Are Universal Potentials Ready for Alkali-Ion Battery Kinetics?

Xingyu Guo, Cheng Gui, Zhenbin Wang · 2026

Read Are Universal Potentials Ready for Alkali-Ion Battery Kinetics? by Xingyu Guo, Cheng Gui, Zhenbin Wang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-10938.txt.

Introduction

Xingyu Guo, Cheng Gui, and Zhenbin Wang benchmark universal machine learning potentials for battery kinetics. Faster alkali-ion battery discovery needs accurate models of atomic-scale motion. Shared potentials promise one model for many materials. How far they can be trusted on steep energy landscapes, the terrain of kinetics, is still unsure. The authors test that trust head on.

They compare M3GNet, CHGNet, MACE, SevenNet, GRACE, and Orb against DFT baselines. The test materials are cathodes and solid electrolytes. Kinetics governs charging speed and lifetime, so errors there mislead design the most. The benchmark tells battery modelers which shared tools are ready for kinetic work and which still lag.

Catalog record

CREATORXingyu Guo, Cheng Gui, Zhenbin Wang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE42 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-10938

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Are Universal Potentials Ready for Alkali-Ion Battery Kinetics? Xingyu Guo, Cheng Gui, Zhenbin Wang 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. † † email: [email protected] † † affiliation: Department of Data Science, City University of Hong Kong, Hong Kong SAR, 999077, China † † affiliation: Hong Kong Institute of AI for Science, Hong Kong SAR, 999077, China † † affiliation: Hong Kong Institute of AI for Science, Hong Kong SAR, 999077, China † † affiliation: Department of Materials Science and Engineering, City University of Hong Kong, Hong Kong SAR, 999077, China † † affiliation: School of Energy and Environment, City University of Hong Kong, Hong Kong SAR, 999077, China 1 Introduction Alkali-ion batteries (AIBs) are central to the sustainable energy transition, with applications spanning portable electronics to grid-scale storage. 12 The performance of these devices is fundamentally governed by the atomic-scale behavior of their electrode and electrolyte materials. Although thermodynamics dictates the open-circuit voltage and theoretical capacity of battery materials 23 , it is kinetics, including ion transport within bulk phases and across interfaces 3 ; 24 , phase boundary dynamics 37 , and electrochemical reactions 52 , that ultimately determines practical power density and operational lifetime. Understanding and predicting these kinetic phenomena is therefore essential for overcoming the rate-limiting steps that currently constrain the development of high-performance battery chemistries. Density functional theory (DFT) has long served as the computational foundation for investigating battery kinetics at the atomic scale. 13 ; 56 However, its substantial computational cost restricts simulations to relatively small system sizes and short timescales, creating a fundamental “scale gap” that prevents effective sampling of the complex diffusion networks governing ion transport in realistic battery materials. This limitation has driven an urgent demand for computational approaches that achieve DFT-level accuracy while enabling large-scale screening and comprehensive kinetic modeling. Recent advances in machine learning interatomic potentials (MLIPs) provide a robust strategy to bridge this gap, offering near-quantum accuracy at a fraction of the computational cost 29 . Specifically, universal MLIPs (uMLIPs), trained on vast datasets spanning the periodic table, have emerged as viable general-purpose surrogates for DFT. The predictive fidelity of uMLIPs for both thermodynamic properties (e.g., formation energies) 20 ; 51 and dynamic properties (e.g., phonon spectra) 38 ; 20 has seen remarkable improvement. As illustrated in Figure 1, this progress is driven by the co-evolution of model architectures and training datasets. On the architectural front, the field has transitioned from invariant graph neural networks (GNNs), such as M3GNet, to sophisticated equivariant frameworks including NequIP 7 , MACE 5 ; 6 , and GRACE 39 . In parallel, training datasets have evolved from the near-equilibrium structures of Materials Project Trajectories 31 ; 17 to the high-diversity Alexandria 49 , and OMat datasets 4 . This synergistic integration has enabled state-of-the-art (SOTA) models to achieve exceptional accuracy, yielding minimal deviations in energies, forces, and stresses compared to DFT benchmarks. Despite these advances, the capability of uMLIPs to predict kinetic properties remains largely unexplored. Kinetic processes inherently involve high-energy transition states, yet most uMLIPs are trained and validated predominantly on near-equilibrium or ground-state structures. Consequently, it remains unclear whether these models can reliably capture diffusion kinetics and related phenomena such as defect migration. Moreover, architectural limitations in many SOTA models, including non-conservative force predictions and discontinuities in predicted potential energy surfaces 21 , may further compromise their accuracy for kinetic property predictions. Addressing this knowledge gap is critical for developing improved uMLIPs suitable for both thermodynamic and kinetic predictions. In this work, we systematically benchmark the predictive capabilities of uMLIPs for modeling ion transport in battery materials using both nudged elastic band (NEB) calculations and molecular dynamics (MD) simulations. For migration barrier predictions, we find that the Orb‑v3 model achieves the lowest errors among all tested models, indicating that architectural improvements contribute more significantly to accuracy gains than expanding training data for static barriers. In contrast, for MD simulations, the GRACE model trained on OMat24 demonstrates the best overall performance, accurately reproducing mean square displacements (MSDs), ion diffusivities, and radial distribution functions (RDFs). Collectively, these results show that current state‑of‑the‑art uMLIPs enable predictive modeling of materials kinetics without system‑specific fine‑tuning. More broadly, our analysis reveals an interplay between model architecture and training data quality: while architectural sophistication is critical for energetic barrier predictions, data scale and diversity, particularly the inclusion of high‑temperature and non‑equilibrium structures, are essential for reliably capturing dynamic transport properties. 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. 2 Results 2.1 Benchmark Dataset Construction A diverse dataset of well-known cathode materials for both Li- and Na-ion batteries was curated to benchmark the accuracy of uMLIPs in predicting ion migration barriers. As detailed in Table 1, this dataset comprises three major electrode structural families: layered, olivine, and maricite. For layered materials, both single- and di-vacancy migration mechanisms (Figure S1) were investigated, as the latter has been reported as the dominant transport pathway due to its significantly lower energy barrier. For olivine and maricite structures, only the single-vacancy mechanism was evaluated. The dataset’s chemical diversity was enhanced by incorporating various transition metals (Ni, Co, Ti, Mn, Fe) and anions (O vs. S) for both Li and Na ions within each structural framework. Additionally, to facilitate a direct comparison between alkali ions, migration barriers were calculated for each material alongside its corresponding Li- or Na-ion counterpart. Materials (A = Li or Na) Structure Prototype Space Group Migration Mechanism \text{ATiS}{\vphantom{\text{X}}}_{\smash[t]{\text{2}}} Layered R \bar{3} m Single vacancy; Di-vacancy \text{ATiS}{\vphantom{\text{X}}}_{\smash[t]{\text{2}}} Layered P \bar{3} m1 Single vacancy; Di-vacancy \text{AMO}{\vphantom{\text{X}}}_{\smash[t]{\text{2}}} (M = Ni, Co, Ti) Layered R \bar{3} m Single vacancy; Di-vacancy \text{AMPO}{\vphantom{\text{X}}}_{\smash[t]{\text{4}}} (M = Mn, Fe, Co, Ni) Olivine Pnma Single vacancy \text{AMPO}{\vphantom{\text{X}}}_{\smash[t]{\text{4}}} (M = Mn, Fe, Co, Ni) Maricite Pnma Single vacancy Table 1: Selected electrode materials for Li + and Na + migration barriers evaluation. The solid electrolyte materials listed in Table 2 were selected to provide a chemically and structurally diverse benchmark set for evaluating uMLIP performance in predicting ion diffusivity. This selection represents key material families currently under active study for all-solid-state batteries, including sulfides, argyrodites and halides. These materials encompass a wide range of bonding environments and

Read the complete text