A Hybrid Physics-Driven Neural Network Force Field for Liquid Electrolytes

Abstract

Electrolyte design plays an important role in the development of lithium-ion batteries and sodium-ion batteries. Battery electrolytes feature a large design space composed of different solvents, additives, and salts, which is difficult to explore experimentally. High-fidelity molecular simulation can accurately predict the bulk properties of electrolytes by employing accurate potential energy surfaces, thus guiding the molecule and formula engineering. At present, the overly simplified classic force fields rely heavily on experimental data for fine-tuning, thus its predictive power on microscopic level is under question. In contrast, the newly emerged machine learning interatomic potential (MLIP) can accurately reproduce the ab initio data, demonstrating excellent fitting ability. However, it is still haunted by problems such as low transferrability, insufficient stability in the prediction of bulk properties, and poor training cost scaling. Therefore, it cannot yet be used as a robust and universal tool for the exploration of electrolyte design space. In this work, we introduce a highly scalable and fully bottom-up force field construction strategy called PhyNEO-Electrolyte. It adopts a hybrid physics-driven and data-driven method that relies only on monomer and dimer EDA (energy deomposition analysis) data. With a careful separation of long/short-range and non-bonding/bonding interactions, we rigorously restore the long-range asymptotic behavior, which is critical in the description of electrolyte systems. Through this approach, we significantly improve the data efficiency of MLIP training, allowing us to achieve much larger chemical space coverage using much less data while retaining reliable quantitative prediction power in bulk phase calculations. PhyNEO-electrolyte thus serves as an important tool for future electrolyte optimization.

Figures (5)

• Figure 1: Overview of the PhyNEO-Electrolyte framework. Starting from electrolyte topology and coordinates, PhyNEO-Electrolyte decomposes total energy into nonbonding and bonding components. Nonbonding interaction are further divided into asymptotic part determined by monomer properties (i.e., multipole, polarizability, dispersion), Slater-type short-range part fitted to energy decomposition analysis, and pairwise neural network (NN) correction. Bonding interaction are modeled via a sub-graphs message-passing neural network. All energy components are summed to yield $\text{E}_\text{PhyNEO}$.
• Figure 2: Accuracy and transferability of PhyNEO-Electrolyte in describing nonbonding and bonding interactions. (a,b) Energy-distance profiles for intermolecular interaction of EC dimers and Li-DEC dimer, comparing PhyNEO’s decomposed components (exchange, electrostatic, polarization, dispersion, direct Hartree-Fock) with SAPT(DFT) benchmarks, with root mean square deviation (RMSD) metrics labeled. Insets illustrate molecular structures. (c) Heatmap of RMSD (kJ/mol) of the physics-driven part ($E_{nb}^{lr} + E_{nb}^{sr}$) of PhyNEO across a diverse set of molecules, demonstrating cross-system transferability. (d) shows the distance scan profile of Li-BOB ion dimer with/without pairwise ML nonbonding correction. (e) comparison between PhyNEO with/without pairwise ML nonbonding correction to true DFT energy, and (f) shows overall bonding energy fitting capability of sGNN.
• Figure 3: Densities predicted by PhyNEO, QRNNdajnowicz_high-dimensional_2022, OPLS4lu_opls4_2021, and BAMBOOgong2025predictive, in comparison with experimental results. Panel (a) shows the density comparison of pure electrolytes solvents and additives, with PhyNEO’s mean prediction error indicated. (b) shows the density comparison of PhyNEO with other ML and conventional force fields, QRNN, OPLS4, BAMBOO and BAMBOO with density alignment (labeled BAMBOO AL), for pure solvents. (c) shows the temperature-dependent density profiles for binary solvent systems, comparing PhyNEO, GAP-MDmagdau_machine_2023, and experiments. (d) shows the density prediction error for electrolyte systems with/without pairwise ML correction in PhyNEO. (e) shows the scatter plot of predicted densities for PhyNEO and OPLS across Li-saltwang2016superconcentrateddave2022autonomous and Na-salt electrolytesmonti2020towards, comparing to experimental results.
• Figure 4: Panel (a) shows temperature dependence of diffusivity of $\text{Li}^{+}$, comparing OPLS4, QRNN, PhyNEO and experimental datahayamizu_temperature_2012 in $\text{LiPF}_6$ + DEC electrolyte. (b) displays the performance relative to experimental resultsuchida_what_2021 for various EC ratios in a LiPF6 + EC + DMC mixture.
• Figure 5: Transferability and data efficiency of PhyNEO electrolyte force field when adding new molecules. (a) illustrates a new molecule (e.g., SL) is added to the library of molecules already covered by PhyNEO (EC, FEC, etc.). (b) shows the RMSD (kJ/mol) as a function of training data amount (number of pairs), showing rapid error reduction with a small increase in training pairs as (a) depicted. (c) is the average RMSD (kJ/mol) vs. number of training pairs from homo-dimers to hetero-dimers within two independent training process. (d) demonstrates the transferability of Pairwise ML nonbonding correction term from Li-EC pair to Li-FEC and Li-PC pairs with similar atomtypes. (e) shows the sGNN could maintain the chemical accuracy prediction of DMC solvent without DMC data, which means sGNN bonding energy prediction could be transferred from the similar bond types. (f) Comparison of covered molecules and dataset size across different electrolyte force fields (APPLE&Pborodin_polarizable_2009,QRNNdajnowicz_high-dimensional_2022, BAMBOOgong2025predictive, PhyNEO)