Enhancing Optimization of Mixed Variables on a Robotic Flow Platform: Integrating Statistical Filtering with Nelder–Mead and Bayesian Methods

摘要

Synthetic chemistry has progressively integrated advanced optimization methods. While traditional approaches often prioritize continuous variables, the significance of discrete variables (e.g., solvent, catalysts, feedstocks, etc.) cannot be underestimated in reaction optimization. Solving mixed-variable optimization problems presents a major challenge, and in addition, there are no available comparative studies assessing the performance of the different available methods. In this study, we conduct a comparative analysis of three mixed-variable optimization approaches: filter-assisted Nelder–Mead and Bayesian optimization and a standard Bayesian optimization approach. Regarding the filter-assisted methods, we build upon our recent sampling–filtering–optimization (SFO) framework, which involves (i) sampling the continuous domain, for all the discrete possibilities, through Design of Experiments (DoE); (ii) filtering relevant discrete variables through statistical analysis; and (iii) optimizing the reaction with the filtered variables. The SFO strategy was tested in formal [3 + 3] cycloadditions of 1,3-cyclohexanedione with citral, conducted in a robotic micromole-scale flow platform. The results not only showcased the performance of different algorithms but also demonstrated the successful development of ultrafast, sustainable, and mild reaction conditions, allowing us to scale up the experimental conditions by a factor of >2400. This work highlights the potential of advanced optimization techniques in synthetic chemistry, particularly in the context of self-optimizing flow reactors.