(544dh) Screening Bimetallic Catalyst for CO2 Reduction Using Machine Learning and DFT Data

Zong Qian Yu, Kevin Tran, Zachary Ulissi

Over 90% of commercial product are achieved through some heterogeneous catalysis. Identifying active and stable electrochemical intermetallic catalysts is extremely costly due to experimental and computational complexity. Further, finding the active site in electrochemical polycrystalline catalysts is difficult due to the sheer number of stable surfaces that exist in each crystal structure, it is very difficult to find suitable surface for CO2 reduction and HER evolution through experimentation. DFT calculations would have to performed on each active site on each unique surface exist on a specific catalyst particle to analytically determine the performance of a catalyst. Due to expensiveness and the time constraints of DFT calculations, active machine learning is implemented to predict adsorption energy based on existing DFT data. Using the arrangement of atoms near the adsorption site, atom radius, bond distances, etc. as parameters, different statistical strategies and non-linear regressions are used to predict the lowest adsorption energy of a given adsorbate. Through machine learning the time of prediction for such problem can be greatly reduced. Currently, the best performing catalyst for CO2 reduction and HER are CuAl alloy and Pt, Ag, and Pd based alloys respectively. Using the machine learning model, minimum energy of adsorption sites with very unique arrangement of atoms can be predicted without using DFT. More attentions are paid to sites with similar minimum adsorption energy with the best performing catalyst in the market. Such predictions will reveal which crystal structure in which bimetallic alloy has many appropriate surface, therefore, greatly reduced the search range compared to before. More studies related can be done on such as methane and ethylene production.