Exploring thermodynamic stability of plutonium oxycarbide using a machine-learning scheme

Abstract

Plutonium oxycarbide plays a crucial role in the fabrication of a carbide fuel and the corrosion of plutonium. In this work, a machine-learning (ML) scheme is used to predict the thermodynamic stability of plutonium oxycarbide PuO_xC_1−x. The training data are generated within the framework of density-functional theory (DFT) and its Hubbard correction. Four ML schemes combined with three structural descriptors are considered and their performance is compared. The optimal ML model for the DFT data set yields remarkably small average errors of approximately 3 meV per atom for mixing energy and 0.003 Å for the lattice parameter, indicating its high prediction accuracy. Utilizing the ML model, we predict the convex hull of PuO_xC_1−x as well as several ordered atomic structures for a specific value of x. The enhanced energetic stability observed in these ordered structures could probably be attributed to the strong hybridization between Pu 5f/6d and C/O 2p, namely robust Pu–C and Pu–O bonds.