Machine Learning Based Multi-Objective Surrogate Optimization of MSMPR Process

Mixed-Suspension Mixed-Product Removal (MSMPR) process is a prominent unit in fine chemical industry as it helps in purification of the desired materials. Mathematical models of MSMPR process, when addressed through high fidelity approaches, are generally time expensive to simulate, rendering its optimization infeasible in real time fashion. While handling such cases, machine learning techniques can be explored as potential alternatives as they are faster to execute. Within the class of machine learning techniques, Support Vector Regressions stand apart due to the quadratic programming problem generated during their formulation leading to quicker and reliable solutions. For accurate approximation of such processes, a proper design of experiment-based approach is needed, which motivated us to propose a sample size estimation technique. Additionally, the MSMPR behavior is approximated using multiple kernel functions, assigning different priority to each input. The tuning parameters of SVR are optimized in an optimization framework using Non-Dominated Sorting Genetic Algorithm. Selecting a SVR model from a set of alternative solutions using Akaike Information Criterion, the SVR model performance is shown on the unseen data. Simulation studies indicate significant savings in computation time, when compared to the MSMPR model as well as the conventional SVR approach.