Multi-objective dynamic optimization of seeded suspension polymerization process

Abstract The temperature profile of the seeded suspension polymerization process was optimized to maximize molecular weight, shell thickness and monomer conversion ratio of core–shell polymer particles. Extreme learning machine radial basis function neural networks with R2 values greater than 0.93 were developed, to predict polymer properties at any point in time, using data generated by a computational fluid dynamics model. The optimal combination of input parameters for each neural network was selected from a pool of 44 variables, by using a weight-based method that uses a support vector regression model, and a global exhaustive search algorithm, consecutively. The neural networks developed were incorporated into a genetic algorithm that maximizes the molecular weight, monomer conversion ratio and shell thickness. The optimum temperature profile generated by the algorithm satisfactorily maximized all target polymer properties. This study also demonstrates that a support vector machine classifier could be reliably used for imposing nonlinear inequality constraints for solving dynamic optimization problems.