Molecular dynamics simulation of nitrogen diffusion in iron and iron nitrides using <i>ab initio</i> data trained machine learning potentials

Ammonia is receiving increasing attention as a hydrogen energy carrier and a green fuel for achieving carbon neutrality. Combustors utilizing ammonia as a fuel are subject to "unwanted" surface nitriding, which hardens and embrittles the metal walls of combustors. Predicting the nitridation rate and nitride layer thicknesses with Fick's diffusion law requires temperature- and concentration-dependent nitrogen atoms' diffusion coefficients in the relevant iron and iron nitrides. Experimentally determining the diffusion coefficients is challenging due to steep N-gradient, and concurrent phase transformations. In this study, we calculated the nitrogen atom diffusion coefficients using molecular dynamics (MD) driven by a machine learning interatomic potential (MLP). The MLP was trained on snapshots extracted from <i>ab initio</i> molecular dynamics (AIMD) trajectories covering different temperatures and nitrogen atom concentrations, and the contribution of each subset to the model's accuracy was quantified. The results show that the MLP reproduces independently the calculated diffusion energy barriers obtained using density functional theory (DFT) in α-, γ-iron and γ'-, ε-iron nitrides. For ε-iron nitride, the MLP-driven MD-computed self-diffusion coefficient was converted into chemical (Fickian) diffusion coefficients, enabling direct comparison with experiments. Our MLP-driven MD simulation delivers diffusion coefficients approaching <i>ab initio</i> accuracy, enabling precise concentration- and temperature-resolved modeling of nitrogen diffusion in iron and its nitrides. Arrhenius fitting to the 830-1500 K chemical diffusion coefficients reproduce the experimental activation energy and pre-exponential factor. Extrapolation to 823 K, the temperature at which most of the experiments are performed, yields a coefficient that falls within the experimental uncertainty. The present MLP-MD model will facilitate the mechanistic understanding of iron surface nitridation and quantitative predictions of "unwanted" nitriding of iron-based combustor wall materials induced by ammonia at high temperatures.