计算机科学
铀
人工智能
机器学习
工艺工程
算法
计算学习理论
高效算法
核工程
材料科学
主动学习(机器学习)
作者
Rui Zhou,Yilin Fang,Guiyun Hang,Zexin Jiang,Huiming Shen,Zhipan Ma,Wanxiao Guo,Jintao Wang,Wenli Yu,Tao Wang,Ning Gao
标识
DOI:10.1016/j.nme.2026.102129
摘要
A machine learning interatomic potential (MLIP) applicable to three representative uranium oxide phases (i.e., UO 2 , γ-UO 3 , and α-U 3 O 8 ) was developed for uranium–oxygen (U-O) systems using the on-the-fly machine learning methodology. The accuracy and reliability of the MLIP were rigorously validated through systematic benchmarking against UO 2 , γ-UO 3 , and α-U 3 O 8 phases. Key structural and thermophysical parameters derived from MLIP simulations exhibit excellent agreement with existing theoretical and experimental studies: Lattice constants deviate by less than 1%, elastic moduli show maximum discrepancies of 10.7% compared to theoretical calculations, and Gibbs free energy values align within 0.277–0.589% deviation at 1000 K. Notably, the thermal conductivities of UO 2 and α-U 3 O 8 match well with reported experimental and computational values. This work presents a comprehensive thermodynamic characterization of UO 3 polymorphs (α, β, γ, δ, η) through phonon dispersion analysis and Gibbs free energy calculations. The results reveal γ-UO 3 as the thermodynamically stable phase, while α-UO 3 presents metastable characteristics, which is in line with recent experimental and theoretical findings. Featuring near-DFT level accuracy, the proposed MLIP lays a preliminary foundation for future studies on structural phase transitions, defect evolution and high-value uranium waste recycling in U-O systems, and establishes a fundamental computational framework for exploring the behaviors of this system under confined environmental conditions.
科研通智能强力驱动
Strongly Powered by AbleSci AI