Machine Learning-Guided Space-filling Designs for High Throughput Liquid Formulations Development

Liquid formulations design typically involves searching a high-dimensional space, owing to the combinatorial selection of ingredients from a larger subset of available ingredients, with a relatively limited experimental budget. Therefore, we need to efficiently select the most informative experiments. These experiments need to optimise the composition of these industrially-manufactured products towards customer defined target-properties. Consequently, we have a mixed discrete-continuous Design of Experiments (DoE) problem, for which there are few computationally efficient solutions, with the exception 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, we explore a system of shampoo formulations, where only stable products are considered feasible. We show a weighted-space filling design, where predictive phase stability classifiers are trained for difficult-to-formulate sub-systems, to guide these experiments to regions of feasibility, whilst simultaneously optimising for chemical diversity through building on MaxProQQ.