acceptodds
Under review as a conference paper at ICLR 2027

Learn the Solid, Not the File: Canonical Inputs for Neural Networks on CAD Boundary Representations

Abstract

Boundary representation (B-rep) is the standard format used by modern CAD systems for parametric 3D models. It turns out, the exact same solid can be represented by different B-reps: for example, two engineers using different operations, a geometry kernel rebuilding the file, and an export setting repartitioning faces will lead to different B-reps even though the underlying solid remains the same. We show that existing B-rep encoders are not robust to variation in the B-rep with the same solid on perturbations applied to standard benchmarks, naturally occurring variations inherent to CAD software, and differences in how designers model the same part via a human dataset we created in FreeCAD. The performance of popular B-rep encoders often collapses catastrophically. We propose the canonical region graph, an input representation whose nodes, features and coordinate frame are derived from the solid itself and show theoretical invariance guarantees on repartitioning and rigid motions. It matches the strongest baseline on standard benchmarks, and is stable under every perturbation we test.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
geometric deep learning, invariant representations, boundary representation (B-rep), computer-aided design (CAD), robustness

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_mcsoOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
1
Soundness
2
Read review
Paper summary

The same solid volume can be represented by many different B-Rep shell topologies. For example, some modelling kernels can split periodic faces. In “LEARN THE SOLID, NOT THE FILE” the authors explore how these topological changes, often introduced by data exchange between modelling kernels, can cause catastrophic loss of robustness on downstream tasks. The lack of rotational invariance is also highlighted.

The authors provide a strategy for mitigating both problems. The B-Rep graph can be canonicalized by removing unnecessary edges between faces which share a geometrically identical underlying surface. The rotation invariance problem can also be fixed by transforming the B-Rep data into a canonical orientation based on the moment of inertia tensor of the shell.

Numerous experiments show these mitigations are effective.

Critical strengths

The authors show an important weakness in many academic B-Rep encoders. These encoders exhibit an over-reliance on the B-Rep topology, making performance on downstream tasks extremely fragile, especially when CAD models are passed between CAD systems.
Encoders often do not provide rotational invariance or equivariance. Consequently, changes in the pose of a solid can cause search and retrieval applications built on the embeddings to fail. The solution of moving the solid to the canonical coordinate system is a pragmatic fix to the problem.

The mitigations suggested by the authors are shown to be effective in all the experiments done. This is a major selling point for the value of the paper. The authors clearly show what is going wrong and how the problems should be fixed.

Why augmentation is not the answer This set of experiments was very interesting.

Critical weaknesses

This paper needs extensive re-writing by a person with good skills in English. Too much of it was impossible to understand due to what I suspect is a language barrier.

It turns out, the exact same solid can be represented by different B-reps: for example, two engineers using different operations, a geometry kernel rebuilding the file

This sentence is very unclear. You mean that the same boundary shape can be modeled by B-Rep shells with very different topologies.

We show that existing B-rep encoders are not robust to variation in the B-rep with the same solid on perturbations applied to standard benchmarks, naturally occurring variations inherent to CAD software, and differences in how designers model the same part via a human dataset we created in FreeCAD.

This sentence is very badly worded.

Appendix D

I’m very interested in the details of how the canonical frame is selected when two eigenvalues are (nearly) equal. The section in the main paper didn’t have enough detail, so I am closely reviewing the appendix section.

We resolve both. Axis signs are fixed by the solid’s third-order moments (its skew along each axis), exactly with a closed-form cubic over the mesh triangles

This is not enough. A reader will need to know exactly how this stuff was computed. What function was used in the CAD system? Did you triangulate and then use your own algorithm? The paper needs to get right down into the integral definition of the moment of inertial tensor and the computation of the third-order moments.

Proposition 1

I understand what the authors are getting at here. The proof makes sense. The weakness is that rational NURBS are capable of defining a single parametric surface which contains multiple primitives. Consider the extrusion of the following 2d NURBS curve.

Degree: p=2p=2

Knot vector: [0,0,0,  1,1,  1+2/2,1+2/2,1+2/2][0,0,0,\;1,1,\;1+\sqrt{2}/2,1+\sqrt{2}/2,1+\sqrt{2}/2]

Control points: (x,y,w)

0, 0, 1

1, 0, 1

2, 0, 1

3, 0, \sqrt{2}/2

3, 1, 1

After extrusion this will build a surface with both a cylindrical and planar section. This kind of surface is actually valid in formats like STEP and it’s quite common from surface modellers which NURBify arc-line sketches and extrude them into single surfaces. CAD software needs to cope with this kind of thing. The definitions leading up to Proposition 1 don’t cover this kind of weird example and consequently the uniqueness to the decomposition suggested by the authors doesn’t really hold. Another wired case to think about is when a T-spline model and then converted to B-Rep.

Questions for authors

Have you read

Vector Neurons: A General Framework for SO(3)-Equivariant Networks

Congyue Deng, Or Litany, Yueqi Duan, Adrien Poulenard, Andrea Tagliasacchi, Leonidas Guibas

This is the only paper I have seen which really addresses rotation invariance issues correctly.

1 comment

Thank you for recognizing the importance of the robustness problem and the effectiveness of the proposed mitigations. We have revised the presentation substantially and address your questions below.

Clarity and presentation. We agree that the original wording obscured the central idea: the same geometric boundary can be represented by B-rep shells with different face and edge decompositions. The revised introduction states this explicitly. The method section now explains the purpose of each construction before introducing its details.

Moment computation and repeated eigenvalues. Appendix D now specifies the boundary-area integrals defining the centroid and inertia tensor, the OpenCASCADE routines used to compute them, and the integration procedure for third-order moments. We triangulate the faces and compute third moments using our own closed-form integration over each triangle. This is exact for the triangles; representing a curved surface by a mesh is still an approximation.

When two axes are ambiguous from repeated eigenvalues, the affected features retain quantities independent of rotation within that plane. For near repeated eigenvalues, directional information in the vector and tensor features is reduced gradually as the eigenvalue gap shrinks. Appendix D now explains these operations and their thresholds, including sign averaging and the separate treatment of region grid features.

Composite NURBS and Proposition 1. Your construction identifies a genuine limitation, which we reproduced. Grouping faces can remove subdivisions within a primitive region, but cannot separate a single face that spans both planar and cylindrical regions. We have revised the result to explicitly restrict the guarantee to primitive-compatible face decompositions: faces may subdivide primitive regions but cannot cross their boundaries.

Vector Neurons. We have now cited and evaluated Vector Neurons. Its rotation-equivariant architecture addresses orientation, while canonical-region construction addresses changes in face partitioning before encoding. Even a rotation-equivariant network can remain sensitive to replacing one face with several because its input graph changes. The VN comparison in our main response shows that canonical regions substantially reduce this sensitivity with an equivariant encoder as well. This supports the compatibility of our representation with different architectures.

We are happy to work with you to ensure that the method, its assumptions, and its contributions are clearly and accurately presented.

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

This paper investigates a robustness problem in learning from CAD B-reps: geometrically identical solids can have different B-rep representations, causing existing B-rep neural networks to produce inconsistent predictions. The paper proposes the Canonical Region Graph, which merges adjacent faces belonging to the same underlying surface and computes geometric features in a canonical coordinate frame derived from the solid itself. The authors also provide theoretical guarantees for invariance to valid face repartitioning and rigid motions, and stability analysis for numerically approximated features. Experiments across multiple datasets and perturbation settings demonstrate that the proposed method substantially improves robustness over existing B-rep encoders.

Critical strengths
  1. Well-motivated and practically relevant problem. The paper identifies an important issue in learning from CAD B-reps: geometrically identical solids can have different face partitions, surface representations, and coordinate frames, potentially leading to inconsistent predictions. This issue is particularly relevant in real-world CAD applications, where models may be created by different designers or processed through different CAD systems.

  2. Principled and intuitive representation design. The proposed Canonical Region Graph addresses representation variability directly at the input level. By merging adjacent faces that share the same underlying surface and expressing geometric features in a solid-derived canonical frame, the method provides a coherent and geometrically motivated solution. Importantly, this representation-level approach is largely independent of the specific graph neural network architecture, making it potentially applicable to different graph-based CAD learning models.

Critical weaknesses
  1. Limited evaluation of the generalizability of the proposed representation across different architectures. The paper presents the canonical region graph as a new input representation for B-rep learning. However, the experiments primarily compare the complete proposed framework with existing methods using different architectures and input encodings. Although Table 6 demonstrates the importance of region merging, it remains unclear whether the observed robustness improvements are mainly attributable to the canonical representation or its combination with the specific graph Transformer and feature design. Evaluating the proposed representation with other compatible architectures, such as AAGNet, UV-Net, or BRepNet, would help establish its generalizability and isolate its contribution.

  2. Insufficient experimental analysis and reproducibility of the proposed framework. Although the paper demonstrates strong robustness under various perturbations, the experimental analysis does not fully explain the contributions of individual components. More detailed ablations, qualitative comparisons, and failure-case analyses would strengthen the empirical evidence. Furthermore, while the authors disclose the use of AI tools for implementation and auditing, the manuscript does not provide an accessible code repository, making independent verification difficult.

Questions for authors

Q1. Generalizability across architectures.

Can the authors clarify whether the proposed canonical representation has been evaluated with other architectures, such as AAGNet, UV-Net, or BRepNet? Would similar robustness improvements be expected when using these architectures with compatible canonical inputs?

Q2. Contribution of individual components.

Table 6 demonstrates the importance of region merging. Could the authors provide additional ablation results to clarify the individual contributions of the canonical frame, feature design, and graph Transformer to the overall performance and robustness?

Q3. Real-world robustness evaluation.

Although the paper includes Rhino import/export experiments and a human study using FreeCAD, the real-world evaluation remains relatively limited. Could the authors provide additional qualitative comparisons or failure-case analyses to demonstrate how the proposed method behaves under different CAD representations?

Q4. Reproducibility and implementation verification.

Given the extensive use of AI tools for implementation and auditing, could the authors clarify the extent of independent code verification and whether the implementation, including the preprocessing and evaluation procedures, will be made publicly available to support reproducibility?

Q5. Evaluation of data augmentation.

Table 5 shows substantial gains from data augmentation. Could the authors provide more details on the augmentation setup? These details would help assess whether augmentation has been sufficiently explored to support the conclusion that it is insufficient.

1 comment

Thank you for recognizing the practical importance of the problem and the motivation. Our main response addresses the shared requests for clearer exposition, additional encoders, and realistic examples.

Q1. Generalizability across architectures. Our main response describes the results across six encoders. We see that across the board, our method is effective.

Q2. Component contributions. These results are now in Appendix B.3 (see Table 12).

Q3. Real-world examples and failures. Appendix E.1 now includes two FreeCAD parts with enlarged comparisons of different face partitions and their canonical regions.

Q4. Reproducibility and verification. We are cleaning up the codebase and will publicly release the implementation. We have already shared the codebase with researchers who contacted us privately.

Q5. Stronger augmentation. We performed a stronger mixed-augmentation experiment in Appendix B.1. Stronger augmentation substantially improves robustness, but generalization and performance gap remains: canonical-region construction achieves markedly lower prediction churn on the tested perturbations. Our method targets the underlying geometric equivalence rather than particular training perturbations, so it can also handle unseen ways of repartitioning a surface; the evidence suggests that augmentation is not consistent when there are examples from unseen perturbations.

We welcome further discussion of the evidence and presentation.

Reviewer_t7vAOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

A method to canonicalize the B-rep of a 3D model by merging face patches, standardizing coordinate frames, and fitting procedural parts to underlying analytic surfaces. This is useful as a way to standardize the B-rep for input to neural net encoders on such shapes. The paper demonstrates that this canonicalization makes predictions on downstream tasks much more stable to representational perturbations that preserve the underlying shape, evaluating against existing B-rep networks and their own standardization approaches.

Critical strengths

The evaluation seems rigorous and the invariance guarantees are welcome.

Critical weaknesses

Even though the idea of the paper seems interesting, the way it is presented is doing both itself and its readers a great disservice, which is a shame.

It is clear that even though there’s been human editing, the paper is poor at doing what a paper should be doing. I think that a paper should have a sense of its reader, a “theory of mind” for the reader’s understanding at any given point in the text, and be able to act on this understanding to guide them through the rest. The paper currently does not respect its audience this way, dedicating detail to the wrong things; it leaves out steps where methodological insight could be gained and instead dwells on minute implementation details in order to emphasize how correct it is, details which are pointless if the reader has not gained any intuition about the method in the first place.

The method exposition (section 3) is only barely a page of dense prose; there are no insightful figures that show the important mechanics of the method (key computation, primitive fitting, example result on realistic models), only very high-level conceptual ones for patch merging and canonical frames. At the very least, a figure showing all the steps composed together in action on a real model would have been illuminating. There are no equations or notation that would better describe those mechanics either (...despite the fact that very formal-looking proofs for correctness and invariance are present). Those paragraphs of dense prose themselves are strange:

  • L153 “then the key is based on the surface type and parameters defining it”. This tells me that it’s based on the surface type, but how exactly is this "based on" relation defined? (Appendix E has some more info, though incomplete still, but I shouldn’t have to scroll to the appendix for this)

  • and then, L154 ”Next, if the face is saved as a spline (i.e. in NURBS, CAD’s generic freeform format that defines a surface using a grid of control points rather than an equation) but actually an analytic surface up to a tolerance, the key is based on the analytic surface.” Isn’t “The key is based on the analytic surface” the same thing that the previous sentence just said ("the key is based on the surface type"), even though this sentence is presented as a contrast to the last one? And again, there are no details on what exactly “based on” entails, and this procedure seems more involved than the one in the first sentence because there seems to be fitting involved ("up to tolerance").

  • The same goes for the rest of that section. L160 “ours is a deterministic least-squares variant with the selection rule and its failure cases in Appendix E“; this seems important and worth maybe some more exposition in the main text. The appendix table row on this ("by fitting sampled points and normals: a fit is accepted below 10^−6 * scale residual [...]") also just dwells on tolerances, and I still don't know what the metrics and procedure are for fitting points and normals to an analytic primitive.

  • There are no figures that show actual CAD models and how this standardization behaves on them. This would provide intuition for how a typical CAD model gets standardized: how the patches are merged, how NURBS curves and procedural parts are transformed into primitives. This is supposed to be a graphics(-adjacent) paper, so it's very bizarre that there are no figures showing it in action on real shapes.

I touched on this above, but all of this would be a bit more tolerable if Appendix E were acceptable. Instead, it is a raw LLM dump with even more impenetrable prose, unhelpful for a human reader. It goes into detail in the wrong ways, has no mathematical details or geometric diagrams even though these are much needed, and is mostly just an info dump on precision and tolerances.

The description of the canonical frame is a bit more thorough, in Appendix D, but more of that really should be in the main text instead.


I will say that it feels very strange for me to be spending effort critiquing the flow of the prose as if it had been all written with intent, when what I'm actually doing is just pointing out obvious LLMisms, Claudisms that will show up on any Claude-generated text, not just this paper (specifically, tics of the Opus 5/Fable 5 family, indeed as the authors have disclosed). I suppose pointing this out will help the authors better tell what to edit out when they use LLMs for paper writing. I do hope that in a revised version of this paper, the authors can put more care into doing this idea more justice.

Questions for authors

As mentioned above:

  • What is the "unique standard form" for the keys for faces, and what invariances are used to compute it and make sure that it's unique?

  • What is the precise "least-squares variant" procedure for fitting to analytic primitives? The appendix only describes tolerances; it doesn't say much about the procedure itself and what is computed from points and normals, what is measured, etc.

  • Figures of this canonicalization on realistic CAD models would be great.

  • It would be helpful to see an evaluation of training the competing models with this proposed canonicalization added (as opposed to the original augmentation/standardizations used in their methods). This would be valuable because the paper doesn't position itself as an innovation in the learning setup (and appears standard in this regard), just data standardization. This means it could be helpful as a preprocess for other B-rep network architectures as well, which would add great value to the proposed method.

1 comment

Thank you for recognizing the interest of the idea and the rigor of the evaluation. We agree that a reader should understand the purpose and mechanics of the method before encountering the details. Our main response addresses the revised presentation, realistic examples, and cross-encoder experiments.

Surface keys and their standard form. The purpose of a key is to answer the following question: do these adjacent faces lie on the same geometric surface, even if the file describes that surface differently? It identifies the supporting surface rather than the boundaries that trim an individual face. The B-rep stores this supporting surface. Native analytic surfaces provide their parameters directly, while spline-stored primitives require recognition and fitting.

Each supported primitive is determined by a small set of geometric parameters. For example, a cylindrical surface has five degrees of freedom: one for its radius, two for its axis direction, and two for its axis position. We construct its key from these parameters. Adjacent faces with matching keys are grouped into connected canonical regions. Disconnected pieces remain separate.

Why spline-stored primitives require fitting. The same surface can be stored in different formats. One file may explicitly store a cylinder and its radius and axis, while another represents the same cylindrical surface as a spline. In the first case, we read the parameters directly. In the second case, we sample points and surface normals, estimate a candidate cylinder, and check whether it matches the samples within tolerance.

Least-squares fitting is a standard approach to estimating geometric primitives from sampled points and normals (e.g.,[1]). In the case of cylinders, we estimate the axis from the normals and fit a circle to the projected points, then check positional and normal agreement. Appendix E now specifies our fitting procedure.

Canonical-frame explanation in the main text. The suggestion to bring more of Appendix D into the method section is also addressed. The detailed computation stays in Appendix D, but its purpose is explained before the reader is referred there.

We welcome your feedback and are happy to work with you to improve the presentation further.

[1] Li, L., Sung, M., Dubrovina, A., Yi, L., & Guibas, L. J. (2019, June). Supervised fitting of geometric primitives to 3d point clouds. In 2019 IEEE/CVF conference on computer vision and pattern recognition (CVPR) (pp. 2647-2655). IEEE.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Overall response

We thank the reviewers for their careful reading and detailed feedback, and for recognizing the motivation, practical value, and significance of our contributions. Many of the concerns were around the clarity and presentation of the paper, along with some suggested additional experiments. The revised manuscript explains the method more directly and includes the additional evidence from the suggested experiments. We are happy to work with the reviewers so ensure that the paper can be suitable for publication.

We address the shared concerns here and the specific questions in the individual replies.

1. Explaining the method more clearly.

The revised method section introduces the intuition before the technical definitions: adjacent faces lying on the same underlying surface are grouped into canonical regions, and their geometry is expressed relative to the whole part. It explains why surface keys are needed, how analytic surfaces are recognized when stored as splines, and how ambiguous frame axes are handled. Appendices D and E have been revised to better explain the method along with providing the detailed procedure.

2. Effect across encoders.

Some reviewers brought up an excellent point that our method should be tested across encoders. We now have comparisons with six encoders.

The table reports prediction churn under composed perturbations: the percentage of paired face predictions that change (lower is better). Entries show without → with canonical regions, as mean ± sample sandard deviation over three seeds.

EncoderFusion360MFInstSeg
EdgeGAT54.53 ± 0.78 → 0.30 ± 0.0726.84 ± 3.33 → 0.008 ± 0.001
AAGNet graph49.82 ± 2.88 → 0.33 ± 0.0522.62 ± 2.61 → 0.007 ± 0.000
UV-Net graph55.93 ± 3.78 → 0.44 ± 0.0133.97 ± 6.12 → 0.015 ± 0.002
GIN58.11 ± 3.14 → 0.55 ± 0.0835.12 ± 0.98 → 0.033 ± 0.002
VN graph60.87 ± 1.10 → 0.82 ± 0.0918.57 ± 0.33 → 0.009 ± 0.001
BRepNet56.98 ± 4.46 → 7.88 ± 1.4520.87 ± 1.12 → 14.00 ± 1.42

Canonical regions improve robustness across all tested encoders, including the rotation-equivariant VN network. These are within-encoder comparisons, not an architecture ranking.

BRepNet also depends on how shared boundaries are divided into edges which is why even after our procedure drastically reduces the churn, it still remains substantial. In the paper, we provide a post-processing extension for encoders similar to BRepNet that brings the churn down to very low levels. These results are shown separately below.

BRepNet inputFusion360MFInstSeg
Canonical regions7.88 ± 1.4514.00 ± 1.42
Canonical regions + boundary-edge combination0.40 ± 0.020.012 ± 0.005

3. Showing the method on real examples.

To address the requests for real examples, Appendix E.1 now shows the method on two human-made FreeCAD parts. Each figure includes the complete part and enlarged views comparing different face partitions with the corresponding canonical regions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.