(514b) Robust Multi-Objective Bayesian Optimization of Continuous Flow Polymer Synthesis

Polymerization conducted in continuous flow offers several advantages over classical batch syntheses, including high consistency, precise synthetic pathways, enhanced safety for exothermic polymerizations using highly toxic components, high pressures and temperatures, and access to chemistries otherwise not accessible by traditional batch-based approaches. Simultaneously, flow polymerization has some disadvantages. For example, although the high surface-to-volume ratios due to the small internal dimensions results in effective heat transfer, the residence time can be decreased relative to those in batch processes. The decreased residence time may affect the reaction rate, conversion, and product selectivity.

In this study, a two-monomer polymer synthesis is conducted through a continuous flow reactor with online monitoring tools such as an inline Raman spectrometer. We aim to control the product molecular weight distribution (MWD) and monomer residuals. MWD impacts material properties such as processability, mechanical strength, and morphological phase behavior. We prefer to have polymers with narrow MWDs. The objective function for MWD is to minimize the difference of the actual and target MW. Residual monomers may represent hazardous compounds, the levels of which must be carefully controlled for health, safety and environmental reasons. The objective functions for residual monomers are to minimize residual monomers A and B.

From an optimization standpoint, all three objective functions are black-box objective functions and measurements of the objectives are resource-intensive. For this type of optimization task, Bayesian Optimization (BO) is a highly sample-efficient optimization method. We aim to simultaneously optimize the multiple objectives described above. For multi-objective optimization, typically there would not exist a particular option for which each objective is fully optimized. The ultimate choice is Pareto optimal, that is to say, an improvement in one objective means deteriorating another objective. In a multi-objective Bayesian Optimization (MOBO) scheme, the goal is to learn the Pareto front. We use the BoTorch library’s MOBO algorithms to identify optimal design parameters that include percent solids, initiator loading percent, pressure, temperature, and residence time. In our real-world application, we realize that implementation of the selected design parameters is subject to input noises. For example, temperature and pressure are subject to uncontrollable random input noise around the nominal values. The optimal designs must be robust with respect to the input noise. For this purpose, we apply BoTorch’s MOBO algorithms based on the MVaR (multivariate value-at-risk) concept to obtain a Pareto front that is robust to input noise.