Deep learning has been becoming increasingly popular in materials science yet there remain a number of challenges to the widespread application for routine estimation of many properties of interest. In this talk I present a potential solution to two of these challenges – (i) estimating the reliability of the model predications at inference time; (ii) developing models in situations where data are scarce. These two challenges are addressed by adding a reliable uncertainty quantification metric to the deep learning model. To represent the material we apply graph neural networks (GNN), which encode physical information about structure in the topology of the graph – the GNN learns a compressed low dimensional representation of the material by coupling to an attention mechanism. This low dimensional representation is then used as the input for a Gaussian process (GP) to learn the relationship between the representation and the property of interest. Unlike most GNNs used in materials modelling this GNN+GP provides a principled uncertainty metric for the prediction in a single inference run. I will demonstrate the effectiveness of this architecture on some common materials property prediction tasks. The uncertainty measure then allows the construction of an active learning (AL) procedure, which facilitates development of deep models in data scarce situations. We begin by training the model on a small initial dataset, then by estimating the quantities and the uncertainties on a larger (un-labelled) dataset we are able to sample the un-labelled data with the largest uncertainty, obtain labels (either from experiment or calculation) and then re-train the model. This so-called entropy sampling approach selects the new targets that will provide the most information to the model from labelling and allows sampling for new labels in a much more efficient manner than a random search. I will demonstrate how this entropy based sampling can decrease the amount of training data required for a model by several fold.

Calibrated deep representations and entropy based active learning for materials property prediction

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.