(624g) Snake: Bayesian Optimization Via Pathwise Exploration

The advent of new tools in flow chemistry and droplet micro-fluidic reactors allow us to carry out many experiments in quick succession, making it perfect ground for the application of automatic experiment design [2-3]. Machine Learning has proven to be a very useful tool for this, and has already been successfully applied in the field [4-6]. For the optimization of black-box functions, Bayesian Optimization (BO) uses a Gaussian Process surrogate model to find a balance between exploitative and explorative experiments [4, 7].

Chemical experiments exhibit peculiar challenges that makes traditional BO methods ineffective [5] or inadequate [6] . In particular, we focus on the fact BO algorithms perform under the assumption that changing an input does not incur any additional experimental cost. However, in practice, large changes to some variables, for example temperature, could mean that the experiment results are less reliable or that there must be a significant waiting time between experiments. This can happen when reactors take a long time to make changes or when we need to wait for a reaction to return to steady state.

Consider a scenario where we can see into the future, so that we know beforehand which points classical BO would query. In this case, we could simply order the queries to attain the smallest input cost. In other words, a good solution would require looking into the future, creating an ordering, and then simply following the path defined by the ordering. We present SnAKe [1], a Bayesian Optimization algorithm that builds on this idea. We further investigate convergence properties and show that if applied naively, the algorithm may get stuck at local optima. Through a sample deletion scheme, SnAKe is able to overcome this and focuses on global optimization.

By nature, the algorithm is also able to perform when the results of the experiments are subject to time delays, which is vital for applications where many experiments are run simultaneously and results require time and effort to be processed. Using synthetic benchmarks, we show the algorithm is able to achieve optimization results comparable with classical BO methods, but uses significant smaller input costs. We finalize by testing on SnAr chemistry benchmark [8] where we achieve the best performance at low cost.

[1] Folch, J.P., Zhang, S., Lee, R.M., Shafei, B., Walz, D., Tsay, C., van der Wilk, M. and Misener, R., 2022. SnAKe: Bayesian Optimization with Pathwise Exploration. arXiv preprint arXiv:2202.00060.

[2] Houben, C. and Lapkin, A.A., 2015. Automatic discovery and optimization of chemical processes. Current opinion in chemical engineering, 9, pp.1-7.

[3] McMullen, J.P. and Jensen, K.F., 2010. Integrated microreactors for reaction automation: new approaches to reaction development. Annual review of analytical chemistry, 3, pp.19-42.

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

[5] Waldron, C., Pankajakshan, A., Quaglio, M., Cao, E., Galvanin, F. and Gavriilidis, A., 2019. An autonomous microreactor platform for the rapid identification of kinetic models. Reaction Chemistry & Engineering, 4(9), pp.1623-1636.

[6] Petsagkourakis, P., Sandoval, I.O., Bradford, E., Zhang, D. and del Rio-Chanona, E.A., 2020. Reinforcement learning for batch bioprocess optimization. Computers & Chemical Engineering, 133, p.106649.

[7] Shahriari, B., Swersky, K., Wang, Z., Adams, R.P. and De Freitas, N., 2015. Taking the human out of the loop: A review of Bayesian optimization. Proceedings of the IEEE, 104(1), pp.148-175.

[8] Felton, K.C., Rittig, J.G. and Lapkin, A.A., 2021. Summit: benchmarking machine learning methods for reaction optimisation. Chemistry‐Methods, 1(2), pp.116-122.