A multiclass classification model for predicting the thermal conductivity of uranium compounds

Abstract

Advanced nuclear fuels are designed to offer improved performance and accident tolerance, with an emphasis on achieving higher thermal conductivity. While promising fuel candidates like uranium nitrides, carbides, and silicides have been widely studied, the majority of uranium compounds remain unexplored. To search for potential candidates among these unexplored uranium compounds, we incorporated machine learning to accelerate the material discovery process. In this study, we trained a multiclass classification model to predict a compound’s thermal conductivity based on 133 input features derived from element properties and temperature. The initial training data consist of over 160, 000 processed thermal conductivity records from the Starrydata2 database, but a skewed data class distribution led the trained model to underestimate compound’s thermal conductivity. Consequently, we addressed the issue of class imbalance by applying Synthetic Minority Oversampling TEchnique and Random UnderSampling, improving the recall for materials with thermal conductivity higher than 15 W/mK from 0.64 to 0.71. Finally, our best model is used to identify 119 potential advanced fuel candidates with high thermal conductivity among 774 stable uranium compounds. Our results underscore the potential of machine learning in the field of nuclear science, accelerating the discovery of advanced nuclear materials.