Algorithms for accelerating the optimization of alloy catalysts

With the fossil fuel reserves being increasingly limited, the world’s energy future is toward exploiting renewable energy sources as an alternative. Late transition metals are known as important catalyst materials for the production of many renewable energy sources. For instance, Ni is the most used metal for catalyzing steam methane reforming, which is responsible for over 95% of hydrogen production worldwide. Alloy catalysts are especially interesting as they can exhibit significantly improved catalytic activity compared to pure metals, but their stability is sometimes a more concerning issue for practical applications. Therefore it is of great significance to optimize the design parameters of an alloy catalyst system to achieve the best performance in terms of activity and stability.
However, the multi-component nature of alloy catalysts brings significant challenges in the computational optimization of the system, especially when the catalyst exists at the nanoscale. The performance of a nanoalloy catalyst, for example, is jointly influenced by many design parameters such as size, shape, elemental composition, chemical ordering and reaction condition. This results in a huge design space of all possible nanoalloy configurations, often known as the “combinatorial explosion”. To make matters worse, the evaluation of the target property of each alloy configuration often requires a time-consuming quantum mechanical calculation. A brute-force approach of enumerating all possible alloy configurations is apparently infeasible, as we will quickly run out of time and computational resources. To get out of this dilemma, we need to design novel algorithms that can guide us through the vast space of alloy catalysts to the few potential candidates with desired properties at manageable cost.
In this thesis, mainly two types of algorithms, namely evolutionary algorithm and machine learning, are used synergistically to accelerate the global optimization of alloy catalysts. The optimization of an alloy catalyst system can be seen as a non-convex global optimization of an expensive black-box function. Evolutionary algorithms are population-based metaheuristics that aim to find sufficiently good solutions to a non-convex function at a reasonable cost by employing strategies inspired by biological evolution. To adapt to realistic size nanoalloy systems, we further introduced symmetry constraints to evolutionary algorithms in order to boost the efficiency and viability. Machine learning algorithms, on the other hand, are data-driven approaches that can significantly reduce the expense of each function evaluation during the global optimization without losing too much accuracy compared to quantum mechanical calculations. Specifically, we used two types of machine learning models, namely neural network and Gaussian process, in conjunction with global optimization algorithms to derive the optimal design parameters of Pt-Ni nanoalloys. In addition, we introduced a novel concept called evolutionary multitasking, which allows us to derive phase diagrams of Pt-Ni nanoalloys under various steam methane reforming conditions. Taken together, our methods can manage the overwhelming complexity involved in the optimization of alloy catalysts, thereby paving the way for high-throughput catalyst screening.