DeepMill: Neural Accessibility Learning for Subtractive Manufacturing

Fanchao Zhong, Yang Wang, Peng-Shuai Wang, Lin Lu, Haisen Zhao

Abstract

Determining machining accessibility—whether a cutting tool can reach a surface without collision—is a fundamental challenge in CNC machining. DeepMill introduces a real-time neural framework using an octree-based convolutional neural network (O-CNN) and "cutter-aware" embeddings. It achieves approximately 95% accuracy while processing complex 3D models in just 0.04 seconds, representing a massive speedup over traditional geometric simulations.

Introduction

In subtractive manufacturing, accessibility is a make-or-break factor. If a designer creates a part with deep cavities or narrow grooves that a standard cutting tool cannot reach without hitting the rest of the part, the design is unmanufacturable. Traditionally, engineers rely on geometric algorithms to simulate tool paths and detect collisions. While accurate, these methods are computationally expensive, often taking minutes or even hours for high-resolution models.

DeepMill addresses this bottleneck by framing accessibility as a 3D segmentation problem. By learning the spatial relationship between tool volumes and part surfaces, it provides a fast, tool-agnostic, and geometry-independent solution for real-time design feedback.

DeepMill segments 3D models into inaccessible (red) and occlusion (green) regions across diverse geometries.

Methodology

The DeepMill architecture treats accessibility as a segmentation task using a novel neural framework:

  • Octree-based CNN (O-CNN): Instead of processing dense 3D grids, DeepMill uses an octree structure. This focuses computation only on the surface of the mesh, allowing the model to handle high-resolution models efficiently without excessive memory usage.
  • Cutter-Aware Embedding: This core innovation embeds the specific parameters of the cutting tool (radii and heights) into the decoder layers. This allows the model to "understand" how different tool sizes interact with the same geometry.
  • Dual-Head Segmentation: The model features two output heads. One predicts inaccessible regions (where the tool cannot reach), and the other identifies occlusion regions (the specific parts of the geometry causing the collision).
  • U-Net Architecture: The framework uses an encoder-decoder structure with skip connections to capture both fine-grained local details and broad global context, such as a tool shaft hitting a distant flange.
The DeepMill architecture features an Octree-based U-Net with a dedicated Cutter Module (CM) for tool-aware feature integration.

Datasets and Training

To train the model, the researchers created a large-scale dataset using over 5,000 CAD shapes from the ABC dataset and freeform shapes from Thingi10K. Ground truth labels were generated using a fast Voronoi-based geometric method to identify inaccessible and occlusion points across a wide variety of randomly generated cutter sizes.

Key Results

DeepMill demonstrates significant improvements in both speed and robustness:

  • High Accuracy: The model achieved 94.7% accuracy in predicting inaccessible regions and 88.7% for occlusion regions.
  • Massive Speedup: For a 15K-vertex CAD model, traditional geometric methods took approximately 224 seconds. DeepMill finished the same task in 0.01 seconds. Even for complex models with over 100,000 vertices, it maintained a processing time of roughly 0.04 seconds.
  • Generalization: Despite being trained primarily on CAD data, the model generalized remarkably well to organic, freeform shapes, proving it learns underlying geometric principles rather than just memorizing standard mechanical features.

Conclusion and Future Work

DeepMill provides a path toward real-time manufacturability feedback within CAD software. By allowing designers to see unmanufacturable regions the moment they edit a shape, it could drastically reduce the iteration cycle between design and production. Future work will focus on expanding the model to support more complex 5-axis machining scenarios and a wider variety of irregular tool shapes.

Summary updated

BibTeX
@inproceedings{zhong2025deepmill,
  author = {Fanchao Zhong and Yang Wang and Peng-Shuai Wang and Lin Lu and Haisen Zhao},
  title = {{DeepMill: Neural Accessibility Learning for Subtractive Manufacturing}},
  booktitle = {SIGGRAPH (Conference Paper Track)},
  year = {2025},
  doi = {10.1145/3721238.3730657},
}
Figure