PREDICTING THERMAL PROPERTIES OF CHEMICAL SUBSTANCES FROM MOLECULAR STRUCTURE AND COMPOSITION

Abstract

Knowing the thermal properties, from molecular structure, can aid in chemical screening, process design and thermodynamic modelling, and help to minimize reliance on expensive experimental measurements. This study has created an interpretable machine-learning model to predict normal boiling points and critical temperatures from the molecular structure and composition. Molecular descriptors used were based on the SMILES representation and were: molecular weight; carbon count; heavy atom count; aromatic atom count; TPSA (topological polar surface area); and the count of rotatable bonds. Random Forest, Extra Trees and Gradient Boosting were trained and tested with 80/20 split respectively to get the separate predictive models. The performance of the model was tested in terms of the coefficient of determination, mean absolute error and root mean squared error, and the importance of the descriptors was determined by permutation importance. Extra Trees performed the best for both properties, R²=0.955, MAE = 10.78 K and RMSE = 18.39 K for normal boiling point; as well as R²=0.908, MAE = 18.41 K and RMSE = 34.43 K for critical temperature. Carbon count was the most influential descriptor for both properties, followed by descriptors representing molecular size, composition, aromaticity, polarity and flexibility. The results show that the descriptor-based ensemble learning is an accurate and efficient procedure to estimate the thermal properties with a high degree of chemical interpretability. This framework may thus be used to provide a quick screening and preliminary evaluation of chemical substances when direct measurement of the thermal properties is not available or is limited.