Research map / Learning theory and algorithms / Combinatorial and search optimization
EngiBench: A Framework for Data-Driven Engineering Design Research
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces EngiBench and EngiOpt, open-source libraries with unified API, physics-based engineering design benchmarks and datasets, plus baseline ML algorithms, and shows cross-domain comparisons of GANs/diffusion models.
Problem
Data-driven engineering design research is slowed by expensive, hard-to-install physics simulators, scarce shared datasets, and heterogeneous problem formulations, making fair cross-domain comparison of ML and optimization methods difficult.
Approach
A unified Python API wrapping simulator-backed engineering design problems (with datasets) plus a CleanRL-style companion library of GAN, diffusion, and surrogate-based optimization baselines, evaluated with design-specific metrics (COG, RVC, MMD, DPP, RF) beyond statistical similarity.
Key results
- On Airfoil, simulation failure ratio was 0.304 for GAN, 0.034 for Diffusion, and 0.014 for BezierGAN, indicating domain-informed generators are more robust.
- Unconditional GAN2D achieved strong COG and MMD on Beams2D and HeatConduction2D while CGAN2D had the best RVC on Beams2D, and CDiffusion2D had the worst RVC on Beams2D.
- For PowerElectronics, surrogate-estimated Pareto fronts diverged sharply from NgSpice, with all squared MMD tests rejecting distributional equality across 10 runs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use unconditional GAN instead of conditional GAN or conditional diffusion | Inverse design models · Beams2D and HeatConduction2D | DPP | Helped |
| Use conditional GAN rather than the other evaluated generative models | Inverse design models · Beams2D and HeatConduction2D | RVC | Helped |
| Use diffusion instead of GAN or conditional GAN | Inverse design models · Photonics2D | DPP | Helped |
| Use domain-informed Bézier airfoil parameterization instead of standard GAN parameterization | Airfoil inverse design · Airfoil | RF: 0.304 ± 0.167 -> 0.014 ± 0.027 | Helped |
| Use BézierGAN instead of diffusion | Airfoil inverse design · Airfoil | RF: 0.034 ± 0.038 -> 0.014 ± 0.027 | Helped |
| Use surrogate-based optimization with MLP surrogates for PowerElectronics | MLP+NSGA-II · PowerElectronics | squared MMD test | Hurt |
| Use domain-informed Bézier airfoil parameterization | Airfoil generative models · Airfoil | RF | Helped |
Surprising findings
- Unconditional GAN may outperform conditional GAN and diffusion models on engineering-relevant metrics such as COG and MMD.
- Surrogate-based Pareto fronts for PowerElectronics diverge sharply from the simulator despite ensembles and tuning.
Limitations
- Does not support unstructured meshes, grammar-based representations, or dynamic/morphing components.
- Does not include multi-part or assembly-level design tasks.
- Only 10 independent runs per algorithm-problem pair; the experiments are proof-of-concept rather than a comprehensive benchmark.
Experimental setup
Datasets: Airfoil dataset (MachAero/ADflow RANS) HeatConduction2D/3D datasets Beams2D ThermoElasticBeams2D Photonics2D (ceviche FDFD) PowerElectronics (NgSpice)
Benchmarks: Beams2D HeatConduction2D Photonics2D Airfoil PowerElectronics
Metrics: COG RVC MMD DPP RF
Baselines: GAN2D CGAN2D CDiffusion2D Bezier-GAN MLP+NSGA-II
Builds on: CleanRL MIDBench Gymnasium NSGA-II Bezier-GAN
Closest papers by meaning
- DrivAerStar: An Industrial-Grade CFD Dataset for Vehicle Aerodynamic Optimization (NeurIPS 2025 · similarity 0.86)
- 3DID: Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization (NeurIPS 2025 · similarity 0.85)
- Optimize Any Topology: A Foundation Model for Shape- and Resolution-Free Structural Topology Optimization (NeurIPS 2025 · similarity 0.84)
- Adaptive Experimental Design to Accelerate Scientific Discovery and Engineering Design (AAAI 2025 · similarity 0.84)
- BikeBench: A Bicycle Design Benchmark for Generative Models with Objectives and Constraints (NeurIPS 2025 · similarity 0.84)
- Learning Surrogates for Offline Black-Box Optimization via Gradient Matching (ICML 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.