DeepGamma: A deep learning model for activity coefficient prediction

Kobi Felton, Hashem B. Safar, Alexei Lapkin

Abstract: Vapor liquid equilibrium is a ubiquitous aspect of designing industrial chemical processes that make products ranging from pharmaceuticals to petrochemicals. To predict vapor liquid equilibrium for a wide variety of industrially relevant mixtures, activity coefficients are often used. However, calculating activity coefficients experimentally is time and labor-intensive, and existing methods for predicting activity coefficients are limited in scope or computationally expensive. Herein, we introduce DeepGamma, a deep learning method for predicting activity coefficients of binary mixtures directly from the molecular structures of their components. DeepGamma is demonstrated to have strong performance on a variety of mixtures with extremely fast prediction times.