Prediction of thermodynamic stability of actinide compounds by machine learning model

Abstract

The thermodynamic phase stability plays a crucial role as it serves as a fundamental parameter governing the synthesizability of materials and their potential for degradation under specific operating conditions. In this study, two machine learning (ML) models, random forest (RF) and neural network (NN), were used to predict the thermodynamic phase stability of actinide compounds using a dataset consisting of 62204 DFT-calculated energies. Our study utilizes a comprehensive range of properties that do not contain structural information, making them applicable to materials composed of any number of constituent elements. Notably, the trained models achieve an approximation that closely aligns with the error obtained from DFT calculations, while drastically reducing computational time by several orders of magnitude. Moreover, we extended our analysis to predict binary phase diagrams of Generation IV nuclear fuels using the trained models. To address the limitations of a single model for predicting certain compounds and enhance model robustness, a simple ensemble learning approach, i.e., the multi-component learner was employed. By synergistically combining prediction outcomes from RF and NN models, the ensemble learning approach excels in accurately predicting phase diagrams of actinide compounds. Utilizing the compound components forecasted by the model as a foundation, we embarked on an extensive series of structural searches and conducted thorough phonon dispersion studies. The outcomes unequivocally affirm the model's efficacy in accurately predicting stable compound compositions.