POD–LSTM Surrogate Modeling for Transient CFD–DEM Prediction of Gas–Liquid–Solid Mixing

Gas–liquid–solid mixing in mechanically agitated tanks is important in chemical, biochemical, and particulate processing, but high-fidelity CFD–DEM simulations remain computationally expensive for repeated analysis and parametric studies. This work proposes a reduced-order surrogate modeling framework for rapid transient prediction of multiphase mixing fields using CFD–DEM data as the reference solution. The simulated system is a baffled cylindrical stirred tank equipped with a 45° pitched blade turbine impeller. Twelve operating cases are considered, including upward and downward pumping directions at impeller speeds from 300 to 800 rpm. The raw unstructured CFD–DEM plane data are preprocessed by interpolating onto fixed structured grids, applying tank-domain masks, and storing the results as plane-wise multi-field tensors. Proper Orthogonal Decomposition is then used to compress each high-dimensional field into energetic spatial modes and time-dependent modal coefficients. Long Short-Term Memory network is trained to predict the temporal evolution of the POD coefficients as functions of impeller speed, pumping direction, previous flow states, and time. Predicted coefficients are projected back into physical space to reconstruct transient multiphase fields. Model performance is evaluated using leave-one-speed-out validation with RMSE, MAE, normalized RMSE, R², and structural similarity. The proposed workflow provides an interpretable and computationally efficient pathway for accelerating CFD–DEM analysis of transient multiphase stirred-tank flows.