CAD-feature enhanced machine learning for manufacturing effort estimation on sheet metal bending parts

Technical Summary: CAD-feature Enhanced Machine Learning for Manufacturing Effort Estimation on Sheet Metal Bending Parts

Problem Statement
Design for Manufacturing (DFM) aims to identify production constraints early to reduce cost and complexity. While Graph Neural Networks (GNNs) operating on Boundary Representations (B-reps) have shown promise for learning directly from CAD models, purely geometric representations often lack the process-specific semantics required for accurate manufacturability prediction. Critical manufacturing factors, such as surface roles, bend intent, or specific feature relationships, are not explicitly encoded in shape alone and are difficult for data-driven models to infer reliably. Furthermore, the field suffers from data scarcity due to intellectual property constraints, making purely data-driven approaches difficult to deploy, while traditional rule-based tools are rigid and labor-intensive. A significant gap exists in validating these methods on genuine industrial production data; most existing literature relies exclusively on synthetic datasets.

Methodology
The authors propose a hybrid approach that integrates rule-based sheet metal feature recognition with graph-based deep learning. The methodology consists of two primary stages:

1. Rule-Based Feature Recognition: A module operates directly on native STEP files to extract manufacturing-relevant information. This process occurs in three stages:

   • Geometric Abstraction: Identification of cylindrical surfaces (radius, length, angle) and labeling of faces as top, bottom, or side.
   • Manufacturing Feature Extraction: Recognition of bends, holes, side holes, and outer contours. Bend features include properties like inner/outer radius, angle, orientation, and aggregated flange lengths.
   • Relational Analysis: Detection of relational features, specifically bend corners where bends meet or intersect.
     The module achieves 100% accuracy on bend surface and property recognition in validation tests.

2. Manufacturing-Enhanced Graph Learning: The recognized features are injected as node attributes into the B-rep face adjacency graph (Attributed Adjacency Graph or AAG).

   • Input Representation: Each face node is represented by a 92-dimensional vector. This combines a 64-dimensional geometric embedding (derived from UV-Net or FoV-Net architectures via CNNs or ray casting) with a 28-dimensional vector of manufacturing features (boolean indicators for bends/holes, continuous parameters, and bend statistics).
   • Architecture: The model utilizes message passing and pooling operations to aggregate face embeddings into a single graph-level representation. Global design attributes (sheet thickness, total area, bounding box volume) are appended before the data is fed into a Fully Connected Neural Network (FCNN) for the final prediction.

Key Contributions

• Hybrid Modeling Strategy: The paper introduces a method to enrich B-rep graphs with recognized manufacturing features, allowing the model to jointly reason over geometric structure and manufacturing semantics. Unlike prior work that discards raw geometry in favor of feature-only graphs, this approach retains the full B-rep topology.
• Real-World Validation: The study validates the approach on a proprietary industrial dataset (KUL-bend) containing 503 sheet metal designs with measured bending times from a production floor. This represents one of the first validations of B-rep learning for manufacturing effort on genuine industrial data, addressing the reliance on synthetic benchmarks.
• Comprehensive Benchmarking: The method is evaluated on both a large-scale synthetic dataset (BenDFM) for tooling collision detection and the real-world KUL-bend dataset for effort estimation.

Experimental Results
The study compares pure B-rep learning models (UV-Net, FoV-Net) against their manufacturing-feature-enhanced counterparts (UV-Net (MF), FoV-Net (MF)).

• Effort Estimation (KUL-bend):

  • The baseline (predicting the mean) yielded an RMSE of 49.8s.
  • Pure geometric models performed better, with FoV-Net achieving an RMSE of 35.10s.
  • Feature Enhancement: Adding manufacturing features consistently improved performance. FoV-Net (MF) achieved the best results with an RMSE of 31.16s and an MAE of 20.38s. UV-Net (MF) showed the lowest MAPE (43.86%).
  • The results indicate that explicit manufacturing knowledge provides complementary information that benefits learning-based CAD analysis, particularly in real-world scenarios with high variability.

• Collision Detection (BenDFM):

  • Pure geometric models significantly outperformed the random baseline (50%), with FoV-Net reaching 80.60% accuracy.
  • Feature enhancement provided consistent but smaller gains (FoV-Net (MF) reached 81.26%). The authors suggest this smaller improvement may be due to the large dataset size allowing models to learn patterns directly from geometry, or that geometric features already capture much of the relevant information for collision detection.

Significance and Claims
The paper claims that hybrid modeling offers a feasible and effective path toward deployable tools for manufacturability assessment and effort estimation in industrial CAD environments. By concentrating the learning signal on process-relevant geometric patterns, the approach improves prediction accuracy across both synthetic and real-world tasks.

The authors maintain a modest tone regarding the study's scope. They acknowledge that the KUL-bend dataset is limited to a single production facility and that bending time targets include irreducible noise from operator and setup variability not encoded in the CAD model. They do not claim universal applicability but suggest that future work should investigate the transfer of this hybrid enrichment strategy to other processes (e.g., milling, injection molding) and larger, more diverse industrial datasets. The work positions itself as a step toward bridging the gap between data-driven learning and domain-specific engineering knowledge.