| Abstract Scope |
In recent years, machine learning interatomic potentials (MLIAPs) have emerged as an alternative to traditional interatomic potentials for molecular dynamics (MD) simulations. While classical potentials lack the accuracy of quantum approaches such as Density Functional Theory (DFT), MLIAPs combine quantum-level precision with the efficiency of classical methods. In this work, we present a new MLIAP for Fe-C systems, using the Neuroevolutional Potential (NEP) method [1]. We curated a database of multiple Fe–C structures, including point defects, carbides, and interfaces. For the short range interactions, needed for studies of collision cascades, the ZBL potential was included. We discuss the requirement of pseudopotentials with more valence electrons o ensure smooth NEP–ZBL transitions for high-energy configurations and proper defect production in collision cascades [2].
[1] Z. Fan, et al, Phys. Rev. B., 2021 vol. 104 pag. 104309
[2] A.Fellman, et al, Computational Materials Science, 2026, vol. 262, p. 114398. |