Surrogate Modeling of Irregular Particle Heating in Gas-Solid Flows Using Deep Operator Networks

Understanding heat transfer in gas-solid flows with irregularly shaped particles is critical for accurate modeling of industrial processes such as fluidized bed reactors and thermal conversion systems. In this study, we develop a data-driven surrogate model using Deep Operator Networks (DeepONets) to predict the transient temperature distribution inside three-dimensional (3D) particles subjected to convective heating. High-fidelity training data are generated using the Glued-Sphere Particle (GSP) model in MFiX, resolving complex intra-particle conduction across a variety of particle geometries and thermal boundary conditions. The DeepONet is trained using exclusively data-driven loss and optimized through distributed parallelism and hyperparameter tuning. The resulting model demonstrates strong agreement with detailed CFD-DEM simulations, capturing spatial and temporal temperature evolution with high accuracy.



This framework enables efficient, high-resolution thermal modeling of particles in gas-solid systems while significantly reducing computational cost, offering a promising pathway for surrogate-assisted multiphase flow simulations.