libNLPBE: An Open-Source Python Package for Solving the Non-Linear Poisson-Boltzmann Equation

Interactions between molecules and their environment play a pivotal role in controlling a wide range of chemical phenomena. For example, it is well known that solvent effects critically influence stability, reactivity, and spectroscopic properties of the solute molecule.[49, 12] These effects become more prominent when the solvent contains electrolytes, where mobile ions in the solution give rise to interesting interfacial phenomena.[27, 42, 53] Thus, elucidating the effects of ions on the solute is crucial for understanding the the shift of its chemical properties.
Theoretical insights into these changes would help design chemistry performed in electrolyte solutions. For example, molecular dynamics simulations could be carried out to understand how electrolytes influence the solute molecule. This approach, however, requires a simulation box densely filled with solvent molecules and electrolytes. Consequently, their degrees of freedom generate a vast number of configurations, resulting in significant computational complexity in obtaining a converged free energy of the system.
In contrast to explicit modeling of electrolyte solutions, implicit solvent models replace the solvent environment with a polarizable continuum, thereby eliminating the complexity of sampling the enormous configurational space generated by explicit models.[38, 49, 35] Among these, the Poisson-Boltzmann equation (PBE) has emerged as an efficient implicit model for describing electrolyte solution environments.[39, 44, 36, 47, 30, 18, 40, 45, 21] The PBE treats ions in the solution as a charge density distributed over space according to Boltzmann statistics, where the associated Boltzmann factor is determined by electrostatic potentials.
Depending on whether the Boltzmann factor is linearized, the PBE is formulated as either the linearized (LPBE) or the non-linear Poisson-Boltzmann equation (NLPBE). The LPBE produces reasonable results under weak electrostatic potentials; however, it inherently fails to accurately describe ion distributions under strong electrostatic potentials. This limitation is evident particularly when describing solute-solution interfaces, where the solute generates a strong electrostatic potential in its vicinity.
Therefore, the NLPBE is necessary to accurately describe ion distributions around the solute molecule. Accordingly, efficient program packages for solving the NLPBE have been developed over the past decades. DelPhi[31] and APBS[25] are among the most popular software packages, both of which employ an atomic charge representation of the solute charge distribution. While this approach produces reasonable results for huge biomolecular systems, it might be insufficient for achieving atomistic accuracy that electronic structure calculations pursue. Some NLPBE solvers support electronic structure calculations such as BigDFT,[18] JDFTx,[30], and DL_MG[54]; however, these are primarily designed for quantum calculations under the periodic boundary conditions, which might be inappropriate for describing molecular systems. To the best of our knowledge, Q-Chem is the only software package that supports the NLPBE solver integrated with molecular calculations.[43, 45]
The limited availability of molecular NLPBE solvers motivated us to develop a library that can be easily integrated with molecular electronic structure codes. Furthermore, we aim to reduce the computational cost of solving the NLPBE, thereby expanding its applicability to a broad range of chemistry performed in electrolyte solutions.
Therefore, we present an open-source Python library for efficiently solving the NLPBE, namely libNLPBE. Our library features the density fitting (DF) approximation for calculating solute electrostatic potentials, the modified Damped Inexact Newton Multigrid developed by Holst (mDINMH), and GPU acceleration, all of which address the major computational bottlenecks in solving the NLPBE.
In the following, we first discuss the expression for the dielectric function and the size-modified ion charge density appearing in the NLPBE (Section II.1). Subsequently, we introduce the finite difference method for numerically solving the NLPBE (Section II.2). Next, we provide the theoretical background of the DF approximation for efficiently calculating solute electrostatic potentials (Section II.3) and the mDINMH method for solving the NLPBE (Section II.4). In Section IV, we show that both the DF approximation and the mDINMH algorithm significantly accelerate solving the NLPBE for molecules. Lastly, we provide hands-on examples of using libNLPBE for electronic structure calculations (Section V).

All density functional calculations[22, 28, 37] were carried out at B3LYP-D3/def2-TZVPPD level of theory[6, 29, 50, 5, 20, 51] as implemented in PySCF.[46] Resolution of the identity approximation with def2-TZVPP-JKFIT auxiliary basis set[52] was generally adopted to accelerate density functional calculations. The analytic and the DF approaches for $\phi^{\text{sol}}_{\text{elec}}(\mathbf{r})$ were implemented using PySCF[46] and GPU4PySCF[32] for CPU and GPU version of libNLPBE, respectively. The algebraic multigrid method in the mDINMH method was implemented using AMGCL library[14, 16, 15] for both CPU and GPU version of libNLPBE. After extensive numerical experiments, we found that 4-level V-cycle multigrid with the sparse approximate inverse relaxation of the zeroth variant (SPAI-0) relaxation method[8] delivers optimal efficiency, During travering the multigrid hierarchy, pre- and post-smoothening were performed once at each level, and the coarsest level was directly solved.
For the dielectric and the ion-exclusion functions (Equation (2) and (7)), we used $\Delta=0.265$ Å, $a=0.44$ Å, and $d_{\text{probe}}=1.4$ Å according to the literature, [45] and the room temperature ($T=298.15$ K) is generally considered throughout this work.
We enforced the convergence threshold of $||\mathcal{G}[\phi^{\text{tot}}_{k}]||<1.0\times 10^{-9}$ for the mDINMH. To demonstrate the performance of the mDINMH, we used the self-consistent (SC) method as the reference method,[18, 10, 9, 45] and, to the best of our knowledge, it is the sole method that has been applied to molecular calculations.[45, 26] Its algorithm[18, 10, 9, 45] and implementation details[26] can be found elsewhere.
We note that the SC method was also implemented using AMGCL library, which allows rigorous performance comparison between the SC and the mDINMH methods. We also note that the convergence scheme for the SC method is different from that of mDINMH. To be specific, the SC method uses the ion charge density $\rho^{\text{ions}}(\mathbf{r})$ and the polarization charge density $\rho^{\text{pol}}(\mathbf{r})$ for convergence test, where the latter is defined as ($\rho^{\text{tot}}(\mathbf{r})=\rho^{\text{sol}}(\mathbf{r})+\rho^{\text{ions}}(\mathbf{r})$)

Convergence of the SC method is examined by testing whether $\Delta\rho^{\text{ions}}(\mathbf{r})$ and $\Delta\rho^{\text{pol}}(\mathbf{r})$ are smaller than their respective threshold value for all grid points. We set $\Delta\rho^{\text{pol}}(\mathbf{r})=1.0\times 10^{-5}$ a.u. and $\Delta\rho^{\text{ions}}(\mathbf{r})=1.0\times 10^{-6}$ a.u. for these threshold values, which are two orders of magnitude tighter than those in the original work.[45] Accordingly, the convergence threshold for the conjugate gradient step on each multigrid level was set to $1.0\times 10^{-8}$ a.u.. The reference data were calculated by Q-Chem 5.4.[43] For consistency with PySCF-calculated DFT results, 75 radial and 302 angular grids for numerical integrations and 50 radial and 194 angular grids for non-linear correlations. All computational results reported in this paper were produced using 8 threads of Intel Xeon Gold 6240 CPU, a NVIDIA GeForce RTX 2080 Ti 11GB GPU card, and 10000 MB of memory.

Density Fitting. Table 1 summarizes the computational time for calculating $\phi^{\text{sol}}_{\text{elec}}(\mathbf{r})$ of vitamin C and 4-nitroaniline using the analytic and the density fitting (DF) approaches on CPU and GPU for 97$\times$97$\times$97 cubic grids. The DF approach on CPU shows dramatic speedups of $\times$81.98 and $\times$68.39 for vitamin C and 4-nitroaniline, respectively, both of which highlight the efficiency of the DF approach.
These accelerations are enabled by a reduced number of ERIs involved in $\phi^{\text{sol}}_{\text{elec}}(\mathbf{r})$ as described in Section II.3 (See Equation (15) and (19)). Comparing the ratio of the numbers of ERIs required by the analytic and the DF approaches (See Table 1), the theoretical maximum speedup is estimated to be $\times$169.60 and $\times$132.20 for vitamin C and 4-nitroaniline, respectively. The actual speedup is less than the theoretical estimate because the entire calculation is divided into smaller batches to reduce the memory cost, resulting in additional computational overhead.
This overhead becomes more noticeable when the calculations are performed on GPU (See Table 1), where the actual speedup is $\times$9.87 and $\times$6.33 for vitamin C and 4-nitroaniline, respectively. The lower speedups observed on GPUs are expected because the analytic approach benefits more from GPU parallelization than the density fitting approach due to a more ERIs for evaluating $\phi^{\text{sol}}_{\text{elec}}(\mathbf{r})$.
For vitamin C and 4-nitroaniline, the CPU-based analytic approach required 1164.88 and 750.23 seconds for $\phi^{\text{sol}}_{\text{elec}}(\mathbf{r})$, whereas the GPU-accelerated DF approach took only 3.86 and 3.74 seconds. These numbers suggest $\times$301.78 and $\times$200.60 acceleration for vitamin C and 4-nitroaniline, confirming the efficiency of both the DF approach and GPU acceleration for computing $\phi^{\text{sol}}_{\text{elec}}(\mathbf{r})$ on grids.

mDINMN. We next evaluate the performance of the mDINMH method by comparing it with the self-consistent (SC) method.[18, 10, 9, 45] Vitamin C and 4-nitroaniline were again chosen as representative examples, and Table 2 summarizes the computational times required to solve the NLPBE on CPU and GPU.
Overall, the mDINMH (10.03 and 10.37 seconds for vitamin C and 4-nitroaniline) outperforms the SC approach (41.84 and 43.87 seconds for vitamin C and 4-nitroaniline) for both molecules in terms of computing time, approximately a 4-fold speedup on the CPU. This speedup clearly demonstrates the efficiency of the mDINMH over the SC approach.
Furthermore, the GPU-accelerated mDINMH shows a remarkable computing time, which only takes 0.57 and 0.59 seconds for vitamin C and 4-nitroaniline, respectively (Table 2). Compared to the SC method on CPU, these results correspond to $\times$73.40 and $\times$74.36 acceleration for vitamin C and 4-nitroaniline, respectively, highlighting the efficiency of GPU parallelization. While we intentionally omitted implementing the GPU-accelerated version of the SC approach due to the efficiency of the mDINMN on CPU, we do not expect that it would outperform the GPU-accelerated mDINMH, given the performance gap observed on CPU.

Convergence Rate. The superior efficiency of the mDINMH over the SC method is primarily attributed to its faster convergence rate. To investigate the convergence rate, we plotted the relative residual ($||\mathcal{G}[\phi^{\text{tot}}_{k+1}]||/||\mathcal{G}[\phi^{\text{tot}}_{k}]||$) of both methods according to their iterative steps as illustrated in Figure 4.
The relative residual of the SC method (blue dots in Figure 4) remains nearly constant until convergence at the 32nd iteration, except for a kink at the second iteration due to a huge residual at the first iteration. This behavior suggests a linear convergence rate for the SC method, which often requires a large number of iterations to converge.
In contrast, the relative residual of the mDINMH (orange dots in Figure 4) shows a plateau until the 9th iteration, followed by an abrupt drop from the 10th iteration until convergence at the 13th iteration. This convergence pattern evidences superlinear convergence of the inexact Newton method enabled by the $Q$-order($1+p$) convergence condition (Equation (26)),[24] which supports faster convergence of the mDINMH than the SC method as observed in Figure 4. As a result, the mDINMH needs fewer iterations to achieve convergence, thereby significantly reducing the computational time as shown in Table 2.

Algebraic Multigrid. We next discuss the applicability of the mDINMH to various numbers of grid points along $x$, $y$, and $z$-axis ($N_{x}$, See Figure 2) by examining their convergence according to iterative cycles, and the results are depicted in Figure 5. Note that $N_{x}$ values in Figure 5 were carefully chosen to demonstrate the versatility of the AMG method across a broad range of grid points (See Section II.4.3).
Specifically, the GMG hierarchy in the original DINMH is contructed from the underlying grid structure, where the grid spacing is doubled at each coarser level and halved at each finer level when traversing the hierarchy (Figure 3).[24, 9, 54] Consequently, the original DINMH achieves optimal efficiency when $N_{x}=2^{n}+1$ ($n\in\mathbb{N}$), a condition satisfied by $N_{x}$ = 65, 129, and 257 in Figure 5. In contrast, $N_{x}$ = 100, 150, and 200 in Figure 5 are even-numbered grid points, which the original DINMH is limited in its ability to handle.
Our mDINMH method, however, is capable of handling both cases by employing AMG as illustrated in Figure 5. The absolute residual successfully reaches convergence within 20 iterations for all $N_{x}$ considered in Figure 5, suggesting the stability of the mDINMH across a broad range of grid points enabled by AMG.

Polarization and Ion Charge Densities. Based on the efficiency of the DF approach (Table 1) and the mDINMH method (Table 2), we consider their combination (DF+mDINMH) as our default method in solving the NLPBE. In this section, we first verify the qualitative accuracy of the DF+mDINMH method. Specifically, we examine $\rho^{\text{pol}}(\mathbf{r})$ (Equation (31)) and $\rho^{\text{ions}}(\mathbf{r})$ (Equation (8)), both of which are derived from the DF+mDINMH-calculated $\phi^{\text{tot}}(\mathbf{r})$. Both quantities were computed for 4-nitroaniline in 1 M NaCl aqueous solution, and the results are illustrated in Figure 6.
The calculated $\rho^{\text{pol}}(\mathbf{r})$ (Figure 6, left) exhibits a positively charged surface around the nitro group (–NO₂). This positive $\rho^{\text{pol}}(\mathbf{r})$ is primarily attributed to the nitro-oxygen atoms, which adopt a partial negative charge in nitro moieties. These negatively charged oxygen atoms are expected to induce a positive charge in the environment to maintain local charge neutrality, resulting in a positive $\rho^{\text{pol}}(\mathbf{r})$ in their vicinity. Our speculation is supported by the restrained electrostatic potential (RESP) atomic charges (Table 3),[4] where the oxygen atoms carry a negative charge of –0.500. Similarly, the calculated $\rho^{\text{ions}}(\mathbf{r})$ (Figure 6, right) shows a positive value near the nitro moiety, suggesting the accumulation of cations caused by the negatively charged oxygen atoms.
In contrast, both $\rho^{\text{pol}}(\mathbf{r})$ (Figure 6, left) and $\rho^{\text{ions}}(\mathbf{r})$ (Figure 6, right) show negative values around the amino group (–NH₂), which result from the positively charged amino-hydrogen atoms. The RESP charge of the these atoms is +0.383 (Table 3), suggesting that the hydrogen atoms develop a negative charge in the environment while increasing the local anion concentration around them.
Overall, the distributions of the mDINMH-calculated $\rho^{\text{pol}}(\mathbf{r})$ and $\rho^{\text{ions}}(\mathbf{r})$ are consistent with those reported by Head-Gordon and coworkers,[45] suggesting the reliability of our mDINMH method.

Solvation Free Energy. While $\rho^{\text{pol}}(\mathbf{r})$ and $\rho^{\text{ions}}(\mathbf{r})$ can demonstrate the qualitative accuracy of the DF+mDINMH method, they are insufficient for validating its quantitative accuracy. Thus, we calculated the free energy of solvation (See Appendix A for its derivation) to demonstrate the quantitative accuracy of the DF+mDINMH method, and the results are summarized in Table 4. Specifically, the electrostatic contribution to the solvation free energy ($\Delta G^{\text{solv}}_{\text{es}}$) was computed for 4-nitroaniline in 1 M NaCl aqueous solution using four different computational schemes (analytic/DF for calculating $\phi^{\text{sol}}_{\text{elec}}(\mathbf{r})$ and SC/mDINMH for solving the NLPBE), and the calculated energies are compared with those from Q-Chem. See Section III for computational details.
We first confirm the accuracy of the DF approach by comparing $\Delta G^{\text{solv}}_{\text{es}}$ computed with the analytic and the DF approaches (Table 4). Both the SC and the mDINMH methods yield $\Delta G^{\text{solv}}_{\text{es}}$ of –4.73 for the analytic approach and –4.70 kcal/mol for the DF approach. Their difference is 0.03 kcal/mol only, which is negligible in practice. These results quantitatively confirm the accuracy of the DF approach.
Furthermore, Table 4 shows that the SC and the mDINMH methods produce the same $\Delta G^{\text{solv}}_{\text{es}}$ up to two digits irrespective to the computational approach for $\phi_{\text{elec}}^{\text{sol}}(\mathbf{r})$. Specifically, both methods produce –4.73 kcal/mol with the analytic approach and –4.70 kcal/mol with the DF approach. These agreements suggest the quantitative accuracy of the mDINMH method. The accuracy of the DF approach and the mDINMH method is further confirmed by $\Delta G^{\text{solv}}_{\text{es}}$ of vitamin C in 1 M NaCl aqueous solution (Table 5), where all four combinations results the same $\Delta G_{\text{es}}^{\text{solv}}$ of –5.26 kcal/mol.
We next evaluate the consistency of the DF+mDINMH method against the NLPBE solver in Q-Chem software (hereafter referred to as reference method).[43, 45] For 4-nitroaniline and vitamin C, the reference method produces $\Delta G_{\text{es}}^{\text{solv}}$ of –4.74 (Table 4) and –5.29 kcal/mol (Table 5), while the DF+mDINMH results –4.70 (Table 4) and –5.26 kcal/mol (Table 5). The deviations (0.04 and 0.03 kcal/mol for 4-nitroaniline and vitamin C, respectively) are practically negligible, suggesting that the DF+mDINMH is capable of producing reliable solvation free energies.
While the NLPBE has been developed primarily for modeling electrolyte solutions, it is also applicable to pure solvents by setting $c^{\text{b}}=0$ M. This enables us to compare the NLPBE-calculated $\Delta G_{\text{es}}^{\text{solv}}$ with those from popular implicit solvent models, such as C-PCM,[3] IEF-PCM,[48] COSMO,[38] and SMD.[35] For their rigorous comparison, we followed the standard PCM convention for the solvent-accessible surface, where atom-specific lengths are represented by scaled van der Waals radii $d_{\text{A}}=1.2\times d_{\text{A,vdW}}$.
Using these scaled van der Waals radii for solvent-accessible surfaces, we computed the solvation free energies of 4-nitroaniline and vitamin C in pure water using various implicit solvent models, and the results are summarized in Table 6. The C-PCM, IEF-PCM, and COSMO models produce $\Delta G_{\text{es}}^{\text{solv}}$ of approximately –15 and –23 kcal/mol for 4-nitroaniline and vitamin C, respectively (first to third row of Table 6). The SMD model (fourth row of Table 6) gives a comparable result of –14.98 kcal/mol for 4-nitroaniline, but predicts a less negative $\Delta G_{\text{es}}^{\text{solv}}$ of –16.18 kcal/mol for vitamin C, an outlier among the energies computed by standard solvation models considered in this study.
In contrast, the NLPBE-calculated $\Delta G_{\text{es}}^{\text{solv}}$ (fifth row of Table 6) is –32.00 and –53.15 kcal/mol for 4-nitroaniline and vitamin C, respectively. Compared to the standard solvent models, the NLPBE negatively overestimate $\Delta G_{\text{es}}^{\text{solv}}$ to a large extent. This overestimation is attributed to, unlike standard solvent models, smooth $\epsilon(\mathbf{r})$ at the solvent-accessible surface as defined in Equation (2), which consequently allows a non-zero $\rho^{\text{pol}}(\mathbf{r})$ (See Equation (31) and Figure 6) inside the solute-accessible surface (See Figure 1). Given the distance from the solute, this non-zero $\rho^{\text{pol}}(\mathbf{r})$ leads to strong electrostatic interactions with the solute charge density, thereby significantly contributing to the solvation free energy.
Therefore, we increased the scaling factor for atom-specific lengths from 1.20 to 1.45 ($d_{\text{A}}=1.45\times d_{\text{A,vdW}}$) to minimize such interactions and recalculated the solvation free energy. As summarized in the sixth row of Table 6, $\Delta G_{\text{es}}^{\text{solv}}$ is computed to be –14.77 and –21.51 kcal/mol for 4-nitroaniline and vitamin C, respectively. These values are which are consistent with the values calculated by the C-PCM, IEF-PCM, COSMO, and SMD models, except for the SMD-calculated $\Delta G_{\text{es}}^{\text{solv}}$ for vitamin C as discussed above.

We move on to ensuring the accuracy of the DF+mDINMH method by evaluating the effect of the electrolyte concentration to the solvation free energy of acetic acid [45] by calculating $\Delta\Delta G_{\text{ions}}$ defined as

We computed $\Delta\Delta G_{\text{ions}}$ with the DF+mDINMH and the reference method, and the results are depicted in Figure 7 (See Table S1 for raw data).

Similar to the work by Head-Gordon and coworkers (Figure 7, Orange line), [45] both the DF+mDINMH and the reference method predict $\Delta\Delta G_{\text{ions}}$ as a non-linear function of electrolyte concentration (Figure 7, Blue line), which consequently breaks the experimental linear trend (Figure 7, Gray line).[33] This deviation is well expected, as our theoretical framework also assumed a concentration-independent Stern layer thickness ($a$) when defining the ion-exclusion function ($\lambda(\mathbf{r})$, Equation (7)).[45] Aside from this deviation, the DF+mDINMH-calculated $\Delta\Delta G_{\text{ions}}$ values (Figure 7, Blue line) are in excellent agreement with those from the reference method (Figure 7, Orange line),[45] confirming the numerical consistency of the DF+mDINMH combination with the reference method.

To further validate the consistency of the DF+mDINMH method with the reference method, we computed $\Delta\Delta G_{\text{ions}}$ for 2-octanone with varying the Stern layer thickness (See Equation (7)).[45] The results are depicted in Figure 8 (See Table S2–S4 for raw data).
Consistent with the reference method (Figure 8b), our DF+mDINMH method predicts not only an increasing $\Delta\Delta G_{\text{ions}}$ as the electrolyte concentration increases but also the trend of $\Delta\Delta G_{\text{ions}}$ vs. electrolyte concentration curve with respect to the Stern layer thickness (Figure 8a). The difference between the DF+mDINMH- and the reference method-calculated $\Delta\Delta G_{\text{ions}}$ is almost negligible as depicted in Figure 8c.
In summary, we have demonstrated that the DF+mDINMH implemented in libNLPBE provides an efficient and accurate framework for predicting electrostatic potential profiles that molecules in electrolyte solutions experience. Therefore, we expect that libNLPBE serves as a powerful library for correcting electrostatic effects in density functional calculations.

We wrap up with providing simple hands-on examples to help users employ libNLPBE library for their own applications. Our libNLPBE library is primarily designed to be fully integrated into PySCF SCF calculations. Specifically, the library computes the solvation correction term that is incorporated into the Fock matrix during each SCF cycle. To facilitate this integration, we provide pbe_for_scf function, which couples the solvent (cm) and the SCF (mf) objects by overriding the latter’s kernel function as shown in Listing 1.

For other quantum chemistry packages, libNLPBE supports single-shot NLPBE calculations to obtain the solvation free energy and the solvation correction term to the Hamiltonian as shown in Listing 2. This requires a molden file containing information about molecular structure, orbital coefficients, and orbital occupancies, which are subsequently used to construct the solute density matrix (dm in Listing 2) for calculating $\phi^{\text{sol}}_{\text{elec}}(\mathbf{r})$. Users can specify an auxiliary basis set when building NLPBE object; otherwise libNLPBE uses def2-universal-JKFIT auxiliary basis set by default. The kernel function of NLPBE object returns the solvation free energy and the Hamiltonian correction term (epbe and vmat in Listing 2), which users can feed into their own code. Considering that most quantum chemistry codes support exporting SCF information into a molden file, we expect libNLPBE to be applicable to a broad range quantum chemical platforms.

We developed libNLPBE, an open-source Python library for solving the non-linear Poisson-Boltzmann equation (NLPBE) for molecular systems. Our library features the density fitting (DF) approximation for calculating solute electrostatic potential, the mDINMH algorithm for efficiently solving the NLPBE, an algebraic multigrid (AMG) method to support various number of grid points, and GPU acceleration.
Using vitamin C and 4-nitroaniline as representative examples, we demonstrated that, compared to the analytic approach, the DF approach accelerates solute electrostatic potential calculations by $\sim$82 times on CPU and $\sim$302 times with GPU acceleration.
To account for the asymmetric Jacobian operator appearing in the Newton equation, we introduced a symmetric preconditioner in the mDINMH, which enables the application of efficient multigrid methods developed for symmetric operators. Unlike the original DINMH that requires symmetrization of $\nabla\epsilon(\mathbf{r})\cdot\nabla$ through discretization, our symmetric preconditioner allows the mDINMH to use of the analytic gradient of $\epsilon(\mathbf{r})$, which is expected to reduce discretization errors. We showed that, compared to the self-consistent method for solving the NLPBE, the mDINMH algorithm achieves $\sim$10 and $\sim$74-fold speedups on CPU and GPU, respectively.
We examined the performance of the mDINMH across various numbers of grid points ($N_{x}$) to highlight AMG as a black-box method for solving the Newton equation. We showed that mDINMH calculations converged within 20 iterations for all $N_{x}$ considered in this study, confirming the applicability of the mDINMH to a broad range of grid counts.
Based on the performance of the DF approach and the mDINMH method, we propose a DF+mDINMH combination as an efficient computational scheme for solving the NLPBE. Its qualitative accuracy was examined by inspecting the distributions of the polarization charge density ($\rho^{\text{pol}}(\mathbf{r})$) and the ion charge density ($\rho^{\text{ions}}(\mathbf{r})$) around 4-nitroaniline. We showed that these are consistent with the calculated atomic charges, where the negatively charged nitro-oxygen atoms develop a positive $\rho^{\text{pol}}(\mathbf{r})$ and $\rho^{\text{ions}}(\mathbf{r})$ whereas the positively charged amino-hydrogen atoms do the opposite. We also compared $\Delta G_{\text{es}}^{\text{solv}}$ of 4-nitroaniline and vitamin C computed by the DF+mDINMH method at zero electrolyte concentration and popular solvent models such as C-PCM, IEF-PCM, COSMO, and SMD. The calculated results demonstrate that, after adjusting the scaling factor for atom-specific lengths, the DF+mDINMH method produces reasonable $\Delta G_{\text{es}}^{\text{solv}}$ compared to that from other solvent models.
We also examined the quantitative accuracy of the DF+mDINMH by calculating the electrostatic contribution to the solvation free energy ($\Delta G^{\text{solv}}_{\text{es}}$) of 4-nitroaniline and vitamin C. The calculated $\Delta G^{\text{solv}}_{\text{es}}$ under the DF approach differs by 0.03 and 0.00 kcal/mol for 4-nitrianiline and vitamine C, respectively, from those under the analytic approach. Furthermore, the mDINMH-calculated $\Delta G^{\text{solv}}_{\text{es}}$ does not show a difference from the SC method up to two decimal points, both of which confirm the accuracy of the DF+mDINMH method.
To further ensure its accuracy, we investigated the effect of electrolyte concentration on the solvation free energy using the DF+mDINMH approach. We found that the calculated free energy vs. electrolyte concentration curve is in great agreement with that produced by the reference method, confirming the accuracy of the DF+mDINMH method.
Our libNLPBE library can be fully integrated with PySCF SCF calculations. For other quantum chemistry code, libNLPBE offers single-shot calculations to obtain the solvation free energy and the solvation correction to the Hamiltonian, provided a molden file that can be prepared by most quantum chemistry program. Therefore, we expect libNLPBE to be applicable to a broad range of quantum chemistry platforms.

The source code of libNLPBE is available at https://github.com/Yang-Laboratory/libNLPBE.