Machine Learning

Machine Learning of Ab-Initio DATA for Accelerated Metal Catalyst Discovery

The ultimate goal of the proposed research is to develop an integrated, experimentally verifiable modeling framework based on quantum-chemical calculations and advanced machine-learning algorithms to accelerate catalyst discovery. In heterogeneous catalysis, chemical reactions are exclusively catalyzed by nanometer-sized particles. Many factors, such as atomic arrangements, elemental compositions, strain/stress, nature of supports, and nanoconfinement, influence physiochemical properties of nanoparticles and so the reaction rate of elementary steps occurring at active sites. Therefore, discovery of high-performance catalysts with atomic-level accuracy requires a deep understanding of the reactivity of atomic ensembles towards molecules or their fragments. A quantum brute-force search for optimal catalysts, although it shows some advantages compared to experimental trial-and-error approaches, is far beyond reach for present-day computing facilities. This project integrates quantum-chemical modeling, electrochemistry, catalysis, database development, and machine learning for accelerated catalyst discovery. Specifically, we will apply this new approach to the design of core-shell alloy nanoparticles for CO2 conversion. Electrochemical reduction of carbon dioxide (CO2) has received considerable attention because of its potential to utilize the abundant greenhouse gas in the Earth’s atmosphere, and the electricity from intermittent renewable sources, to yield fuels and chemicals that are traditionally derived from petroleum. Copper, silver, and gold are, arguably, the only known metals that produce appreciable amounts of hydrocarbons and oxygenates, or the fuel precursor carbon monoxide, although at high overpotentials (~1.0 V). We hypothesize that coinage metal alloy nanoparticles with atomically tailored size/shape, structure/composition, and point/planer defects can considerably reduce the energy loss for electroreduction of CO2 and improve the Faradaic efficiency to produce valuable C2 species (e.g., ethylene and ethanol). The project, as illustrated in Figure above, consists of three intertwined core activities: 1) development of machine-learning models with the dataset from quantum-chemical calculations of nanoparticle properties; 2) development of a web interface that leverages available materials properties databases for catalyst prediction; 3) verification and improvement of the modeling framework through experimental testing.