Gas–solid reactor optimization based on EMMS-DPM simulation and machine learning

Abstract

Design, scaling-up, and optimization of industrial reactors mainly depend on step-by-step experiments and engineering experience, which is usually time-consuming, high cost, and high risk. Although numerical simulation can reproduce high resolution details of hydrodynamics, thermal transfer, and reaction process in reactors, it is still challenging for industrial reactors due to huge computational cost. In this study, by combining the numerical simulation and artificial intelligence (AI) technology of machine learning (ML), a method is proposed to efficiently predict and optimize the performance of industrial reactors. A gas–solid fluidization reactor for the methanol to olefins process is taken as an example. 1500 cases under different conditions are simulated by the coarse-grain discrete particle method based on the Energy-Minimization Multi-Scale model, and thus, the reactor performance data set is constructed. To develop an efficient reactor performance prediction model influenced by multiple factors, the ML method is established including the ensemble learning strategy and automatic hyperparameter optimization technique, which has better performance than the methods based on the artificial neural network. Furthermore, the operating conditions for highest yield of ethylene and propylene or lowest pressure drop are searched with the particle swarm optimization algorithm due to its strength to solve non-linear optimization problems. Results show that decreasing the methanol inflow rate and increasing the catalyst inventory can maximize the yield, while decreasing methanol the inflow rate and reducing the catalyst inventory can minimize the pressure drop. The two objectives are thus conflicting, and the practical operations need to be compromised under different circumstance.