Are Universal Potentials Ready for Alkali-Ion Battery Kinetics?

TL;DR

This work benchmarks state-of-the-art universal machine-learning interatomic potentials (uMLIPs) for alkali-ion battery kinetics, addressing whether these models can faithfully reproduce atomic-scale migration barriers and diffusion without material-specific fine-tuning. By evaluating NEB-predicted barriers and MD-derived transport properties across cathode and solid-electrolyte chemistries, the authors show that architectural advances (notably Orb-v3) yield the best static barrier accuracy ($MAE \approx 75$–$111$ meV), while training data diversity—especially non-equilibrium, high-temperature samples (OMat24)—drives dynamic diffusion fidelity, with GRACE trained on OMat24 delivering the best MD predictions. The study reveals a robust interplay between model architecture and data quality: sophisticated equivariant architectures excel in energetic landscapes, but large, diverse datasets dominate when capturing finite-temperature kinetics. Collectively, the results establish modern uMLIPs as reliable zero-shot surrogates for high-throughput kinetic screening in next-generation energy-storage materials, enabling efficient exploration of diffusion networks and barrier-controlled processes at scale.

Abstract

Accelerating alkali-ion battery discovery requires accurate modeling of atomic-scale kinetics, yet the reliability of universal machine learning interatomic potentials (uMLIPs) in capturing these high-energy landscapes remains uncertain. Here, we systematically benchmark state-of-the-art uMLIPs, including M3GNet, CHGNet, MACE, SevenNet, GRACE, and Orb, against DFT baselines for cathodes and solid electrolytes. We find that the Orb-v3 family excels in static migration barrier predictions (MAE $\approx$ 75--111 meV), driven primarily by architectural refinements. Conversely, for dynamic transport, the GRACE model trained on the OMat24 dataset demonstrates superior fidelity in reproducing ion diffusivities and structural correlations. Our results reveal that while architectural sophistication (e.g., equivariance) is beneficial, the inclusion of high-temperature, non-equilibrium training data is the dominant driver of kinetic accuracy. These findings establish that modern uMLIPs are sufficiently robust to serve as zero-shot surrogates for high-throughput kinetic screening of next-generation energy storage materials.

Figures (5)

• Figure 1: Overview of the uMLIP development ecosystem. The schematic illustrates the progression from training datasets, categorized by scale and thermodynamic state, through a range of training strategies to distinct model architectures. Architectures are grouped by their geometric symmetry properties (invariant, equivariant, and non‑equivariant), with representative models and defining structural features highlighted for each category.
• Figure 2: Number of NEB calculation failures across various uMLIP models. Failures are categorized as (top) Type I: computational divergence and (bottom) Type II: unphysical barriers (where $E_{\mathrm{TS}} < E_{\mathrm{initial/final}}$).
• Figure 3: Mean absolute error (MAE) of energy barriers predicted by various uMLIP models relative to DFT benchmarks. The models are grouped by architecture, with bar patterns indicating the specific training dataset used.
• Figure 4: Comparison of ion migration barriers predicted by uMLIPs, $E_{\mathrm{a}}^{\mathrm{uMLIP}}$ with reference DFT values, $E_{\mathrm{a}}^{\mathrm{DFT}}$. Panels (a–i) show results for different uMLIP architectures and training data, including invariant, equivariant, and non‑equivariant models trained on distinct datasets. Each point corresponds to a unique ion migration pathway obtained from NEB calculations. The dashed line indicates perfect agreement between uMLIP and DFT predictions. Mean absolute errors (MAEs) and coefficients of determination (R$^2$) are reported in each panel.
• Figure 5: Performance of uMLIP molecular dynamics (MD) simulations compared with ab initio molecular dynamics (AIMD) in predicting ion transport properties. (a) Mean absolute error (MAE) of log(MSD(Å$^2$)) of Li+/Na+ ions over 50 ps of MD simulations at 800 K and 1200 K;(b) MAE of the calculated Li+/Na+ diffusivity (10$^{-3}$ cm$^2$/s) at 800 K and 1200 K. (c) MAE of radial distribution functions (RDF), averaged over MD trajectories between 30 and 50 ps at 800 K and 1200 K.