Cost-Informed Bayesian Reaction Optimization

Jan Weinreich, Alexandre A. Schoepfer, Rubén Laplaza, Jérôme Waser, Clémence Corminboeuf

Bayesian optimization (BO) of reactions becomes increasingly important for advancing chemical discovery. Although effective in guiding experimental design, BO does not account for experimentation costs. For example, it may be more cost-effective to measure a reaction with the same ligand multiple times at different temperatures than buying a new one. We present Cost-Informed BO (CIBO), a policy tailored for chemical experimentation to prioritize experiments with lower costs. In contrast to BO, CIBO finds a cost-effective sequence of experiments towards the global optimum, the “mountain peak” (see Fig. 1). We envision use cases for efficient resource allocation in experimentation planning for traditional or self-driving laboratories.

Figure 1. Cost-Informed Bayesian Optimization (CIBO) versus Bayesian Optimization (BO). BO suggests a direct and steep path with expensive climbing equipment and a higher chance of costs for suffering injuries. CIBO suggests a slightly longer but safer path with lower equipment costs needed for the ascent.