Simulation of Mixing Processes Using Particle Methods with Surrogate-Based Flow Prediction
Please login to view abstract download link
In industrial applications, predicting the diffusion and mixing of small amounts of gas or liquid within large spaces is often required. Typical examples include indoor air conditioning in buildings and homes, dust dispersion in cleanrooms, and mixing of minor components in chemical plants. Particle-based methods [1] offer high accuracy for predicting such phenomena by employing a Lagrangian approach, which avoids excessive numerical diffusion commonly observed in Eulerian mesh-based methods. However, when the computational domain is significantly larger than the dispersed phase volume, the calculation of large-scale flow becomes computationally expensive. For steady flows, tracer-based approaches combined with mesh-based simulations are widely used. In contrast, for periodically varying flows that evolve over time with increasing velocity, tracer-based methods are difficult to apply. To address this, our previous work developed a hybrid approach combining mesh-based simulations for large-scale periodic flows with particle-based simulations for small-scale transport [2], enabling efficient and accurate predictions. In this study, we further accelerate the computation by constructing a surrogate model of large-scale flow using a neural network, which has recently gained widespread application. By coupling the surrogate flow predictions with particle-based simulations, we achieved substantial reductions in computational time. The proposed method showed good agreement with mesh-based results while reducing the computation time for large-scale flow by more than about a factor of 15. [1] Koshizuka, S. and Oka, Y., “Moving-particle Semi-implicit Method for Fragmentation of Incompressible Fluid”, Nuclear Science and Engineering, Vol. 123 (1996), pp. 421-434. [2] Ishii, E., Sano, T., Hosaka, T., Koshizuka, S. and Matsunaga T., “Application of particle method to mixing process simulation”, 16th World Congress on Computational Mechanics (WCCM-XVI).
