EngiBench offers a collection of engineering design problems, datasets, and benchmarks to facilitate the development and evaluation of optimization and machine learning algorithms for engineering design. Our goal is to provide a standard API to enable researchers to easily compare and evaluate their algorithms on a wide range of engineering design problems.

Installation

⚠️ Some problems run under Docker or Singularity. Others require native installation of dependencies, please consult the documentation of the specific problem.

pip install engibench

You can also specify additional dependencies for specific problems:

pip install "engibench[beams2d]"

Or you can install all dependencies for all problems:

pip install "engibench[all]"

API

from engibench.problems.beams2d.v0 import Beams2D

# Create a problem
problem = Beams2D(seed=0)

# Inspect problem
problem.design_space  # Box(0.0, 1.0, (50, 100), float64)
problem.objectives  # (("compliance", "MINIMIZE"),)
problem.conditions  # (("volfrac", 0.35), ("forcedist", 0.0),...)
problem.dataset # A HuggingFace Dataset object

# Train your models, e.g., inverse design
# inverse_model = train_inverse(problem.dataset)
desired_conds = {"volfrac": 0.7, "forcedist": 0.3}
# generated_design = inverse_model.predict(desired_conds)

random_design, _ = problem.random_design()
# check constraints on the design, config pair
violated_constraints = problem.check_constraints(design=random_design, config=desired_conds)

if not violated_constraints:
   # Only simulate to get objective values
   objs = problem.simulate(design=random_design, config=desired_conds)
   problem.reset(seed=42)
   # Or run a gradient-based optimizer to polish the design
   opt_design, history = problem.optimize(starting_point=random_design, config=desired_conds)

You can also play with the API here: . We also provide good quality implementations of algorithms compatible with the API in EngiOpt.

Development

Both EngiBench and EngiOpt are open source projects and we welcome contributions! If you want to add a new problem, please reach out to us first to see if it is a good fit for EngiBench.

Installation

To install EngiBench for development, clone the repo, install the pre-commit hooks, and install all dev dependencies:

git clone git@github.com:IDEALLab/EngiBench.git
cd EngiBench
pre-commit install
pip install -e ".[dev]"

Our pre-commit config also includes a hook to ensure compliance with conventional commit formatting (see CONTRIBUTING.md).

Also worth installing ruff and mypy in your editor as we are checking the code style and type safety on our CI.

Adding a new problem

See docs/tutorials/new_problem.md.

License

The code of EngiBench and EngiOpt is licensed under the GPLv3 license. See the LICENSE file for details. All the associated datasets are licensed under the CC-BY-NC-SA 4.0 license.

Citing

If you use EngiBench in your research, please cite the following paper:

@inproceedings{felten_engibench_2025,
	title = {{EngiBench}: {A} {Framework} for {Data}-{Driven} {Engineering} {Design} {Research}},
	url = {https://openreview.net/forum?id=YowD33Q89V},
	urldate = {2025-10-07},
	author = {Felten, Florian and Apaza, Gabriel and B\¨aunlich, Gerhard and Diniz, Cashen and Dong, Xuliang and Drake, Arthur and Habibi, Milad and Hoffman, Nathaniel J. and Keeler, Matthew and Massoudi, Soheyl and VanGessel, Francis G. and Fuge, Mark},
	booktitle = {Proceedings of the 39th Conference on Neural Information Processing Systems ({NeurIPS} 2025)},
	year = {2025},
}