AI3SD Video: Preserving Structural Motifs in Machine-Learning Approaches to Modeling Water Clusters

Chemical structures are naturally viewed as collections of atoms connected through bonds, and graph theory provides a natural tool for capturing that intuition in a concrete mathematical fashion. Over the past several years, graph neural networks have become increasingly popular for modeling chemical systems. To build on this work, the multi-laboratory ExaLearn project, part of the DOE Exascale Computing Project, is developing novel capabilities that combine state-of-the-art machine-learning techniques with high-performance computing to enable the rapid exploration of chemical space on exascale-class systems. Water clusters offer an interesting use case for the development of machine-learning approaches that preserve intermolecular interactions and structural motifs. We apply a dataset of ~5 million hydrogen-bonded water clusters that display interesting long-range structural patterns to explore unique challenges in property prediction and molecular generation.[1] J. A. Bilbrey, J. P. Heindel, M. Schram, P. Bandyopadyay, S. S. Xantheas, S. Choudhury. “A look inside the black box: Using graph-theoretical descriptors to interpret a Continuous-Filter Convolutional Neural Network (CF-CNN) trained on the global and local minimum energy structures of neutral water clusters,” J. Chem. Phys., 2020, 153, 024302.[2] S. Choudhury, J. A. Bilbrey, L. Ward, S. S. Xantheas, I. Foster, J. P. Heindel, B. Blaiszik, M. E. Schwarting. “HydroNet: Benchmark Tasks for Preserving Long-range Interactions and Structural Motifs in Predictive and Generative Models for Molecular Data,” Machine Learning and the Physical Sciences workshop at NeurIPS, 2020