Breaking the Boundaries of Bayesian Optimization Utilizing Continuous Chemistry Digital Twins

A novel chemistry optimization methodology has been developed and applied to digital twins for a nucleophilic aromatic substitution reaction and a two-step process to produce an intermediate for the antibiotic ciprofloxacin. Bayesian optimization (BO) has been shown to be a highly effective algorithm for the optimization of a multitude of problems from machine learning to chemistry; however, all current methodologies require expert knowledge about the problem of interest for BO to be efficient. In this work, we propose a data analysis strategy, Breaking-the-Boundaries (BtB), to enhance the strengths of BO while diminishing the requirement for expert knowledge by allowing for the design space to be expanded autonomously. BtB analyzes the data set in between each BO iteration to determine which parameter would require a boundary expansion to move the design space toward the true global optimum. We apply our BtB-BO methodology to digital twins and provide evidence that it obtains optimal conditions that traditional BO can never reach when given a weak initial design space as small as 0.55% of the full design space. BtB-BO maintains the BO optimization rate and finds optimal solutions for complex optimization problems with weak priors in fewer than 30 iterations in all case studies.