/BO_GP

Bayesian optimization based on Gaussian processes

Primary LanguagePython

BO-GP

Bayesian optimization based on Gaussian processes (BO-GP) for CFD simulations.

The BO-GP codes are developed using GPy and GPyOpt. The optimizer is non-intrusive and can be linked to any CFD solver.

Reference:

Y. Morita, S. Rezaeiravesh, N. Tabatabaeia, R. Vinuesaa, K. Fukagata, P. Schlatter, Applying Bayesian Optimization with Gaussian Process Regression to Computational Fluid Dynamics Problems, Journal of Computational Physics, 2021.

Exmaple: Turbulent boundary layer (TBL) with non-zero pressure gradient.

See Section 5 in the above reference. The flow is simulated using OpenFOAM.

Questions/Remarks:

Questions can be forwarded to salehr@kth.se, morita@kth.se, and pschlatt@kth.se.

List of included files and folders:

  • driver_BOGP.py: main driver for running the example, i.e. BO-GP of pessure-gradient TBL simulated by OpenFOAM.

  • gpOptim/: Bayesian optimization codes based on Gaussian processes, using GPy and GPyOpt.

    • workDir/
      • gpList.dat
    • gpOpt.py

  • OFcase/: OpenFOAM case folder

    • system/
      • yTopParams.in (written in main_pre.py, used by blockMeshDict & controlDict).
      • blockMeshDict
      • controlDict
      • decomposeParDict
      • fvSchemes
      • fvSolution
    • 0/
      • U,p,k,omega,nut
      • *_IC files (use inflow.py to make these files).
    • constant/
      • polyMesh/ (not included)
      • transportProperties
    • jobscript
    • OFrun.sh

  • OFpost/: Post-processing the results of OFcase.

    • main_post.py

  • OFpre/: Pre-processing the OFcase

    • main_pre.py: creating yTopParams.in using the latest parameter sample.
    • inflow/inflow_gen.py: Creating inflow conditions for RANS of TBL with pressure gradient using DNS data for the TBL with zero-pressure gradient.

  • figs/: To save figures produced when running the optimization.

    • make_movie.sh: make movie in png/ from pdf files.

  • data/: Created when running the BO-GP.

  • storage/: Created when running the BO-GP.

Settings & inputs (to run the example):

  • In driver_BOGP_example.py: U_infty, delta99_in, Nx, Ny, Nz, t, loop params, path, beta_t etc.
  • /gpOptim/gpOpt.py: number of parameters, range of parameters, tolerance, GP kernel, xi, etc.

Requirements:

  1. python3.X
  2. numpy
  3. matplotlib
  4. GPy
  5. GpyOpt
  6. OpenFOAM v.7 (or v.6)
  7. bl_data/ in OFpre/inflow/ (DNS data from here)

How to test the example for different settings:

  • To change the structure of the geometry

    • create the new inflow from precursor using OFpre/inflow/inflow_gen.py (precursor results required)
    • update the blockMeshDict
    • update the driver accordingly

  • To change the number of prosessors used for the OpenFOAM simulation

    • update nProcessors in the driver
    • update decomposeParDict
    • update jobScript

  • To change the parameterization of the upper wall

    • change qBound in gpOpt.py
    • update blockMeshDict

  • To change beta_t (target pressure-gradient parameter beta)

    • change beta_t in the driver

  • When you clone this repository and get errors, please try run:

    • mkdir data
    • mkdir storage
    • mkdir OFcase/constant/polyMesh/