High-Throughput Electrocatalyst Screening and Machine Learning for Feature Selection and Prediction of Alkaline Fuel Cell Catalysts

Today's polymer fuel cells use a strong acid environment which prevents the use of cheap, earth abundant catalysts. This has led to a strong desire to discover electrocatalysts that are highly active in an alkaline environment and don't contain rare, platinum group metals (PGM). In this study, we modified our previous method of screening electrocatalysts to discover new, highly active alloys for alkaline hydrogen oxidation (HOR). 1584 alloy samples were prepared from a selection of 12 elements and screened, and the most active alloys were tested in an alkaline polymer fuel cell. We examine catalysts with activities higher than platinum for alkaline HOR and alloys with similar activity to platinum but lower concentrations of PGM metals. The onset potential of each alloy was joined with a dataset of elemental descriptors and used to train three machine learning models: a neural network, a gradient boosted decision tree regressor, and an ARD Baysian ridge regressor. The models were evaluated for their accuracy to predict new catalysts and calculate the importance of each feature from the dataset to identify features with interesting trends and significance. I will also discuss how these results can be used to understand alkaline HOR and what are the next steps going forward to identify even more promising catalysts.