Simultaneous optimization of reactor design and operating conditions using geometry-aware physics-informed neural network

Although design of reactor geometry is a crucial task, most existing optimization studies have considered only operating parameters on a fixed geometry, due to the high computational cost of computational fluid dynamics (CFD). To address this, we present an inverse-design framework that does not require CFD data by incorporating a geometry-aware learning concept into a physics-informed neural network (PINN), which can generalize flow fields across any arbitrary geometry. The PINN predicted velocity profiles with a relative root-mean-squared error (RMSE) below 4 % across six geometries. Additionally, a compartment model was employed to efficiently couple the velocity profiles with chemical reactions, and the PINN-based compartment model successfully reproduced ethylene distribution in a polymerization reactor within 3 min at a given blade design. To further leverage this, we jointly optimized the blade geometry and operating parameters to produce a desired molecular weight distribution (MWD). The optimized blade design and condition achieved an MWD matching the target, with an R² of 0.994. It does not require CFD data for training and is generally applicable to any non-ideal reactor design problems, confirming the potential of generalizable PINNs as a practical tool for reactor geometry optimization without repeated CFD simulations.