 100D (blue), CMIP6
pre-training without fine-tuning (purple), CMIP6 pre-training with fine-tuning (orange), and CMIP6 pre-training with model-ID embedding finetuned (red). Hollow bars show the
performance of the direct approach trained on the full 120k-sample dataset. Error bars indicate $\pm 1$ standard deviation across 10 seeds.}
\label{fig:strategies}
\end{figure}
The PCA-based approach (blue) achieves high matched correlation (above 0.85) but low skewness pattern correlation (below 0.75), low projected variance ratio, and a low memorisation ratio (both below 0.6), showing clear memorisation of the training set. Results are similar regardless of whether PCA is computed over the full or scarce ERSSTv5 NG-LIM dataset (not shown). The CMIP6 pre-trained model without fine-tuning (purple) shows the opposite pattern: low matched correlation (below 0.7), moderate skewness pattern correlation (above 0.75), and high projected variance and memorisation ratios (1.2 and 1.3, respectively). These results suggest that CMIP6 pre-training avoids overfitting but leads the model to learn an erroneous variability structure, reflecting known biases in CMIP6 simulations. This is further supported by the comparison of NG-LIM EOFs across CMIP6 models (Figures S5–6).
Turning to fine-tuning strategies, direct fine-tuning improves on or matches the direct, PCA, and CMIP6 pre-trained approaches for the EOF cross-correlation (above 0.85), the skewness pattern correlation (above 0.7), and projected variance ratio (~0.7). Yet it remains far from optimal and shows clear memorisation (~0.6), indicating overfitting to the scarce ERSSTv5 NG-LIM data. In contrast, the approach of fine-tuning only the model embedding to the ERSSTv5 NG-LIM data brings substantial improvements: with an EOF cross-correlation above 0.85, a skewness pattern correlation above 0.9, a projected variance ratio above 0.8, and a memorisation ratio close to 0.95. This indicates that the method recovers bias-adjusted modes of variability without memorising the training data. These results indicate that the 8-dimensional embedding is sufficiently small to regularise the network, preventing overfitting, while sufficiently large to capture meaningful inter-model differences in their representation of variability and identify relevant directions for improvement. A 16-dim embedding vector was also tested, and similar results were found (not shown).
Finally, as a demonstration on scarce real-world observations, we fine-tune the CMIP6 NG-LIM pre-trained model on raw ERSSTv5 SST fields rather than synthetic NG-LIM fields. Figure 4 compares PC1–PC2 scatter plots (a–d) and skewness maps (e–h) across the LIM (a, e), NG-LIM (b, f), our diffusion model fine-tuned on ERSSTv5 via CMIP6 pre-training and model embedding (c, g), and the raw ERSSTv5 observations (d, h). Our approach captures the skewness pattern better than the NG-LIM (pattern correlation: 0.95 vs. 0.80; RMSE: 0.14 vs. 0.27), particularly the extension of positive skewness from the eastern into the central Pacific and the elongated pattern of negative skewness south-east of Indonesia. The PC1–PC2 scatter plots further show that the diffusion model better captures the vertical asymmetry (along the EP+/CP- axis), whereas the NG-LIM displays a more symmetric pattern.
\begin{figure}
\noindent\includegraphics[width=\textwidth]{fig4_summary.png}
\caption{(a--d) Scatter plots of the two dominant modes of sea surface temperature variability (PC1 vs PC2), as generated by: (a) a linear model (LIM), (b) a non-Gaussian linear
model (NG-LIM), (c) our diffusion model leveraging CMIP6 simulations and fine-tuned to ERSSTv5, and (d) the observed record (ERSSTv5, 1948--2022). (e--h) Maps of skewness---a measure of warm--cold asymmetry,
where positive values (red) indicate that warm anomalies (El Ni\~no) reach larger anomalies than cold anomalies (La Ni\~na).}
\label{fig:summary}
\end{figure}
\section{Summary and Discussion}\label{sec:discussion}
In this study, we trained a diffusion model on monthly synthetic tropical Pacific SST fields derived from LIMs \cite{martinez-villalobosLowOrderDataDrivenModel2025} and assessed whether it recovers its underlying low-dimensional structure (10 PCs). The main findings are:
\begin{itemize}
\item Diffusion models accurately recover the low-dimensional structure of both Gaussian and non-Gaussian synthetic datasets given sufficient training data.
\item Only the $\mathbf{x}$-prediction parameterisation recovers all target statistics under a low-capacity architecture (a simple multi-layer perceptron), consistent with \citeA{liBackBasicsLet2025}.
\item The $\sim$700 monthly samples in ERSSTv5 fall well below the $\sim$7,000 sample threshold at which the diffusion model converges to optimal performance, confirming that current observational records alone are insufficient.
\item Pre-training on CMIP6 with a learned model embedding, followed by the fine-tuning of the embedding on scarce observations, yields the best performance on both synthetic and raw ERSSTv5 observations. Alternative strategies --- including CMIP6 pre-training with or without fine-tuning and dimensionality reduction --- yield suboptimal results.
\end{itemize}
Our results confirm that diffusion models are effective tools for modelling complex, non-Gaussian distributions in climate applications such as ENSO. Yet not all parameterisations recover the low-order structure of the data, as noted by \cite{liBackBasicsLet2025}. Besides, we quantify that these models require thousands of samples to achieve optimal performance. The success of the learned model embedding aligns with the findings of \citeA{panImprovingSeasonalForecast2022}, demonstrating that leveraging multi-model diversity can compensate for observational scarcity when properly exploited. However, our experiments also highlight important caveats: removing extremes from the training data prevents the model from reproducing them, and commonly used summary statistics in climate evaluation can fail to detect overfitting and memorisation, highlighting the need for more sensitive diagnostics. Controlled synthetic experiments, such as those presented here, offer a principled complement to skill-based benchmarks for diagnosing ML model capabilities and failure modes across climate applications.
Several limitations should be acknowledged. Our experiments use a controlled, synthetic setup that excludes non-stationary behaviour such as long-term trends or regime shifts, which is far from the true complexity of the climate system. Furthermore, our evaluation framework is purely statistical and unconditional (random distribution sampling): it assesses whether diffusion models match the low-dimensional manifold of the training distribution. However, it does not evaluate temporal dynamics or the predictive skill that would require a conditional (on an initial state) modelling framework. Accordingly, the model does not serve as an ENSO prediction system in its current form (unlike the LIM and NG-LIM). Besides, the state of ENSO is solely characterised by SSTs, in line with \cite{martinez-villalobosLowOrderDataDrivenModel2025}, which is a strong simplification of the coupled nature of it.
%Some recent approaches instead seek to unify the two \cite{cachayDYffusionDynamicsinformedDiffusion2023,cachayElucidatedRollingDiffusion2025}, motivated by parallels between diffusion time and predictability horizons
Conditional generation constitutes the most natural extension of this work, opening the door to the emulation of ENSO dynamics and its prediction. Turning this framework into a prediction system would require careful evaluation, as the short observational ENSO record induces skill sampling uncertainties \cite{verjansLargePotentialPerformancebased2026}. Approaches such as classifier-free guidance (Ho \& Salimans, 2022), diffusion posterior sampling (Chung et al., 2023), or coupled flow matching \cite{lipmanFlowMatchingGenerative2023} offer promising avenues in this regard. Besides, several studies have already started to apply conditional diffusion approaches for weather and climate emulation \cite{priceGenCastDiffusionbasedEnsemble2024,stockSwiftAutoregressiveConsistency2025,cachayElucidatedRollingDiffusion2025}. Furthermore, exploring the connections between such conditional setups and dynamical systems theory could yield more mechanistic interpretations of the learned relationships. For ENSO applications specifically, richer state characterisation, including subsurface ocean heat content and surface wind stress forcings, would provide a more physically grounded conditional setup \cite{newmanEmpiricalModelTropical2011}. Finally, considering the success of the diversity captured by the model embedding, incorporating higher-resolution simulations \cite{rackowMultiyearSimulationsKilometre2025, doblas-reyesDestinationEarthDigital2026} as pre-training data (beyond the current CMIP6 ensemble) would likely enhance the fidelity of the learned distribution.
%Text here ===>>>
%%
% Numbered lines in equations:
% To add line numbers to lines in equations,
% \begin{linenomath*}
% \begin{equation}
% \end{equation}
% \end{linenomath*}
%% Enter Figures and Tables near as possible to where they are first mentioned:
%
% DO NOT USE \psfrag or \subfigure commands.
%
% Figure captions go below the figure.
% Acronyms used in figure captions will be spelled out in the final, published version.
% Table titles go above tables; other caption information
% should be placed in last line of the table, using
% \multicolumn2l{$^a$ This is a table note.}
% NOTE that there is no difference between table caption and table heading in the final, published version
%
%----------------
% EXAMPLE FIGURES
%
% \begin{figure}
% \includegraphics{example.png}
% \caption{caption}
% \end{figure}
%
% Giving latex a width will help it to scale the figure properly. A simple trick is to use \textwidth. Try this if large figures run off the side of the page.
% \begin{figure}
% \noindent\includegraphics[width=\textwidth]{anothersample.png}
%\caption{caption}
%\label{pngfiguresample}
%\end{figure}
%
%
% If you get an error about an unknown bounding box, try specifying the width and height of the figure with the natwidth and natheight options. This is common when trying to add a PDF figure without pdflatex.
% \begin{figure}
% \noindent\includegraphics[natwidth=800px,natheight=600px]{samplefigure.pdf}
%\caption{caption}
%\label{pdffiguresample}
%\end{figure}
%
%
% PDFLatex does not seem to be able to process EPS figures. You may want to try the epstopdf package.
%
%
% ---------------
% EXAMPLE TABLE
%
% \begin{table}
% \caption{Time of the Transition Between Phase 1 and Phase 2$^{a}$}
% \centering
% \begin{tabular}{l c}
% \hline
% Run & Time (min) \\
% \hline
% $l1$ & 260 \\
% $l2$ & 300 \\
% $l3$ & 340 \\
% $h1$ & 270 \\
% $h2$ & 250 \\
% $h3$ & 380 \\
% $r1$ & 370 \\
% $r2$ & 390 \\
% \hline
% \multicolumn{2}{l}{$^{a}$Footnote text here.}
% \end{tabular}
% \end{table}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% SIDEWAYS FIGURES and TABLES
% AGU prefers the use of {sidewaystable} over {landscapetable} as it causes fewer problems.
%
% \begin{sidewaysfigure}
% \includegraphics[width=20pc]{figsamp}
% \caption{caption here}
% \label{newfig}
% \end{sidewaysfigure}
%
% \begin{sidewaystable}
% \caption{Caption here}
% \label{tab:signif_gap_clos}
% \begin{tabular}{ccc}
% one&two&three\\
% four&five&six
% \end{tabular}
% \end{sidewaystable}
%% If using numbered lines, please surround equations with \begin{linenomath*}...\end{linenomath*}
%\begin{linenomath*}
%\begin{equation}
%y|{f} \sim g(m, \sigma),
%\end{equation}
%\end{linenomath*}
%%% End of body of article
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% Optional Appendices go here
%
% The \appendix command resets counters and redefines section heads
%
% After typing \appendix
%
%\section{Here Is Appendix Title}
% will show
% A: Here Is Appendix Title
%
%\appendix
%\section{Here is a sample appendix}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Optional Glossary, Notation or Acronym section goes here:
%
% Glossary is only allowed in Reviews of Geophysics
% \begin{glossary}
% \term{Term}
% Term Definition here
% \term{Term}
% Term Definition here
% \term{Term}
% Term Definition here
% \end{glossary}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Acronyms
%% NOTE that acronyms in the final published version will be spelled out when used in figure captions.
% \begin{acronyms}
% \acro{Acronym}
% Definition here
% \acro{EMOS}
% Ensemble model output statistics
% \acro{ECMWF}
% Centre for Medium-Range Weather Forecasts
% \end{acronyms}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Notation
% \begin{notation}
% \notation{$a+b$} Notation Definition here
% \notation{$e=mc^2$}
% Equation in German-born physicist Albert Einstein's theory of special
% relativity that showed that the increased relativistic mass ($m$) of a
% body comes from the energy of motion of the body—that is, its kinetic
% energy ($E$)—divided by the speed of light squared ($c^2$).
% \end{notation}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% DATA SECTION and ACKNOWLEDGMENTS
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section*{Open Research Section}
All datasets used in this publication are publicly available at \url{https://zenodo.org/records/18773512}. Code and instructions to reproduce the results of this manuscript can be found in \url{https://gitlab.earth.bsc.es/es/enso_diff}.
\section*{Conflict of Interest declaration}
The authors declare there are no conflicts of interest for this manuscript.
\acknowledgments
We thank the reviewers for their useful comments and suggestions. V.V. acknowledges funding from the European Union Horizon project EXPECT (grant 101137656).
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% REFERENCES and BIBLIOGRAPHY
%
% \bibliography{<name of your .bib file>} don't specify the file extension
% don't specify bibliographystyle
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\bibliography{DIFF_MANIFOLD.bib}
%Reference citation instructions and examples:
%
% Please use ONLY \cite and \citeA for reference citations.
% \cite for parenthetical references
% ...as shown in recent studies (Simpson et al., 2019)
% \citeA for in-text citations
% ...Simpson et al. (2019) have shown...
%
%
%...as shown by \citeA{jskilby}.
%...as shown by \citeA{lewin76}, \citeA{carson86}, \citeA{bartoldy02}, and \citeA{rinaldi03}.
%...has been shown \cite{jskilbye}.
%...has been shown \cite{lewin76,carson86,bartoldy02,rinaldi03}.
%... \cite <i.e.>[]{lewin76,carson86,bartoldy02,rinaldi03}.
%...has been shown by \cite <e.g.,>[and others]{lewin76}.
%
% apacite uses < > for prenotes and [ ] for postnotes
% DO NOT use other cite commands (e.g., \citet, \citep, \citeyear, \nocite, \citealp, etc.).
%
\end{document}
More Information and Advice:
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% SECTION HEADS
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Capitalize the first letter of each word (except for
% prepositions, conjunctions, and articles that are
% three or fewer letters).
% AGU follows standard outline style; therefore, there cannot be a section 1 without
% a section 2, or a section 2.3.1 without a section 2.3.2.
% Please make sure your section numbers are balanced.
% ---------------
% Level 1 head
%
% Use the \section{} command to identify level 1 heads;
% type the appropriate head wording between the curly
% brackets, as shown below.
%
%An example:
%\section{Level 1 Head: Introduction}
%
% ---------------
% Level 2 head
%
% Use the \subsection{} command to identify level 2 heads.
%An example:
%\subsection{Level 2 Head}
%
% ---------------
% Level 3 head
%
% Use the \subsubsection{} command to identify level 3 heads
%An example:
%\subsubsection{Level 3 Head}
%
%---------------
% Level 4 head
%
% Use the \subsubsubsection{} command to identify level 3 heads
% An example:
%\subsubsubsection{Level 4 Head} An example.
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% IN-TEXT LISTS
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Do not use bulleted lists; enumerated lists are okay.
% \begin{enumerate}
% \item
% \item
% \item
% \end{enumerate}
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% EQUATIONS
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Single-line equations are centered.
% Equation arrays will appear left-aligned.
Math coded inside display math mode \[ ...\]
will not be numbered, e.g.,:
\[ x^2=y^2 + z^2\]
Math coded inside \begin{equation} and \end{equation} will
be automatically numbered, e.g.,:
\begin{equation}
x^2=y^2 + z^2
\end{equation}
% To create multiline equations, use the
% \begin{eqnarray} and \end{eqnarray} environment
% as demonstrated below.
\begin{eqnarray}
x_{1} & = & (x - x_{0}) \cos \Theta \nonumber \\
&& + (y - y_{0}) \sin \Theta \nonumber \\
y_{1} & = & -(x - x_{0}) \sin \Theta \nonumber \\
&& + (y - y_{0}) \cos \Theta.
\end{eqnarray}
%If you don't want an equation number, use the star form:
%\begin{eqnarray*}...\end{eqnarray*}
% Break each line at a sign of operation
% (+, -, etc.) if possible, with the sign of operation
% on the new line.
% Indent second and subsequent lines to align with
% the first character following the equal sign on the
% first line.
% Use an \hspace{} command to insert horizontal space
% into your equation if necessary. Place an appropriate
% unit of measure between the curly braces, e.g.
% \hspace{1in}; you may have to experiment to achieve
% the correct amount of space.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% EQUATION NUMBERING: COUNTER
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% You may change equation numbering by resetting
% the equation counter or by explicitly numbering
% an equation.
% To explicitly number an equation, type \eqnum{}
% (with the desired number between the brackets)
% after the \begin{equation} or \begin{eqnarray}
% command. The \eqnum{} command will affect only
% the equation it appears with; LaTeX will number
% any equations appearing later in the manuscript
% according to the equation counter.
%
% If you have a multiline equation that needs only
% one equation number, use a \nonumber command in
% front of the double backslashes (\\) as shown in
% the multiline equation above.
% If you are using line numbers, remember to surround
% equations with \begin{linenomath*}...\end{linenomath*}
% To add line numbers to lines in equations:
% \begin{linenomath*}
% \begin{equation}
% \end{equation}
% \end{linenomath*}