Accuracy and limitations of machine-learned interatomic potentials for magnetic systems: A case study on Fe-Cr-C

Machine-learned interatomic potentials (MLIPs) are rapidly becoming the standard for atomistic simulations, but their application to magnetic materials remains challenging because spin fluctuations must be treated either explicitly or implicitly. Here, we investigate this problem for the technologically important Fe-Cr-C system by constructing two deep machine learning potentials within the DeePMD framework: one trained on nonmagnetic DFT data (DP-NM) and one on spin-polarized DFT data (DP-M). Extensive validation against experiments reveals that DP-NM accurately reproduces dynamic, collective properties such as viscosity and melting temperatures, while DP-M excels in describing static, local properties such as density, especially for Fe-rich alloys. We rationalize these findings by noting that, at high temperatures, local magnetic moments self-average over space and time, making explicit spin treatment unnecessary for transport properties but essential for equilibrium volumes. To mitigate the high cost of spin-polarized DFT sampling, we employ a transfer-learning strategy, pretraining on nonmagnetic data followed by fine-tuning on a small spin-polarized dataset, which reduces computational expense by more than an order of magnitude. Furthermore, we benchmark several state-of-the-art foundation models: MACE, GRACE, DPA3 and fine-tuned variants of the latter against our specialized potentials. We find that foundation models offer competitive accuracy after fine-tuning but remain significantly slower in molecular dynamics simulations, limiting their practicality for large-scale transport property calculations. Our results establish clear design principles for MLIPs targeting magnetic alloys: (i) nonmagnetic training data suffice for dynamic properties of paramagnetic melts, and (ii) spin-polarized training is essential only for precise static properties of ferromagnetic phases. These insights provide a roadmap for efficient development of transferable MLIPs for magnetic systems and clarify the role of foundation models and transfer learning in accelerating this process.