Machine learning-guided space-filling designs for high throughput liquid formulation development

Abstract

Liquid formulation design involves using a relatively limited experimental budget to search a high-dimensional space, owing to the combinatorial selection of ingredients and their concentrations from a larger subset of available ingredients. This work investigates alternative shampoo formulations. A space-filling design is desired for screening relatively unexplored formulation chemistries. One of the few computationally efficient solutions for this mixed nominal-continuous design of experiments problem is the adoption of maximum projection designs with quantitative and qualitative factors (MaxProQQ). However, such purely space-filling designs can select experiments in infeasible regions of the design space. Here, stable products are considered feasible. We develop and apply weighted-space filling designs, where predictive phase stability classifiers are trained for difficult-to-formulate (predominantly unstable) sub-systems, to guide these experiments to regions of feasibility, whilst simultaneously optimising for chemical diversity by building on MaxProQQ. This approach is extendable to other mixed-variable design problems, particularly those with sequential design objectives.