Machine Learning Assisted Metaheuristic Based Optimization of Mixed Suspension Mixed Product Removal Process

Mixed-Suspension Mixed-Product Removal (MSMPR) process plays key role in pharmaceutical industry as it is vital for separation and purification stages. A first principles-based detailed model is required to capture physical phenomena occurring in the process. However, such models are time consuming, and sometimes numerically unstable preventing their online implementation. Concurrent optimization can be achieved using data-based modelling. Finite high-fidelity data obtained from the MSMPR model was used to develop surrogates using Support Vector Regression (SVR). The model development is achieved in an optimization framework which results in multiple Pareto optimal SVR models. Generic model development approach is facilitated by assigning each input a different kernel parameter and exploring multiple kernel functions. To ensure model development with limited data points, sample size estimation algorithm is proposed specifically curated for the SVR technique. An evolutionary algorithm is used to find the optimal SVR models, which are fast and accurate, thereby enabling faster optimization of such processes. The study indicates many-fold decrease in optimization time of the MSMPR process thereby facilitating realization of real-time optimization. Comparison with baseline techniques like ridge regression and multi-linear regression techniques establishes the competitiveness of the proposed algorithm.