Unlocking Potential Catalysts: A Machine Learning Approach with Bayesian and Regression Models

Abstract Due to their excellent catalytic efficiency, endurance, adaptability, and unusual structure, single‐atom alloys are an important category of materials with huge potential for efficiently utilising rare and costly metals in catalytic applications. Since no two of these materials are alike, designing a catalyst for each presents its own set of special challenges. The development of catalysts can be sped up with the use of machine learning, however conventional machine learning approaches frequently necessitate large datasets and costly feature engineering. In addition, these approaches frequently have difficulty juggling competing aims and constraints as they expand their search space. This research makes use of Bayesian optimisation (BO) to help in the identification of effective catalysts. Even with as few as 5 starting data points from costly density functional theory (DFT) computations results, our BO workflow is able to quickly and accurately discover the best single‐atom alloy surfaces. Not only that for the applicability of our model on other systems, we have chosen dataset comprises transition metal surfaces as well and found suitable performances which further validates the universality of our model. Our BO model outperforms a random search technique on many different adsorption systems by making use of simple, easily accessible features. Apart from BO, we have also designed other regressor models for searching the best catalyst and interestingly we found that for a small sample size where generating data is very difficult, K‐Nearest Neighbour regressor (KNR) outperforms BO. This research not only unlocks the potential of this BO as well as regressor models in catalysis research but also lays down a robust foundation for future work aiming to optimize material selection based on adsorption characteristics.