Interpretable Machine-Learning Surrogate Model for Accelerating CFD-DEM Simulations of Fluidized-Bed Biomass Gasification: From the Viewpoint of 1D Intraparticle Heat Transfer

Abstract Intraparticle heat transfer plays a key role in fluidized-bed biomass gasification reactors. The isothermal (ISO) model commonly used in CFD-DEM neglects this effect, while the more accurate one-dimensional (1D) nonisothermal model is computationally intensive, limiting its large-scale applications. To address these challenges, this study introduces a machine-learning approach that replaces complex intraparticle heat-transfer calculations with a surrogate model. A 1D model accounting for intraparticle thermal resistance was developed and evaluated against the ISO model. The results demonstrate that the 1D model accurately resolves transient intraparticle temperature gradients, providing a much more physically rigorous prediction of syngas components. However, this high fidelity comes at the expense of nearly doubling the computational time. To improve efficiency, three surrogate models (MLP, PINN, and LSTM) were trained on 1D simulation data. Efficiency analysis shows that these surrogate models achieve a computational speedup of approximately 1.5–1.6 times over the baseline 1D model, while maintaining high predictive accuracy. Further evaluation of particle temperature, heat transfer, and reaction rates shows that the PINN model markedly improves physical consistency by enforcing energy-conservation constraints. Consequently, the PINN model retains the 1D model’s high accuracy while achieving substantial computational acceleration. Finally, as particle count increases, the PINN model achieves a 1.2× to 1.6× speedup and saves up to 80 h, demonstrating its significant value for accelerating large-scale reactor simulations. The PINN model bridges the gap between physical accuracy and computational efficiency, facilitating the robust transition from single-particle physics to large-scale fluidized-bed simulations.