Predicting Compound Melting Temperature from Computationally Derived Properties via Machine Learning

Melting temperature is a fundamental material property used in a wide variety of scientific disciplines. To eliminate the need for experimentally measuring melting temperature, often costly and potentially dangerous, we developed a new computational approach that predicts a compound’s melting temperature from its material properties, computed by Density-functional theory or other computational methods, using machine learning models. The proposed machine learning models achieved promising results, with a Mean Percentage Absolute Error of 15%-20%. We also identified important material properties in predicting melting temperature across various groups of compounds containing metals, transition metals, post transition metals, alkaline earth metals, halogens, and chalcogens.