A machine-learning interatomic potential for Si-C-N-H with application to polymer-to-ceramic processing of polysilazanes

Abstract

We developed a machine-learning interatomic potential (MLIP) based on Moment Tensor Potentials for atomistic simulations in the Si-C-N-H system. The MLIP was trained on ordered and disordered configurations—including crystalline phases, polymers, amorphous models, and high-temperature ab initio molecular dynamics trajectories of chemical reactions—and achieves Density–Functional–Theory-level accuracy for energies, forces, and stresses. We apply the MLIP to complex processes that were previously inaccessible at scale: self-diffusion in Si3N4, negative thermal expansion in Si(NCN)2, fracture in SiC/Si3N4 composites, and the polymer-to-ceramic transformation of polysilazanes. Large-scale simulations over nanoseconds capture structural evolution during pyrolysis and reproduce experimental observations, such as the formation of nanometer-sized graphitic segregations in SiCN ceramics. Our results demonstrate that the MLIP extends quantum-level accuracy to technologically relevant problems, providing new opportunities for predictive modeling of ceramic materials and their transformations.