(104a) Design of Flow Chemistry Experiments Using Batch Bayesian Optimization

Advances in chemical engineering laboratory technologies are transforming traditional “trial-and-error” experimental paradigms [1,2]. Specifically, experiments can be miniaturized and run extremely quickly in the form of microreactors/microdroplets. New measurement techniques further expedite laboratory processes by enabling rapid analyses of experimental outcomes. Given the above, experiments can be performed in high-throughput and must be designed quickly; at times samples must be taken even before receiving full results from previous experiments (e.g., due to measurement time delay).

To this end, recent works, such as [3-5], have focused on “self-optimizing,” or “closed-loop” experimental platforms, which update the design of future experiments continuously as experiments are performed. In most cases, these self-optimizing platforms seek solely to optimize a given performance criterion, rather than to develop a predictive mathematical model as in traditional model-based design of experiments. Therefore, a common strategy is to combine mathematical optimization with a Gaussian Process surrogate model that predicts the outcomes of future experiments in a Bayesian Optimization framework.

This work investigates how high-throughput flow-chemistry experiments can be continuously and systematically designed, while accounting for (uncertain) varying experimental timeframes and heterogenous measurements. We propose a new approach that leverages developments in multi-fidelity [6] and batch [7] Bayesian Optimization. These techniques provide, respectively, statistical frameworks for optimizing experimental regimes with different available measurements and with multiple experiments (i.e., an “asynchronous batch”) occurring simultaneously due to associated time delay(s). Our computational studies on black-box optimization of both industrially relevant and standard test instances show that the proposed methods perform favorably compared to random sampling and traditional Bayesian Optimization approaches.

[1] Houben, C., & Lapkin, A. A. (2015). Automatic discovery and optimization of chemical processes. Current Opinion in Chemical Engineering, 9, 1-7.

[2] Selekman, J. A., Qiu, J., Tran, K., Stevens, J., Rosso, V., Simmons, E., Xiao, Y., & Janey, J. (2017). High-throughput automation in chemical process development. Annual Review of Chemical and Biomolecular Engineering, 8, 525-547.

[3] Schweidtmann, A. M., Clayton, A. D., Holmes, N., Bradford, E., Bourne, R. A., & Lapkin, A. A. (2018). Machine learning meets continuous flow chemistry: Automated optimization towards the Pareto front of multiple objectives. Chemical Engineering Journal, 352, 277-282.

[4] Felton, K. C., Rittig, J. G., & Lapkin, A. A. (2021). Summit: Benchmarking machine learning methods for reaction optimization. Chemistry-Methods, 1(2): 116-122.

[5] Coley, C. W., Eyke, N. S., & Jensen, K. F. (2020). Autonomous discovery in the chemical sciences part I: Progress. Angewandte Chemie International Edition, 59(51), 22858-22893.

[6] Kandasamy, K., Dasarathy, G., Oliva, J., Schneider, J., & Poczos, B. (2019). Multi-fidelity Gaussian process bandit optimisation. Journal of Artificial Intelligence Research, 66, 151-196.

[7] González, J., Dai, Z., Hennig, P., & Lawrence, N. (2016). Batch Bayesian optimization via local penalization. In Artificial Intelligence and Statistics (pp. 648-657). PMLR.