MECHBench: A Set of Optimization Benchmarks inspired by Structural Mechanics

Overview

This repository provides an easy-to-use framework for evaluating optimization algorithms on structural mechanics problems using OpenRadioss. The optimization problem setup is defined in main.py, which controls simulation calls, design variables, objectives, and constraints.

We aim to provide a framework for testing algorithms and models on various structural optimization problems. It includes modules for defining optimization problems, generating finite element meshes, solving Explicit Finite Element Method (FEM) models, and post-processing results.

Mechanical Test Cases

The repository currently contains three mechanical benchmark problems: a star-shaped crash box, a three-point bending problem, and a crash tube with trigger optimization. Each of these numerical models provides multiple outputs, enabling the definition of various optimization tasks (single/multi-objective, unconstrained/constrained scenarios). This diversity allows researchers to evaluate optimization algorithms on problems with different characteristics, objectives, and constraints within structural mechanics.

Starbox (Problem 1)	Three-Point Bending (Problem 2)	CrashTube (Problem 3)

Repository Structure

The project is structured as follows:

MECHBench/
├── main.py                      # Main optimization problem script
├── results/                     # Example output results (from Three Point Bending Case)
├── src/
│   └── sob/                        # Directory containing the core modules of the project
│       ├── __init__.py                # Initializes the Python package
│       ├── problems.py                # Defines optimization problem classes
│       ├── fem.py                     # Implements finite element analysis functionality
│       ├── mesh.py                    # Handles mesh generation and manipulation
│       ├── solver.py                  # A function to perform calls to OpenRadioss (starter and engine)
│       ├── post_processing.py         # Provides tools for post-processing FEM simulation results
│       ├── utils/                     # A directory containing utilities for setting up the simulations
│       │   ├── run_openradioss.py       # Contains RunOpenRadioss class to generates the commands to run OpenRadioss as well as external executables (TH_TO_CSV,ANIM_TO_VTK)
│       │   └── solver_setup.py          # Contains RunnerOptions class for simulation setup configuration. Extension of a dictionary to set up the runs.
│       └── lib/                       # A directory containing supplementary files and libraries
│           ├── py_mesh.py                   # Mesh functions for starbox model (Deprecated)
│           ├── py_mesh_v2.py                # Modified py_mesh for crash tube model, trigger generation variables added (Deprecated)
│           ├── gmsh_base_meshes.py          # Base class to generate GMSH meshes via __call__
│           ├── starbox_gmsh.py              # Mesh generation for starbox model with GMSH framework
│           ├── crashtube_gmsh.py            # Mesh generation for crash tube model with GMSH framework
│           └── three_point_bending_gmsh.py  # Mesh generation for three-point bending model with GMSH framework
├── tests.py                      # Contains multiple functions for checking the module functionality
├── README.md                     # Project documentation (this file)
└── requirements.txt              # Python dependencies

Getting Started

1. Clone the repository

git clone https://github.com/BayesOptApp/MECHBench.git
cd MECHBench

2. Install dependencies

Create a new environment (recommended) and install required packages:

pip install -r requirements.txt

⚠️ Ensure that OpenRadioss is installed and accessible from your environment. For installation instructions, refer to OpenRadioss Documentation. There's also an official YouTube video for building it from source code. You can either download the stable version or build the binaries yourself. You should have the tools TH_TO_CSV and ANIM_TO_VTK in the directory ../OpenRadioss/exec as well as the starter and engine executables. Also, the main.py should point the path to the OpenRadioss directory in order to run the code (further instructions later).

Usage

The optimization setup is controlled via main.py. Modify this file to:

• Set the path to your local OpenRadioss installation.
• Select the test case and the dimensionality of your problem.
• Configure objectives and constraints as needed for your problem.
• Specify Runner options and designated output folder.

Running the benchmark

python main.py

This will:

1. Initialize the optimization problem.
2. Evaluate the objective function(s) by running OpenRadioss simulations.
3. Store the results in the designated output folder.

Problem Setup in main.py

Below is a detailed walkthrough of the execution module main.py.

Operating System Detection

if linux_system:
    orss_main_path = "/home/ivanolar/Documents/OpenRadioss2/OpenRadioss_linux64/OpenRadioss/"
else:
    orss_main_path = "C:/Users/iolar/Downloads/OpenRadioss_win64_2/OpenRadioss/"

This code snippet sets the path to your OpenRadioss installation folder based on the operating system.

Runner Options

runnerOptions = {
    "open_radioss_main_path": orss_main_path,
    "write_vtk": True,
    "np": 4,
    "nt": 1,
    "h_level": 1,
    "gmsh_verbosity": 0,
}

Configures how the simulation is run:

Option	Description
open_radioss_main_path	Path to OpenRadioss binaries.
write_vtk	Whether to output VTK files for visualization.
np	Number of processors for parallel runs.
nt	Number of threads per process.
h_level	Mesh refinement level (if used).
gmsh_verbosity	Verbosity level for GMSH mesh generation.

Main function

Usage example of the main function:

def main():
    sim_id = 253
    vector = [2.5, -1]
    f = sob.get_problem(1, 2, "intrusion", runnerOptions, sequential_id_numbering=False)
    obj_value = f(vector, sim_id)
    print(obj_value)

Explanation:

• sim_id = 253: Unique identifier for this simulation run. Results will be saved in folders named accordingly.
• vector = [2.5, -1]: Design variable vector to evaluate (must match problem dimensionality).
• sob.get_problem(1, 2, "intrusion", runnerOptions, sequential_id_numbering=False): Retrieves objective function handle f:
  • 1: Problem index or category ID.
  • 2: Number of design variables.
  • "intrusion": Target metric (e.g., intrusion depth).
  • runnerOptions: Simulation configuration.
  • sequential_id_numbering: Uses sequential numbering if True.
• obj_value = f(vector, sim_id): Evaluates the objective function by:
  • Running OpenRadioss simulation with the design vector.
  • Storing results linked to sim_id.
  • Returning the computed objective value(s).
• print(obj_value): Prints the output for inspection.

Metrics to Extract

In the following table you may find all the possible metrics you can extract for the problems. You can either input a sole target metric or define a list (or tuple) with the names of the metrics to extract multiple. With this in mind, if you want to define different optimization problems (extra constraints, or multi-objective) you can extract the variables/metrics in order to tailor your own problem.

Metric	Requires FEM Simulation	Notes
mass	No
absorbed_energy	No
intrusion	Yes	Requires running FEM simulation
specific-energy-absorbed	No	Computed from mass and absorbed_energy
mean_impact_force	Yes	Requires running FEM simulation
peak-impact-force	Yes	Requires running FEM simulation
usage_ratio	Yes	Requires running FEM simulation, Computed from mean_impact_force and peak-impact-force
load_uniformity	Yes	Requires running FEM simulation, Computed from mean_impact_force and peak-impact-force
penalized_sea	Yes	Requires running FEM simulation, Computed from mass and absorbed_energy, only available for Problem 1
penalized_mass	Yes	Requires running FEM simulation, Computed from mass and absorbed_energy, only available for Problem 2

Contact constributors

• Ivan Olarte Rodriguez (ivan.olarte.rodriguez@liacs.leidenuniv.nl)
• Maria Laura Santoni (maria-laura.santoni@lip6.fr)
• Elena Raponi (e.raponi@liacs.leidenuniv.nl)

Other contributors

• Feifan Li (feifan.li@tum.de)