Machine Learning-aided Process Design for Formulated Products

Robotic experiments were coupled with the previously published Thompson Sampling Efficient Multiobjective Optimization (TS-EMO) algorithm, using a batch sequential design approach, in order to optimize the composition and the process conditions of a commercial formulated product. The algorithm was trained with a previously collected data set used to optimize the formulation without taking into account the influence of the process conditions. The target was to obtain a clear homogeneous formulation within a certain viscosity range, minimizing the cost of the adopted ingredients. The GP surrogate models used in the algorithm were found suitable to model the complex unknown relationship between the input space and the outputs of interest, identifying suitable samples with a general decrease in the formulation price, needed mixing power, and process time. The proposed methodology can lead to quicker product design and therefore can generate considerable profit increase with an early product release time.