Design Green Chemicals by Predicting Vaporization Properties Using Explainable Graph Attention Networks

Abstract

Computational predictions of vaporization properties aid the de novo design of green chemicals, including clean alternative fuels and working fluids for efficient thermal energy recovery. Here, we developed chemically explainable graph attention networks to predict five physical properties pertinent to performance in utilizing renewable energy: heat of vaporization (HoV), critical temperature, flash point, boiling point, and liquid heat capacity. The predictive model for HoV was trained using ~150,000 data points, with considering their uncertainties and temperature dependence. Next, this model was expanded to the other properties through transfer learning to overcome the limitations due to fewer data points (700-7,500). Chemical interpretability of the model was then investigated, demonstrating that the model explains molecular structural effects on vaporization properties. Finally, the developed predictive models were applied to the design of chemicals that have desirable properties as efficient and green working fluids and fuels, enabling fast and accurate screening before experiments.