Inverse design of the membrane reactors enabled by an inverse-forward physics-informed learning framework

Membrane reactors offer enhanced efficiency and operability over traditional fixed-bed reactors, but their geometric and operating optimization are complicated by the lack of kinetic parameters. This study presents an inverse-forward computational framework that leverages the parameter inversion and forward modeling capabilities of physics-informed machine learning. The parameter inversion model estimates unknown kinetic parameters using a temperature partitioning strategy and noisy data. These parameters are linked to temperature, and the resulting correlations are used in forward modeling to create a parametric model. To reduce the training costs, the framework integrates parameter-based transfer learning with forward modeling. The source and target models are employed to develop gPC-based surrogate models, facilitating multi-objective optimization and decision-making. Using the methane dehydroaromatization membrane reactor as an example, we demonstrate the application of this framework for multi-scenario inverse design (high-yield, equilibrium, and low-coking) to efficiently determine the corresponding operating and geometric conditions.