Active-Learning Framework for the Discovery of Novel Materials for Catalysis

The discovery of abundant and affordable materials for sustainable energy applications is a significant challenge of this decade. The space of possible material compositions and crystal structures is so vast that brute-force screening approaches are unfeasible. However, density functional theory (DFT) combined with machine-learning (ML) techniques have a huge potential to speed up the search and hold a great promise for the discovery of novel materials.

Here, we present a ML-based active-learning framework to search for stable and meta-stable inorganic materials. The workflow includes the generation of candidate structures, which is followed by automated DFT job submission, ML-training, and acquisition. When using a search space of experimentally observed crystal structures, this approach was recently applied to the discovery of new stable polymorphs of IrO_x, predicted to have an enhanced stability and catalytic activity for the oxygen evolution reaction (OER). In this talk, we extend the search to hypothetical crystal structures that are generated from a bottom-up prototype enumeration scheme, based on spacegroup and Wyckoff positions. This is applied to the search for cheaper and more abundant materials. Last, we envisioned how the screening approach can be extended to surface and chemisorption properties, on order to extend the search to catalytic applications.