acceptodds
Under review as a conference paper at ICLR 2027

Clarifying the Concepts and Terminologies for representation of AI-Driven Parametric CAD Datasets

Abstract

A CAD model can be encoded either as an explicit geometric representation, such as a point cloud, a mesh, or a boundary representation (B-Rep), or as an implicit procedural representation of its modeling procedure. Explicit geometric representations primarily support tasks such as 3D visual recognition, segmentation, classification, visualization, rendering, analysis, and machining. However, they capture the geometric shape and visual characteristics of a CAD model rather than its design procedure. Implicit representations may encode either simple Boolean operations procedure (solid modeling, simple sketch-extrude) or feature-based modeling procedure (parametric feature-based modeling, which involves sketches, constraints, feature trees, topological naming, and selection mechanisms). Because the “solid modeling” lacks the core component for parametric modeling, editing and reuse of a CAD model still demand substantial manual effort in industrial scenarios, limiting its ability to support rapid product development in globally competitive markets. Consequently, industrial-grade CAD paradigm has progressively evolved from “solid modeling” to “parametric feature-based modeling.” The latter has become the dominant CAD paradigm in contemporary industrial CAD applications, including CATIA, Inventor, Onshape, SolidWorks, Creo, NX and so on. Dataset is the foundation of the AI-driven CAD. Based on two milestone books in the development history of CAD area, this paper clarifies the concepts and terminologies, which are conflated between DeepCAD dataset (the first publicly available dataset for solid modeling) and WHUCAD dataset (the first publicly available dataset for parametric feature-based modeling). Our clarification aligns with the theory researches of CAD area as well as the real-world industrial CAD applications.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
CAD Dataset, Representation Learning, CAD Data Representation, Solid Modeling, Parametric Feature-based Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_wCBWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper clarifies terminology for AI-driven CAD datasets by distinguishing geometric representations from executable modeling procedures, and adjustable parameters from retained feature relationships. Drawing on established CAD literature, it compares the documented representations of DeepCAD and WHUCAD in terms of sketches, constraints, feature organization, selection, and topological naming. An illustrative editing example explains why matching geometry or preserving construction order does not necessarily preserve design intent.

Critical strengths
  1. The paper explains why geometric similarity and executable construction histories do not by themselves establish meaningful parametric editing. This is useful for assessing what CAD datasets actually support.

  2. The comparison distinguishes what a representation can express from what individual dataset models actually contain (Secs. 4.1 and 4.5). It also acknowledges documentation limits, such as explicit sketch constraints not being separately specified in WHUCAD’s original schema (Table 2). These qualifications make the comparison more credible and help avoid overstating the datasets’ capabilities.

Critical weaknesses
  1. Although the paper explicitly focuses on DeepCAD and WHUCAD, it aims to guide the description of AI-driven CAD datasets more broadly. The two examples do not sufficiently show how the proposed distinctions apply to datasets that preserve only some constraints or feature relationships. This limits the practical guidance the paper offers beyond the selected comparison.

  2. The paper mainly organizes established CAD concepts and compares representations already described in the original dataset papers. Figure 2 also adapts an existing example. While the discussion is useful, it does not clearly identify specific misleading claims in AI-CAD research or explain how correcting them changes existing conclusions. The contribution beyond a useful summary therefore remains limited.

Questions for authors
  1. Using the proposed criteria, how would you classify SketchGraphs, Fusion 360 Gallery, and ABC? Would these cases require any changes to the distinctions presented in the paper?

  2. Which components are necessary to classify a representation as parametric feature-based?

  3. How should representations with feature references but no explicit sketch constraints, or constraints but no model-level references, be described?

Reviewer_xEWcOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This is a conceptual paper in the AI‑CAD research area. It uses classic CAD theory documents to sort out different CAD data formats. It tells the difference between explicit geometric representations (point cloud, mesh, B‑Rep), which only store final shape results, and implicit procedural representations that save modeling steps. Further, it distinguishes solid modeling from industrial parametric feature‑based modeling. Having adjustable parameters and operation sequences does not mean real industrial parametric feature‑based modeling. With this analysis framework, the paper compares two popular public datasets, DeepCAD and WHUCAD. DeepCAD only stores solid‑modeling workflows like sketch, extrude and boolean operations, without feature association dependencies. WHUCAD adds engineering features, topological selection and entity references to support associative editing. However, this work only does theoretical analysis based on original public documents. It has no large‑scale quantitative statistics on real dataset samples, so it cannot show how complete the information is inside real data points. It only looks at these two datasets and does not test more similar CAD datasets. It also ignores later extended versions of these datasets. Besides, it proposes no new algorithms and brings no performance gain for CAD generation models. Overall, this meta‑analysis helps standardize AI‑CAD paper writing and reduces concept confusion. But it lacks real‑world experimental support and cannot directly improve CAD model generation performance.

Critical strengths

Many existing AI‑CAD papers misuse key terms. They treat any procedural CAD data with adjustable parameters as “parametric feature‑based modeling”. This paper draws clear boundaries from classic CAD theory. It separates explicit geometry, solid‑modeling procedures and true parametric feature‑based modeling. Clear terminology helps readers correctly judge what capabilities a CAD dataset actually has. This reduces misleading claims in future related work.

Critical weaknesses
  1. All conclusions come from reading original papers and repository documents. The authors do not run statistics on actual samples inside DeepCAD and WHUCAD. Some real‑world dataset entries may deviate from documented schema. Without sample‑level verification, we cannot confirm how many real models follow these theoretical distinctions. This weakens the factual ground of the whole conceptual argument, which weighs negatively for acceptance.
  2. The work only examines DeepCAD and WHUCAD. It does not extend the proposed conceptual framework to other popular public CAD datasets in this field. Also, it ignores later extended versions of these two datasets. Readers cannot know whether these conceptual rules generalize to other resources. Since this paper aims to provide community‑wide guidance, such narrow coverage limits its practical impact.
Questions for authors
  1. Your current analysis draws conclusions only from official documents of the two datasets. Did you perform any sample‑level check on real model files from DeepCAD and WHUCAD? If yes, could you report statistics about how many real samples strictly follow the documented schema? If no, what risks should readers keep in mind when applying your conceptual conclusions to actual dataset samples?
  2. You focus only on DeepCAD and WHUCAD. Have you tried applying your proposed conceptual framework to other widely‑used public CAD datasets? If not, what main barriers stop you from doing such cross‑dataset evaluation?
  3. Your analysis excludes later extended versions of DeepCAD and WHUCAD. Do the extended variants break or change the representational distinctions you describe in this paper? Can you give concrete examples?
  4. In your view, what simple measurable metrics can researchers use in practice to judge whether a given CAD dataset supports true associative parametric‑feature‑based editing?
Reviewer_UkJEOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The core idea of this paper is that the final geometry of a CAD model is insufficient to fully represent its design intent and editability. It is also important to preserve the modeling history and the relationships between features that enable subsequent modifications. To explain this, the paper first distinguishes explicit geometric representations, such as meshes and B-Reps, from implicit procedural representations, such as solid modeling and parametric feature-based modeling. It then argues that modern industrial CAD systems predominantly use parametric feature-based modeling because construction sequences with adjustable parameters alone do not necessarily preserve the dependencies required for reliable editing and model reuse. Finally the paper compares the DeepCAD and WHUCAD datasets to illustrate why preserving a construction sequence with adjustable parameters is not equivalent to preserving the feature relationships necessary for associative editing. Overall the paper primarily aims to clarify existing concepts and terminology in the CAD literature, particularly in the context of AI-driven CAD datasets.

Critical strengths

The strengths are:

  1. The paper is well written and organized, with clear figures and explanations that help illustrate the key concepts.
  2. I agree with the central argument that modern CAD systems rely heavily on parametric feature-based modeling and that this is an important distinction from simply representing a procedural sequence with adjustable parameters.
  3. The paper provides useful terminological distinctions between related but non-equivalent concepts, such as parameterization, feature-based modeling, topological naming, and associative editing. These distinctions are clearly explained.
Critical weaknesses

The weakness are:

  1. My primary concern is the limited novelty compared with existing literature. The distinctions discussed in the paper have been known for a long time in CAD literature. The original DeepCAD and WHUCAD papers also discuss the modeling information captured by their respective datasets. I do not believe the paper provides any substantially new insights.
  2. The paper also does not propose a new algorithm, dataset, or any new understanding that would contribute to machine learning research.
  3. Another major weakness is that the Fusion 360 Gallery dataset is not compared with DeepCAD and WHUCAD. Including this dataset would help illustrate the paper's central argument more clearly.
  4. The paper does not explain how the proposed terminological distinctions should influence model architectures, training objectives, benchmarks, or evaluation protocols.
  5. Several relevant works are missing from the discussion, including UV-Net, BRepNet, and SolidGen.
Questions for authors

My questions are:

  1. How would the proposed distinctions classify the Fusion 360 Gallery dataset and other existing CAD datasets?
  2. What specific new knowledge or conceptual insights does this work contribute beyond the distinctions already established in the CAD literature and documented in the original DeepCAD and WHUCAD papers?
  3. How can these distinctions inform future machine learning research? In particular, do they suggest concrete changes to model architectures, training objectives, benchmark design, or evaluation protocols?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.