Anubis: Bayesian optimization with unknown feasibility constraints for scientific experimentation

Model-based optimization strategies, such as Bayesian optimization (BO), have been deployed across the natural sciences in design and discovery campaigns due to their sample efficiency and flexibility. The combination of such strategies with automated laboratory equipment and/or high-performance computing in a suggest-make-measure closed-loop constitutes a self-driving laboratory (SDL), which have been endorsed as a next-generation technology for autonomous scientific experimentation. Despite the promise of early SDL prototypes, a lack of flexible experiment planning algorithms prevents certain prevalent optimization problem types from being addressed. For instance, many experiment planning algorithms are unable to intelligently deal with failed measurements resulting from a priori unknown constraints on the parameter space. Such constraint functions are pervasive in chemistry and materials science research, stemming from unexpected equipment failures, failed/abandoned syntheses, or unstable molecules or materials. In this work, we provide a comprehensive discussion and benchmark of BO strategies to deal with a priori unknown constraints, characterized by learning the constraint function on-the-fly using a variational Gaussian process classifier and combining its predictions with the typical BO regression surrogate to parameterize feasibility-aware acquisition functions. These acquisition functions balance sampling parameter space regions deemed to be promising in terms of optimization objectives with avoidance of regions predicted to be infeasible. In addition to benchmarking feasibility-aware acquisition functions on analytic optimization benchmark surfaces, we conduct two realistic optimization benchmarks derived from previously reported studies: inverse design of hybrid organic-inorganic halide perovskite materials with unknown stability constraints, and the design of BCR-Abl kinase inhibitors with unknown synthetic accessibility constraints. We deliver intuitive recommendations to readers on which strategies work best for various scenarios. Overall, this work contributes to advancing the practicality and efficiency of autonomous experimentation in SDLs. All strategies introduced in this work are implemented as part of the open-source Atlas Python library.