Machine learning assisted material identification of photocatalysts for CO2 reduction reaction

Abstract

As the awareness of climate change and global warming rises, carbon capture, utilization, and storage (CCUS) has become an appealing solution to achieve net zero carbon emission globally. The technology of direct air capture (DAC) is more readily as compared to utilization, which hindered the wide application of CCUS, and thus the net-zero target. CO2 Reduction (CO2R) by electrochemical cell, a kind of CCUS, using sunlight as energy source – namely artificial photosynthesis – offers an alternative route for carbon intensive chemical fuel / product production. However, one of the major challenges of electrochemical cell development for CO2R is the performance of catalyst, e.g. sunlight absorbance, productivity, selectivity, stability, scalability etc. The traditional material discovery / screening progress relies very much on trial and error. Because of the recent technical advancement in computational power and data transparency, open access databases containing computational and experimental results are available for further analysis. This research aims to set up a rational and efficient workflow for material identification of CO2R reaction with the aid of machine learning (ML) technology. This research developed an ML model of a more accurate band gap energy prediction by GeoCGNN using data from SNUMAT for training and adopted Pre-trained ML model developed by OCP for more efficient approximation to the quantum mechanical simulations (density functional theory, DFT) as an alternative approach for material identification. The ML technology process forms a part of the material screening workflow, and the accuracy of ML model can be further enhanced as more data is generated. With the aid of ML technology, a consolidated dataset with 5,633 potential catalysts containing parameters related to thermodynamic stability, aqueous stability, solar spectral responsivity, and selectivity from various open-source databases was compiled. Pre-trained ML model was adopted to predict adsorption energies of CO and OH, describers of product selectivity in CO2R reaction, forming a new selection criterion for material screening. Material screening based on the consolidated dataset was performed with selection criteria set in different critical aspects of CO2R reaction with reference to literature. Seven potential photocatalysts for photoelectrocatalytic CO2R were identified, i.e. Si3Os2, ZnSe, Cs2SiAs2, Cs3Sb5Se9, RbAuSe, MoSe2 and Rb2Ni3Se4. The workflow set out would be able to identify catalysts for CO2R reaction that have been reported in literature and some with potential to be adopted. With the more efficient workflow for material identification, the development of electrochemical cells would be facilitated and, hopefully, the wide application of CCUS to help in achieving net-zero target.