
Data-driven large-step simulations of transport in particulate multiphase flows
Please login to view abstract download link
Many complex, fluid-mechanical systems such as particulate or other types of multiphase flows exhibit a broad range of spatial and temporal scales. This makes the simulation of slow, long-term processes like heat transfer or chemical conversion using standard computational fluid dynamics (CFD) and/or discrete element method (DEM) techniques infeasible. To reduce computation times and ultimately reach real-time capability, one can accelerate prediction algorithms with high-fidelity data obtained from slow but detailed calculations. We present two points of view on the evolution of complex flows and transport processes in them. Sufficiently simple dynamics such as recurrent or pseudo-steady behavior can be time-extrapolated with an iterated nearest-neighbor approach [1], which allows for fast simulations of, e.g., species or heat transfer in such systems with relatively little data. In contrast to this simple, easily interpretable strategy, deep neural operators can be applied more flexibly and to broader range of problems but require significantly more data. Of specific interest are methods like universal physics transformers (UPTs) which first map a state into latent space before propagation takes place in a highly non-linear, purely data-driven fashion. With NeuralDEM [2], we have recently demonstrated how well this strategy works for particle dynamics in various regimes. We have applied our methodologies to a variety of problems ranging from dense, steady granular flow in hoppers to dilute, highly dynamic fluidized beds with a specific emphasis on the temporal multiscale nature of these systems. The power of our data-driven simulations is demonstrated with investigations of residence times, mixing indices and transport behavior. Results agree very well with standard CFD-DEM algorithms but come with runtimes that are several orders of magnitude smaller so that real-time simulations are feasible. REFERENCES [1] T. Lichtenegger, P. Kieckhefen, S. Heinrich, and S. Pirker, “Dynamics and long-time behavior of gas–solid flows on recurrent-transient backgrounds”, Chem. Eng. J., 364, 562-577 (2019). [2] B. Alkin, T. Kronlachne, S. Papa, S. Pirker, T. Lichtenegger, and J Brandstetter, “NeuralDEM – Real-time Simulation of Industrial Particulate Flows”, arXiv preprint arXiv:2411.09678 (2024).