AI that reads 3D models and designs processing strategies
Fusion CAM automation: Geometry analysis → AI-based machining feature recognition → Machining strategy design → Visualization of digital twin results.
PythonOpenCascadeTensorFlowAutodesk Fusion APICAM
Recognizing machining features in 3D models to automate Fusion CAM operations
Product View Simulation of a Fusion material processed into a standard specimen (302 variant) via an automatically generated manufacturing process. Green areas indicate regions within tolerance.
9 typesMachining feature to identify · Include stock
±0.15 mmCriteria for determining over-machining based on design tolerances (MDT example)
1 caseAccepting applications for registration in the labeling program.
1
Geometry analysis
Read the faces and edges of the 3D model (STEP) and create a graph of relationships such as type of curved surface and convex/concave connections.
2
AI machining feature recognition
Identifies 9 types of machining features—such as stock, holes, pockets, chamfers, and fillets—through face classification and shape combination analysis using a graph neural network.
3
Designing Machining Strategies
Establish the machining direction, clamping, process sequence, tooling, and cutting conditions; then, adjust path tolerances to meet surface roughness targets to create a Fusion CAM process.
4
Visualization of Digital Twin Results
Based on MDT simulation results provided by a Korea Institute of Machinery and Materials (KIMM), we created a 3D visualization that displays graphs of cutting forces and chip thickness, as well as the discrepancies between the actual machining results and the original design.
Assigned role
We conducted R&D on the entire process—from geometry analysis and machining strategy design to implementation in Fusion—while the digital twin component was used exclusively for visualizing simulation results.
3D model shape analysis and graph input
AI for Machining Feature Recognition and Labeling Software for Training
Designing Machining Strategies and Generating Fusion CAM Processes
Visualize MDT analysis results from Korea Institute of Machinery and Materials (KIMM) using graphs and 3D models
The digital twin (MDT) simulation was handled by a Korea Institute of Machinery and Materials (KIMM). I was responsible solely for the visualization of the provided simulation results.
Results
We developed AI that reads 3D CAD shapes, a learning labeling tool, a machining strategy design engine, and Fusion CAM linkage. The digital twin was only responsible for visualizing the results of simulations performed by Korea Institute of Machinery and Materials (KIMM), creating a screen where cutting forces and machining results could be viewed in graphs and 3D.
Verification extends to Fusion/MDT simulations; it does not constitute approval of actual machining quality
AIMachining feature to identify 9
StockStock/reference geometry
Blind holeHole with a closed bottom
Through-holeHole passing through a component
OpenPocketSide-open pocket
Close panelClosed-perimeter pocket
ChamferChamfering of edges.
FilletRounded corners
CounterboreCylindrical hole with a step
CountersinkHole with a conical entrance
AI classifies the seven basic types by surface and identifies counterbore and countersink by analyzing the shape combination of hole, step, and chamfer. Below is an example of six actual learning STEP models colored with correct answer labels.
The color coding corresponds to the features listed above. This data displays the geometries and labels used for training; it does not indicate prediction accuracy for new, unseen parts.
Fusion CAM operations built from recognized features
1–3 are representative Fusion CAM processes for the front setup. Figure 4 shows the flow of changing the processing direction and continuing the process on the opposite side. The blue line on the Fusion screen is the cutting path, and the yellow line is the connection/movement between paths.
1. Roughing of outer material
Product ViewRemoves remaining material around the part by processing it at multiple depths.
2 · Hole drilling
Product ViewSelect a drill based on the position and depth of the hole, and generate a machining path along the axial direction.
3. Roughing internal multi-stage pockets
Product ViewRemoves material from surrounding pockets while leaving the central cylindrical boss intact. The dense blue lines represent the cutting path, repeated at various depths.
4. Machining the back side after changing the processing direction
Product ViewThis screen shows the roughing path for the front internal pocket in Fusion. Processes 1–3 are performed in this +Z direction setup.
Setup 2 · Back top surface roughing (−Z)
Product ViewA view in Fusion showing the same part from the opposite side. After flipping the part, the process continues with the roughing of the top surface on the reverse side (in the −Z direction).
Setup 2 · Back contour finishing
Product ViewThe finishing toolpath for the outer contour is shown from the same reverse-side perspective. The finishing process follows the roughing of the top surface and outer contour on the reverse side.
Once the holes and pockets on the front side are machined, the part is flipped and secured for the next setup. The tool approach direction—relative to the model coordinate system—is changed from +Z to −Z, followed by roughing the top surface of the back side, roughing the outer profile, and finally finishing the contours.
Charts and 3D views of digital twin results
The digital twin (MDT) simulation was performed by a Korea Institute of Machinery and Materials (KIMM). I was responsible for the graphs and 3D visualizations that display these results. Shown below is a visualization of the cutting loads and machined geometry derived from the stored MDT analysis data.
Cutting Force, Chip Thickness, and NC Segment
Product ViewSelect specific machining segments to compare cutting forces and chip thickness; examine the corresponding NC block, tool, feed rate, and spindle speed at the point where these values reach their maximum. Translated reference · View Korean original
Machined Shapes and Deviations from the Design
Product ViewExamine the machining results with 3D rotation/enlargement and direction change. Design surface, remaining material, and overcut mark are distinguished by tolerance standard color and area ratio. Translated reference · View Korean original
Surface Waviness and Shape Error Maps
Product ViewAnalyses with this output display surface waviness and geometric deviations using color maps, shown separately from actual measured roughness values. Translated reference · View Korean original
Expected changes from this technology
Process design to reduce repetitive tasks
By reducing the repetitive tasks of selecting individual faces or holes and configuring machining processes, experts can focus more on tool selection and reviewing machining conditions.
Sharing Expert Know-how
By capturing the shape analysis and process design procedures of experienced experts into software, it becomes possible to assist operators with limited CAM experience and small-scale manufacturing facilities in leveraging technical expertise.
Resource conservation through virtual verification
Analyzing cutting loads and deviations from the design prior to actual machining can help reduce the waste of materials, tools, and time caused by trial and error.
The above points represent the expected benefits of utilizing this technology. Actual time and cost savings upon implementation in the field would require separate verification.
3DConcept model, step by step
A conceptual model illustrating the process sequence and tooling generated by the actual engine for this component. It shows the steps for modeling, feature recognition, and vise machining.
A-01Machined Parts
Concept model
Assigning machining features to each face
Face-level predictions are grouped into pockets, counterbores, holes, steps and chamfers. The rounded outer corners are not fillets; they belong to the outer profile together with the side walls.
1Pocket9 faces
2Counterbore2 places
3Through-hole4 places
4Step (offset/ledge)2 faces
5Outer profile8 side faces
6ChamferC2 · 6 faces
Drawing
A-01
Name
Pocket/stepped block
Step
02 Feature recognition
Unit
mm
Author
Gyu-tae Kim
Updated
Details · Actions taken at each step, issues and solutions, verification
01Actions taken at each step
1
Geometry analysis
STEP surface, edge, and curved surface type extraction
Graph of convex and concave relationships between adjacent faces
Inputting a hierarchical graph that groups faces and triangular facets
2
AI machining feature recognition
Labeling program for creating training data
Data splitting based on geometry families
Composite shape involving pockets, fillets, and chamfers
Comparison of effects: input, augmentation, and label changes
3
Designing Machining Strategies
Organizing predicted faces into groups of machining features
Defining machining direction, setup, process sequence, tooling, and cutting conditions.
Path tolerance and feed rate adjusted for surface finish targets
Setup and process creation with Fusion API
Maintaining existing operations even when tools are changed or an application fails.
4
Visualization of Digital Twin Results
Graphs showing cutting force and chip thickness for each process/NC segment.
Display maximum, average, and P95 values, along with the corresponding tool, feed rate, and spindle speed.
Visualization of deviations from the design, surface waviness, and geometric errors resulting from the machining process.
The simulation was handled by Korea Institute of Machinery and Materials (KIMM), while I was responsible for the visualization of the results.
Closed pocket
Open pocket
Blind hole
Extruded boss · Material
Chamfer
Stock ledge/step
Research render Composite shape pairs with only the distinguishing features altered. Verified that labels remain consistent even when re-imported via STEP.Hole, pocketChamferFilletStock (Workpiece)
02Problem and solution
ProblemMisclassifying even a few faces changes how machining features are grouped.
ResolveWe calculated scores based on individual faces, as well as the number of parts for which every face was correctly identified.
ProblemThe reason for the change in score—whether it was due to label modification or training—was ambiguous.
ResolveScores resulting from manual label corrections and scores resulting from machine learning-based improvements were calculated separately.
ProblemRecognition results must be integrated with Fusion’s setups, tools, and operations to form a complete machining process.
ResolveWe categorized scenarios into those requiring only a tool change and those requiring the creation of a new operation.
ProblemLooking at one face alone, closed and open pockets were hard to tell apart, so 4 of every 10 closed pockets were missed.
ResolveI expanded the face input from 5 to 43 columns so the model also sees the surrounding boundaries and whether cross-sections are open; it now finds every closed pocket.
ProblemThe MDT results showed the tool load spiking the moment the tool retracted (chip thickness 1.67 mm).
ResolveI changed the retract so the tool rises to the safe height at feed rate before the rapid move, cutting the tool load in step roughing from 8,191 N to 251 N.
03Verification
Verified items
Labels are maintained even when saved and read again in STEP
Segmentation by series, and alignment of training and inference inputs
Average metrics, errors by class, and count of parts with zero errors
Independent test set of 25 parts and 1,346 faces: face accuracy 99.33%, macro F1 99.58%
Korea Institute of Machinery and Materials (KIMM)
Closed pocket
Open pocket
Blind hole
Extruded boss · Material
Chamfer
Stock ledge/step