acceptodds
Under review as a conference paper at ICLR 2027

Hist2Hex: Construction-History-Guided CAD Decomposition with a Language Model for Hexahedral Meshing

Abstract

Hexahedral meshing of CAD models is central to simulation-driven design. It requires first decomposing the solid into hex-meshable blocks, which remains a long-standing bottleneck. Existing methods operate on the final boundary representation. This representation omits the construction history and the feature boundaries it records. We observe that replaying the Boolean construction history deterministically associates each final body with its body-specific construction subsequence. We present Hist2Hex, a two-stage framework that exploits this structure for hexahedral meshing. In the first stage, deterministic history replay reconstructs the final bodies and establishes these associations between geometry and history without learning. In the second stage, a language model decomposes each sub-solid into hex-meshable leaf blocks. It selects cut planes conditioned on three sources of information: the serialized history, the boundary-representation (B-rep) geometry, and a table of enumerated candidates. Hist2Hex achieves the highest final-mesh success rates across all three datasets, while a controlled Fusion360 ablation shows that serialized history improves candidate selection with Stage 1 and the candidate pool fixed.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
hexahedral meshing

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_mMj7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Hist2Hex, a two-stage framework for CAD block decomposition and hexahedral meshing. It first replays CAD construction histories to associate each solid with its modeling operations. An LLM then recursively selects cutting planes from geometry-derived candidates, conditioned on the construction history and current B-rep.

Critical strengths
  1. I think this paper is interesting. Using CAD construction history to guide block decomposition is a reasonable and interesting idea. The history provides additional structural information that may be difficult to recover from the final B-rep alone.
  2. Selecting cuts from predefined geometric candidates avoids directly predicting continuous cutting parameters. The combination of recursive decomposition and executable verification also makes the predictions more reliable.
  3. The experimental results are encouraging. Hist2Hex achieves the highest meshing success rates across all three datasets. The history ablation on Fusion360 is particularly useful, showing a drop from 60% to 40% when the serialized history is removed while keeping the candidate pool fixed.
Critical weaknesses

I want to clarify that I am not an expert of this field, so the following points are mainly questions that I am curious about.

  1. I find this paper relatively simple and thin in technical content. It mainly focuses on automatically guiding block decomposition, rather than learning the hexahedral meshing process itself. For example, in surface quadrangulation, a closely related research area, NeurCross learns cross fields and uses libQEx to extract quad meshes, while SeamGPT predicts seams (which play a somewhat similar structural role to block decomposition in hexahedral meshing) and further trains an autoregressive model to generate quad-dominant meshes guided by these seams. Although these works address a different meshing problem, they explore more substantial learning-based approaches to mesh generation. In contrast, Hist2Hex mainly uses an LLM to select cutting planes from predefined candidates. Stage 1 relies on deterministic history replay, while the actual hexahedral meshing is handled by existing meshing tools. The training pipeline also largely consists of standard SFT, rejection sampling, and DPO, without introducing a substantially new learning objective or architecture. Therefore, while using CAD construction history is an interesting idea, I find the overall technical contribution relatively incremental and are more suitable for a open-source project.

  2. The comparison with learning-based baselines is not fully convincing. The RL baseline uses a different cutting-action space, while Classifier-hist is a much smaller model (1.6M parameters versus 1.5B). It is therefore difficult to separate the benefits of history conditioning and the LLM architecture from differences in model capacity and action representation.

  3. Figure 8 reports relatively low history-replay success rates for several operations, such as Cut (45.2%) and Mirror (43.8%). Since Stage 2 depends on successfully reconstructed solids and their histories, it is unclear how these failures affect the reported end-to-end performance. I would like the author to exhibit some failure cases of this method.

  4. Some accepted meshes in Table 1 have negative minimum scaled Jacobians, which can indicate inverted elements. It is unclear whether the final verifier explicitly checks for element inversion. This matters because the paper emphasizes generating meshes suitable for numerical simulation.

Questions for authors

Since I am not the expert of this field, I will adjust my score according to the rebuttal from the authors and try to converge the decision of this paper with other reviewers. The reason why I propose weak reject right now is because this paper is not in a good layout (especially some tables and some subtitles) and meanwhile I think it is sort of incremental and requires more explanations or verifications.

Reviewer_3dfJOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes using a CAD part’s construction history to guide a language model in selecting the next cut from a set of candidates, decomposing the part into smaller blocks that are easier to mesh with hexahedral elements.

Critical strengths

The proposed approach is intuitive: the final geometry loses some cues from the modeling process, and the construction history can supply this missing information. The experimental results also support the effectiveness of this approach.

Critical weaknesses

Unless I have overlooked something, the main paper repeatedly refers to supplementary material, but I could not find it. I therefore consider this submission incomplete and would recommend that the authors resubmit.

Questions for authors

If the supplementary material can be found?

Reviewer_3N95OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies block decomposition for hexahedral meshing. Its main idea is to use the CAD construction history, which contains information that is no longer explicit in the final B-rep. Hist2Hex first replays the construction sequence and groups operations by target body. It then uses a fine-tuned Qwen2-1.5B model to recursively select cuts from a set of executable candidates, conditioned on the construction history, current B-rep, and current decomposition state. Training consists of SFT, rejection-sampling bootstrap, and DPO, with an OCC-based meshing pipeline used to verify the generated decompositions.

Critical strengths
  1. The use of construction history is well motivated. The final B-rep loses information about how different features and Boolean operations were created, while this information can be useful for selecting decomposition cuts. The paper turns this observation into a concrete method rather than using history only as additional text.

  2. The predicted actions are executable and are evaluated through the full meshing pipeline. This is considerably more convincing than reporting cut-prediction accuracy alone. The Fusion360 ablation, where Stage 1 and the candidate pool are fixed, also provides useful evidence that the serialized history itself helps.

Critical weaknesses
  1. It is unclear whether Table 1 measures the complete two-stage pipeline. The evaluation set is described in terms of Stage-1-produced sub-solids, while Figure 8 shows that Stage 1 replay can fail fairly often for several feature types. If replay failures are excluded before constructing the 1,500 evaluation instances, then the reported success rates are conditional on successful replay and do not represent performance from an original CAD model to a final mesh. This could materially change the practical conclusion.

  2. Training and evaluation rely heavily on the same fixed meshing/verifier setup. This verifier is used for filtering training data, accepting rejection-sampling results, constructing DPO pairs, deciding when recursion stops, and computing the final success rate. The results therefore show that the method works well with this particular stack, but it is less clear whether the decompositions remain good under another mesher or under different tolerances and timeouts. The matched history ablation is also only reported on Fusion360.

Questions for authors

Can the authors report the history-removed ablation on DeepCAD and WHUCAD while keeping Stage 1 and the candidate pool unchanged? It would also be useful to know whether the ranking remains similar when final meshes are checked using another meshing configuration or different verification tolerances.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.