Abstract: Optimizing chemical reaction parameters is an expensive optimization problem. Each experiment takes a long time and the raw materials are expensive. High-throughput methods combined with the parallel Efficient Global Optimization algorithm can effectively improve the efficiency of the search for optimal chemical reaction parameters. In this paper, we propose a multi-objective populated expectation improvement criterion for providing multiple near-optimal solutions in high-throughput chemical reaction optimization. An l-NSGA2, employing the Pseudo-power transformation method, is utilized to maximize the expected improvement acquisition function, resulting in a Pareto solution set comprising multiple designs. The approximation of the cost function can be calculated by the ensemble Gaussian process model, which greatly reduces the cost of the exact Gaussian process model. The proposed optimization method was tested on a S_NAr benchmark problem. The results show that compared with the previous high-throughput experimental methods, our method can reduce the number of experiments by almost half. At the same time, it theoretically enhances temporal and spatial yields while minimizing by-product formation, potentially guiding real chemical reaction optimization.

Key words: Algorithm, Chemical reaction, Computer simulation, Efficient global optimization, Machine learning

摘要： Optimizing chemical reaction parameters is an expensive optimization problem. Each experiment takes a long time and the raw materials are expensive. High-throughput methods combined with the parallel Efficient Global Optimization algorithm can effectively improve the efficiency of the search for optimal chemical reaction parameters. In this paper, we propose a multi-objective populated expectation improvement criterion for providing multiple near-optimal solutions in high-throughput chemical reaction optimization. An l-NSGA2, employing the Pseudo-power transformation method, is utilized to maximize the expected improvement acquisition function, resulting in a Pareto solution set comprising multiple designs. The approximation of the cost function can be calculated by the ensemble Gaussian process model, which greatly reduces the cost of the exact Gaussian process model. The proposed optimization method was tested on a S_NAr benchmark problem. The results show that compared with the previous high-throughput experimental methods, our method can reduce the number of experiments by almost half. At the same time, it theoretically enhances temporal and spatial yields while minimizing by-product formation, potentially guiding real chemical reaction optimization.

关键词: Algorithm, Chemical reaction, Computer simulation, Efficient global optimization, Machine learning