Catalytic Cracking of Vacuum Gas Oil: Mechanistic-Based Kinetics versus AI-Based Approaches

This study develops and compares kinetic and machine learning (ML) approaches to model the catalytic cracking of vacuum gas oil (VGO). Experiments were conducted using FCC-type catalysts in a fluidized bench-scale batch Riser Simulator (RS), invented and developed at the Chemical Reactor Engineering Centre (CREC). The RS replicates the industrial FCC conditions, such as temperature, pressure, and reaction time. A 5-lump phenomenological kinetic model is proposed to assess the conversion of VGO and the formation of light cycle oil, gasoline, light gases, and coke. This phenomenologically based model (PBM) incorporates catalyst deactivation via coke deposition and employs a temperature-centered Arrhenius formulation to reduce parameter cross-correlations. In parallel, an ML model is trained on the same experimental data to predict product yields. Both modeling approaches are evaluated based on predictive accuracy, generalization, and computational efficiency. Results show that both the PBM and the ML models provide robust frameworks for optimizing FCC operations and improving product selectivity. It can be observed, however, that the reported ML model, even with its intrinsic limitations in terms of its phenomenological basis, gives a better overall product lump fitting than the PBM, with this being particularly true for the coke lump.