PaperlayerView exact revision on arXiv

Source-linked AI summary

Automatic assembly modeling through attention-based assembly feature mate prediction and assembly feature pattern recognition

Ziheng Dai, Wanbin Pan, Zhihong Tang, Shuming Gao

arXiv:2610.14633v1cs.CVcs.CEcs.GR

TL;DR

Assembly modeling remains labor-intensive, while existing data-driven approaches either fail to produce editable B-Rep constraints or rely on prescribed part pairs or faces. This paper predicts assembly feature mates from unassembled B-Rep parts and constructs an annotated dataset of 529 assemblies, 43,646 parts, and 106,493 assembly features.

  • Problem

    Assembly modeling requires repeated interactive constraint selection, while data-driven methods either do not yield editable B-Rep constraints or rely on prescribed pairs or faces.

    Source: §1 · p. 1

  • Method

    The framework predicts assembly feature mates from unassembled B-Rep parts, using an attention-based network and assembly feature patterns, and builds an annotated dataset.

    Source: §1 · p. 3, §3.1.1 · p. 5

  • Results

    The automatic annotation pipeline produces a dataset of 529 assemblies, 43,646 parts, and 106,493 assembly features.

    Source: §1 · p. 3

  • Takeaways & Limitations

    Assembly feature mates associate mating features with complementary constraints, providing explicit feature semantics for prediction modules.

    Source: §3.1.1 · p. 5

    The study focuses on hole-shaft and hole-hole assembly feature mates.

    Source: §4.1 · pp. 9–10

Abstract

from arXiv · show

Assembly modeling plays an important role in constructing product models, yet it is still mainly performed via interactive operations that are experience-dependent and labor-intensive. Data-driven strategies hold promise for automating this process; however, existing approaches either predict poses that cannot directly generate editable B-Rep assembly constraints or require prescribed part pairs or predefined assembly faces. To reduce the large candidate space of constraint prediction without predefined assembly faces, we propose a data-driven framework based on assembly feature mate prediction. The framework performs automatic assembly modeling from unassembled B-Rep parts, using assembly features to group the geometric entities of local mating interfaces and restrict candidate formation. Given the lack of suitable assembly feature annotations in existing datasets, we construct a large-scale dataset through an automatic annotation pipeline. Using the annotated dataset, we train an attention-based network to predict assembly feature mates, with Engineering Compatibility Fusion (ECF) incorporating explicit engineering compatibility into learned matching scores. To improve the accuracy of automatic assembly modeling, we use assembly feature patterns (AFPs) to guide the matching process. Finally, a post-processing search selects assembly feature mate and AFP mate candidates, whose associated constraints are generated and solved to complete assembly modeling. Experimental results show that the framework achieves assembly constraint accuracies of 74.79% on the six-assembly comparison set and 64.09% on the extended test set of 16 assemblies, which includes the comparison set. With ground-truth face labels supplied as additional input, the corresponding accuracies increase to 81.78% and 80.93%, respectively.

1. Introduction

Assembly modeling is labor-intensive, and existing data-driven approaches either fail to produce editable B-Rep constraints or rely on prescribed part pairs or faces. This paper targets assembly feature mates, uses AFPs to guide matching, and constructs a dataset with assembly feature annotations.

1. Introduction

  • Complementary constraint groups define feasible assembly relations, whereas an isolated plane-coincidence constraint can produce geometric interference.The figure contrasts a coaxial and plane-coincidence group with a lone plane-coincidence constraint.Source: Figure 1 · p. 2
  • Prior data-driven methods either predict poses rather than editable B-Rep constraints or depend on specified part pairs or predefined assembly faces.These restrictions reduce the candidate space, but leave assembly prediction dependent on prescribed inputs.Source: §1 · p. 2
  • The framework predicts assembly feature mates to address prediction feasibility, and uses assembly feature patterns to improve automatic assembly modeling accuracy.Existing public datasets lack the annotations required for this formulation, motivating an automatic annotation pipeline.Source: §1 · p. 3
  • The proposed framework models assemblies directly from unassembled B-Rep parts without predefined assembly faces.An attention-based network predicts assembly feature mates, with AFPs incorporated to improve accuracy.Source: §1 · p. 3
  • The automatic annotation pipeline produces a dataset containing 529 assemblies, 43,646 parts, and 106,493 assembly features.The dataset supports training with assembly feature annotations.Source: §1 · p. 3

2. Related work

Prior work represents assembly semantics at different levels, recovers relations from positioned assemblies, and learns from mesh, point-cloud, or B-Rep geometry. The paper identifies a gap in predicting assembly relations from unassembled B-Rep parts without predefined assembly faces.

2.1. Representation of Assembly Semantics in CAD

  • Assembly features and assembly feature mates make higher-level semantics explicit by organizing low-level constraints into part-side and cross-part representations.The paper uses assembly features for potential connection regions and assembly feature mates for mated interfaces, kinematic relations, and constraints.Source: §2.1 · p. 3

2.2. Recovery of Assembly Semantics from Existing Assembly Configurations

  • Methods that recover semantics from existing assemblies rely on prior part poses and cannot provide interface information for parts without relevant assembly history.Assembly semantics may be lost or withheld even when geometry and assembled configuration remain available.Source: §2.2 · p. 3

2.3. Learning-Based Assembly from Meshes and Point Clouds

  • Mesh- and point-cloud methods support configuration and kinematic reasoning, but predict poses or articulation models rather than editable B-Rep constraints.Their outputs do not directly yield constraint-based CAD assembly models.Source: §2.3 · p. 4

2.4. Learning-Based Prediction of Assembly Semantics from B-Rep Part Models

  • B-Rep methods infer part-side interfaces, predict relations for prescribed part pairs, or use faces from prior assemblies; none identified here predicts from input B-Rep parts without predefined faces.The paper frames inference over unassembled, unordered parts as requiring geometry and topology alone.Source: §2.4 · p. 4

3. Basic concepts and approach overview

The paper represents mating interfaces and their relations as assembly features, feature mates, and higher-level feature patterns. Its framework combines semantic dataset construction, attention-based feature-mate prediction, AFP correspondence generation, and constraint solving for assembly modeling from B-Rep parts without predefined assembly faces.

3.1. Basic concepts

  • The basic concepts organize assembly semantics as mating interfaces on individual parts and assembly relations between parts.Assembly features represent interfaces, while their mates represent relations between features from different parts.Source: §3.1 · p. 4, §3.1.1 · p. 5

3.1.1. Assembly feature and assembly feature mate

  • An assembly feature mate links two features from different parts with complementary constraints that jointly realize their kinematic relation.The framework predicts these feature mates rather than treating their constituent constraints as independent decisions.Source: §3.1.1 · p. 5

3.1.2. Assembly feature pattern and assembly feature pattern mate

  • An AFP is a group of repetitive features sharing high-level assembly semantics, represented by its constituent features, in-plane pose, and pairwise-distance layout.The pairwise-distance matrix represents intrinsic shape independently of absolute in-plane position and orientation.Source: §3.1.2 · pp. 5–6
  • An AFP mate relates two patterns through layout-consistent feature correspondences, allowing group-level matching to reject locally plausible but layout-inconsistent pairs.The correspondence maps every feature of the smaller AFP, while unmatched features may remain in the larger pattern.Source: §3.1.2 · p. 6

3.2. Overall approach

  • The framework models B-Rep parts without predefined assembly faces, combining semantic dataset construction, assembly feature mate prediction, AFP mate generation, and constraint solving.Its assembly stage ranks assembly feature mate and AFP mate candidates, generates associated constraints, and solves them to obtain an assembly configuration.Source: §3.2 · pp. 6–7, Figure 4 · p. 8
  • The approach constructs semantically annotated training data by extracting geometric constraints from CAD assemblies and transforming them into assembly feature and pattern annotations.Part-mate pretraining supplies the encoder used in assembly feature mate prediction; Engineering Compatibility Fusion combines learned scores with explicit engineering compatibility.Source: §3.2 · p. 7

4. Dataset construction

The annotation pipeline addresses omitted constraints and geometrically valid but semantically misleading entity selections, converting assembly constraints into standardized feature and pattern annotations. The study focuses on hole–shaft and hole–hole mates, including hole correspondences where fasteners are omitted.

4. Dataset construction

  • Constraint annotations can omit physical relations or reference the wrong entity: simplified constraints may leave holes unannotated, and a valid coaxial selection may not represent physical contact.Figure 5 illustrates omitted fasteners and an external cylindrical surface selected instead of the mechanically correct inner bore.Source: §4 · pp. 8–9, Figure 5 · p. 9

4.1. Assembly feature and assembly feature mate types

  • The annotation targets include hole–shaft mates and hole–hole correspondences, the latter representing coaxial holes when standard fasteners are omitted from CAD models.The pipeline’s pattern-level annotations are retained for future AFP mate prediction, though they are not used to train the neural network in this study.Source: §4.1 · pp. 9–10, §4.2 · p. 10

4.2. Automatic dataset annotation pipeline

  • The pipeline reconstructs incomplete or misaligned constraints, translates constraints into face-based features, and extracts assembly feature patterns and their mates.Constraint reconstruction checks entity coincidence and partners, using adjacency-based search to retrieve correct entities; feature translation checks physical contact, and pattern extraction groups kinematically complete mates in a shared spatial context.Source: §4.2 · p. 10

5. Part encoder pretraining

A part encoder is pretrained on part-mate prediction to provide geometric and topological part embeddings for later assembly feature mate prediction. The part-mate relation is predicted symmetrically and trained with binary cross-entropy.

5.2. Part mate prediction

  • Part-mate prediction refines part embeddings through attention-based relation layers, then decodes pairwise scores into symmetric probabilities that parts are mated.The predicted relation satisfies c_ij = c_ji; the encoder is pretrained on this assembly-related task and frozen during feature-level training.Source: §5 · p. 10, §5.2 · pp. 11–12, §5.3 · p. 12

5.1. Part embedding extractor

  • The UV-Net-based encoder produces a 128-dimensional embedding for each B-Rep part, capturing geometric and topological information for subsequent prediction modules.The encoder combines curve, surface, and graph modules.Source: §5.1 · p. 10

5.3. Loss function

  • The part-mate predictor is trained with binary cross-entropy between its symmetrized probabilities and the ground-truth part-mate matrix.The decoder’s pairwise probabilities are the outputs compared with the ground-truth matrix.Source: §5.3 · p. 12

6. Assembly feature mate prediction

The module predicts assembly feature mates by combining local feature geometry with parent-part context, then applying attention-based scoring and explicit engineering compatibility. Its training objective compares compatibility-adjusted candidate scores with a one-hot target.

6.1. Assembly feature vector extraction

  • Each hole or shaft is represented by a 21-dimensional vector containing its type, geometric parameters, and position and orientation in the part coordinate system.The vector includes part and type identifiers, cylinder and plane parameters, and zero-padding; the part identifier, feature type, and primary mating dimension also inform ECF.Source: §6.1 · p. 13

6.2. Context-guided semantic reweighting mechanism

  • Context-guided reweighting uses parent-part embeddings to guide weighted aggregation of local candidate feature representations before mate prediction.Each local feature vector supplies a query and value, while its parent-part embedding supplies a key.Source: §6.2 · pp. 13–14

6.3. Attention-based feature mate predictor

  • Multi-head attention scores a query against competing candidates, using sets of 20 during assembly modeling because encoding source parts imposes substantial memory requirements.Training pairs each query with one positive and 19 negative candidates; larger inference sets require more memory or simpler part geometries.Source: §6.3 · p. 14

6.4. Engineering Compatibility Fusion

  • Engineering Compatibility Fusion multiplies learned matching scores by explicit compatibility based on cross-part eligibility and type-conditioned dimensional compatibility.Same-part candidates receive zero compatibility; hole–shaft candidates are weighted by radius agreement, while hole–hole candidates bypass dimensional reweighting.Source: §6.4 · pp. 14–15
  • ECF’s compatibility-adjusted scores are used for candidate ranking and pattern-level score aggregation without posterior normalization.The paper describes the multiplicative score as a qualitative posterior reweighting, with learned scores proportional to prior probabilities and engineering compatibility proportional to evidence likelihood.Source: §6.4 · p. 15

6.5. Loss function

  • Training minimizes mean squared error between ECF-adjusted candidate scores and a one-hot vector marking the correct feature mate.Although ECF has no learnable parameters, its compatibility weights affect gradients propagated to the predictor through the fused scores.Source: §6.5 · p. 16

7. AFP recognition and mate candidate generation

AFP recognition and candidate generation evaluate feature mates as groups, retaining possible pattern correspondences rather than relying only on independent local matches. The method recognizes canonical layouts in coplanar feature groups and uses metric-consistency search for groups without those layouts.

7. AFP recognition and mate candidate generation

  • Groupwise AFP matching addresses cases where individually plausible feature mates do not form a valid pattern mate, while accommodating multiple mappings and non-canonical patterns.Rectangular and circular hypotheses generate admissible feature mappings; metric-consistency search handles groups that fit neither layout.Source: §7 · p. 16, §7.2 · p. 17

7.1. Canonical AFP recognition from coplanar feature groups

  • Canonical recognition searches coplanar feature groups for rectangular grids and circular layouts, retaining hypotheses that satisfy geometric residual and layout-structure checks.Rectangular hypotheses require at least two rows and columns and can represent partially populated layouts; circular recognition checks equal-angle structure as well as radial residuals.Source: §7.1 · p. 16

7.2. Parametric matching of canonical AFPs

  • Parametric matching enumerates grid symmetries or circular phase and direction mappings, then retains all type- and metric-consistent correspondences and removes duplicates.If no valid groupwise correspondence is found, individual assembly feature mate candidates remain available for downstream selection.Source: §7.2 · p. 17

8. Assembly constraint generator

The constraint generator ranks assembly-feature mates and AFP mates on a shared score, then selects candidates for constraint generation and solving.

8. Assembly constraint generator

  • The generator selects assembly-feature mate and AFP-mate candidates, instantiates their associated constraints, and passes them to the geometric constraint solver.Its search resembles the Best-First strategy of Vergez et al. and evaluates results by cumulative assembly-feature-mate scores.Source: §8 · p. 17

8.1. Unified scoring

  • An individual assembly-feature mate receives the average of the two directional ECF scores, with each feature serving once as query and once as candidate.This makes the score for a feature pair symmetric across query direction.Source: §8.1 · p. 17
  • An AFP mate is scored by summing the averaged scores of its constituent assembly-feature mates.The combined score lets AFP mates and individual feature mates be compared within the same search.Source: §8.1 · p. 17

S FMi. (38)

The search maximizes cumulative compatibility-adjusted mate scores while requiring each selected candidate to be assemblable and consistent with constraints already applied.

S FMi. (38)

  • An AFP mate’s score accumulates the matching evidence of its layout-consistent feature correspondences.Its constituent feature-mate scores are counted once in the cumulative objective.Source: S FMi. (38) · p. 17, §8.2 · p. 18

8.2. Search procedure

  • Candidates are ranked by score, and the search maximizes the cumulative compatibility-adjusted scores of applied feature mates, counting each mate once.The total includes constituent feature mates of AFP mates.Source: §8.2 · p. 18
  • A candidate is accepted only if its feature types are compatible, its parts do not interfere, and it conflicts with no established constraints.Accepted cylindrical-face pairs generate coaxial constraints, while planar-face pairs generate plane-coincidence constraints for the solver.Source: §8.2 · p. 18
  • The search builds a greedy solution, then explores selections within Hamming distance three and returns the feasible explored assembly with the highest cumulative score.An AFP mate counts as one candidate regardless of its number of constituent feature mates.Source: §8.2 · p. 18

9. Results

The framework is evaluated on a 16-assembly test set and a six-assembly comparison set, with results reported for automatic operation and ground-truth face-label input.

9.1. Dataset

  • The annotation pipeline yields 529 assemblies, 43,646 parts, 106,493 canonical features, and 42,618 AFPs; holes and shafts account for approximately 99.89% of extracted features.The 16 assemblies used for assembly-level evaluation and feature-level ablation were excluded from network training.Source: §9.1 · p. 18

9.2. Evaluation metrics

  • Evaluation measures constraint recovery (ACCc), recovered constraint cycles (ACCℓ), non-intersecting part pairs (ACCv), and top-1 feature-mate prediction accuracy.The final assembly metrics compare the generated constrained assembly with the target assembly.Source: §9.2 · pp. 18–19

9.3. Experimental results

  • With face labels, ACCc = 81.78%, ACCℓ = 65.38%, and ACCv = 100%, versus Vergez et al.’s 72.94%, 32.86%, and 81.81%; without labels, ACCc = 74.79%.The comparison set comprises six corresponding original assemblies; the reported baseline results are from Vergez et al..Source: §9.3 · p. 19, Table 2 · p. 20
  • On the 16-assembly extended test set, ACCc = 64.09%, ACCℓ = 32.14%, and ACCv = 82.27%; with face labels, the values are 80.93%, 65.48%, and 100%.The ten additional assemblies were evaluated without fine-tuning, and all 16 assemblies were excluded from network training.Source: §9.3 · p. 20, Table 3 · p. 20
  • Removing AFP recognition and mate generation lowers ACCc from 74.79% to 38.93% on the comparison set and from 64.09% to 31.77% on the extended set.ACCℓ also falls from 34.61% to 7.69% and from 32.14% to 9.68%, respectively; the authors attribute the benefit to layout-constrained correspondences.Source: §9.3 · p. 20
  • With ECF, feature-mate ACCf is 76.27%, compared with 55.79% when ECF is removed, a difference of 20.48 percentage points.The model without ECF was retrained and evaluated with ECF removed on the 16-assembly extended test set.Source: §9.3 · p. 20, Table 4 · p. 22
  • Supplying ground-truth face labels raises extended-set ACCc from 64.09% to 80.93%; filtering by labels excludes plausible but nonparticipating features that create match ambiguity.The feature-extraction process remains unchanged; extracted features are filtered according to the labels of their constituent faces.Source: §9.3 · pp. 20, 22
  • Without face labels, no assembly is completely correct, while five (31.25%) need corrections involving at most two parts; with labels, nine (56.25%) are completely correct and 68.75% need corrections involving at most two parts.The at-most-two-parts category includes completely correct assemblies.Source: §9.3 · p. 22, Table 5 · p. 23

10. Discussion and limitations

The discussion quantifies how local interface associations reduce candidate formation, then describes unresolved matching ambiguity, representation limits, and sensitivity to initial poses during constraint solving.

10.1. E

  • 76.55% is the mean per-part reduction across 139 parts in 16 assemblies; paired local restrictions reduce cross-part candidates, but counts do not measure runtime or feasible complete states.For each part pair, the retained fraction is the product of the two parts’ retained fractions; the per-part mean does not determine aggregate reduction across pairs.Source: §10.1 · p. 23

10.2. Multi-solution ambiguity

  • Locally compatible features can yield competing matches, and symmetric parent parts or layouts can still prevent the framework from consistently identifying the target relations.Examples include inserting a handle through a hole from either side and connecting similar links in multiple ways.Source: §10.2 · p. 24, Figure 11 · p. 24

10.3. Limitations of the assembly feature representation

  • The representation omits interfaces formed across several parts and individual constraints outside grouped complementary interfaces; pre-assembly extraction also misses features that emerge as parts are assembled.Some isolated coaxial relations are not directly covered by the grouped-constraint prediction target.Source: §10.3 · pp. 24–25, Figure 12 · p. 25

10.4. Influence of initial part poses on constraint solving

  • For moving parts, SolidWorks’ default initial poses can lead the solver to an interfering configuration, causing the interference check to reject a correct relation even when a noninterfering configuration exists.Thus, the configuration reached from a given initialization can limit recovery of the target assembly.Source: §10.4 · p. 25

11. Conclusion

The work presents an assembly-feature-mate framework and a 529-assembly dataset for automatic modeling from unassembled B-Rep parts. The workflow reports 74.79% and 64.09% assembly constraint accuracy on the comparison and extended test sets, respectively, while ambiguity and representation coverage remain challenges.

11. Conclusion

  • The framework predicts assembly feature mates from unassembled B-Rep parts without predefined assembly faces, and AFPs further improve automatic assembly modeling accuracy.A dataset of 529 assemblies was constructed through automatic annotation to provide assembly feature annotations for network training.Source: §11 · p. 25
  • 74.79% on the assembly comparison set and 64.09% on the extended test set of 16 assemblies are the complete workflow’s assembly constraint accuracies.Ablations support the role of AFP layout information in modeling accuracy and ECF in assembly feature mate prediction accuracy.Source: §11 · pp. 25–26
  • Multi-solution ambiguity and limited representation coverage remain challenges for automatic assembly modeling with local interface associations and AFP layout information..Source: §11 · p. 26
Loading 2610.14633v1…