Physical pooling functions in graph neural networks for molecular property prediction

Graph neural networks (GNNs) are emerging in chemical engineering for the end-to-end learning of physicochemical properties based on molecular graphs. A key element GNNs is pooling function which combines atom feature vectors into fingerprints. Most previous works use a standard to predict variety properties. However, unsuitable functions can lead unphysical that poorly generalize. We compare and select meaningful GNN methods physical knowledge about learned The impact demonstrated with calculated from quantum mechanical computations. also our results recent set2set approach. recommend using sum prediction depend size size-independent. Overall, we show significantly enhances generalization.