ABC Dataset: A Million-Scale Parametric CAD Benchmark
New1. Overview
ABC-Dataset is a collection of one million Computer-Aided Design models released at CVPR 2019 for research on geometric deep learning, hosted by the Deep Geometry organisation and published under an MIT licence for the dataset as a whole. Its distinguishing property is that every model is stored as a collection of explicitly parametrized curves and surfaces rather than as a triangle soup, so the underlying B-rep topology and analytic geometry are available alongside a sampled mesh. That representation makes the dataset a source of ground truth for differential quantities such as normals and curvature, for patch segmentation, for geometric feature detection and for shape reconstruction. Because any resolution or file format can be resampled from the parametric description, different geometric learning algorithms can be compared on identical geometry, which is why the dataset became a standard reference for CAD reconstruction and generative CAD work. For a physics-AI programme it matters because industrial simulation starts from CAD geometry, and a dataset that carries exact surface semantics is the only way to train and fairly evaluate models that must recover design intent rather than merely fit a point cloud.
2. How It Works
Models were harvested from the public Onshape model library over roughly four months using the platform's public interface, downloading STEP and Parasolid files, and then passed through a four-step processing pipeline. The first step loads the B-rep with Open Cascade and translates each face and edge into an explicit parametric form, so every surface patch and curve has an analytic description that can be sampled at any resolution. The second step uses Gmsh to tessellate, either with a uniform maximum edge length or with curvature-adaptive refinement. The third step walks the topology tree to establish correspondence between the continuous and discrete representations, computes differential quantities analytically as ground truth, and exports per-patch and sharp-edge labels into YAML. The fourth step post-processes the results to filter by quality and patch count, compute statistics and resample meshes to a target vertex count. Four automatic filters are applied: models with empty or tiny source STEP files are dropped, models containing only a single primitive are dropped, models needing shape healing during translation or meshing are dropped, and meshes whose output size suggests over-refinement are dropped.
3. Scope and Assumptions
The dataset covers CAD models of man-made mechanical parts, which is its explicit scope and also its acknowledged limitation: it carries no semantic labels and no category annotations, so a model is a geometric object rather than a labelled instance of a class. Each model is distributed in up to ten formats, of which meta, STEP, Parasolid and merged STL are guaranteed to be present; object meshes with per-vertex normals and curvature, feature YAML, statistics and Onshape FeatureScript are available for roughly three quarters of the models, and processing errors mean some models lack some formats. Data is distributed in chunks of ten thousand models per format, with URL lists, MD5 and size manifests provided for verification. The dataset assumes users will sample the parametric description themselves for their target resolution, and it assumes a research use of the CAD models, whose copyright remains with the original Onshape creators. It is not a semantic or physical dataset: there are no material properties, no simulation results and no assembly relations, and the models come from an unfiltered public library, so broken boundaries, self-intersections, duplicate vertices, duplicate models and trivially simple beginner models all remain.
4. Getting Started
The practical entry point is the project page, which links per-format download lists such as all_v00.txt and one list per format, and documents parallel download with wget or curl over the listed chunks. A first session should fetch the meta list only, verify MD5 and sizes against the manifests, and inspect the YAML metadata to understand what a record looks like before pulling the much heavier geometry. The next step is to download the STEP and Parasolid chunks for one or two thousand models, run them through the provided containerised pipeline to produce OBJ meshes with per-vertex ground-truth normals and curvature, and confirm that the patch labels in the feature YAML align with the mesh faces. A useful validation is to resample one model at two different vertex counts and check that the analytic curvature is unchanged while the discrete estimate converges, because that is exactly the property the dataset is built to expose. Teams should start with a small chunk, fix the seed and record the exact chunk identifiers, since results are not comparable across different subsets and the dataset is still growing.
5. Comparison with Alternatives
Against mesh and point-cloud corpora such as ShapeNet and ModelNet, the difference is the representation: those provide sampled geometry without parametric semantics, whereas ABC provides analytic surfaces and curves together with the exact differential quantities derived from them. Against smaller CAD collections such as Thingi10K or machining-feature datasets, ABC's advantage is scale and the presence of explicit parameterisation, at the cost of curation. Against construction-sequence datasets used for generative CAD, the relevant comparison is DeepCAD, which derives about one hundred seventy-eight thousand command sequences from ABC and couples the dataset with a generative model, and CADFit, which targets the mesh-to-CAD direction; ABC itself supplies geometry and ground-truth derivatives but not a modelling history for every model, so it is the substrate those methods build on rather than a replacement for them. For simulation-oriented work it also differs from engineering flow datasets such as AirfRANS and CFDBench, which carry physics fields rather than CAD topology.
6. Benchmarks and Evidence
The paper's own use case is a large-scale surface-normal estimation benchmark. It defines two tasks, local patches grown by breadth-first search from random vertices and full-model prediction, at three mesh resolutions of five hundred twelve, one thousand twenty-four and two thousand forty-eight vertices, and at three training-set sizes of ten thousand, fifty thousand and one hundred thousand models, with an eighty-twenty train-test split published for reproducibility. Twelve methods are compared, seven learned and five classical, using a cosine loss averaged over models and a median angular deviation in degrees. The headline result is counter-intuitive: when classical methods are allowed to use mesh connectivity, the simplest uniform-weighted normal estimate reaches a loss of zero point zero zero nine and a median deviation of zero point zero zero degrees on the full-model, highest-resolution, hundred-thousand setting, outperforming every learned method by an order of magnitude, including a Laplacian network that also uses connectivity. The authors conclude that current graph architectures do not exploit connectivity well. Gains from ten thousand to one hundred thousand samples are described as modest, and learning runs were capped at seventy-two hours, so several entries report results from partially trained models. These figures are the authors' own benchmark on their published splits.
7. Role in Physics-AI Workflows
ABC sits at the front of the physics-AI pipeline, where geometry enters. Every industrial simulation begins with a CAD model, and a dataset that carries exact analytic surfaces and ground-truth differential quantities lets a team train and evaluate geometry-aware models before any solver is involved: surface reconstruction that must recover patch structure, segmentation that must align with real B-rep faces, normal and curvature estimation that must match analytic values, and mesh generation whose quality depends on knowing where the sharp edges are. It is the substrate behind the CAD generation line represented by DeepCAD and the reverse-engineering line represented by CADFit, and its ground truth also makes it a stress test for robust geometry processing, since even a trivial algorithm produced invalid normals on roughly one hundred models. Downstream, the geometry it supplies feeds meshing and visualisation tooling such as PyVista and solver ecosystems such as OpenFOAM, while the physics simulation primer explains why the difference between parametric and tessellated geometry propagates all the way to simulation error. The physics-AI learning path places this geometry stage before the surrogate-modelling stage covered by tools like NeuralOperator and Poseidon, and the classical solver cost-boundary guide frames why geometry fidelity, not model capacity, often decides whether a learned surrogate is worth deploying.
8. Limitations and Failure Modes
The first limitation is the absence of semantics: ABC has no category labels and no semantic annotations, so it cannot be used directly for class-conditional tasks without building your own labelling. The second is curation quality: models come from an unfiltered public library and the four automatic filters are purely geometric and topological, leaving broken boundaries, self-intersections, duplicate vertices, duplicate models and many trivial single-primitive models that were introduced by beginners. The third is format completeness: only meta, STEP, Parasolid and merged STL are guaranteed, so pipelines that depend on feature labels, per-vertex curvature or statistics must handle missing files. The fourth is an indexing trap that reliably bites newcomers: vertex indices in the feature YAML are zero-based while the OBJ file is one-based, so a naive join silently shifts every label. The fifth is that the benchmark trains only a single normal per point, leaving multi-normal detection at sharp edges and corners as explicitly future work, and the cosine loss itself does not penalise a flipped normal. The sixth is scope: the dataset covers CAD parts only, so organic shapes, scanned meshes and simulation outputs are out of scope.
9. Version and Source Notes
ABC-Dataset was published at CVPR 2019 in the paper "ABC: A Big CAD Model Dataset For Geometric Deep Learning" by Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin and Daniele Panozzo, with the arXiv preprint 1812.06216 first submitted on 15 December 2018 and revised on 30 April 2019. The authors are affiliated with TU Berlin, Skoltech, the Institute for Information Transmission Problems and New York University; "Deep Geometry", the name of the GitHub organisation that hosts the project, is not an institution and should not be read as one. The models were collected from the public Onshape library, the dataset as a whole is distributed under an MIT licence, the CAD models remain the copyright of their original creators and are recorded in the per-model metadata, and the processing pipeline is released under the GPL in containerised form. Release history on the project page runs from 24 April 2019, when the first million models shipped with STEP, Parasolid, STL and meta, through 25 May 2019 adding OBJ, feature and statistics formats for about seven hundred fifty thousand models, 29 August 2019 adding the normal-estimation benchmark, to 29 September 2019 adding FeatureScript files. All formats currently carry the v00 suffix, and the site states that the dataset continues to grow, so the one-million figure is a 2019 snapshot rather than a fixed final count. Size and per-model patch statistics are published as manifests and statistics files rather than as totals in the paper, and this entry makes no inference about them beyond what those sources state.
10. References
- ABC-Dataset project page, deep-geometry.github.io/abc-dataset, format documentation, download lists, verification manifests and release history.
- Koch, Matveev, Jiang, Williams, Artemov, Burnaev, Alexa, Zorin, Panozzo, "ABC: A Big CAD Model Dataset For Geometric Deep Learning", CVPR 2019, arXiv 1812.06216.
- ABC-Dataset repository, github.com/deep-geometry/abc-dataset, download lists and processing toolkit.
- Onshape public model library, the source of the CAD models used to build the dataset.
- Open Cascade, the open-source B-rep kernel used to translate each face and edge into explicit parametric form.
- Gmsh, the meshing tool used for uniform and curvature-adaptive tessellation.
- Normal-estimation benchmark archive at New York University, archive.nyu.edu handle 2451/59958, the ten-thousand, fifty-thousand and hundred-thousand splits.
- DeepCAD entry on this site, a generative CAD model trained on roughly one hundred seventy-eight thousand ABC-derived command sequences.
- CADFit entry on this site, a mesh-to-CAD approach that complements the geometry-level representation ABC provides.
- ShapeNet and Thingi10K, the mesh and CAD corpora used as scale comparisons in the ABC paper's dataset table.