A Hybrid Physics–Machine Learning Approach for Modeling Plastic–Bed Interactions during Fluidized Bed Pyrolysis

The axial mixing/segregation behavior of single plastic particles in a bubbling fluidized bed reactor has been investigated by noninvasive X-ray imaging techniques in the temperature range of 500–650 °C and under pyrolysis conditions. Experimental results showed that the extent of mixing between the plastic particle and the fluidized bed increases as both the temperature and fluidization velocity increase. Three modeling approaches were proposed to describe the axial mixing/segregation behavior of the plastic particle, i.e., a purely mechanistic model, a physics-informed neural network (PINN), and an augmented PINN (augPINN). The former model is based on the second law of motion. The second model is a standard PINN, built by simply embedding the second law of motion in the loss function. The third approach involves the introduction of a new interphase distribution parameter, P, into the model. This parameter represents the relative importance of the effects of the emulsion and bubble phases on the plastic particle. This parameter was obtained by training the neural network using the X-ray axial displacement data. The augPINN has been shown to outperform both the mechanistic and the standard PINN models in describing the axial mixing/segregation of polypropylene particles. Moreover, the obtained parameter P was found to be physically interpretable. The main novelty of this work is to show how different frameworks based on the concept of physics-informed machine learning can be successfully applied to complex and real-world hydrodynamic data sets.