Autonomous Online Optimization in Flash Chemistry Using Online Mass Spectrometry

Abstract

The goal of sustainable and efficient chemical production has led to an increased focus on continuous processes, especially in the production of fine chemicals or active pharmaceutical ingredients. However, developing and optimizing continuous processes can be challenging. Autonomous online optimization can help facilitate this process. This article presents a fully automated flow chemistry platform for optimizing flash chemistry (reaction times less than 1 s), using online mass spectrometry and global optimization algorithms, such as Bayesian optimization and SNOBFIT, for autonomous online optimization of chemical reactions. The algorithms were tuned and statistically evaluated using simulated optimization runs. Subsequently, they were applied in a practical case study, using a mixing-sensitive example of flash chemistry as a model system to investigate online reaction optimization. Automated response factor fitting was used to obtain quantitative data directly during reaction monitoring. This approach allowed the extraction of meaningful data without the need for postprocessing. The use of an initial design of experiments (DoE) approach was advantageous as it provides a well-discovered experimental space and often leads to a minimal number of subsequent experiments for optimization. Although random starting points may require fewer total experiments, the DoE approach offers greater reliability in achieving optimal results. Comparative analysis between Bayesian optimization and SNOBFIT indicates that Bayesian optimization outperforms SNOBFIT, achieving better results with fewer experimental iterations. Thus, Bayesian optimization has proven to be a powerful tool for autonomous optimization of chemical processes.