Abstract
The global mission to reduce fossil fuel consumption has led to the escalating demand for electrochemical energy storage (EES) devices such as fuel cells and batteries. Computational techniques like Density Functional Theory (DFT) have recently been coupled with Machine Learning (ML) for high-throughput material screening and discovery. Transition metal surfaces are popular electrocatalyst candidates, but predictive ML regression models have only been applied to select metals such as Pt and Cu. Additionally, characterizing the contributions of each feature is challenging, especially on black-box models. In this work, regression models were trained to predict the adsorption energies of carbon, hydrogen, and oxygen on 27 fcc (111) monometallic surfaces and applied model-agnostic interpretation methods to evaluate feature importance. Over 200 adsorption energies on transition metal surfaces were collected from Catalysis-hub.org, a surface reaction database. A dataset was constructed for each adsorbate, and was composed of ten surface atomic, surface electronic, and adsorbate properties collected from online databases and DFT calculations on adsorbate-free surfaces. Then, the fine-tuned random forest regression, Gaussian process regression, and artificial neural network models predicted atomic adsorption energies while permutation feature importance calculated feature contributions. All ML model accuracies were found to be competitive with those from literature, with Gaussian process regression reporting the lowest errors of the three models. Coordination number was also found to have the largest contributions for all models. ML-DFT methodologies such as this can be expanded to accommodate alloys and more adsorbates for a wider screening of potential EES materials.
