Data-driven machine learning for scale-up of bubbling and turbulent fluidized beds: flow hydrodynamics and reactor performance

Fluidized beds provide excellent solids mixing and high heat and mass transfer rates, making them widely used across various industries. However, conventional experimental and computational fluid dynamics (CFD) approaches are often time-consuming and computationally expensive. To address these challenges, this study presents a machine learning approach using multi-layer feedforward neural networks to predict flow hydrodynamics and reactor performance in fluidized beds with diameters of 76 mm and 152 mm. Based on experimental datasets for bubbling and turbulent regimes, models with five input and two output parameters were developed. Three artificial neural network (ANN) architectures were trained using the backpropagation algorithm, with hyperparameters tuned via 5-fold cross-validation. The generalization ability of the models was evaluated using axial and radial distributions of the outputs, leading to the identification of the optimal architecture. Additionally, a hold-out validation was performed to assess performance on unseen data. The Wide ANN methodology, combined with Shapley additive explanations (SHAP), achieved high accuracy (R2 = 0.92 on unseen data for both outputs) with significantly lower computational cost than CFD, demonstrating its potential as an efficient alternative to fluidized bed analysis and scale-up.