Open Catalyst Project: Finding Similarity Among Adslabs

The Open Catalyst Project (OCP) is a collaborative research project between Facebook Artificial Intelligence Research and Carnegie Mellon University’s Department of Chemical Engineering. Currently, an open challenge in the development of renewable energy is finding low-cost catalysts to drive reactions at high rates. The use of AI or machine learning may provide a method to efficiently approximate specific catalyst characteristics, leading to new approaches in finding effective catalysts. This research project focuses on finding similarity between adsorbate-surface configurations, i.e. adslabs. An adslab include information of the surface, such as the catalyst, the adsorbate, miller index of the surface and site type. A pretrained Dimenet++ model was applied to each adslab of interest to obtain embeddings for all atoms in the adslab. Each atom will have a 256-length vector embedding. Squared exponential kernel trick was then applied to embeddings of each pairwise atoms as an indicator of their similarity. Adslabs were treated as node in a Graph Neural Network model and the Cuthill-McKee algorithm is applied to group the similar surfaces together to ease future training and prediction. Such similarity among adslabs can be applied to improve efficiency and accuracy in future model training, such as Gaussian Processes during active learning. Fewer parent DFT calls are need and accuracy of the initial stages would be greatly improved.