Synergy Between Design of Experiments and Machine Learning for Enhanced Domain Expertise

The aim of the study was to explore a simplex mixture of three raw materials using a design of experiment (DOE) and to characterize it in terms of price, viscosities, and stabilities at different conditions (temperatures). Due to a difficulty in postulating an a priori model and the presence of a possible area of ​​instability of the formulas, which could compromise the success of the DOE and the subsequent analysis of the results (no measurable response in the event of instability), a space filling design type with excluded zone was conducted. A first modelling with different machine learning type models (SVM, Gaussian process) was carried out, but certain areas of the experimental space were poorly described due to missing values ​​for viscosity (e.g., too low viscosity or instability of some formulations). Using information from domain expertise, and with the help of a local data imputation method ​​by K-nearest neighbors, the modelling was corrected and provided satisfactory results, thus giving a better representation and understanding of the experimental space and enabling the identification of a promising formulation candidate.