📝 SummarXiv

Abstract

Machine-learned force fields (MLFFs), particularly pre-trained foundation models, promise to bring ab initio-level accuracy to the length and time scales of molecular dynamics. Yet this shift raises a central question: is it better to build a specialist model from scratch or adapt a generalist foundation model for a specific system? The trade-offs in data efficiency, predictive accuracy, and risks of out-of-distribution (OOD) failure remain unclear. Here, we present a benchmarking framework that contrasts bespoke (from scratch) and fine-tuned foundation models in a test case of a technologically relevant 2D material, Cr-intercalated Sb2Te3, using the MACE architecture. Our framework employs migration pathways, evaluated through nudged elastic band (NEB) trajectories, as a diagnostic probe that tests both interpolation and extrapolation. We assess accuracy for equilibrium, kinetic (atomic migration), and mechanical (interlayer sliding) tasks. While all models capture equilibrium structures, predictions for non-equilibrium processes diverge. Task-specific fine-tuning substantially improves kinetic accuracy compared with both from-scratch and zero-shot models, but can degrade learned representations of long-range physics. Analysis of internal representations shows that training paradigms yield distinct, non-overlapping latent encodings of system physics. This work offers a practical guide for MLFF development, highlights migration-based probes as efficient diagnostics, and suggests pathways toward uncertainty-aware active learning strategies.