Methodology for the Interpretation of a Neural Network Model in Comparison to a Physical Model: A Fluid Catalytic Cracking Application

• Published in: Industrial & engineering chemistry research 2024-10, Vol.63 (39), p.16736-16752
• Main Authors: Rodríguez-Fragoso, Martín, Elizalde-Solis, Octavio, Ramirez-Jimenez, Edgar
• Format: Article
• Language: English
• Subjects: Kinetics, Catalysis, and Reaction Engineering
• ISSN: 0888-5885; 1520-5045
• DOI: 10.1021/acs.iecr.4c02002

A black-box model based on a machine learning (ML) strategy is developed to predict the behavior of fluid catalytic cracking (FCC) experimental facility data. The results are then compared against an 8-lump discrete model. An artificial neural network (ANN) is set up for this study, where it was found that the best-performing configuration identified through a systematic trial and error approach encompassed 270 neurons in the hidden layer balancing complexity and generalization, where one of the main metrics was the residual sum of squares (RSS) of the experimental data. The transfer function of the hidden layer is chosen as the sigmoid function, and the transfer function of the output layer was selected as a linear function to directly map the network’s predictions to the required output range. Six descriptors were chosen (time on stream, temperature, molecular weight, and mass fractions of paraffins, naphthenes, and aromatics from freed), which are critical parameters influencing the conversion and selectivity of reactions within the FCC reactor. The results are in good agreement with the experimental results, with a maximum RSS value of 1.215 × 10–3 in the testing stage for gasoline prediction. The parity plots of the predicted values closely match the experimental values, and no overfitting is observed, as the difference between the training and testing stage RSS is 12.7%. On the other hand, a SHAP analysis was used to help the interpretability of the results and compared to a sensitivity analysis to demonstrate the validity of the study. Moreover, a gradient-based methodology was developed for the interpretability of intermediate products. This allowed us to simulate the FCC facility and to generate heat maps where it was possible to find the regions where the highest yield for products of interest like gasoline or coke are obtained with respect to some main operating variables like temperature and time on stream.