Electrochemical CO2 reduction is a promising route to produce fuels and chemicals using renewable electricity. However, an abundance of different molecules can form during the reaction and so finding catalysts that selectively make certain products is a key challenge. In recent years, machine learning (ML) methods have been employed to help identify promising catalysts, yet product selectivity is not always considered, and the design spaces investigated have been limited. Now, Kun Jiang, Seoin Back and colleagues at Sogang University and Shanghai Jiao Tong University report a ML approach to identify promising catalysts, taking into account activity and product selectivity without the need for time-consuming density functional theory calculations or surface structure modelling.
Building on their own work and that of others, the researchers combine a ML model that predicts adsorbate binding energies of active motifs with maps that use adsorbate binding energies as descriptors to predict product selectivity. The ML model includes information on composition and coordination number of the active motifs and therefore can be used to glean insight on how stoichiometry and morphology influences catalysis. The team explore 465 binary combinations of elements and predict their selectivity to four different products: hydrogen, formate, CO and C1+, where the latter corresponds to products that are more reduced than CO (for example, CH4 and multicarbon products). They find that Cu-Pd and Cu-Ga binary alloys have high selectivity to C1+ and formate, respectively, and validate their predictions experimentally. ML methods such as that reported by Jiang, Back and colleagues should help to speed catalyst discovery in the future.
This is a preview of subscription content, access via your institution