Safe multi‐objective optimization of stirred reactor scale‐up with a scale‐aware shared–expert CFD surrogate

Scaling up stirred‐tank reactors alters hydrodynamics, affecting mixing time ( θ 95 ), heat transfer ( h ), and power ( P ). Traditional scale‐up methods often use single‐criterion rules (e.g., constant P / V ) that overlook trade‐offs, especially when fluid properties change. High‐fidelity CFD is too slow for exploring high‐dimensional design spaces, and single‐scale models often fail to generalize across scales. To overcome these issues, a multi‐scale 3D CFD database spanning diverse geometries, scales, and fluids was constructed and a multi‐objective scale‐up framework (SSEN–SOW–NSGA‐II) was developed. The core is a Scale‐aware Shared–Expert Network (SSEN) that captures regime‐dependent scaling with specialized neural network experts and safety‐focused training (penalizing underprediction of θ 95 or overprediction of h ). SSEN is integrated with NSGA‐II optimization, and a Safe Operating Window (SOW) constraint enforces macroscopic thermal feasibility through a heat‐removal requirement. The optimized design outperforms constant P / V scaling, reducing mixing time by 41.7% and boosting heat transfer by 48.4%.