U-Net-based deep learning framework for transient and time-averaged predictions in gas–solid fluidized beds

This work proposes a U-Net-based convolutional neural network (CNN) with skip connections to predict transient and time-averaged flow characteristics in a two-phase fluidised bed. Conventional CFD simulations for gas-solid fluidised beds are computationally expensive, and prior ML models often lack generalisation or have high training complexity. A key novelty lies in introducing time as an input channel, enabling the U-Net to handle transient flow prediction without the complexity of recurrent structures. Additionally, the model simultaneously predicts multiple physical and turbulence-related quantities—such as gas and particle volume fractions, velocities, pressure, Reynolds stresses, and turbulent kinetic energy (TKE)—within a unified framework, offering both spatial and temporal fidelity. Trained on 135 CFD simulations with 85 snapshots each, the model generates high-resolution (0.001 s) predictions and shows strong agreement with OpenFOAM CFD results, with deviations below 10 % for volume fractions and under 1 % for pressure, including unseen test cases. The model accurately replicates turbulence features and passes domain-wide error analysis. Computationally, it achieves a 50× speed-up over traditional CFD, covering time-series generation, time-averaging, and turbulence quantification. This establishes the U-Net model as an efficient tool for real-time analysis and large-scale parametric studies in complex multiphase systems. Future work may extend the framework to include heat and mass transfer for thermally reactive or catalytic fluidised beds.