Polymer graph neural networks for multitask property learning

Abstract

The prediction of a variety of polymer properties from their monomer composition has been a challenge for material informatics, and their development can lead to a more effective exploration of the material space. In this work, PolymerGNN, a multitask machine learning architecture that relies on polymeric features and graph neural networks has been developed towards this goal. PolymerGNN provides accurate estimates for polymer properties based on a database of complex and heterogeneous polyesters (linear/branched, homopolymers/copolymers) with experimentally refined properties. In PolymerGNN, each polyester is represented as a set of monomer units, which are introduced as molecular graphs. A virtual screening of a large, computationally generated database with materials of variable composition was performed, a task that demonstrates the applicability of the PolymerGNN on future studies that target the exploration of the polymer space. Finally, a discussion on the explainability of the models is provided.