Scaling Up Deep Learning for Computational Materials Discovery

Deep learning models are often evaluated on validation sets sampled from the same distribution as their training sets. In the natural sciences and engineering, however, models are evaluated for their ability to generalize beyond their "training set," either in applications of discovery or theoretical modeling. This dichotomy has caused confusion in deep learning, where methods like active learning and curriculum learning do not improve performance on independent and identically distributed (IID) academic datasets such as ImageNet, while being indispensable tools in real-life applications such as autonomous driving. With the increasing interest in using machine learning in the physical sciences, this dichotomy poses an obstacle to making meaningful progress.

I will provide specific examples of this problem in the context of computational materials discovery, where neural networks that can predict the formation energy of inorganic crystals with unprecedented accuracy have been shown not to improve the efficiency of stable materials discovery at 0K. I will present our progress in addressing this challenge and discuss future work.