A curated list of engineering design datasets and benchmarks by the DeCoDE lab.
Dataset
DrivAerNet++ - 39 TB and 8,000 car designs with full CFD fields. Published at NeurIPS 2024, winner of an MIT Open Data Prize, and commercially licensed through MIT TLO.
Dataset
Ship-D - 30,000 parametric ship hulls with wave-resistance labels, downloaded more than 126,000 times and the foundation of the ShipGen generative model.
Dataset
LINKS - 100 million planar linkage mechanisms and 1.1 billion coupler curves from kinematic simulation.
Dataset
BIKED - 4,500+ community-designed bicycles with parametric data and images. Seeded the BikeCAD AI tool.
Dataset
OpenTO - 2.2 million topology optimization samples, the largest open TO corpus and the training foundation of our OAT foundation model.
Dataset
VideoCAD - 41,000+ annotated CAD UI interaction videos for training UI agents (NeurIPS 2025).
Dataset & Benchmark
Dataset & Benchmark
DesignQA benchmark - A multimodal benchmark of 1,451 questions grounded in real engineering documentation for evaluating large language models, built from the Formula SAE competition rules.
Benchmark
CarBench - A comprehensive benchmark for neural surrogates on high-fidelity 3D car aerodynamics, evaluating eleven architectures on DrivAerNet++.
Dataset & Benchmark
BikeBench - A bicycle design benchmark for generative models with objectives and constraints, combining parametric designs, images, and simulation-based performance evaluations.
Benchmark
CADBench - A multimodal benchmark of 18,000 samples across five input modalities and six metrics for evaluating AI-assisted CAD program generation.
Benchmark
TREDBench - A curated collection of 83 real-world tabular regression datasets with engineering labels for studying domain adaptation in tabular foundation models.
Benchmark
Engineering Classification Benchmark - Eight tabular engineering design classification problems for evaluating new classification algorithms, from our "Untrained and Unmatched" paper.
Dataset
LINKS dataset - A dataset of 100 million planar linkage mechanisms and 1.1 billion coupler curves obtained from kinematic simulations. The dataset also contains curated curves, 100 million negative samples, and a publicly available simulation software.
Dataset
BIKED dataset - A dataset of 4,512 bicycle models with parametric data, images, and segmented component images.
Dataset
BIKED++ dataset - A multimodal dataset of 1.4 million bicycle image and parametric CAD designs.
Dataset
FRAMED dataset - A dataset of 4,500 bicycle frames and ten performance metrics obtained from structural simulations.
Dataset
Aircraft Lift and Drag dataset - A dataset of lift and drag performance values of 4,045 3D aircraft models from Shapenet.
Dataset
Turbo-compressors dataset - A dataset of 22 million turbo-compressors and their performance under different operating conditions.
Dataset
VideoCAD dataset - A large-scale dataset of 41,000+ annotated CAD user-interface interaction videos for learning long-horizon UI interactions and 3D reasoning from CAD software.
Dataset
GenCAD-Code dataset - Over 163,000 CAD-model image and CadQuery code pairs derived from the DeepCAD dataset, built for training vision-language models that generate editable CAD code. Hosted on Hugging Face.
Dataset
SynthBal dataset - A 109 GB synthetically balanced CAD dataset from GenCAD-3D, with CAD programs, point clouds, meshes, and STLs, designed to correct complexity imbalance in CAD training data. Hosted on Hugging Face.
Dataset
DrivAerNet++ - A large-scale multimodal dataset of 8000 detailed 3D car meshes and aerodynamic performance data comprising of full 3D pressure, velocity fields, and wall-shear stresses, point clouds and parts annotation.
Dataset
DrivAerNet dataset - A dataset of 4000 detailed 3D car meshes and aerodynamic performance data comprising of full 3D pressure, velocity fields, and wall-shear stresses.
Dataset
Autosurf aircraft dataset A dataset of 1,050 airplane models with segmentation labels, created using NASA's Open Vehicle Sketch Pad (OpenVSP).
Dataset
Car Drag Coefficient - A dataset of 4,948 3D car meshes, their renderings, and their drag coefficients.
Dataset
BlendedNet dataset - An open blended-wing-body aircraft aerodynamics dataset and surrogate model developed with MIT Lincoln Laboratory (IDETC 2025).
Dataset
Topodiff topology optimization dataset - A dataset of 33,000 images corresponding to optimal topologies for diverse boundary conditions. The dataset also contains their physical fields, compliance values, and an additional 42000 non-optimal topologies.
Dataset
OpenTO dataset - The largest open-source topology optimization corpus - 2.2 million SIMP-solved samples spanning varied domain shapes, resolutions, boundary conditions, and loads. Hosted on Hugging Face.
Dataset
NITO-3D dataset - 122,000 optimized 3D topologies across varied resolutions and aspect ratios, generated with our GPU-based high-throughput topology optimization solver - an order of magnitude larger than prior public 3D datasets.
Dataset
SHIP-D dataset - A dataset of 30,000 ship hulls each with design and functional performance information, including parameterization, mesh, point-cloud, and image representations, as well as 32 hydrodynamic drag measures under different operating conditions.
Dataset
Airfoil dataset - A synthetic dataset of 48,503 airfoils and their aerodynamic performance computed using OpenFOAM.
Dataset
Milk Frother dataset - A multimodal dataset of 1,126 milk frother sketches and their text descriptions. The dataset is derived from a milk frother dataset collected at the Brite lab.