From Molecules to Climate Solutions: Exploring the Spider Neural Network Architecture for Representing Interacting Graphs

Abstract

In materials science, a natural way to represent substances is as a graph, where atoms serve as nodes and the bonds between them are depicted as edges. Such a graph typically has chemically significant sub components, such as a subgraph of nodes representing metallic atoms. This senior thesis project introduces the Spider model, an E(3) invariant neural network that encompasses such interacting graphs, capitalizing off of structural knowledge while learning predictive tasks. We explore the application of the Spider framework to two materials science applications relevant in the context of climate change: energy projections for renewable energy catalysts and property predictions for metal organic frameworks. We also explore modifications of the Spider architecture such that it is better suited for these applications. Training the Spider framework on over a million structures representing heterogeneous catalysts, representing just a small fraction of the entire Open Catalyst 20 training set, we obtain prediction results that already outperform median baseline metrics, and are projected to outperform the strongest baselines using the full dataset. In another application, training Spider on the structural representations of 1,000 metal organic frameworks, we demonstrate its capacity to predict density, largest included sphere, and largest free sphere. These results illustrate capabilities of Spider to learn the underlying data distribution of various materials and generate predictions that are plausible under this learned distribution. Future work may include representation learning for molecular crystals or generative modeling for carbon capture materials.