Skip to content

02 Examples and data

Jump to bottom
Lucie Kubíčková edited this page Sep 30, 2026 · 3 revisions

Examples and data

Examples of CFDNNetAdapt usage contain:

  • benchmark examples for ZDT and LZ test functions;
  • a CFD-based shape-optimization example using OpenFOAM; and
  • input, output, and indicator data discussed in the accompanying paper in data/.

The example scripts add src/ and thirdParty/ to sys.path themselves. Run them from their example directory rather than from the repository root.

Running the benchmark examples

The benchmark cases evaluate analytical test functions, so they do not need OpenFOAM. Each case contains:

  • Allrun.py: full optimization run (computationally expensive);
  • partialRun.py: partial run with reduced number of iterations and seeds;
  • testRun.py: small smoke-test configuration with minimal settings;
  • computeIndicators.py: calculation of inverted generational distance and hypervolume indicators; and
  • postProcess.py: visualization of optimization run results in the objective space.

All scripts write optimization results to 01_algoRuns/, including log file, trained neural networks, and data from individual MOEA optimizations.

Example of running a smoke test for ZDT1 with 10 parameters from the case directory:

cd example/ZDTS/ZDT1_10
python3 testRun.py

Other ZDT cases are ZDT2_10, ZDT3_10, ZDT4_02, and ZDT6_02. LZ cases are under example/LZS/, with functions LZ1 through LZ5, LZ7, LZ8, and LZ9; each is provided with 3 and 10 parameters.

Running the CFD shape-optimization example

The CFD example is in example/convDifShapeOptim/. It couples the optimizer to OpenFOAM case generation and simulation.

The directory includes script for smote test testRun.py, reduced optimization run partialRun.py and full optimization run Allrun.py. In smoke test, the CFD simulations are only created and not evaluated. In the full optimization run, CFD simulations are evaluated in full and the whole optimization is computationally expensive.

All scripts write optimization results to 01_algoRuns/ and the generated OpenFOAM cases are placed in ZZ_cases/.

To create the OpenFOAM simulations, additional python scripts are placed in 01_pyCodes and an OpenFOAM base case is saved in 10_baseCase/. To calculate cost function value, a custom OpenFOAM post-processing utility has to be compiled:

cd example/convDifShapeOptim/02_ofCodes
source /opt/openfoam8/etc/bashrc
./Allwmake

Included data

The committed data directory contains data from runs discussed in the accompanying paper (DOI):

Benchmark results

Each of the five ZDT directories contain:

  • *_optSols.dat: reference optimal solutions;
  • *_dnn_nondoms.dat, *_gpr_nondoms.dat, *_nsgaii_nondoms.dat, and *_socemo_nondoms.dat: non-dominated solutions from CFDNNetAdapt, CFDGPRAdapt, NSGA-II, and SOCEMO methods;
  • matching *_hvValues.dat files: hypervolume histories; and
  • matching *_igdValues.dat files: inverted-generational-distance histories.

The available ZDT datasets are ZDT1_10, ZDT2_10, ZDT3_10, ZDT4_02, and ZDT6_02.

CFD data

data/convDifShapeOptim/ contains results for the CFD-based case:

  • default.dat: default-geometry objective values;
  • baseline_common_all.dat and baseline_common_nondoms.dat: baseline optimization results;
  • baseline_new_lastgens.dat: results from next generations calculated with only MOEA; and
  • cfdnnetadapt_lastgens.dat: final CFDNNetAdapt generations.

The original CFD sampling and scaled training data used by the example are also kept under example/convDifShapeOptim/00_prepCFDData/. These files include all sampled solutions, non-dominated solutions, allowed mesh deformation (min and max angles), and raw optimization results with un-evaluatable CFD solutions.

Clone this wiki locally