Advancing Vapor Pressure Prediction: A Machine Learning Approach with Directed Message Passing Neural Networks

Abstract

The knowledge vapor pressure of a chemical as a function of temperatures is important in many chemical and environmental engineering applications. This study introduces a novel approach utilizing a machine learning model based on the directed message passing neural network (D-MPNN) architecture to predict the vapor pressure of organic molecules over a broad temperature spectrum. We investigate various strategies for incorporating temperature effects into our models, a key factor for accurate vapor pressure predictions. Our results show that the D-MPNN model markedly surpasses the traditional PR + COSMOSAC method, achieving a significantly lower average absolute relative deviation (AARD) of 0.617 (from D-MPNN vs. 1.36 from PR + COSMOSAC) for an extensive dataset of 19,081 molecules. This improvement is notable as it does not require additional critical property measurements or quantum mechanical calculations for the molecules. This study underscores the potential of machine learning to accurately capture complex molecular features for reliable vapor pressure prediction, presenting a robust alternative to traditional methods dependent on critical property data or quantum mechanical calculations. This breakthrough is especially advantageous for assessing the properties of a novel or under-characterized chemical species.