Enhancing cyclone separator performance via computational fluid dynamics and intelligent optimization: synergizing design of experiments, machine learning, and multi-objective genetic algorithms

Abstract

Cyclone separators are extensively utilized for the efficient separation of solid particles from fluid flows, where their operational effectiveness is intrinsically linked to the equilibrium between pressure drop and collection efficiency. However, in extreme industrial environments, such as fluidized catalytic cracking processes, severe wall erosion poses a significant challenge that compromises equipment lifespan. The present study aims to identify an optimal trade-off among separation efficiency, energy consumption, and erosion rate through the optimization of geometric ratios in cyclone separators. By adjusting specific key dimensions, erosion can be mitigated, extending the separator’s lifespan in harsh conditions. The relationships between six geometric dimension ratios and inlet gas velocity with respect to performance indicators are systematically investigated using design of experiments and computational fluid dynamics simulations. To develop a robust performance prediction model that accounts for multiple influencing factors, an auto machine learning approach is employed, incorporating ensemble learning strategies and automatic hyperparameter optimization techniques, which demonstrate superior performance compared to traditional artificial neural network methodologies. Furthermore, pareto-optimal solutions for maximizing separation efficiency while minimizing pressure drop and erosion rate are derived using the nondominated sorting genetic algorithm II, which is well-suited for addressing complex nonlinear optimization problems. The results show that the optimized cyclone separator design enhances separation efficiency from 76.19% to 87.95%, reduces pressure drop from 1698 to 1433 Pa, and decreases the erosion rate from 8.06 × 10⁻⁵ to 7.32 × 10⁻⁵ kg·s⁻¹, outperforming the conventional Stairmand design.