(560bz) Multi-Task Machine Learning to Predict ORR Catalyst Descriptors and Performance across Surface Composition

Due to the growing global population and industrialization of many countries, demand for energy for social and economic development is increasing. The oxygen reduction reaction (ORR) is an important reaction in fuel cell devices and the lack of sufficiently good ORR catalysts limits the overall efficiency of such devices. Despite the search for novel catalysts and new bimetallic catalysts are regularly discovered or optimized, the ORR activity of materials have not improved significantly. Large-scale screens for optimal ORR composition have focused on single descriptors using the assumption of linear scaling relations between key intermediates. In this work we demonstrate the high-throughput generation of adsorption energy datasets for *O, *OH, and *OOH intermediates on over 3000 possible surfaces including 820 different crystal bulk structure across 44 unique elements. We show that these quantities can be predicted with coordination-based descriptors to approximately 0.2 eV accuracy, sufficient to screen for potential activity across dozens of elements. Further, we show that these predictions can be improved through multi-task non-linear regression to simultaneously include and improve upon the well-known scaling relations. We also identify adsorption configurations that deviate from linear scaling relations by using machine-learning guided high throughput calculations. Finally, we use these methods to predict promising catalysts for experimental ORR validation.