Gerico Vidanes

Deep neural networks are able to achieve high accuracy in semantic segmentation of geometries used in computational engineering. Being able to recognise abstract and sometimes hard to describe geometric features has applications for automated simulation, model simplification, structural failure analysis, meshing, and additive manufacturing. However, for these systems to be integrated into engineering workflows, they must provide some measures of predictive uncertainty such that engineers can reason about and trust their outputs. This work presents an empirical study of practical uncertainty estimation techniques that can be used with pre-trained neural networks for the task of boundaryrepresentation model segmentation. A point-based graph neural network is used as a base. Monte-Carlo Dropout (MCD), Deep Ensembles, testtime input augmentation, and post-processing calibration are evaluated for segmentation quality control. The Deep Ensemble technique is found to be top performing and the error of a human-in-the-loop system across a dataset can be reduced from 3.8% to 0.7% for MFCAD++ and from 16% to 11% for Fusion360 Gallery when 10% of the most uncertain predictions are flagged for manual correction. Models trained on only 5% of the MFCAD++ dataset were also tested, with the uncertainty estimation technique reducing the error from 9.4% to 4.3% with 10% of predictions flagged. Additionally, a point-based input augmentation is presented; which, when combined with MCD, is competitive with the Deep Ensemble while having lower computational requirements.

Geometry understanding is a core concept of computer-aided design and engineering (CAD/CAE). Deep neural networks have increasingly shown success as a method of processing complex inputs to achieve abstract tasks. This work revisits a generic and relatively simple approach to 3D deep learning – a point-based graph neural network – and develops best-practices and modifications to alleviate traditional drawbacks. It is shown that these methods should not be discounted for CAD tasks; with proper implementation, they can be competitive with more specifically designed approaches. Through an additive study, this work investigates how the boundary representation data can be fully utilised by leveraging the flexibility of point-based graph networks. The final configuration significantly improves on the predictive accuracy of a standard PointNet++ network across multiple CAD model segmentation datasets and achieves state-of-the-art performance on the MFCAD++ machining features dataset. The proposed modifications leave the core neural network unchanged and results also suggest that they can be applied to other point-based approaches.