BenDFM: A Taxonomy and Synthetic CAD Dataset for Manufacturability Assessment in Sheet Metal Bending¶

Conference: CVPR 2026 arXiv: 2603.13102 Code: github.com/UGent-CVAMO/bendfm Area: CAD / Intelligent Manufacturing / Geometric Deep Learning Keywords: Manufacturability Assessment, Sheet Metal Bending, Synthetic CAD Dataset, DFM Taxonomy, B-rep Learning

TL;DR¶

This paper proposes a two-dimensional taxonomy of manufacturability metrics (configuration dependence × feasibility/complexity) and introduces BenDFM, the first synthetic CAD dataset for sheet metal bending (20,000 parts, covering both manufacturable and non-manufacturable designs). Benchmark results show that topology-aware graph representations (UV-Net, AUC 0.896) consistently outperform point cloud methods (PointNext, AUC 0.844) across all four task categories.

Background & Motivation¶

Background: Design for Manufacturability (DFM) requires predicting production feasibility and difficulty during the design phase. While deep learning has made notable progress in manufacturing process selection, advances in Intra-Process Manufacturability Assessment (IPMA) have been limited.

Limitations of Prior Work: (1) The definition of "manufacturability" is inconsistent in the literature—some metrics depend on specific equipment configurations (e.g., available punch sizes), while others represent universal geometric constraints (e.g., unfolding self-intersection); some involve discrete judgments while others are continuous measures. This semantic confusion hinders method comparison and knowledge transfer. (2) Industrial datasets suffer from severe survivorship bias—only optimized, manufacturable designs are retained, while failed non-manufacturable cases are discarded, preventing models from learning the geometric features associated with production failures. (3) Existing synthetic datasets primarily serve subtractive processes (drilling, milling), rely on simple cuboid geometries, and have been shown to generalize poorly.

Key Challenge: Sheet metal bending requires simulating sequential bending operations and complex part–tool interactions (collision detection), making it far more intricate than simple aperture-ratio judgments. Yet no dedicated dataset or unified definitional framework exists.

Goal: To systematically define the conceptual space of manufacturability and to create the first synthetic CAD dataset for sheet metal bending that includes both manufacturable and non-manufacturable designs.

Key Insight: Establish a taxonomy to unify terminology, then generate richly annotated data via parametric modeling combined with physical bending simulation.

Core Idea: A two-dimensional taxonomy (configuration dependence × feasibility/complexity) combined with a process-aware synthesis pipeline (bending trajectories + collision detection) yields the first sheet metal bending dataset capable of supporting learning-based DFM.

Method¶

Overall Architecture¶

Three primary contributions: (1) a manufacturability metric taxonomy that systematically defines the semantic space; (2) the BenDFM dataset generated via parametric modeling and bending simulation; and (3) benchmark evaluations on two state-of-the-art 3D learning architectures to validate the insights derived from the taxonomy.

Key Designs¶

Two-Dimensional Manufacturability Taxonomy:
- Function: Systematically defines the semantic space of manufacturability metrics.
- Mechanism: Metrics are partitioned into four quadrants along two axes—the horizontal axis distinguishes configuration-independent (geometrically intrinsic, e.g., unfolding self-intersection) from configuration-dependent (tool/equipment-specific, e.g., punch collision) metrics; the vertical axis distinguishes feasibility (discrete judgment: "can it be manufactured?") from complexity (continuous measure: "how difficult is it to manufacture?").
- Quadrant examples: Geometric feasibility = unfolding overlap (infeasible regardless of equipment); Configuration feasibility = punch collision (potentially resolved by changing tools); Geometric complexity = unfolded area (pure geometric property); Configuration complexity = number of flips (dependent on bending sequence).
- Design Motivation: Explains the root cause of definitional inconsistencies across the existing literature; provides a unified language for DFM research applicable to any manufacturing process; and clarifies the boundaries of model generalizability.
BenDFM Dataset Generation Pipeline:
- Function: Generates a large-scale synthetic sheet metal bending dataset containing both manufacturable and non-manufacturable parts.
- Mechanism:
  - Parametric bend generation: Built on PythonOCC, a five-step procedure (edge selection → bend face construction → bend extrusion → flange extrusion → repetition) generates 20,000 parts. Weighted sampling favors edges closer to the base and longer edges; bend reliefs, beveled/rounded flanges, and symmetric offsets are supported.
  - Tool geometry parameterization: Punch and die geometries are modeled explicitly, with tool profiles defined via trigonometric functions.
  - Dynamic bending simulation: Intermediate states are generated at 5° increments along the bending trajectory; bend allowance is computed; collisions are detected at each intermediate angle.
  - Collision detection augmentation: (a) Punches/dies are instantiated at three positions (left/center/right) to reflect real-world alignment flexibility; (b) post-hoc detection re-checks collisions after all bends are completed to capture interference of subsequent flanges with previously formed bends.
  - Unfolding and overlap detection: Self-intersections in the flat-pattern unfolding are detected via B-rep Boolean operations.
- Dataset scale: 20,000 parts in STEP format, each accompanied by an unfolded model, full bending sequence parameters, and four-quadrant manufacturability labels. The BenDFM subset contains 14,000 parts (collision task, 50/50 balanced); the BenDFM-U subset contains 6,000 parts (unfolding overlap task, 50/50 balanced).
- Design Motivation: Forty percent of bends are forced to be collision-free to prevent the dataset from being dominated by obviously non-manufacturable designs.
Dual-Architecture Benchmark:
- Function: Evaluates two state-of-the-art 3D learning architectures across four categories of manufacturability metrics.
- UV-Net (GNN on B-rep AAG): Operates on the attributed adjacency graph; each face is sampled with a 10×10 UV grid and each edge with a 10-point sequence; 2D/1D CNNs encode these representations before feeding into a GCN.
- PointNext (PointNet family): Operates on 1,024 sampled surface points using coordinate and normal features.
- Design Motivation: Validates the importance of topological information for manufacturability prediction while identifying the particular challenges posed by configuration-dependent metrics.

Loss & Training¶

Classification tasks: BCE loss; regression tasks: MSE loss. Adam optimizer, lr = 0.0005, batch size = 32, early stopping with patience = 20. An 80/10/10 split is used; results are averaged over 5 runs with different random seeds, with standard deviations reported.

Key Experimental Results¶

Main Results: Feasibility Classification¶

Model	Task	AUC	Acc (%)	F1 (%)	Type
UV-Net	Punch collision (config. feasibility)	0.840	76.07	75.41	Classification
PointNext	Punch collision	0.827	73.83	71.74	Classification
UV-Net	Unfolding overlap (geometric feasibility)	0.896	81.80	81.31	Classification
PointNext	Unfolding overlap	0.844	76.13	76.55	Classification

Main Results: Complexity Regression¶

Model	Task	MAE	RMSE	MAPE (%)	Type
UV-Net	No. of flips (config. complexity)	0.54	0.86	35.52	Regression
PointNext	No. of flips	0.59	0.90	39.33	Regression
UV-Net	Unfolded area (geometric complexity)	14.60	20.68	5.90	Regression
PointNext	Unfolded area	20.24	28.82	8.28	Regression
Baseline (training set mean)	No. of flips	0.984	1.180	67.67	—
Baseline (training set mean)	Unfolded area	89.81	111.35	46.01	—

Ablation Study¶

Analysis Dimension	Finding	Notes
UV-Net vs. PointNext	UV-Net outperforms point cloud methods on all four tasks	B-rep topology is critical for manufacturability prediction
Geometric vs. configuration metrics	AUC 0.896 vs. 0.840	Geometric metrics are systematically easier to learn
40% collision-free bends	Effectively prevents trivially non-manufacturable samples from dominating	Creates more nuanced decision boundaries
Folded vs. unfolded geometry as input	Unfolded geometry slightly superior	Directly exposes collision constraints

Key Findings¶

UV-Net consistently outperforms PointNext across all four tasks, indicating that preserving B-rep topological relationships is critical for manufacturability prediction.
Geometric (configuration-independent) metrics are systematically easier to learn than configuration-dependent metrics—configuration dependence introduces additional tool-specific complexity.
Configuration complexity (number of flips) still yields a MAPE of 35.52%, underscoring the core challenge of inferring operation-sequence information from final geometry alone.
Both models substantially surpass the baseline (mean prediction), yet a considerable gap remains before practical deployment.

Highlights & Insights¶

The taxonomy is incisive: it substantively explains why different papers define "manufacturability" so differently, providing the community with a unified language and a research roadmap.
The dataset design is thorough: dynamic bending trajectory simulation, three-position alignment flexibility, and post-hoc collision detection are all rarely seen in synthetic datasets.
The performance gap between configuration-independent and configuration-dependent metrics empirically validates the taxonomy's central insight: there is a fundamental trade-off between generalizability and configuration specificity.
The work provides urgently needed infrastructure (dataset + benchmark + taxonomy) for the industrial AI community, representing a foundational contribution.

Limitations & Future Work¶

Only a single fixed punch/die configuration is used; cross-configuration generalization remains unvalidated.
Models observe only the final CAD geometry without encoding bending operation sequence information—sequential architectures (e.g., RNN/Transformer) may better capture operation-order dependencies.
The domain gap between synthetic data and real industrial parts requires physical validation (phenomena such as springback and elastic recovery are not modeled).
Coverage is limited to sheet metal bending; other forming processes (deep drawing, stamping, tube bending) await extension.
The inverse problem—recommending optimal configurations from a candidate tool set, thereby recasting IPMA as a configuration selection task—is a promising direction for future work.

vs. Ghadai et al. (2018): Uses simple cuboids with drilled holes for subtractive-process DFM; BenDFM offers far greater geometric complexity and physical fidelity, and is the first to cover forming processes.
vs. Peddireddy et al. (2021): Focuses on internal cavities and undercuts in subtractive processes; BenDFM requires simulating dynamic bending trajectories and part–tool interactions.
vs. SMCAD (Ma & Yang, 2024): SMCAD adds features to a small set of fixed base geometries for feature recognition; BenDFM is generated from a parametric bending process, yielding greater geometric diversity and DFM relevance.
Taxonomic implication: The configuration-dependence dimension provides guidance for transferability analysis in any data-driven DFM system.

Rating¶

⭐⭐⭐⭐ (3.5/5)

Novelty ⭐⭐⭐⭐: The taxonomy systematically fills a definitional gap; the dataset is the first to cover forming processes.
Experimental Thoroughness ⭐⭐⭐: The two architectures are representative, but broader method comparisons and cross-configuration experiments are lacking.
Writing Quality ⭐⭐⭐⭐⭐: Structure is clear, narrative logic is complete, and the taxonomy is articulated with precision.
Value ⭐⭐⭐⭐: Provides urgently needed infrastructure (taxonomy + dataset + benchmark) for the manufacturing AI community.