Application of interpretable group-embedded graph neural networks for pure compound properties

The ability to evaluate pure compound properties of various molecular species is an important prerequisite for process simulation in general and in particular for computer-aided molecular design (CAMD). Current techniques rely on group-contribution (GC) methods, which suffer from many drawbacks mainly the absence of contributions for specific groups. To overcome this challenge, in this work, we extended the range of interpretable graph neural network (GNN) models for describing a wide range of pure component properties. The new model library contains 30 different properties ranging from thermophysical, safety-related, and environmental properties. All of these have been modeled with a suitable level of accuracy for compound screening purposes compared to current GC models used within CAMD applications. Moreover, the developed models have been subjected to a series of sanity checks using logical and thermodynamic constraints. Results show the importance of evaluating the model across a range of properties to establish their thermodynamic consistency.