On the importance of catalyst-adsorbate 3D interactions for relaxed energy predictions

• DOI: 10.48550/arxiv.2310.06682
• Publication Date: 2023-01-01
• Authors: Alvaro Carbonero; Alexandre Duval; Victor Schmidt; Santiago Miret; Álex Hernández-Ga...; Yoshua Bengio; David Rolnick
• Keywords: Pooling; Position (finance)

The use of machine learning for material property prediction and discovery has traditionally centered on graph neural networks that incorporate the geometric configuration all atoms. However, in practice not this information may be readily available, e.g.~when evaluating potentially unknown binding adsorbates to catalyst. In paper, we investigate whether it is possible predict a system's relaxed energy OC20 dataset while ignoring relative position adsorbate with respect electro-catalyst. We consider SchNet, DimeNet++ FAENet as base architectures measure impact four modifications model performance: removing edges input graph, pooling independent representations, sharing backbone weights using an attention mechanism propagate non-geometric information. find site impairs accuracy expected, modified models are able energies remarkably decent MAE. Our work suggests future research directions accelerated materials where reactant configurations can reduced or altogether omitted.