{"id":"23a538e3-0369-461d-a31b-5363e85afa6d","arxiv_id":"2506.05417","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A new HDF5-based B-rep format and dataset, with a Python library, that removes the dependency on proprietary CAD kernels in point-cloud learning pipelines.","lead":"This paper introduces a new open file format based on HDF5 for storing CAD boundary representation (B-rep) data, along with a Python library and converted versions of the ABC and Fusion 360 datasets. It aims to let machine-learning pipelines read and process CAD models without proprietary STEP kernels or per-node licenses.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The conversion pipeline is the weak link: the HDF5 format is only as equivalent as OpenCascade's STEP parser, which is never checked against an independent kernel, and the use cases would not detect silently corrupted geometry or topology.","rationale":"The contribution is a replacement for STEP that removes license barriers; the whole value proposition is that the HDF5 file contains the same B-rep with no loss. That claim has two parts: (i) the format is expressive enough, and (ii) the converter fills it correctly. The paper gives a thorough specification for (i), so the weakest link is (ii). The parser is OCC only. OCC is open-source and widely used, but it is not a certified STEP implementation; it has known issues with some STEP constructs and its own meshing fails on a non-negligible fraction of models. The paper's Section 6 sentence that different STEP files require different kernels is an admission that OCC coverage is incomplete, yet the paper neither measures how many files OCC misreads nor compares with another kernel. The use cases are not a substitute: they evaluate on data produced by the same pipeline, and downstream accuracy can be robust to small geometric errors (e.g., normal estimation with a 30-degree threshold is forgiving; reconstruction F-score is computed against the same parametric surfaces, not the original STEP). An error in OCC's parse would propagate into the HDF5 file and the benchmark would not detect it. This is exactly the reader's weakest_assumption. The proposed independent double-parsing test would settle the question: if mismatches are zero, the concern is resolved; if nonzero, the paper needs to report and characterize discrepancies or weaken the equivalence claim. Since the reader already identified this issue and assigned CONDITIONAL, my read does not change the verdict.","tokens_in":8709,"tokens_out":8483,"duration_ms":92844,"concrete_test":"Select 200 models (100 from ABC, 100 from Fusion 360). Parse each original STEP file twice: once with the paper's OpenCascade pipeline and once with an independent, non-OCC STEP kernel/parser (e.g., Parasolid, ACIS, or the originating CAD application's API). Serialize both parses into the HDF5 schema and compare per part and per face: surface type and parameters, 2D and 3D curve definitions, loop/half-edge/edge adjacency, shell/face orientation, and point clouds sampled from the parametric surfaces (using Chamfer/Hausdorff distance). Report mismatch counts, per-entity discrepancy rates, and maximum deviation relative to the model's bounding box. If the two parses agree within a stated numerical tolerance on all 200 models, the OCC-correctness concern is resolved; if not, the paper must either fix the converter, document the failure modes, or weaken the equivalence claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the HDF5 dictionary is an equivalent representation of the STEP B-rep. Section 3.1 states that STEPTOHDF5 uses OpenCascade (OCC) to parse and extract all geometric and topological information; no independent STEP parser or CAD kernel is used anywhere in the pipeline. Section 5's four use cases only show that off-the-shelf models achieve roughly published accuracies on data produced by this pipeline. That does not validate equivalence: if OCC silently repairs a surface, misparameterizes a trim curve, drops a seam edge, or changes a face adjacency, the HDF5 file stores that error, and downstream learning tasks can still produce plausible normals, denoising, reconstructions, and primitive labels. The paper's own Section 6 acknowledges that different STEP files require different kernels, an admission that OCC coverage is incomplete, but the paper never measures how often OCC misparses a STEP file or how large the resulting geometric/topological deviations are. The reported meshing failure rates (1.56% to 8.82%) also show OCC is not fully robust, making untested parsing fidelity a live risk. Since the dataset's entire value is that it removes the need for a CAD kernel without losing B-rep fidelity, this unverified single-point dependence is the most load-bearing assumption in the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a new HDF5-based dictionary format for boundary representation (B-rep) CAD geometry, together with a converter (STEPTOHDF5) based on OpenCascade, a Python processing library (ABS), and a large converted dataset derived from ABC and Fusion 360. The authors argue that this format is an equivalent, fully specified, open alternative to STEP, removing the need for a proprietary CAD kernel in learning pipelines. They demonstrate the library on four use cases: normal estimation, denoising, surface reconstruction, and segmentation, reporting accuracy numbers that are broadly in line with published results.","tokens_in":8962,"tokens_out":4257,"duration_ms":43271,"significance":"If the equivalence claim is established, the format and dataset would be a valuable community resource for machine learning on CAD geometry, enabling large-scale processing without per-node kernel licenses. The paper's strengths include a detailed format specification in Appendix A, released source code, a substantial conversion effort, and a library that provides parametric sampling, normals, and curvature. However, the validation is currently insufficient to support the central claim: the conversion is never checked against an independent CAD kernel, and the use-case experiments are circular with respect to conversion fidelity. The dataset itself, if properly verified, would be a meaningful contribution; the current manuscript does not yet provide that verification.","major_comments":[{"comment":"The central claim that the HDF5 format is 'equivalent' to the original STEP B-rep is not validated. Section 3.1 states that STEPTOHDF5 uses OpenCascade to parse and extract all geometric and topological information, with no independent verification against another kernel. The use cases in Section 5 only compare off-the-shelf model accuracies to published values; they never compare the HDF5 geometry or topology against the original STEP files. A silent OpenCascade parsing error (e.g., a misparameterized trim curve, a dropped seam edge, or a changed face adjacency) would be propagated faithfully into the HDF5 file and would remain undetected. Section 6 even acknowledges that 'different STEP files require different kernels,' which is an admission of incomplete coverage, but no attempt is made to measure how often OpenCascade misparses a file or how large the resulting deviations are. The authors should add a direct comparison between the HDF5 representation and the original STEP geometry/topology, ideally using an independent kernel or a set of analytic ground-truth shapes, with quantitative error metrics.","section":"§3.1, §5, §6"},{"comment":"The ground-truth normals, points, and labels used in all four use cases are computed by ABS from the HDF5 files produced by the same OpenCascade-based pipeline, so those experiments are circular with respect to conversion fidelity. They cannot catch a conversion error: if OpenCascade silently distorts a surface, the sampled points and normals will be consistent with the distorted surface, and the downstream model may still produce plausible outputs. The comparison with BPNet's published accuracy in Table 2 provides some external anchoring, but it is not a test of geometric equivalence. At least one experiment should use an independent source of ground truth, such as evaluating the analytic parametric surfaces directly or reading the original STEP file with a different kernel, to confirm that the HDF5 data matches the STEP geometry.","section":"§5, Listings 3–5"},{"comment":"The surface-reconstruction results in Table 1 are reported without error bars or any measure of variance, and several entries are non-monotonic and counterintuitive. For example, at 4,000 samples the Chamfer distance is 3.39 at σ=0 but 1.91 at σ=0.005, and at 8,000 samples it is 1.63 at σ=0 but 2.56 at σ=0.025. Such values suggest either high variance across the 1,000 selected models or sensitivity to the specific random sampling. Without means and standard deviations over multiple runs or bootstrap resamples, the table does not provide convincing evidence that the format produces data consistent with the method's expectations. This is load-bearing because the table is presented as evidence of the format's integrity.","section":"Table 1"},{"comment":"The reported meshing failure rates (1.56% for ABC, 8.82% for Assembly, 1.04% for Joint, 0.02% for Reconstruction, 0.07% for Segmentation) show that OpenCascade's meshing is not fully robust, yet the paper does not explain how these failures affect dataset usability. It is unclear whether a model with a failed mesh is skipped entirely, whether only some faces are missing, and whether the failure rate correlates with model complexity or face type. This matters because the format stores a mesh per face (Appendix A.3), and a failed mesh means the HDF5 file contains an empty dataset, which downstream users must handle. The authors should clarify the handling of failed meshes and quantify the impact on the dataset's completeness.","section":"§4, Table 2"},{"comment":"The introduction claims 'we convert the Fusion 360 and ABC datasets and add another million models from OnShape,' but the dataset section (§4) lists only one million ABC models and the four Fusion 360 subsets; the OnShape portion is never described in terms of size, selection criteria, or conversion statistics. This discrepancy makes the dataset's actual composition unclear. Please clarify whether the released dataset includes OnShape models and, if so, provide the corresponding details and statistics.","section":"§1, §4"}],"minor_comments":[{"comment":"The comment above Listing 1 says 'Example of computing normal at every point,' but the code computes a binary label and returns 1 or 0, not a normal. The comment should be corrected, or the listing should be replaced with the normal-computing code from Listing 3.","section":"Listing 1"},{"comment":"Listing 5 contains Python syntax errors: 'elif :' on line 11 has a misplaced colon, and line 13 'degree [(2,3),(3,2)]' is missing an equals sign. Since the paper demonstrates the library's ease of use, the code should be syntactically valid.","section":"Listing 5"},{"comment":"The caption of Figure 2 says 'Point-cloud for a model where OpenCascade fails to generate a mesh,' while the text says 'approximately 5% of the models fail to produce a mesh.' It is unclear whether the failure is per model or per face, and whether the point cloud in Figure 2 is generated by an alternative fallback or by sampling the parametric surfaces. Please clarify.","section":"§3.1 and Figure 2"},{"comment":"In the description of b-spline curves and surfaces, the text says 'We also track if the curve is periodic or if it closed' and later 'if it is u_closed/v_closed.' The first should read 'if it is closed' for grammatical consistency.","section":"Appendix A.1"},{"comment":"The dataset section says 'Our dataset includes one million models from ABC' and the Fusion subsets, but the conclusion says 'several million models' have been converted. The total count should be made consistent, and the OnShape contribution, if any, should be stated explicitly.","section":"§4, Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern in the reader's report is well-founded: the absence of any independent geometric comparison between the HDF5 output and the original STEP files is the key weakness. The use-case experiments are circular for validating conversion fidelity, and the paper's own admission that different STEP files require different kernels compounds the risk. The format specification and released code are solid enough that the paper could become acceptable after a substantive validation addition; I do not see grounds for outright rejection. I would also flag that the paper's scope is closer to a dataset/benchmark contribution than a traditional CV method paper, so the editor may want to consider fit with the journal's aims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Best quick take: this is a useful infrastructure paper, not a research breakthrough. The HDF5 half-edge format is fully specified (Appendix A), the STEPTOHDF5 converter and ABS library are public, and the converted ABC and Fusion 360 datasets are a real resource. If you work on CAD or geometric deep learning, this removes a genuine bottleneck: you can now sample points, normals, and curvatures directly from parametric surfaces without a proprietary kernel.\n\nWhat it does well: the format is simple and well-documented, the library's callback design for sampling is clean, and the four use cases show the pipeline actually runs end-to-end with off-the-shelf models. The authors are honest that OpenCascade's mesher fails on 1–9% of models depending on dataset, and they note the kernel dependence in the conclusion.\n\nThe soft spot is real but not fatal: the 'compliance with the original STEP files' claim is not actually tested. The use cases compare model accuracy against published numbers on meshes, which would not catch silent geometry or topology corruption introduced by OpenCascade. If OCC misparameterizes a trim curve or drops a seam edge, the HDF5 stores that error faithfully and the downstream learning task still looks plausible. The paper's own Section 6 admits different kernels handle different STEP files, so coverage is incomplete. Tables 1 and 2 also have no error bars, which would help given the small model selection. These are addressable: run a direct comparison of sampled points, normals, and topology counts against a second kernel or against OCC's own direct output from the original STEP files, and report per-model deviations.\n\nFor a reader in CAD/ML, this is worth reading and citing. It deserves a serious referee, but the reviewer should ask for that compliance test before acceptance, or the equivalence language needs to be dialed back to 'a usable B-rep representation.'\n\nRecommendation: send to peer review, require a direct geometric fidelity check or a revised claim.","headline":"Useful infrastructure paper—a well-specified HDF5 B-rep format and public converted datasets—but the equivalence claim is under-validated because nothing checks OpenCascade's parsing fidelity.","tokens_in":9477,"tokens_out":2315,"would_cite":true,"duration_ms":23475,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that any STEP B-rep can be preserved exactly as an open HDF5 dictionary that any program can read.","keywords":["boundary representation","STEP files","HDF5 format","CAD datasets","parametric surfaces","half-edge topology","point cloud sampling","machine learning on CAD"],"falsifier":"Convert a set of STEP files with a second CAD kernel that implements its own STEP reader, then compare per-face surface types, parameter domains, and dense sampled point positions to those stored in the HDF5 files; any systematic discrepancy beyond numerical tolerance in surface type or trim curves would falsify the claimed equivalence.","tokens_in":8515,"feed_emoji":"📐","tokens_out":4892,"duration_ms":42141,"temperature":0.7,"pith_summary":"The paper is trying to establish that boundary-representation CAD files in STEP format can be repackaged, without loss, into an openly specified HDF5 dictionary that any program can read. Such a format would remove the need for proprietary CAD kernels in large machine-learning pipelines, where per-node licenses are costly. The authors converted the Fusion 360 and ABC datasets, plus a million OnShape models, and built a Python library that samples points, normals, and curvature directly from the parametric surfaces. They verify the format's integrity by running four standard learning tasks (normal estimation, denoising, surface reconstruction, and segmentation) with existing models and getting accuracies close to published values.","feed_headline":"STEP CAD files repackaged as open HDF5 any program can read","feed_subtitle":"A new dataset and library let ML pipelines sample parametric surfaces without a CAD kernel license.","key_machinery":"The central object is the HDF5 dictionary whose topology section is a standard half-edge data structure: each solid stores shells, each shell stores faces with orientation flags, each face stores loops and a surface index, each loop stores half-edges, each half-edge stores mates, an edge, and a 2D curve index, and each edge stores a 3D curve index and start and end vertices. This hierarchy is what lets the format preserve adjacency and trimming information exactly while remaining readable by any HDF5 client. The conversion tool uses OpenCascade to parse STEP files into this dictionary, and the processing library navigates it like a half-edge structure and samples points, normals, and derivatives from the continuous parametric surfaces.","core_discovery":"The paper claims that a STEP file's B-rep is fully determined by two layers: a geometry layer holding parametric curves and surfaces (with control points, knots, weights, transforms, and trim domains) and a topology layer holding a half-edge hierarchy of solids, shells, faces, loops, half-edges, edges, and vertices, with orientation flags and indices that tie every topological entity to its geometric counterpart. Storing these layers as an HDF5 dictionary, with an optional per-face mesh, gives an equivalent format to the original STEP file. The authors state this format is fully specified, cross-platform, and cross-language, and that their conversion pipeline and processing library make the parametric information directly accessible, something earlier mesh-based pipelines had to reconstruct heuristically.","pith_inferences":["Editorial inference: if the equivalence is exact, the format could serve as a neutral interchange layer between CAD kernels, so a model stored once could be read by any kernel that implements the dictionary.","Editorial inference: a natural stress test would be to convert a random sample of STEP files with a second independent kernel and compare face-level surface types, trim curves, and sampled geometry; the paper's validation does not yet include such a comparison.","Editorial inference: because the paper found OpenCascade's mesher fails on roughly 1 to 9 percent of models depending on dataset, downstream users should check which faces have empty meshes; learning on point clouds is unaffected, but mesh-based methods may need fallbacks."],"forward_implications":["Machine-learning pipelines can sample a CAD shape's smooth parametric surfaces directly, obtaining exact point positions, normals, and curvature without meshing the whole model.","Datasets converted once (Fusion 360, ABC, OnShape) can be distributed as HDF5 files, so a cluster node needs no CAD kernel license to read or process them.","Tasks that need parametric information, such as primitive or degree segmentation, can get it directly from the topology and geometry instead of recovering it from meshes with heuristics.","Because the format is fully specified and HDF5-based, other languages and platforms can interoperate with the same files, reducing the version-incompatibility problem of STEP across kernels."],"supporting_citations":[{"why":"OpenCascade is the parser that extracts geometric and topological information from STEP files, so the entire format's fidelity rests on it.","marker":"[2]"},{"why":"ABC supplies one million STEP models that the paper converts into the HDF5 dataset.","marker":"[9]"},{"why":"Fusion 360 Gallery supplies the Assembly and Reconstruction subsets used in the use cases.","marker":"[13]"},{"why":"Fusion 360 Joinable supplies the Joint and Segmentation subsets used in the use cases.","marker":"[14]"},{"why":"BPNet is the segmentation model whose reported accuracy the paper compares against when testing the dataset.","marker":"[6]"},{"why":"NKSR is the surface-reconstruction model used to evaluate whether sampled data preserves geometry.","marker":"[7]"},{"why":"PathNet is the denoising model used to check that point clouds sampled from the new format behave like mesh-sampled data.","marker":"[12]"},{"why":"DeepFit is the normal-estimation model used to test the accuracy of normals sampled from the format.","marker":"[1]"}],"fun_headline_variants":["CAD B-rep data freed from STEP kernel licensing","HDF5 replaces STEP for CAD boundary representation","No CAD kernel needed: open HDF5 B-rep dataset","Parametric CAD geometry now accessible without STEP kernels","Open HDF5 format lifts CAD B-rep data out of STEP"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that OpenCascade parses every STEP file correctly and completely, so the HDF5 dictionary inherits the true geometry and topology; if OpenCascade silently alters a surface or drops a trim curve, the format stores that error faithfully and the paper's validation would not detect it.","fun_headline_variants_meta":{"raw":{"variants":["CAD B-rep data freed from STEP kernel licensing","HDF5 replaces STEP for CAD boundary representation","No CAD kernel needed: open HDF5 B-rep dataset","Parametric CAD geometry now accessible without STEP kernels","Open HDF5 format lifts CAD B-rep data out of STEP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000308,"raw_usage":{"total_tokens":1720,"prompt_tokens":864,"completion_tokens":856,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":480,"completion_tokens_details":{"reasoning_tokens":775}},"tokens_in":480,"tokens_out":856,"duration_ms":22991,"temperature":1.0,"reasoning_tokens":775,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:40:32.208123+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Convert a set of STEP files with a second CAD kernel that implements its own STEP reader, then compare per-face surface types, parameter domains, and dense sampled point positions to those stored in the HDF5 files; any systematic discrepancy beyond numerical tolerance in surface type or trim curves would falsify the claimed equivalence.","supporting_citations":[{"cited_title":"Open cascade.https://dev.opencascade.org/","cited_arxiv_id":null,"evidence_quote":"OpenCascade is the parser that extracts geometric and topological information from STEP files, so the entire format's fidelity rests on it."},{"cited_title":"Abc: A big cad model dataset for geometric deep learning","cited_arxiv_id":null,"evidence_quote":"ABC supplies one million STEP models that the paper converts into the HDF5 dataset."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Fusion 360 Gallery supplies the Assembly and Reconstruction subsets used in the use cases."},{"cited_title":"BPNet: B\\'ezier Primitive Segmentation on 3D Point Clouds","cited_arxiv_id":"2307.04013","evidence_quote":"BPNet is the segmentation model whose reported accuracy the paper compares against when testing the dataset."},{"cited_title":"Neural kernel surface reconstruction","cited_arxiv_id":null,"evidence_quote":"NKSR is the surface-reconstruction model used to evaluate whether sampled data preserves geometry."},{"cited_title":"Pathnet: Path-selective point cloud denoising.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(6):4426–4442, 2024","cited_arxiv_id":null,"evidence_quote":"PathNet is the denoising model used to check that point clouds sampled from the new format behave like mesh-sampled data."},{"cited_title":"Deepfit: 3d surface fitting via neural network weighted least squares","cited_arxiv_id":null,"evidence_quote":"DeepFit is the normal-estimation model used to test the accuracy of normals sampled from the format."}],"review_version":1}