Accuracy and Limitations of Machine-Learned Interatomic Potentials for Magnetic Systems: A Case Study on Fe-Cr-C

Abstract

Machine-learned interatomic potentials (MLIPs) have become the gold standard for atomistic simulations, yet their extension to magnetic materials remains challenging because spin fluctuations must be captured either explicitly or implicitly. We address this problem for the technologically vital Fe-Cr-C system by constructing two deep machine learning potentials in DeePMD realization: one trained on non-magnetic DFT data (DP-NM) and one on spin-polarised DFT data (DP-M). Extensive validation against experiments reveals a striking dichotomy. The dynamic, collective properties, viscosity and melting temperatures are reproduced accurately by DP-NM but are incorrectly estimated by DP-M. Static, local properties, density, and lattice parameters are captured excellently by DP-M, especially in Fe-rich alloys, whereas DP-NM fails. This behaviour is explained by general properties of paramagnetic state: at high temperature, local magnetic moments self-average in space and time, so their explicit treatment is unnecessary for transport properties but essential for equilibrium volumes. Exploiting this insight, we show that a transfer-learning protocol, pre-training on non-magnetic DFT and fine-tuning on a small set of spin-polarised data, reduces the computational cost to develop magnetic MLIPs by more than an order of magnitude. Developing general-purpose potentials that capture static and dynamic behaviors throughout the whole composition space requires proper accounting for temperature-induced spin fluctuations in DFT calculations and correctly incorporating spin degrees of freedom into classical force fields.