Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Perspective on Training Machine Learning Force Fields for Solid-State Electrolyte Materials

Zihan Yan, Shengjie Tang, Yizhou Zhu

Abstract

Machine learning force fields enable high-accuracy modeling of solid-state electrolytes (SSEs). This perspective evaluates dataset size, reference quality, and model architectures. We show that rigid SSE frameworks favor efficient learning, prioritizing data quality over quantity. Crucially, force RMSE does not reliably predict transport performance. By analyzing locality and benchmarking frameworks, we provide practical guidelines to accelerate the development of next-generation solid-state batteries.

A Perspective on Training Machine Learning Force Fields for Solid-State Electrolyte Materials

Abstract

Machine learning force fields enable high-accuracy modeling of solid-state electrolytes (SSEs). This perspective evaluates dataset size, reference quality, and model architectures. We show that rigid SSE frameworks favor efficient learning, prioritizing data quality over quantity. Crucially, force RMSE does not reliably predict transport performance. By analyzing locality and benchmarking frameworks, we provide practical guidelines to accelerate the development of next-generation solid-state batteries.
Paper Structure (6 sections, 1 equation, 5 figures)

This paper contains 6 sections, 1 equation, 5 figures.

Table of Contents

  1. Introduction
  2. How large a dataset is needed to train MLFFs for SSEs?
  3. What quality of DFT reference data is necessary?
  4. How Long-range Interactions Affect Force Field Accuracy and Performance?
  5. How much impact does including long-range interactions in MLFF actually have on SSE?
  6. Summary and Outlook

Figures (5)

  • Figure 1: Dataset-size sensitivity tests on MLFF performance in Li7La3Zr2O_12 (LLZO), Li3YCl6 (LYC), and Li_10GeP2S_12 (LGPS). The validation root mean square errors of (a) energy and (b) force. (c) The diffusivities of Li-ions (1000K for LLZO, 500K for LYC and LGPS). (d) The activation energies of Li-ions.
  • Figure 2: (a-c) Force error distributions for LLZO, LYC, and LGPS comparing $\Gamma$-point DFT ($\Gamma$-DFT) and two NEP models. NEP is trained on high-accuracy data and $\Gamma$-NEP is trained on $\Gamma$-DFT data. The high-accuracy DFT data are used as reference. (d-f) Arrhenius plots of Li-ions diffusivities calculated by NEP and $\Gamma$-NEP models.
  • Figure 3: (a-c) Relative RMSE of energies and forces for LLZO, LYC, and LGPS, predicted by different models. (d-f) Arrhenius plots of Li-ions diffusivities for the corresponding materials, calculated using different models.
  • Figure 4: Scaling of (a) simulation speed and (b) total wall-clock time required for $10^6$ MD steps as a function of the number of atoms. All tests were performed on the LLZO system using a single NVIDIA V100 GPU (32 GB).
  • Figure 5: (a) Schematic of force locality in atom-centered many-body potentials (adapted from Ref deringer2021gaussian): the displacement of atom A influences the force on atom C via their shared neighbor B, extending the force locality to twice the cutoff radius ($2\times R_{cut}$). (b) Schematic of the locality test: atoms within $2\times R_{cut}$ of the central atom are fixed, while atoms in the outer region are subjected to MD perturbations. (c–e) Locality errors versus $2\times R_{cut}$ for LLZO, LYC, and LGPS, respectively.