Comparing underutilized derivative-free algorithms for their usage in automated chemical process optimization

Description

Optimizing chemical processes has traditionally been carried out using trial-and-error laboratory work, for example the one-variable-at-a-time (OVAT) approach. Recently, however, more advanced strategies, such as the Design of Experiments (DoE) approach and even more complex machine-learning algorithms have been reported in literature. Combining these advanced optimization strategies with automated continuous processes and in-line analytics has led to the so-called field of self-optimization of chemical processes. The focus of the optimization algorithms employed in the field is to find sets of process parameters to optimize a specific process step outcome, such as the yield of chemical reactions. Currently, two of the most widely used methods found in literature are the Nelder-Mead Simplex Algorithm and the Stable Noisy Optimization by Branch and FIT (SNOBFIT) method.
While the extant literature in the field of self-optimization has shown the successful application of both mentioned methods, a wide variety of conceivably suitable algorithms exist in computational sciences, that have not yet been used in the self-optimization of chemical process steps.
In this work, 16 derivative-free optimization algorithm implementations were compared both theoretically and via a real-life model reaction. A benchmarking procedure was carried out, meaning that the algorithms were applied on a set of test functions or test problems. All algorithms were tested in three categories: single-optimum test problems with and without noise, as well as multi-optimum test problems without noise. After the benchmarking, all algorithms were then compared via a Suzuki-Miyaura cross-coupling reaction in continuous flow. A fully automated flow setup was developed capable of performing experiments and analyzing the reaction yield in real-time. This was combined with in-house developed Python script, capable of hosting the different optimization algorithms on one platform
Our results show that multiple novel algorithm implementations outperformed the Nelder-Mead and SNOBFIT algorithms in both the theoretical and practical cases, using average number of iterations needed and percentage of successful optimizations as ranking criteria.

Period	24 Aug 2022
Event title	ACHEMA 2022
Event type	Conference
Location	Frankfurt, GermanyShow on map
Degree of Recognition	International