Manufacturing Software Development Notes · Sean Kim
Autodesk Fusion CAM automation: AI feature recognition
Even if the average performance score of an AI for machining feature recognition improves, there is no guarantee that the resulting CAM processes will be generated correctly. Based on previous R&D, we outline the key verification points across shape recognition, process generation, and result validation.
Study render: An example showing the machining features of a composite part, broken down by face. This is not a photograph of a customer's part.
Turning CAD faces and their connections into AI input
STEP The types of surfaces, edges, and curves were read from the model, and the convex/concave relationships of the adjacent surfaces were expressed in a graph. Dealing with features created by multiple faces together, such as pockets or holes, requires not only the shapes of the individual faces, but also their connection relationships.
My scope of work included geometric analysis, development of a labeling program for training data, AI for recognizing machining features, design of machining strategies, and their application in Fusion. For the Digital Twin (MDT), I was responsible solely for the visualization of simulation results generated by a Korea Institute of Machinery and Materials (KIMM).
View average scores alongside errors for the entire part
Misclassifying even a few faces can alter the grouping of machining features. We evaluated results based on face-level metrics, the number of parts with all faces correctly identified, errors by class, and outcomes by part family. Even if the overall average improved, we rejected candidates that showed an increase in errors for specific part families.
We have separated the improvements gained from label modifications from those resulting from changes to the training process. This is because if inputs, augmentations, and labels are all altered simultaneously, it becomes difficult to determine which specific change led to the improvement.
Define data splits and evaluation scopes based on geometry families.
Verify that the input used during training and the input used for inference follow the same criteria.
Check whether the label is maintained even if you save and reread the STEP.
Turning recognition results into CAM operations that run in Fusion
Recognition results alone do not constitute a complete machining process. Machining orientation, clamping, process sequence, tools, and cutting conditions were determined and then integrated into Fusion’s setups, tools, and operations.
We distinguished between cases requiring only a tool change and those requiring the creation of a new operation. We then verified the application process, tool changes, and error recovery using automated regression testing—ensuring that results remain consistent not only upon initial generation but also after any subsequent changes.
Visualize simulation results from partner organizations
AI's basic surface classification and shape combination analysis identifies 9 types of stock, blind hole, through hole, open pocket, closed pocket, chamfer, fillet, counterbore, and countersink. The full example introduces six labels of STEP models for actual learning and the Fusion machining path for each feature.
In the result viewer, cutting force and chip thickness can be examined by process and NC segment, and the design model can be compared with the machined result in 3D. The tolerance for the public example is ±0.15 mm against the design. Surface waviness and form-error maps appear only in analyses that include that output.
I verified face accuracy on an independent test set of 25 parts and 1,346 faces (99.33%), the stock simulation and the comparison with the reference STEP file. On-site evaluation with real customer-order parts, measured surface roughness, reading drawing tolerances and approval of actual CNC machining quality were not part of the verification.
Korea Institute of Machinery and Materials (KIMM)