Fast Prediction of FCC Regenerator Temperature Fields Using Digital – Twin Based on Convolutional Autoencoder Dimensionality Reduction and Machine Learning

Fluid catalytic cracking (FCC) regenerators are central to petroleum refining, yet the high computational cost of high-fidelity computational fluid dynamics (CFD) simulations precludes their use in real-time monitoring, optimisation, and digital-twin deployment in maximising the regeneration efficiency of FCC catalysts with least carbon emission. To bridge this gap, we propose a surrogate modelling framework that couples CFD-generated temperature snapshots with dimensionality reduction and machine-learning to reconstruct full two-dimensional regenerator temperature fields instantaneously from operating conditions. One hundred and thirteen CFD cases were simulated using ANSYS Fluent, spanning inlet oxygen concentration (18–24 mol%), inlet gas temperature (523–623 K), and spent catalyst coke content (0.5–2.0 wt%), with 80% of cases used for training and 20% for testing. Intensive efforts were made to assess two reduced-order modelling approaches, including linear proper orthogonal decomposition (POD) and nonlinear convolutional autoencoder (AE), as well as various machine learning regression algorithms (i.e., random forest, Gaussian process regression, and artificial neural network) for their propensity on the mapping between operating conditions and reduced latent space, as well as prediction performance. More importantly, apart from field and zone-averaged temperatures, localised hot spots were selected for the comparison across different approaches. As confirmed, the non-linear autoencoder is superior at reducing the dimension with minimal loss in reconstruction accuracy. All three machine learning regression algorithms demonstrated an acceptable accuracy on the prediction of field and zone average temperature profile, while ANN was the only one that was able to capture the localised hot spots, reflecting its largest propensity in coping with the high non-linearity in the FCC regenerator. The surrogate model validated also successfully reduced the online prediction time from 6–8 hours per CFD run down to seconds, demonstrating a digital-twin-ready methodology for real-time FCC regenerator thermal-state estimation, enabling rapid scenario screening, combustion control, and Industry 4.0–aligned operational decision support in industrial refineries.