Cross-scale prediction of oxidation behavior of uranium alloys based on deep learning

Uranium alloys are irreplaceable structural and functional materials in strategic equipment. However, the strong correlation characteristics of 5f electrons unique to actinide elements result in high chemical reactivity and environmental sensitivity, making the investigation of their corrosion behavior critically important. This study focuses on the oxidation kinetics of the U-2.5Nb alloy and proposes a deep learning-based cross-scale prediction method for the oxidation performance of uranium alloys. Microscopic features of the uranium alloy were extracted through molecular dynamics (MD) simulations and combined with experimental oxidation weight gain data to construct a predictive model bridging microscopic structures and macroscopic oxidation behavior. In the MD simulations, a supercell model of U-2.5Nb was first built, and oxygen molecules were introduced on its surface. Molecular dynamics simulations were then performed to obtain key microscopic parameters. To enable cross-scale modeling, a random forest algorithm was employed to evaluate the importance of multiple MD-derived features. Multi-model comparative experiments were designed and implemented to train these features, and finally, an Long Short-Term Memory (LSTM) neural network was utilized to predict the oxidation behavior curves from limited microscopic data. The model's performance was comprehensively evaluated using metrics such as root mean square error, coefficient of determination, and mean absolute error. • A cross-scale framework correlates atomistic MD features with U–2.5Nb alloy macroscopic oxidation behavior. • MD-derived features (diffusion, potential energy, VACF) are filtered via random forest for deep learning. • LSTM accurately predicts oxidation kinetics from limited microdata, enabling robust actinide alloy oxidation modeling.