cs.CVMar 13, 2026

A2Z-10M+: Geometric Deep Learning with A-to-Z BRep Annotations for AI-Assisted CAD Modeling and Reverse Engineering

Authors: Pritham Kumar Jena, Bhavika Baburaj, Tushar Anand, Vedant Dutta, Vineeth Ulavala, Sk Aziz Ali

Organizations: 3D Vision Group (3DVG) · BITS Pilani, Hyderabad, India

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Reverse engineering and rapid prototyping of computer-aided design (CAD) models from 3D scans, sketches, or simple text prompts are vital in industrial product design. However, recent advances in geometric deep learning techniques lack a multi-modal understanding of parametric CAD features stored in their boundary representation (BRep). This study presents the largest compilation of 10 million multi-modal annotations and metadata for 1 million ABC CAD models, namely A2Z, to unlock an unprecedented level of BRep learning. A2Z comprises (i) high-resolution meshes with salient 3D scanning features, (ii) 3D hand-drawn sketches equipped with (iii) geometric and topological information about BRep co-edges, corners, and surfaces, and (iv) textual captions and tags describing the product in the mechanical world. Creating such carefully structured, large-scale data, which requires nearly 5 terabytes of storage to leverage unparalleled CAD learning/retrieval tasks, is very challenging. The scale, quality, and diversity of our multi-modal annotations are assessed using novel metrics, GPT-5, Gemini, and extensive human feedback mechanisms. To this end, we also merge an additional 25,000 CAD models of electronic enclosures (e.g., tablets, ports) designed by skilled professionals with our A2Z dataset. Subsequently, we train and benchmark a foundation model on a subset of 150K CAD models to detect BRep co-edges and corner vertices from 3D scans, a key downstream task in CAD reverse engineering. The annotated dataset, metrics, and checkpoints will be publicly released to support numerous research directions.

Figures & tables

DatasetCAD Models#ModelsBRep LabelsText Caption3D Scans3D SketchDesign History
DeepCAD [ 67 ]Simple ✓170,000+✗✗✗✗✓
Fusion360 [ 65 ]Simple ✓∼\sim 8,000✗✗✗✗✓
CADParser [ 73 ]Simple ✓41,000✗✗✗✗✓
CAD-MLLMSimple ✓453,220✗✓✗✗✓
ABC [ 29 ]Cmplx. ✓1,000,000✗✗✗✗✗
CC3D-Ops [ 12 ]Cmplx. ✓∼\sim 50,000✓✗✓✗✗
Table 1: Comparison of open-source CAD datasets w.r.t A2Z
Figure 2: Low-poly CAD meshes are compared to high-quality scans after a series of geometry processing steps ( Sec. 3.1 ).
Figure 3: BRep annotations on 3D scans of A2Z dataset.
Figure 4: Data and annotation analytics of A2Z
Figure 5: Five million sketches with BRep annotations
Figure 6: ∼\sim 5K models of charging points, serial bus etc.
MethodSketch Artist LevelGeminiGPTHuman
Face IDs—8.438.798.37
Face Type—7.888.718.05
Sketches28.217.478.36
37.877.719.12
47.978.119.03
58.318.439.61
Table 2: Visual quality scores (higher is better).
Level MLTD Unigram Bigram Low 59.32 856 11355 High 70.52 908 12181 (a) Avg. MLTD [ 55 ] and uni/bigrams. Method ID Type Loop Boundary 99.37 97.67 99.99 Face 99.93 99.83 — (b) Annotation coverage (%).
Table 3: Two tables highlights the lexical diversity ( left ) on captions and 3D scans’ annotation coverage statistics ( right ).
Figure 7: Qualitative Comparison of state-of-the-art methods in BRep Boundary and Junction classification as a downstream task.
Figure 8: Realism of Sketch.: We compute a CLIP cosine-similarity score of 30.58 between generated captions and the corresponding 2D sketches, confirming that textual descriptions are well-grounded in visual content. Additionally, a tag-level similarity score of 64.7% (CAD-derived tags vs. sketch-derived tags) demonstrates that the hierarchical tag vocabulary remains consistent across the BRep and sketch modalities.
EvaluatorModalityLevel-2Level-3Level-4Level-5
GPT-5sketch7.65±0.347.65\pm 0.347.20±0.487.20\pm 0.485.53±0.385.53\pm 0.385.05±0.395.05\pm 0.39
Geminisketch9.01±0.28\mathbf{9.01\pm 0.28}7.33±0.527.33\pm 0.526.06±0.586.06\pm 0.583.92±0.473.92\pm 0.47
Humansketch7.20±0.327.20\pm 0.327.86±0.38\mathbf{7.86\pm 0.38}7.85±0.447.85\pm 0.447.67±0.717.67\pm 0.71
Table 6: Realism scores (mean ±\pm std, 1–10 scale) for multi-level 2D sketches evaluated by automated VLM raters and human engineering students. Higher is better. Level-3 achieves the strongest evaluator consensus, balancing abstraction with recognizability.
EvaluatorModalityStep-IStep-IIStep-IIIStep-IV
GPT-5scan6.51±0.846.51\pm 0.846.54±1.016.54\pm 1.017.58±0.587.58\pm 0.589.41±0.60\mathbf{9.41\pm 0.60}
Geminiscan7.73±1.487.73\pm 1.487.91±0.927.91\pm 0.928.54±0.788.54\pm 0.789.28±0.81\mathbf{9.28\pm 0.81}
Humanscan6.32±0.296.32\pm 0.296.58±0.806.58\pm 0.807.20±0.517.20\pm 0.519.83±0.22\mathbf{9.83\pm 0.22}
Table 7: Realism scores (mean ±\pm std, 1–10 scale) for synthesized scans at each pipeline stage, evaluated against Artec3D ground-truth scans of CC3D models. Refer to Figure. 12 for visuals. Step IV achieves near-perfect human consensus.
Figure 9: Hierarchical categorization of 1 million CAD models for retrieval and causal mining. The two examples illustrate how – (on the right) one can fetch some CAD models when asked to find – ‘A mechanical component associated to Gear type transmission systems supporting Axial rotation’.
CategoryTag CountPercentage
Geometric Features12,23045.13%
Unclassified6,06822.39%
Mechanical Components4,60116.98%
Structural Elements2,72410.05%
Electronic Components8012.96%
Fasteners6782.50%
Table 8: Distribution of the 27,10227{,}102 unique tags over top-level categories and their Percentage ( %\% ) of CAD models.
Figure 10: This figure visually describes how evaluation and scoring mechanisms from GPT-5, Gemini, and the Human feedback-based evaluation process for different types of annotations on 10K randomly selected samples.
Figure 11: Boundary and Corner detection on randomly selected real-world 3D scan data from the Grab-CAD repository.
Figure 12: Visual comparison between CC3D [ 6 ] CAD models, its real scan pairs, and synthesized scans from CAD (ours)
Figure 13: Visual elaboration of the textual captions on more complex CAD models.
Figure 14: Visual elaboration of the textual captions on more complex CAD models
Figure 15: Visual elaboration of the textual captions on more complex CAD models
Figure 16: Visuals of some randomly selected CAD models of charging ports under A2Z. Color codes define face-Types.
Figure 17: Visuals of same CAD models of charging ports under A2Z where color codes define face-IDs.
Figure 18: Visuals of some randomly selected samples of CAD models for Tablet type enclosures (from Approx. 20K Electronic enclosure models) with varying corner geometry and topology ( at bottom ). A closer look at the BRep faces, edges, and vertices near the corner parts ( on top ) shows how many faces are adjacent to each edge.
Figure 19: More visual examples of BRep Annotations for reverse engineering from 3D scans/sketches.
ChunkBoundaryJunction
 RecallPrecisionRecallPrecision
00.9720.8850.6900.740
10.9720.9050.7100.774
20.9800.8830.6810.703
30.9770.8830.6880.739
40.9640.8410.6980.701
Table 9: Precision-recall measures for all 100 chunks of A2Z using our foundation model.
More in this topic

Boundary Representation Modeling

See all 9 papers

Explore similar work

CardsList
Mar 5, 2026

DreamCAD: Scaling Multi-modal CAD Generation using Differentiable Parametric Surfaces

Computer-Aided Design (CAD) relies on structured and editable geometric representations, yet existing generative methods are constrained by small annotated datasets with explicit design histories or boundary representation (BRep) labels. Meanwhile, millions of unannotated 3D meshes remain untapped, limiting progress in scalable CAD generation. To address this, we propose DreamCAD, a multi-modal generative framework that directly produces editable BReps from point-level supervision, without CAD-specific annotations. DreamCAD represents each BRep as a set of parametric patches (e.g., Bézier surfaces) and uses a differentiable tessellation method to generate meshes. This enables large-scale training on 3D datasets while reconstructing connected and editable surfaces. Furthermore, we introduce CADCap-1M, the largest CAD captioning dataset to date, with 1M+ descriptions generated using GPT-5 for advancing text-to-CAD research. DreamCAD achieves state-of-the-art performance on ABC and Objaverse benchmarks across text, image, and point modalities, improving geometric fidelity and surpassing 75% user preference. Code and dataset will be publicly available.

Mohammad Sadil Khan, Muhammad Usama, Rolandos Alexandros Potamias +4

DFKI, Germany · RPTU Kaiserslautern, Germany · Imperial College London, United Kingdom +1

Text-to-CAD GenerationParametric CAD Modeling
AskPDF
Feb 25, 2026

BrepCoder: A Unified Multimodal Large Language Model for Multi-task B-rep Reasoning

Recent advancements in deep learning have actively addressed complex challenges within the Computer-Aided Design (CAD) domain.However, most existing approaches rely on task-specifi c models requiring structural modifi cations for new tasks, and they predominantly focus on point clouds or images rather than the industry-standard Boundary Representation (B-rep) format. To address these limitations, we propose BrepCoder, a unifi ed Multimodal Large Language Model (MLLM) that performs diverse CAD tasks from B-rep inputs. By leveraging the code generation capabilities of Large Language Models (LLMs), we convert CAD modeling sequences into Python-like code and align them with B-rep. We then adopt a two-stage training strategy: First, pre-training on reverse engineering to learn geometric features and design logic. Second, eff ectively extending the model to various downstream tasks such as completion, error correction, and CAD-QA. Consequently, by interpreting B-rep as structural code, BrepCoder achieves superior generalization across diverse tasks, demonstrating its potential as a general-purpose CAD agent.

Mingi Kim, Yongjun Kim, Jungwoo Kang +1

Chungnam National University, Daejeon, Republic of Korea

Multimodal Large Language Models
AskPDF
Jun 3, 2026

BRepCLIP: Contrastive Multimodal Pretraining on BRep Primitives for CAD Understanding

Learning representations of CAD models is a largely open problem. While 3D representation learning has flourished around point clouds and meshes, the native format of CAD - boundary representations BReps, which encodes exact parametric surfaces, curves, and their topology, has received little attention as a representation learning substrate. We introduce BRepCLIP, the first framework to align BRep geometry with language and image embeddings through contrastive pretraining. We model each CAD object as a sequence of face and edge tokens with separate discrete vocabularies for surface and curve geometry, augmented with spatial and semantic descriptors that capture surface types (e.g., cylindrical, torus, NURBS) and curve primitives (e.g., line, arc, B-spline). A transformer encoder aggregates these tokens into a global BRep embedding, aligned with CLIP's text and image encoders via a joint contrastive objective. BRepCLIP generates more discriminative and semantically grounded embeddings than existing point-based alternatives, improving Top-1 retrieval over OpenShape by 40.4%, 22.0%, and 23.9% on ABC, CADParser, and Automate, respectively, and improving zero-shot classification on FabWave by 15% in Top-1 score. We further demonstrate its utility as a CAD-aware similarity metric for evaluating text and image-conditioned CAD generation, establishing the importance of structure-aware pretraining for multimodal CAD understanding. Project page is available at https://muhammadusama100.github.io/BrepClip2026/

Muhammad Usama, Didier Stricker, Mohammad Sadil Khan +1

1DFKI, Germany · 2RPTU Kaiserslautern-Landau, Germany

3D Shape RetrievalMultimodal PretrainingBoundary Representation ModelingMultimodal Contrastive Learning3D VLMs
AskPDF