Comparing fine-tuning strategies of MACE machine learning force field for modeling Li-ion diffusion in LiF for batteries

TL;DR

This work benchmarks the MACE-MPA-0 foundation model against a DeePMD benchmark for Li interstitial diffusion in LiF, focusing on data-efficient fine-tuning strategies. Using MD-based diffusivity and Arrhenius analyses, the authors show that fine-tuning with a small DeePMD-derived dataset (around a few hundred configurations) or even training from MACE-MPA-0-data alone can achieve activation energies near the reference value, with D0 and diffusivity converging toward DeePMD-like behavior. The results highlight the practical potential of foundation-model–driven MLIPs for battery SEI materials, where carefully designed training sets yield accurate transport properties with substantially reduced data requirements. Overall, the study provides actionable insights into dataset design, active learning, and fine-tuning strategies that enable efficient, accurate modeling of Li diffusion in LiF for improved battery performance and safety.

Abstract

Machine learning interatomic potentials (MLIPs) are transforming materials science and engineering by enabling the study of complex phenomena, such as those critical to battery operation. In this work, we benchmark the MACE machine learning model against a well-trained DeePMD potential for predicting interstitial lithium diffusivity in LiF, a key component in the solid electrolyte interphase in Li ion batteries. Our results demonstrate that the MACE-MPA-0 foundational model achieves comparable accuracy to well-trained DeePMD, in predicting key diffusion properties based on molecular dynamics simulation, while requiring minimal or no training data. For instance, the MACE-MPA-0 predicts an activation energy Ea of 0.22 eV, the fine-tuned model with only 300 data points predicts Ea = 0.20 eV, both of which show good agreement with the DeePMD model reference value of Ea = 0.24 eV. In this work, we provide a solid test case where fine-tuning approaches - whether using data generated for DeePMD or data produced by the foundational MACE model itself - yield similar robust performance to the DeePMD potential trained with over 40,000 actively learned data, albeit requiring only a fraction of the training data.

Figures (4)

• Figure 1: Arrhenius plot for activation energy $E_a$ calculated with MACE-MPA-0 and our fine-tuned model (incorporating 200 DeePMD data points deangelis2025 and 100 MPtraj pre-training data points, ft200-pt100) from 9 ns MD trajectories. For reference, we include DeePMD results found in Ref. deangelis2025.
• Figure 2: The diffusivity at 400 K and 450 K computed from 3 ns trajectories for different fine-tuned MACE-MPA-0 models trained using number of fine-tuning data points between 110 to 800 and fixed pre-training dataset of 1000 configuration for all the cases. The dashed lines are the DeePMD reference value deangelis2025.
• Figure 3: The diffusivity at 400 K and 450 K computed from 3 ns trajectories for different fine-tuned MACE models trained with different number of pre-training data ranging from 0 to 10,000 configurations and fixed number of fine-tuning data of (a) 800 data points and (b) 710 data points from DeePMD model generated dataset deangelis2025. Note that the 800 data points contains the same 710 datapoints with 90 additional LiF bulk structures. The dashed lines are the DeePMD reference value deangelis2025.
• Figure 4: The diffusivity at 400K and 450K calculated from 3 ns trajectories for FT1 and FT2 models. The dashed lines are the DeePMD reference value deangelis2025.