Active learning-enhanced neuroevolution potential for predictive modeling of UO2 thermophysical properties

Abstract

Accurately characterizing the temperature dependence of UO2 thermal conductivity is crucial for evaluating its performance under nuclear reactor operating conditions. However, experimental measurements are costly, density functional theory (DFT) calculations are constrained by small spatiotemporal scales, and traditional empirical potentials struggle to capture strong anharmonic effects. To this end, we developed a machine-learned neuroevolution potential (NEP) with near-DFT accuracy using an active learning strategy, and we systematically evaluated and cross-validated the thermal conductivity of UO2 within equilibrium molecular dynamics (EMD), homogeneous nonequilibrium molecular dynamics (HNEMD), and nonequilibrium molecular dynamics (NEMD). The results demonstrate that HNEMD delivers a high signal-to-noise ratio, low uncertainty, and rapid convergence, exhibiting superior computational efficiency and robustness. At 500 K, the spectral phonon mean free path span approximately one order of magnitude, and heat-transport channel lengths exceeding about 5 μm approach the bulk thermal conductivity limit. In the 800–1500 K range, the NEP reproduces the experimental temperature dependence of UO2 thermal conductivity, while at lower temperatures (300–800 K), it achieves predictive accuracy comparable to DFT+U. Through systematic validation of UO2 fundamental properties including the equation of state, phonon dispersion relations, elastic constants, heat capacity, and linear thermal expansion coefficient, demonstrates that the constructed NEP is reliable and broadly applicable. This work provides methodological support for multiscale thermal transport modeling of nuclear fuels and reactor safety assessment.