(290f) Machine Learning Electronic Structure for Heterogeneous Catalysis

The electronic structure of a material plays a key role in its surface chemistry and catalytic performance, but the process of utilizing this information for quantitative predictions is often challenging. One such approach is to develop descriptors, such as the d-band center, which condenses this information into individual features which are more amenable for property prediction. However, identifying valid descriptors via human intuition can be time consuming and quickly becomes intractable for chemically diverse systems, thus necessitating a more flexible and automated approach. Here we describe how machine learning can be used to obtain the electronic structure features via a trainable process, which can furthermore efficiently leverage existing high-throughput datasets for catalysis. We find this approach can easily outperform conventional descriptors in prediction accuracy while maintaining a high degree of physical meaning and interpretability in the models. The electronic structure space can consequently be effectively accessed and explored with significant implications for catalyst discovery and design.