In silico high throughput screening of bimetallic and single atom alloys using machine learning and ab initio microkinetic modelling

Abstract

The advent of machine learning (ML) techniques in solving problems related to materials science and chemical engineering is driving expectations to give faster predictions of material properties. For heterogeneous catalysis applications, relying on the age-old Sabatier principle, an ab initio in silico high throughput screening of catalyst materials is envisaged, wherein ML based methods show potential to significantly reduce the experimental as well as computation cost. The availability of ML algorithms (in open source libraries like Scikit-Learn) and materials database (like CatApp and Materials Project) further augments this realization. By using these resources, ML models are developed to predict the binding energies of oxygen and carbon on bimetallic alloys and Cu-based single atom alloys (SAAs) using the features of metals that are readily available in the periodic table. Several ML models for predicting oxygen binding energy for AA terminated A₃B alloys are analysed and gradient boosting regression (GBR) is observed to give superior performance with a root mean square error of 0.31 eV in the test. In addition, GBR based ML models are demonstrated to predict the oxygen and carbon binding energies of AB terminated A₃B alloys with a test error of 0.38 eV and 0.35 eV respectively. The binding energy of oxygen and carbon on Cu-based SAAs is predicted with a test error of 0.36 eV and 0.37 eV respectively. Moreover, the computational time for predicting the binding energy using ML is 0.0006 s on a dual-core laptop which is significantly less than the time required for DFT calculations. DFT and ML calculated carbon and oxygen binding energies for the bimetallic A₃B alloys are further used in an ab initio microkinetic model to calculate the turn over frequency (TOF) for ethanol decomposition and non-oxidative dehydrogenation reactions. The TOFs over bimetallic alloys obtained using the ML calculated binding energies follow the same trend as that obtained using the DFT energies, with the TOF values being the same or within an order of magnitude range. This shows that catalyst screening using binding energy as a descriptor can be performed using ML models, bypassing time and resources consuming DFT calculations. This is likely to speed up the process of novel catalyst discovery.