Active learning and explicit electrostatics enable accurate modeling of electrolytes

TL;DR

Active learning-based training set generation with D-optimality and MaxVol enables system-specific MLIPs to model EC/EMC electrolytes and LiPF6 solutions with near-DFT accuracy at efficient computational cost. The study compares MTP and MTP-QRd to assess explicit electrostatics, finding that QRd can reach comparable or better accuracy with fewer parameters in EC/EMC mixtures, though stability may vary for LiPF6 solutions. AL yields diverse training sets automatically, producing stable densities within about 6% of experiment for pure solvents and 6–11% MAE for ionic conductivity across temperatures and compositions. The results establish a practical, transferable MLIP pipeline for electrolyte screening and highlight the potential of environment-aware electrostatics to improve data efficiency in MLIPs for complex electrolytes.

Abstract

Machine learning interatomic potentials (MLIPs) offer near-ab initio accuracy with the efficiency of classical force fields, making them attractive for modeling electrolytes. Collecting a diverse training set is essential for their accuracy and reliability, and explicit treatment of strong electrostatic interactions may be necessary. In this work, we demonstrated that active learning can automatically generate diverse training sets for moment tensor potentials (MTPs), enabling reliable molecular dynamics simulations of pure ethylene carbonate (EC), ethyl methyl carbonate (EMC), their mixtures, and LiPF6 solutions. The resulting MTPs exhibit excellent transferability across various EC/EMC compositions, producing ionic conductivities within 11% mean deviations from experiments. In addition, we assessed the impact of explicitly incorporating electrostatics by extending MTP with a charge redistribution scheme. Our results show that this extended MTP achieves accuracy comparable to MTP for EC/EMC mixtures with fewer parameters and reproduces ionic conductivity with only a 6% mean deviation from experiment.

Figures (16)

• Figure 1: a) Mean densities predicted by an ensemble of 3 MTPs, together with 1-$\sigma$ confidence intervals, compared to literature values magduau2023machine obtained by interpolating experimental data johnson1985properties; b) energy-density diagram for the training set of pure EC molecular liquid. The color coding represents index number $N$ of the sample in the training set. First 200 samples were extracted from AIMD, after that samples were selected in the active learning loop.
• Figure 2: Mean densities predicted by an ensemble of three MLIPs with 1-$\sigma$ confidence intervals, compared to literature values magduau2023machine obtained by interpolating experimental data johnson1985properties.
• Figure 3: Total, intra- and intermolecular energies per atom and forces, obtained with ensembles of three MTPs and MTP-QRd models of levels 12, 16, and 20 compared with PBE-D3 calculations for 7EC:3EMC validation set configurations.
• Figure 4: Intermolecular force error norms averaged over configurations taken from the 3EC:7EMC validation set.
• Figure 5: a) Ionic pair types; b) Ionic pair composition of training set; c) number of ligands in Li$^+$ solvation shells present in the training set; d) coordination number (CN) distributions in ionic pairs present in training set; d) Force error magnitudes averaged over validation set configurations; e) radial distribution function (RDF) Li-O for 8 LiPF$_6$ in 80 EC, predicted by MTP compared to the AIMD and ReaxFF ones ong2015lithium.