Active Discovery of Catalysts for Sustainable Energy Storage

Consumption of fossil fuels has caused climate to change at unmanageable rates, making adaptation difficult both environmentally and economically. This is shown by an orders-of-magnitude increase in global extinction rates, signaling environmental destabilization. It is also shown by an increasing rate and severity of extreme weather events such as droughts, floods, or wildfires, straining our economies and endangering our food & water supplies. We may be able to slow these changes by transitioning

from fossil fuels to solar energy, but solar energy’s inconsistent availability makes implementation difficult. This could be addressed by storing the energy in solar fuels,

which are fuels created from solar energy, CO2, and H2O. Unfortunately, solar fuels are hindered by a lack of commercial viability. We could solve this issue by finding

catalysts that produce solar fuels more quickly, selectively, and efficiently. This thesis comprises several projects aimed at discovering catalysts for solar fuel production. We first show how recent advances in computation and data science

can accelerate the catalyst discovery process. We illustrate this point by creating a software framework that performs high-throughput density functional theory (DFT)

calculations, which can be used to predict catalyst performance. Then we combine our framework with a heuristic method for “active discovery”. Active discovery is

the automated process of: using a dataset to choose next experiments; adding the experiment results to the dataset, and repeating this process iteratively. We used active discovery to identify several catalysts for CO2 reduction and H2 evolution. We then work with collaborators to find experimental evidence showing that one of the candidates, CuAl, can reduce CO2 to ethylene with selectivity of up to 80%. Next, we improve our active discovery process by creating a multiscale model that predicts macro-scale catalyst performance from atomic-scale DFT functional theory and machine-learned predictions. Lastly, we combine this model with a multiscale sampling strategy for selecting calculations, and we show how this strategy can be

used to discover catalysts even more efficiently.