CAD Defeaturing Using Machine Learning
New machine learning-based tools to reduce tedious human operations to prepare models for simulation are under development. Impact on shortening design turn-around for mechanisms and components for LEP is promising.
A key step in the design and development process is having engineering analysts review CAD models and assemblies that are developed from manufacturing specifications. These CAD models contain a variety of irrelevant details, such as the head of a bolt, which, while are necessary for manufacturing, have little to no effect on the outcome of simulations and decrease overall simulation speed.
CAD simulations require software that creates a mesh; however, current technology requires defeaturing before mesh creation. Currently, there are two methods in practice to “defeature” a CAD model which are time intensive, tedious, or partially automated, but are too corrective and remove key features for physics simulations. Sandia is creating user-guided, smart defeaturing solutions that can learn from previous models for use in its Cubit software. The solutions use machine learning predictions of mesh quality for geometric features of a CAD model prior to meshing to identify potential problem areas and provide a prioritized list of suggested resolutions.
Technology Advancement: To improve overall meshing outcome, Sandia’s solution uses machine learning models that are trained using a combination of geometric and topological features from the CAD model and local quality metrics for ground truth. CAD defeaturing predicts local mesh quality without meshing, the best CAD operation for potential use, and allows the user to choose the appropriate operation.
Impact: The machine learning algorithms identify options previously unidentified by the CAD analysts and have speed-up evaluation processes through shortening defeaturing time. While the process has yet to be applied to multiple parts or fully automated, continual improvements are being made to the approach.
2018: Initial experiments to predict meshing outcomes and drive defeaturing on CAD parts using ensembles of decision trees
2019: Implementation in CAD tool and initial results published. Initial experiments with reinforcement learning methods.
2020: ML methods to drive improvement of CAD assemblies to correct gaps, overlaps and misalignments.
2021: Improvement of reinforcement learning methods to automatically drive defeaturing.
1. CAD Defeaturing using Machine Learning. SAND2019-6988C
2. Machine Learning-based CAD Defeaturing for Tetrahedral Mesh Generation. SAND2019-0835C