Deep learning for boundary representation CAD models

Abstract

This thesis explores utilising deep learning methodologies for tasks relating to learning from boundary representation (B-Rep) CAD models. The ambition is to use deep learning for an automatic feature recognition algorithm to identify geometric features to help automate the CAD to analysis pre-processing task of defeaturing, as it is a necessary but time-consuming operation. The three main activities to achieve this ambition were: i) construct a shape representation that could both encode information carried by the B-Rep, while also being suitable as a direct input to a deep learning algorithm, ii) develop a deep learning algorithm that could take advantage of and learn from this shape representation, and iii) generate a dataset that can be used for the automatic feature recognition task. In response to these objectives a new shape representation called a hierarchical B-Rep graph was constructed that encodes the geometry and topology of the B-Rep CAD model through a hierarchical graph, denoting information about a 2D surface mesh and B-Rep face topology. The neural architecture developed to learn from the hierarchical B-Rep graphs is called Hierarchical CADNet. It is composed of two spatial graph convolutional networks that operate on either level of the hierarchical B-Rep graph, while facilitating information sharing between the networks. Lastly, a new automatically generated CAD dataset with specific geometric features was created called MFCAD++. This dataset was used to train the Hierarchical CADNet architecture for the automatic feature recognition task. The new approach was compared with other state-of-the-arts methods on the MFCAD++ dataset, and other related datasets, and showed comparable or better performance to these methods. For example, on the MFCAD++ dataset, Hierarchical CADNet achieved a test accuracy of 97.63% compared to the next highest tested method’s accuracy of 85.98%.