Fixed-bed reactors are widely employed in industrial processes such as ethylene oxide production. Traditional multiobjective optimization typically requires repeated evaluations of a large number of candidate designs, making it time-consuming and computationally expensive. This study introduces a physics-informed DeepONet framework for multi-objective prediction. The model encodes catalyst geometric parameters via a multilayer perceptron (branch net) and processes 3D spatial coordinates using a convolutional neural network (trunk net) to extract spatial features. To ensure physical consistency, mass conservation constraints are embedded in the loss function. The proposed framework effectively captures the nonlinear mapping between 3D structural parameters and flow fields, achieving much higher accuracy compared to purely data-driven models. Coupled with a genetic algorithm, it enables the efficient identification of optimal geometries, resulting in improvements of 23.46% in ethylene conversion and 1.46% in ethylene oxide yield compared with the base case operating point, thereby facilitating practical advancements in industrial reactor design and optimization.
Meng et al. (2026) studied this question.