Influence of Exchange-Correlation Functionals and Neural Network Architectures on Li$^+$-Ion Conductivity in Solid-State Electrolyte from Molecular Dynamics Simulations with Machine-Learning Force Fields

Abstract

With the rapid advancement of machine learning techniques for materials simulations, machine-learned force fields (MLFFs) have become a powerful tool that complements first-principles calculations by enabling high-accuracy molecular dynamics simulations over extended timescales. Typically, MLFFs are trained on data generated from density functional theory (DFT) using a specific exchange-correlation (XC) functional, with the goal of reproducing DFT-level properties. However, the uncertainties in MLFF-based simulations--arising from variations in both MLFF model architectures and the choice of XC functionals--remain insufficiently understood. In this work, we construct MLFF models of different architectures trained on DFT data from both semilocal and hybrid functionals to describe Li$^+$ diffusion in the solid-state electrolyte Li$_6$PS$_5$Cl. We systematically investigate how different XC functionals influence the Li$^+$ diffusion coefficient. To reduce statistical uncertainty, the mean squared displacements are averaged over 300 independent molecular dynamics (MD) trajectories of 70 ps each, yielding statistical variations below $1\%$. This enables a clear assessment of the respective influences of the functional and the MLFF model. Due to its tendency to underestimate band gaps and migration barriers, the semilocal functional predicts consistently higher Li$^+$ diffusion coefficients, compared to the hybrid functional. Furthermore, comparisons among various neural network methods reveal that the differences in predicted diffusion coefficients arising from different network architectures are of the same order of magnitude as those caused by different functionals, indicating that the choice of the network model itself substantially influences the MLFF predictions. This observation calls from an urgent need for standardized protocols to minimize model-dependent biases in MLFF-based MD.