Graph neural networks for the prediction of infinite dilution activity coefficients

Abstract

The use of predictive methods for physicochemical properties is of special interest given the difficulties involved in the experimental determination of large chemical spaces. In this work, we focus on the prediction of infinite dilution activity coefficients γ_ij^∞ of organic systems using graph neural networks (GNNs). Our proposed method involves the use of one GNN that extracts the relevant solvent information and one GNN for doing so for the solute. The vectorial representations of these chemical species are then combined into a binary-system fingerprint which is used as the input to a supervised learning framework. We compare our approach to the 8 most commonly employed phenomenological/mechanistic methods for predicting γ_ij^∞. Our method is able to predict γ_ij^∞ with competitive performance to the state-of-the-art mechanistic methods, achieving a lower mean absolute error (MAE) compared to the broadly used COSMO-RS and UNIFAC-Dortmund methods. We also present a series of parallel residual hybrid models that combine both mechanistic and GNN-based approaches. These hybrid models overall improve the performance of the individual model instances.