Uncertainty Estimation and Robust Training of Materials Graph Neural Network Models

Graph Neural Network (GNN) models have become a promising new tool for materials simulations. One key advancement of GNN models is the invention of universal interatomic potentials (UIPs) that are transferable to the whole periodic table. Such GNN UPs have trainable parameters at the scale of millions, and they achieve reasonable training and test accuracies for millions of training and test structures spanning the space of known chemistry. Such UIPs open the opportunity to explore and discovery hypothetical materials at the unprecedented scales of tens to hundards of millions. However, even UIPs suffer from poor reliability for extrapolation, and their prediction accuracy for unseen structures are worse than that for training structures. The usefulness of UIPs would be questionable if their reliability cannot be confirmed. In our work, we tackle this challenge by introducing a method to estimate the prediction uncertainty of GNN models for materials simulations, with which we also come up with a training set specially aimed for ambient-temperatures simulations of all existing structures in the Materials Project.