Fluid Catalytic Cracking (FCC) units are critical yet operationally complex systems characterized by severe non-linearities, strong phase interactions, and harsh thermodynamic environments, making real-time monitoring and product yield prediction challenging. While traditional first-principles kinetic and hydrodynamic models provide mechanistic fidelity, their high computational cost limits real-time closed-loop control. Conversely, purely data-driven soft sensors often suffer from poor generalization and lack interpretability. This study presents a comprehensive state-of-the-art review and comparative framework integrating multi-lump kinetic models, multiphase hydrodynamic intensification, and advanced deep learning architectures—including spatial-temporal neural networks, physics-guided neural networks (PGNN), and explainable artificial intelligence (XAI). By synthesizing mechanistic multi-phase interactions in the reactor-regenerator loop with data-driven soft sensing methodologies, we outline critical design trade-offs in prediction accuracy, physical consistency, and computational overhead. Furthermore, feature attribution frameworks utilizing Shapley Additive Explanations (SHAP) are evaluated to ensure transparent, trustworthy decision-making for plant-wide optimization. This work establishes a robust blueprint for developing next-generation, physics-informed digital twins and intelligent soft sensors in sustainable petroleum refining.

Fluid Catalytic Cracking (FCC); Soft Sensor; Physics Guided Neural Networks; Kinetic Modeling; Explainable AI (XAI).