(640c) Bayesian Optimization Using Dynamic Experiments

Optimizing chemical reactions is in general a lengthy process. This is especially true when the chemical design space is large or the reaction mechanism is complex. Data-rich experimentation minimize time and resource requirements to study and optimize chemical reactions by means of data-driven experiments or by using the collected data to obtain chemical kinetics information. Different techniques have been developed to explore a chemical design space using data-rich experimentation, to ultimately optimize reactions. Such methods can be coupled with machine learning (ML) to guide the optimization process, possibly achieving a completely automatic procedure. Among the various ML models, Bayesian optimization (BO) methods leverage collected experimental data and statistics to specifically suggest new experiments to be conducted to efficiently converge on a desired optimum. Such methods rely on a series of independent experiments (applied to batch or continuous reactors) to gradually reduce the uncertainty of the optimum location in the chemical design space. As the experiments are independent, they can become extremely time-consuming and may require considerable resources.

To overcome these issues, a BO method was developed using continuous tubular reactors operated under dynamic conditions. Contrarily to methods based on steady-state data, the dynamic experiments provide more information by modifying the operating conditions over time and analyzing the effluents of the reactor continuously. Under ideal conditions, a plug-flow reactor (PFR) operated in dynamic regime is equivalent to a series of batch reactors (BR) having a reaction time corresponding to the residence time of each pocket of fluid in the PFR. On the other hand, each BR is equivalent to a steady PFR. Consequently, a single dynamic experiment collects information of different operating conditions, sampling the design space over a parameter trajectory and obtaining results that are equivalent to steady experiments. This information can be used for both data-driven optimization and chemical kinetics analysis. Compared to steady ones, data-rich dynamic experiments provide the same amount of information over shorter times (reductions up to 90% of experimental time) and with fewer chemicals (saving up to 80% of reactants).

The proposed algorithm computes new dynamic experiments (trajectories) to optimize the reaction conditions and locate the desired optimum with subsequent experiments. The algorithm initially reconstructs the design space using the information collected during the previous experiments. Afterwards, exploitation and exploration of the design space are balanced in a BO framework, using an ad-hoc acquisition function for trajectories, and finally a new dynamic experiment is suggested.

The method was successfully applied experimentally to the optimization of a regioselective Suzuki-Miyaura cross-coupling reaction using an automated platform. The automated system employed a tubular reactor to study the effect of residence time and catalyst speciation on the yield of the desired product. On-line HPLC was used to analyze samples of the reactor outflow together with in-line NMR using a benchtop spectrometer.