Predicting thermal conductivity of transition metals using machine learning gaussian process regression models

Abstract

The lattice thermal conductivity of transition metals is fundamental to their applications in thermal management, high-temperature alloys, and energy-related technologies, as well as to understanding their lattice dynamics and heat-transport behavior. However, reliable and accurate prediction of thermal conductivity remains challenging due to anharmonic lattice effects, complex structure-property relationships, and limited experimental data. This study aims to predict the thermal conductivity of transition metals using a dataset of nearly 1000 individual records that include their physical and chemical properties, with the goal of designing advanced materials. To achieve the goal, this study employed several Gaussian Process Regression (GPR)-based models: Rational Quadratic GPR (RQGPR), Squared Exponential GPR (SEGPR), Matern 5/2 GPR (MGPR), Exponential GPR (EGPR), and Optimizable GPR (OGPR), which were investigated on the transition metals dataset. Among the implemented models, the OGPR showed excellent predictive performance, achieving R² values of 98.47% and 99.89% in the validation and tests for thermal conductivity prediction, respectively. Furthermore, the RReliefF algorithm is employed to reveal the significance of input features to the target property. Moreover, there is a strong relationship between the target attribute and the input feature that has the greatest influence. Furthermore, the partial dependence plots are also employed to interpret feature influence and validate model predictions. These findings demonstrate the effectiveness of machine-learning GPR models in predicting the thermophysical properties of transition metals.