Nonlinear Reduced-Order Modeling for Three-Dimensional Turbulent Flow by Large-Scale Machine Learning

This script was used for the paper "Nonlinear Reduced-Order Modeling for Three-Dimensional Turbulent Flow by Large-Scale Machine Learning".

Setting up execution environment on Fugaku

1. Install Python environment

   mkdir -p /data/$(id -g -n)/$(id -u -n)
   tar xvfz ../pytorch.tar.gz -C /data/$(id -g -n)/$(id -u -n)
   ln -s /data/$(id -g -n)/$(id -u -n)/pytorch $HOME/
   

2. Place required data
   • Place flow field snapshot data (na_nn_*.ary) in the directory flowfield
   • Place grid file (ex. grid.dat) in the directory grid

Execution on Fugaku

1. Go to run directory

   cd run/
   

2. Launch interactive job

   pjsub --interact -L "node=3" -L "rscunit=rscunit_ft01" -L "rscgrp=int" -L "elapse=6:00:00" --sparam "wait-time=600" --mpi "proc=12"
   

3. Activate Python environment

   source pytorch_setting.sh
   

4. Create time-averaged flow field data

   python3 ../src/xmean.ary ../param/xmean.toml
   

5. Training (Mode docomposition)

   mpiexec -np 12 python3 ../src/pytorch/train_mdcnnae_modelpar.py ../param/train_mdcnnae.toml
   

   The number of MPI processes must be (number of modes + 1) x (number of MPI processes for the data parallelism). Batchsize is divided by the number of MPI process for the data parallelism.
6. Inference (Output time series of latent vector)

   mpiexec -np 3 python3 ../src/pytorch/infer_mdcnnae_modelpar.py ../param/infer_mdcnnae.toml
   

   The number of MPI processes must be (number of modes + 1).
7. Inference (Flow filed reconstruction)

   mpiexec -np 3 python3 ../src/pytorch/infer_decoded_field.py ../param/infer_decoded_field.toml
   

   The number of MPI processes must be (number of modes + 1).

Setting up execution environment on GPU server

1. Install Python environment (Anaconda)

   wget https://repo.anaconda.com/archive/Anaconda3-2021.05-Linux-x86_64.sh
   sh Anaconda3-2021.05-Linux-x86_64.sh
   

2. Create virtual environment of Anaconda

   conda create --name pytorch-gpu python=3.8
   

3. Install required Python modules

   conda install -y pytorch cudatoolkit=11.3 -c pytorch
   conda install -y toml
   conda install -y matplotlib
   conda install -y scipy
   conda install -y mpi4py
   

Execution on GPU server

1. Go to run directory

   cd run/
   

2. Activate virtual environment of Anaconda

   conda activate pytorch-gpu
   

3. Create time-averaged flow field data

   python3 ../src/xmean.ary ../param/xmean.toml
   

4. Training (Mode docomposition)

   python3 ../src/pytorch/train_mdcnnae.py ../param/train_mdcnnae.toml
   

5. Inference (Flow filed reconstruction)

   python3 ../src/pytorch/infer_mdcnnae.py ../param/infer_mdcnnae.toml
   

References

• T. Murata, K. Fukami, K. Fukagata, Nonlinear mode decomposition with convolutional neural networks for fluid dynamics. Journal of Fluid Mechanics. 2020;882:A13. doi:10.1017/jfm.2019.822
• T. Nakamura, K. Fukami, K. Hasegawa, Y. Nabae, K. Fukagata, Convolutional neural network and long short-term memory based reduced order surrogate for minimal turbulent channel flow. Physics of Fluids 1 February 2021; 33 (2): 025116. https://doi.org/10.1063/5.0039845
• K. Ando, K. Onishi, R. Bale, A. Kuroda, M. Tsubokura, Nonlinear reduced-order modeling for three-dimensional turbulent flow by large-scale machine learning, Comput. Fluids 266 (2023) 106047.
• K. Ando, R. Bale, A. Kuroda, M. Tsubokura, Neural network-based parametric model reduction for predicting turbulent flow for different vehicle geometries, Comput. Fluids (2026) 107196.

License

This project is licensed under the MIT License. See LICENSE for details.