Machine learning for the experimental and computational development of heterogeneous catalysis

Machine learning techniques have emerged as a useful tool for identifying complex patterns and correlations in large datasets. These techniques could be particularly useful in heterogeneous catalysis research for enabling the correlation of the catalyst performance to its physicochemical properties. So far in the chemistry and material science communities, machine learning models have mostly been built on high-throughput quantum chemistry calculations, and only selected case studies have led to the experimental discovery of improved catalyst materials. The slow pace and limited number of scientific breakthroughs may be attributed to simplistic assumptions about catalyst structure in quantum chemistry calculations and the incomplete experimental data available. Therefore, we believe that the development of high-throughput approaches closely coupled with machine-learning-based approaches could help accelerate experimental catalysis research. To aid the community, we bring together the available body of work applying high-throughput approaches and machine learning to the development of solid heterogeneous catalysis. We offer an objective view of the trends in the field by performing a detailed and systematic comparison of papers based on the (1) the ML method, the features used as model input and output, (3) the material, device or reaction investigated, (4) the dataset size, and (5) the overall achievement. Furthermore, for models reporting unitless R2 values, we quantitatively analyze the model performance as a function of the features used, the reaction type and the dataset size.