Atomistic modeling of uranium monocarbide with a machine learning interatomic potential

TL;DR

This paper addresses the need for accurate, scalable atomistic modeling of uranium monocarbide (UC) under reactor-relevant conditions. It develops a machine learning interatomic potential (MLIP) based on the HIP-NN framework with tensor sensitivity, trained via an active-learning loop on DFT+U data to cover diverse UC configurations, including defects. The resulting MLIP achieves close agreement with DFT+U predictions for lattice parameters, elastic constants, defect formation energies, and diffusion barriers, and enables large-scale molecular dynamics and diffusion analyses inaccessible to direct DFT+U. While some defect-diffusion activation energies are modestly overestimated, the MLIP captures essential trends, such as faster carbon diffusion relative to uranium and defect-assisted diffusion pathways, offering a practical tool for UC qualification in advanced nuclear fuels. Overall, this work provides a robust, efficient pathway to explore UC behavior at high temperatures and defect concentrations, accelerating reactor-fuel design and safety analyses.

Abstract

Uranium monocarbide (UC) is an advanced ceramic fuel candidate due to its superior uranium density and thermal conductivity compared to traditional fuels. To accurately model UC at reactor operating conditions, we developed a machine learning interatomic potential (MLIP) using an active learning procedure to generate a comprehensive training dataset capturing diverse atomic configurations. The resulting MLIP predicts structural, elastic, thermophysical properties, defect formation energies, and diffusion behaviors, aligning well with experimental and theoretical benchmarks. This work significantly advances computational methods to explore UC, enabling efficient large-scale and long-time molecular dynamics simulations essential for reactor fuel qualification.

Figures (9)

• Figure 1: Perspective view of the UC crystal structure in the AFM-I phase and magnetic moments for uranium along the (001) direction (red arrows). The U atoms and C atoms are represented by blue and gray spheres, respectively.
• Figure 2: Phonon band structures for (a) NM ordering and (b) AFM ordering using DFT, SCAN and DFT+U with different $U_{\mathrm{eff}}$ values. Experimental data from Ref. jackman1986 included as black points. AFM structure with $U_{\mathrm{eff}}$=1.25 eV calculations capture experimental optical modes better than the other values.
• Figure 3: Parity plots comparing ensemble HIP-NN predictions with DFT+U reference data for (a) Energy per atom correlation plot showing ensemble-mean predictions versus DFT+U, and (b) Force parity plot comparing ensemble-mean component forces to DFT+U values. Colors indicate the logarithm of the point density.
• Figure 4: Phonon band structure as predicted by the MLIP (red). For comparison, the phonons from DFT+$U_{\mathrm{eff}}=1.25$ eV calculation on AFM structure is shown (blue). Experimental data from Ref. jackman1986 included as black points.
• Figure 5: Comparison of lattice parameter as a function of temperature obtained from HIP-NN, DFT+U-MD, and experimental data from Ref. meddez1964.