DeepMill: Neural Accessibility Learning for Subtractive Manufacturing

TL;DR

DeepMill addresses the need for fast and accurate accessibility analysis in subtractive manufacturing by predicting inaccessible and occlusion regions on arbitrary CAD and freeform surfaces under varying cutter parameters. It introduces a cutter-aware, dual-head octree-based CNN (O-CNN) with an embedded cutter module, operating on point clouds, and trained via a joint cross-entropy loss to produce both inaccessible and occlusion maps in real time. A Voronoi-based geometric data-generation pipeline accompanies the model, enabling scalable labeling across diverse cutter sizes and complex geometries. The approach achieves high accuracy ($94.7\%$ inaccessible, $88.7\%$ occlusion) and ultra-fast inference ($\approx 0.04\text{s}$ per shape), with substantial efficiency gains over traditional geometric methods, and is validated across CAD, freeform, and volume-occupying datasets. The work also provides a dataset and ablation studies, illustrating cutter-embedding benefits and promising extensions to symmetry-aware design and volume-based accessibility.

Abstract

Manufacturability is vital for product design and production, with accessibility being a key element, especially in subtractive manufacturing. Traditional methods for geometric accessibility analysis are time-consuming and struggle with scalability, while existing deep learning approaches in manufacturability analysis often neglect geometric challenges in accessibility and are limited to specific model types. In this paper, we introduce DeepMill, the first neural framework designed to accurately and efficiently predict inaccessible and occlusion regions under varying machining tool parameters, applicable to both CAD and freeform models. To address the challenges posed by cutter collisions and the lack of extensive training datasets, we construct a cutter-aware dual-head octree-based convolutional neural network (O-CNN) and generate an inaccessible and occlusion regions analysis dataset with a variety of cutter sizes for network training. Experiments demonstrate that DeepMill achieves 94.7% accuracy in predicting inaccessible regions and 88.7% accuracy in identifying occlusion regions, with an average processing time of 0.04 seconds for complex geometries. Based on the outcomes, DeepMill implicitly captures both local and global geometric features, as well as the complex interactions between cutters and intricate 3D models.

Figures (13)

• Figure 1: DeepMill's network architecture. The input mesh is converted into a point cloud with normals, with each point corresponding to a Voronoi cell's site. Features are progressively extracted through the encoder, and the decoder, embedded with a cutter module (CM), restores spatial resolution. Both encoder and decoder are stacked with several Octree-based residual blocks. Finally, each site is subjected to dual-task binary segmentation through two header layers. Red and green represent the inaccessible and occlusion regions, respectively.
• Figure 2: Comparisons with GraphSAGE. $M$ is converted into a graph, with nodes representing mesh vertices and edges representing topological connections of them. Similar to O-CNN, initial node features include vertex coordinates and normals. Node features are propagated and updated through successive convolutions on neighboring nodes.
• Figure 3: Illustration of inaccessible point detection. Orange represents the cutter, and the gray points represent sampled Voronoi sites. (a) A ball-end cutter can be simplified using four parameters. Note that above cutter is a non-accessible shaft space, and $PD$ is set to infinity. (b) For collision detection with red points, the mesh is first rotated, and points are quickly filtered by checking whether they lie within the detection box (red) of radius FR+$\sigma$, which eliminates most points far from the cutter. $\sigma$ is set to 5 in our experiments. (c) A finer collision check is performed for the points inside the box. (d) To prevent the cutter from penetrating the mesh without detection, the spacing between adjacent sites must be smaller than the cutter’s ball-end radius ($CR$).
• Figure 4: Illustration of occlusion point calculation. (a)$\sim$(c) Perform collision detection for three inaccessible points, recording the points that collide with the cutter in each cutter direction and counting the total number of collisions for each point. (d) The top 10$\%$ of points with the highest total collision counts are labeled as occlusion points ($l_O$).
• Figure 5: The gallery of DeepMill prediction results. On the left are CAD shapes, and on the right are freeform shapes. For each row of shapes, the first and third columns show the inaccessible and occlusion regions predicted by DeepMill. In the second and fourth columns, darker shades represent under-predicted areas, while lighter shades indicate over-predicted areas.