Spin Neural Network Potential for Magnetic Phase Transitions in Uranium Dioxide

TL;DR

This paper introduces a spin neural network potential for uranium dioxide that models magnetic phase transitions by incorporating spin and spin-orbit effects, enabling accurate large-scale simulations of its magnetic behavior.

Contribution

The study develops a novel machine learning potential that explicitly includes spin degrees of freedom and spin-orbit coupling for uranium dioxide, facilitating accurate modeling of magnetic phase transitions.

Findings

Accurately reproduces DFT energies, forces, and lattice constants.

Successfully captures the antiferromagnetic-paramagnetic transition.

Predicts transition temperature with correct order of magnitude.

Abstract

Uranium dioxide (UO2) is a prototypical nuclear fuel material, yet predicting its thermophysical properties across a wide temperature range remains challenging. One factor contributing to this difficulty is the complex magnetic ordering at low temperatures, where spin-orbit coupling produces strong coupling between spin and lattice degrees of freedom. Direct DFT simulations of magnetic phase transitions at finite temperatures are computationally prohibitive. Here, we develop a spin neural network potential (SpinNNP) that explicitly incorporates spin degrees of freedom together with spin-orbit coupling to describe the magnetic states of UO2.Reference datasets were generated using magnetic constrained DFT+U calculations with spin-orbit coupling, covering a wide range of non-collinear spin configurations. The SpinNNP accurately reproduces DFT energies, atomic forces, spin forces, and lattice constants. Machine learning molecular dynamics simulations with spin dynamics successfully capture the antiferromagnetic-paramagnetic transition. Although the predicted magnetic ground state differs from experiment due to known limitations of the underlying DFT description, the transition temperature obtained is of the correct order of magnitude compared with experiment. These results demonstrate that machine-learning potentials can enable large-scale spin-lattice simulations of actinide oxides and provide a practical route toward predictive modeling of complex magnetic materials.