Continuous modeling of a particle flow: CFD/AI coupling

  • Darves-Blanc, Quentin (École des Mines de Saint-Étienne)
  • Bonnefoy, Olivier (École des Mines de Saint-Étienne)
  • Martin, Sylvain (École des Mines de Saint-Étienne)
  • Gavet, Yann (École des Mines de Saint-Étienne)
  • Kauric, Guilhem (Orano)
  • Macqueron, Corentin (Orano)
  • William, Johann (Orano)
  • Ndiaye, Abibatou (Orano)

Please login to view abstract download link

The discrete element method (DEM) has made possible to simulate the dynamic behavior of granular media very accurately. Unlike experiments, it allows us to follow the evolution of physical quantities of interest everywhere and at any time. These qualities allowed it to become a first rank tool for simulation in many fields involving powders: geophysics, nuclear or pharmacy... However, despite its advantages, this method presents significant problems. Consumption of computational resources is high, geometries are simplified, the number of particles is much lower than what could be encountered at industrial scales. As for the continuous methods, they suffer from the absence of a constitutive equation to describe the different flow regimes. Therefore, we propose to explore an approach which consists in using Neural Networks trained from DEM data and then coupled to CFD solvers to simulate granular flows. This solver will be trained using pre-existing homogenized DEM results, and will provide the continuous solver (CFD) with information that were previously inaccessible. One of the interests of this approach lies in the capacity of generalization of certain algorithms and in their execution speed: A ML model, once it has been trained, can make almost instantaneous predictions and thus go much faster than "classical" numerical modeling. The objective is to take advantage of the benefits of each of these methods (accuracy, speed...) to accelerate the simulations, to go to the macroscopic scale, while keeping the critical information of the microscopic scales. More precisely, the work we propose consists in determining the dynamic viscosity of a granular fluid by AI trained on homogenized DEM results, then injecting it into a CFD solver during the resolution, thus using the flexibility and the generalization capacity of this type of algorithm. We propose here a presentation of the general approach, the tools used and the results.