Atomistic modeling of uranium monocarbide with a machine learning interatomic potential

Abstract

Uranium monocarbide (UC) is an advanced ceramic fuel candidate due to its high uranium density and thermal conductivity. Here we develop a machine learning interatomic potential (MLIP) for rocksalt UC using the HIP-NN architecture trained to DFT+U reference data generated via an active learning workflow. An ensemble of eight independently trained models is used for uncertainty-guided data selection and to quantify model-to-model variability in predicted properties. The resulting MLIP provides a computationally efficient surrogate for DFT+U within its training domain, reproducing the DFT+U-MD thermal expansion over 300–2500 K and capturing key elastic and vibrational features of UC, including the acoustic phonon branches. The model also reproduces qualitative trends in defect energetics and defect-assisted diffusion, while we identify specific limitations: optical phonon frequencies are underestimated and selected interstitial migration barriers are non-quantitative due to limited transition-state coverage in the current training set. This MLIP provides a practical framework for large-scale molecular dynamics studies of near-stoichiometric rocksalt UC and intrinsic point-defect processes at reactor-relevant temperatures.