An advanced reduced-order model for real-time and high-accuracy computation of industrial gas–solid flow systems

Gas–solid flow systems are ubiquitous in science and engineering. To accurately simulate their phenomenon, the discrete element method coupled with computational fluid dynamics (DEM–CFD) method has become the standard numerical approach. However, the traditional high-fidelity physical DEM–CFD method (i.e., the full-order model, FOM) suffers from prohibitive computational cost. To solve the problem regarding the computational cost of FOM, the data-driven reduced-order model (ROM) has been proposed. Nevertheless, the conventional ROM developed for accelerating the DEM–CFD method cannot predict the aperiodic transient gas–solid flow systems, which is essential to understand plenty of scientific phenomena. This is because the conventional ROMs are commonly based on regression approaches, often requiring periodic or quasi-periodic patterns of considered systems to be learned. To resolve the essential problem for the existing ROMs, an advanced ROM approach, namely, Flattened Proper Orthogonal Decomposition (POD) coefficients ROM (FlaP-ROM), is proposed for parametrized real-time DEM–CFD simulation of aperiodic gas–solid flow systems. Specifically, to avoid the traditional regression, the POD coefficients are flattened to one feature and directly mapped by a fully connected neural network in the parameter space. Validation tests are performed in two large-scale aperiodic transient gas–solid flow systems: a jet fluidized bed and a die-filling. Through the validation tests, FlaP-ROM exhibits excellent accuracy and efficiency in characterizing the flow dynamics and macroscopic properties in a wide range of aperiodic transient gas–solid flow systems.