S-GNN: State-Dependent Graph Neural Networks for Functional Molecular Properties

Property models are an integral part of many chemical engineering applications and have been the subject of a lot of interest, especially with recent advancements in deep learning such as graph neural networks. Despite being of major importance, little effort has been dedicated to functional properties where the property dependency goes beyond the molecular structural information and depends on the state variables such as temperature and pressure. In this work, we demonstrate a flexible framework to extend graph neural networks to account for such use cases. A total of 13 different temperature-dependent properties were modeled covering enthalpy of vaporization, various heat capacities, densities, and thermal conductivities as well as surface tension and vapor pressure. While many were successfully modeled with some reaching an average absolute relative error below 6%, some still require further attention such as surface tension and thermal conductivity to achieve good accuracy.