A mPOD-based reduced-order modelling approach for fast gas-solid flow simulations
Huiting Chen, Wangyan Li, Jie Bao, Yansong Shen
Abstract
Gas-solid flow systems are widely practised in chemical engineering and known for their multi-scale and intricate complexity, and thus their simulations are usually time-consuming and computationally demanding. To address this long-standing challenge, a novel non-intrusive reduced-order modelling (ROM) approach for gas-solid flow simulations is introduced and applied to a fluidised bed for demonstration. For the first time, a multi-scale proper orthogonal decomposition (mPOD) is applied to decompose the gas-solid data into a set of spatial and temporal bases that are spectrally cleaner and energetically more relevant than those produced by other decomposition methods. To tackle the complexities of high-dimensional data and long-term dependencies encountered in prior approaches, a hybrid deep learning framework, i.e., the transformer encoder–long short-term memory decoder model, is employed for the ROM prediction. The proposed method demonstrates excellent performance in balancing accuracy and efficiency for capturing the complex dynamics of gas-solid flows. The mean absolute percentage error between the proposed ROM and full-order model is as small as 10% in this fluidised bed case, demonstrating high accuracy. Additionally, it achieves a 1000-fold speedup in efficiency, providing a cost-effective tool for practical applications in real-time monitoring and control of gas-solid systems. • A pioneer in applying multi-scale POD to gas-solid flow simulations. • Effectively handles high-dimensional data and long-term dependencies using a hybrid deep learning framework. • Achieves high accuracy for velocity, pressure, and particle void fraction in gas-solid flows. • Provides an approximate speedup of 448 times over full-order simulations.