M2P 2023

Nonlinear Reduced-Order Modelling for Three-Dimensional Turbulent Flow around Complex Geometry using Large-Scale Distributed Machine Learning

  • Ando, Kazuto (RIKEN Center for Computational Science)
  • Bale, Rahul (RIKEN Center for Computational Science)
  • Kuroda, Akiyoshi (RIKEN Center for Computational Science)
  • Tsubokura, Makoto (RIKEN Center for Computational Science)

Please login to view abstract download link

A large-scale machine learning-based nonlinear reduced-order modeling method was applied to a three-dimensional turbulent flow field around a complex geometry shape, namely vehicle body. This method utilizes two types of neural networks. First is a convolutional-autoencoder-like neural network for decomposition the three-dimensional turbulent flow into the arbitrary number of modes [1]. Second is a long short-term memory neural networks to predict the time evolution of the flow field. Due to this method, it was demonstrated that the computational cost decreases several orders of magnitude without a major loss in accuracy. Neural-network-based mode decomposition for the three-dimensional flow field requires a huge computational cost in terms of calculation and memory usage. Therefore, a distributed machine-learning method was implemented using a hybrid parallelism scheme tailored to the network structure [2]. Thus, it was possible to reproduce about 1 million cells of the three-dimensional turbulent flow (Re=6.5×104) around complex shape into tens of modes with sufficient accuracy. In this study, a uniform flow around a vehicle boy model was used as a test case. To validate the method, the reduction performance of the proposed mode decomposition method was compared with the proper orthogonal decomposition (POD) method. Furthermore, the target flow field was reproduced using our method, and the reconstruction accuracy was evaluated in terms of various criteria compared with the accuracy based on POD in conjunction with Galerkin projection method.