Bayesian optimization of single-atom alloys and other bimetallics: efficient screening for alkane transformations, CO2 reduction, and hydrogen evolution

Abstract

Single-atom alloys form an important class of material that has shown great potential in maximizing the use of rare and expensive metals in catalysis due to their high catalytic performance, robustness, tunability and unique structure. Single-atom alloys present particular challenges for screening as they can deviate from some traditional catalyst design frameworks. While machine learning (ML) can be quite useful in accelerating catalyst design, most traditional ML methods require relatively large datasets and/or high-level, expensive featurization. Additionally, most of these ML methods are incapable of handling multiple objectives and constraints over the intended search space. In this work, we leverage Bayesian optimization (BO) to guide our search for high-performing catalysts. We show that our BO workflow can be initialized with as few as 2 to 8 data points, and often identifies the optimal single-atom alloy surface in just a few iterations. Our workflow was used to efficiently search across multiple adsorption systems and datasets and significantly outperformed a random search method, using simple, off-the-shelf features. For applications, we used BO to identify potential high-performing catalysts for alkane transformations, CO₂ reduction, and hydrogen evolution. Our BO workflow identified Hf₁Cu for alkane transformations; Y₁Au, Y₁Cu and Y₁Ag for CO₂ reduction; and an Ag–Ir bimetallic alloy for hydrogen evolution. Simple stability tests indicate all three single-atom alloys for CO₂ reduction are stable and most likely synthesizable. The workflow developed here can also be used for experiments or high-level theory calculations as well as other classes of materials.