Combining machine learning and atomistic simulations through active learning, this project will explore and rank a very large composition space of multicomponent oxides according to their stability and activity. The predictions of these models will then be validated in the lab and scaled up through collaborations with academics and industry.
Promising technologies such as solar fuels, fuel cells, electrolyzers, and metal-oxygen batteries are key for sustainable and independent energy generation and storage. However, they are being held back by a common challenge: the need to interconvert electricity and chemical energy in the formation and consumption of oxygen molecules. We lack efficient and cheap electrocatalytic materials for the so-called oxygen evolution reaction and oxygen reduction reaction (OER and ORR). Precious metals are effective but have limited practical applications because of their high cost and low abundance. Non-platinum group metal oxides such as Ni, Fe, and Co are active for OER and ORR in alkaline solutions, but are not stable in the acidic environment encountered in water electrolyzers or have low activity. Multi-component oxide materials can combine diverse earth-abundant elements to achieve the desired properties. Perovskites, for example, have an ABO3 composition and have shown much promise as OER/ORR catalysts because they can be doped with a rich variety of A and B site cations, and these changes in elemental composition allow fine-tuning of the electrocatalytic properties. Despite the promise of these materials, and the success of atomistic simulations in explaining and predicting catalytic activity of surfaces, the large combinatorial space of substitutions hinders design of new materials. There are billions of 5-element combinations and, even with computers, they cannot be evaluated one by one in search for the best. This project will build an autonomous computational activelearning loop to perform global optimization of multicomponent oxide catalysts for OER/ORR. Combining machine learning and atomistic simulations through active learning, this project will explore and rank a very large composition space of multicomponent oxides according to their stability and activity. The predictions of these models will then be validated in the lab and scaled up through collaborations with academics and industry.
The project will utilize atomistic simulations based on quantum mechanics to rank candidate materials for their stability and their catalytic properties. Machine learning models will play two important roles. One in replacing the expensive simulations with fast surrogate functions that reproduce the accuracy at a fraction of the cost. Neural network architectures fine-tuned for materials will be trained on the continuously acquired data. The second use of machine learning is inspired in the recent successes in games like chess or go, where machines learn to play a game based on discrete decisions. In our case, the game is designing a material, and the discrete “moves” are selecting which elements to use in which positions in the material to get stable and active catalysts. By rapidly exhausting a large design space, this project will generate highly promising and diverse materials for OER/ORR that will then be validated experimentally and scaled up towards devices and products. The project will also produce a complete “map” of the chemical space of oxides for OER/ORR which can guide further material searches for other applications.