An AI-ready fine-tuning framework for accurate machine-learning interatomic potentials in solid-solid battery interfaces

TL;DR

The paper tackles the data- and cost-intensive challenge of achieving quantum-accurate interatomic potentials for solid-solid battery interfaces. It introduces FIRE, a fine-tuning framework that combines replay-augmented continual learning with an efficient sampling pipeline to adapt a foundation interatomic potential to task-specific interfacial systems. Across six representative interfaces, FIRE achieves energy RMSE below 1 meV/atom and force RMSE near 20 meV/Å using roughly 10% of the data, while faithfully reproducing electrochemical and mechanical properties in agreement with experiments. The approach offers a general, data-efficient pathway to predictive atomistic modeling of complex interfaces, enabling simulations beyond the reach of first-principles methods and extending applicability to diverse interfacial materials.

Abstract

Atomistic modeling of solid-solid battery interfaces is essential for understanding electro-chemo-mechanical coupling, but the complex interfacial chemistry and heterogeneous environments pose major challenges for quantum-accurate, data-efficient modeling. Herein, we propose an approach of fine-tuning with integrated replay and efficiency (FIRE), a general framework for universal machine-learning interatomic potentials by combining efficient configurational sampling with a replay-argumented continual strategy, achieving quantum-level accuracy at moderate cost. Across six solid-solid battery interface systems, FIRE consistently achieves root-mean-square errors in energy below 1 meV/atom and in force near 20 meV/angstrom, marking an order-of-magnitude improvement over existing models while requiring only 10% of the original datasets. In addition, the fine-tuned model successfully reproduces key mechanical and electrochemical properties of the materials, in close agreement with experimental data. The FIRE offers a generalizable and data-efficient approach for developing accurate interatomic potentials across diverse materials, enabling predictive simulations beyond the reach of first-principles methods.

Figures (4)

• Figure 1: Overview of the FIRE framework: (A) comparison between from-scratch training and fine-tuning paradigms; (B) replay-augmented continual fine-tuning strategy that leverages both pretraining and task-specific datasets; (C) efficient sampling workflow, including a pre-fine-tuned model for robust MD generation, dimensionality reduction, and diversity-based selection of representative configurations.
• Figure 2: Model evaluation of the FIRE approach: (A) atomic configurations of the six battery materials studied in this work, from left to right: Na/Na3SbS4, LiCl/GaF3, Li2CO3/LiF, Li3PS4/Li3B11O18, Li/Li6PS5Cl, and Li_7La_3Zr_2-xM_xO_12 ($M = \mathrm{Ta}, \mathrm{Nb}$); (B) the corresponding force predicted by fine-tuned model with respect to that of DFT calculation; (C) model evaluation via RMSE for energy and force, respectively. For comparison, results collected from literature and vanilla fine-tuning are shown.
• Figure 3: Data efficiency of the FIRE approach: (A–B) PCA projections of sampled configurations for Na/Na3SbS4 and Li_7La_3Zr_2-xM_xO_12 ($M = \mathrm{Ta}, \mathrm{Nb}$), showing broader structural coverage with increasing dataset size; (C–D) data-efficiency benchmarks: replay-argumented fine-tuning achieves the lowest energy and force errors with reduced computational cost, outperforming vanilla fine-tuning, random sampling, and from-scratch training.
• Figure 4: Property predictions from FIRE-generated fine-tuned MLIP models. (A) First column: Optimized structures of T-Na3SbS4, LLZTO, and LPSCl. Second column: Mean square displacement (MSD) curves from MACE-MD simulations at relevant temperatures. Third column: Room-temperature impedance spectra with equivalent circuit fits. (B) Comparison of predicted Young's modulus and Poisson’s ratio against experimental values. Young’s modulus was obtained by spatially averaging AFM nanomechanical maps. Reference values of Poisson’s ratios are taken from: T-Na3SbS4 Awais2024Na3SbX3Poster, LLZTO yuElasticPropertiesSolid2015, and LPSCl AEL_Chen. (C) Scaling of MD wall time per simulation step. MLIP-based MD maintains nearly constant cost up to 3456 atoms, in contrast to the cubic scaling behavior of VASP-based MD.