Dynamic Multi-Fidelity Decision-Making Based on Extended Bayesian Optimization

State-of-the-art Bayesian optimization algorithms have the shortcoming of relying on a rather fixed experimental workflow. The possibility of making on-the-fly decisions about changes in the planned sequence of experiments is usually excluded and the models often do not take advantage of known structure in the problem or of information given by intermediate proxy measurements [1-3]. We hypothesize that an extended Bayesian optimization procedure, with surrogate models and acquisition functions that can flexibly choose to modify the workflow on the fly, will improve the performance of state-of-the-art methods for optimization in SDLs, in terms of time, materials consumption, and quality of discovered optima.<br/>To address these limitations, we developed a surrogate model composed of a sequence of Gaussian processes, that can take advantage of the modular structure of experimental processes and that is capable of handling sparse datasets where only partial information (proxy measurements) is available for certain or even the majority of data points [4]. We implemented an acquisition function, based on a mixture of expectation improvement and upper confidence bound, that allows the optimizer to selectively sample from individual sub-processes. Finally, we devised a synthetic dataset generator to simulate multi-step processes with tunable function complexity at each step, to evaluate the efficiency of our model compared to standard BO under various scenarios.<br/>We conducted experiments to evaluate our model across nine distinct scenarios involving two-step processes, optimizing 20 randomly generated processes in each scenario. In all scenarios our multi-step optimizer outperformed the benchmark methods, demonstrating superior performance in terms of both the quality of the optimum found within the allotted budget and in terms of convergence speed. This advantage is particularly evident in scenarios where the complexity of the first step exceeds that of the second step. We are currently in the process of validating our results on real-world datasets.<br/><br/>References<br/>[1] Wu, J., Zhang, J., Hu, M., Reiser, P., Torresi, L., Friederich, P., Lahn, L., Kasian, O., Guldi, D.M., Pérez-Ojeda, M.E. and Barabash, A., 2023. Integrated system built for small-molecule semiconductors via high-throughput approaches. Journal of the American Chemical Society, 145(30), pp.16517-16525.<br/>[2] Seifermann, M., Reiser, P., Friederich, P. and Levkin, P.A., 2023. High-Throughput Synthesis and Machine Learning Assisted Design of Photodegradable Hydrogels. Small Methods, 7(9), p.2300553.<br/>[3] Jenewein, K.J., Torresi, L., Haghmoradi, N., Kormányos, A., Friederich, P. and Cherevko, S., 2024. Navigating the unknown with AI: multiobjective Bayesian optimization of non-noble acidic OER catalysts. Journal of Materials Chemistry A, 12(5), pp.3072-3083.<br/>[4] Luca Torresi, Pascal Friederich, 2024, submitted.