[Paper Review] Learning Localized Geometric Features Using 3D-CNN: An Application to Manufacturability Analysis of Drilled Holes

Aditya Balu, Sambit Ghadai+4 more|arXiv (Cornell University)|Dec 7, 2016

3D Surveying and Cultural Heritage19 references3 citations

TL;DR

This paper proposes a 3D-CNN framework augmented with surface normals to learn localized geometric features in CAD models for manufacturability analysis. By leveraging 3D gradient-weighted class activation mapping (3D-GradCAM), the method identifies critical features—such as hole depth-to-diameter ratios and proximity to edges—that determine whether a drilled hole is manufacturable, achieving accurate, interpretable classification without manual feature engineering.

ABSTRACT

3D convolutional neural networks (3D-CNN) have been used for object recognition based on the voxelized shape of an object. In this paper, we present a 3D-CNN based method to learn distinct local geometric features of interest within an object. In this context, the voxelized representation may not be sufficient to capture the distinguishing information about such local features. To enable efficient learning, we augment the voxel data with surface normals of the object boundary. We then train a 3D-CNN with this augmented data and identify the local features critical for decision-making using 3D gradient-weighted class activation maps. An application of this feature identification framework is to recognize difficult-to-manufacture drilled hole features in a complex CAD geometry. The framework can be extended to identify difficult-to-manufacture features at multiple spatial scales leading to a real-time decision support system for design for manufacturability.

Motivation & Objective

• To develop a deep learning framework that automatically identifies localized geometric features critical for manufacturability in CAD models.
• To address the limitations of 2D projections and basic voxel representations in capturing local geometric details for manufacturing constraints.
• To create an interpretable, real-time decision support system for design for manufacturability (DFM) by learning complex DFM rules from data.
• To extend deep learning applications beyond object recognition to localized feature detection in mechanical design.
• To enable integration of manufacturability feedback directly into CAD workflows through a GPU-accelerated, end-to-end learning pipeline.

Proposed method

• Voxelizing CAD models into 3D occupancy grids and augmenting them with surface normal information to enrich geometric representation.
• Training a 3D-CNN on a dataset of manufacturable and non-manufacturable parts to classify manufacturability based on local features.
• Applying 3D gradient-weighted class activation mapping (3D-GradCAM) to localize and visualize the specific 3D regions responsible for classification decisions.
• Using a semi-supervised learning approach to identify relevant local features without requiring explicit feature labeling.
• Employing GPU-accelerated voxelization pipelines to enable real-time processing of CAD models for interactive design feedback.
• Validating the model on diverse geometries, including blocks with single/multiple holes, L-shaped parts, and cylindrical components.

Experimental results

Research questions

• RQ1 Can a 3D-CNN effectively learn localized geometric features in CAD models when augmented with surface normals?
• RQ2 How well can 3D-GradCAM localize the specific features that determine manufacturability in drilled holes?
• RQ3 Can the framework generalize to complex geometries with varying external shapes while maintaining accurate manufacturability prediction?
• RQ4 To what extent does surface normal augmentation improve feature learning compared to voxel-only representation?
• RQ5 Can the method outperform traditional rule-based DFM tools by learning complex, implicit DFM rules from data?

Key findings

• The 3D-CNN with surface normal augmentation successfully learned and localized key manufacturability features such as hole depth-to-diameter ratios and proximity to edges.
• 3D-GradCAM accurately highlighted non-manufacturable features, such as holes too close to a face or corner, even in complex geometries like L-shaped blocks and cylinders.
• The framework achieved higher accuracy than conventional rule-based DFM tools by learning implicit, data-driven DFM rules without manual feature engineering.
• The model generalized well to unseen geometries, correctly classifying parts with multiple holes on the same or different faces based on manufacturability criteria.
• The method enabled interpretable feedback by visualizing which 3D regions influenced the classification, supporting design modification.
• The GPU-accelerated pipeline allowed for real-time manufacturability analysis, suitable for integration into CAD systems.