Revising Message Passing Graph Neural Networks for Catalyst Design

Applying deep learning to search for catalysts is an important challenge for solving energy storage problems and the conversion of greenhouse gases into more valuable products caused by global warming. We present several graph neural networks (GNNs) in our work, including convolutional and message passing architectures with physically informed node and edge attributes for atomistic systems. We demonstrate an improvement in adsorption energy predictions on the OC20 dataset using our proposed architecture in terms of the mean absolute error of predicted energy and energy within threshold metrics. Proposed architectures are stable to overfitting and can be used to predict experimental and quantum chemical properties of a wide spectrum of materials and molecules. We propose the use of two GNN architectures (EdgeUpdateNet and OFMNet) together with an advanced featurization method of nodes and edges. We represent edge fingerprints as elements of interatomic interaction matrices (Coulomb matrix, Ewald sum matrix, Sine matrix). For node fingerprints, we use elements of the orbital field matrix (OFM), a one-hot representation of the electronic state of atoms with surrounding atomic orbitals. Also, we propose and implement the representation of catalytically active atoms as a subgraph. Proposed methods and architectures demonstrate improvement in the accuracy of adsorption energy predictions. Especially significant improvements are observed in out of domain examples for both adsorbates and catalysts. The generalizability and extrapolation capabilities on out of domain examples of the proposed architectures also make proposed GNNs feasible for use in catalyst screening in the vast chemical space.