A physics‐informed neural network for solving gas–solid two‐phase flow

Abstract

Physics‐informed neural networks (PINNs) provide a promising framework for inferring flow fields from indirect measurements, yet their application to gas–solid two‐phase systems remains limited. This work develops a PINN‐based approach to reconstruct gas‐ and solid‐phase velocities directly from gas‐phase volume fraction without requiring initial conditions. The method is first validated using numerical simulation data over a fluidized bed, achieving relative errors on the order of 10 −1 for all gas‐ and solid‐phase velocity components. It is then applied to experimental concentration fields obtained from high‐speed imaging. Using 300 image samples for training, the model simultaneously predicts concentration and velocity fields. The predicted time‐averaged axial velocity profiles agree well with simulations, and cross‐sectional distributions show strong consistency. Furthermore, key flow features, including bubble‐induced structures and voidage patterns, are accurately captured. The proposed framework demonstrates the potential of PINNs for non‐intrusive, data‐driven reconstruction of multiphase flow fields.