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

Area of Science:

• Materials Science
• Nuclear Engineering
• Computational Physics

Background:

• Accurate thermal conductivity of uranium dioxide (UO2) is vital for nuclear reactor performance and safety.
• Experimental measurements are expensive, while density functional theory (DFT) is limited by scale.
• Traditional empirical potentials fail to capture complex anharmonic effects in UO2.

Purpose of the Study:

• To develop a machine-learned potential (NEP) for accurate UO2 thermal conductivity prediction.
• To evaluate and validate the NEP using various molecular dynamics techniques.
• To provide a reliable tool for multiscale thermal transport modeling in nuclear fuels.

Main Methods:

• Developed a neuroevolution potential (NEP) using active learning for near-DFT accuracy.
• Employed equilibrium molecular dynamics (EMD), homogeneous nonequilibrium molecular dynamics (HNEMD), and nonequilibrium molecular dynamics (NEMD) for validation.
• Systematically validated UO2 fundamental properties: equation of state, phonon dispersion, elastic constants, heat capacity, and thermal expansion.

Main Results:

• HNEMD demonstrated superior efficiency, robustness, and low uncertainty in thermal conductivity calculations.
• The NEP accurately reproduced experimental UO2 thermal conductivity temperature dependence (800-1500 K) and matched DFT+U accuracy (300-800 K).
• The validated NEP showed reliability and broad applicability for UO2 fundamental properties.

Conclusions:

• The developed NEP offers a computationally efficient and accurate method for UO2 thermal conductivity.
• This approach overcomes limitations of existing experimental and computational techniques.
• The study provides essential methodological support for nuclear fuel performance and reactor safety assessments.