A physics-informed neural network (PINN)-based model investigation for hydrodynamics and conversion in a gas-solid fluidized bed reactor
In the field of catalysis, the gas-solid fluidized bed has become the reactor of choice for processes such as ozone decomposition due to its highly efficient heat and mass transfer properties. To demonstrate the intrinsic impacts and generate high-quality predictions for this fluidization system, it is necessary to develop an efficient correlation model between the diverse impacting factors and the reactor fluidization performance. To enhance the applicability of the model and reduce its dependence on experimental data quality, a Physics-informed Neural Network (PINN) is employed in this study to model the catalytic decomposition of ozone in a gas-solid fluidized bed. In addition to the experimental data, the PINN was trained using an ozone transport equation derived in this work for the fluidization system. The experimental data was divided to train and test the PINN model, and the resulting model was compared with traditional, purely data-driven artificial neural networks. The results indicated that the PINN model's predictions maintained a greater consistency with the rational laws due to the introduced physical constraints, while still ensuring high consistency with the experimental observations. Compared to traditional models, the PINN model is more robust and less impacted by the quality of the experimental data. The introduced physical bias was proved to enhance the model's generalization capability and interpretability. The importance of incorporating physical bias in developing high-quality machine learning models is highlighted for performance prediction in fluidization systems. It underscores the need for continued efforts to advance physics-informed machine learning methods in engineering.
