Boosting heterogeneous catalyst discovery by structurally constrained deep learning models

The discovery of new catalysts is one the significant topics computational chemistry as it has potential to accelerate adoption renewable energy sources. Recently developed deep learning approaches such graph neural networks (GNNs) open opportunity significantly extend scope for modelling novel high-performance catalysts. Nevertheless, representation particular crystal structure not a straightforward task due ambiguous connectivity schemes and numerous embeddings nodes edges. Here we present embedding improvement GNN that been modified by Voronoi tesselation able predict catalytic systems within Open Catalyst Project dataset. Enrichment was calculated via tessellation corresponding contact solid angles types (direct or indirect) were considered features edges volumes used node characteristics. auxiliary approach enriching intrinsic atomic properties (electronegativity, period group position). Proposed modifications allowed us improve mean absolute error original model final equals 651 meV per atom on dataset 6 intermetallics Also, consideration additional dataset, show sensible choice data can decrease values above physically-based 20 threshold.