Deep Learning for Geometry Understanding

This presentation was made at the 2019 NAFEMS World Congress in Quebec Canada

Resource Abstract

Design engineers typically lack the tools to answer important engineering and analysis questions concerning the safety, performance or cost of complex products. Instead, these questions are dealt with by experts who rely on powerful, but complicated software tools such as simulation packages. This process often adds a significant overhead both in terms of required engineering hours as well as required process time. More recently, highly automated, application-specific software tools (a.k.a. engineering apps or simulation apps) have emerged that can be directly used by design engineers. However, many engineering workflows still require a significant amount of manual work, thus rendering this method of tool democratization impossible. Often, the manual work arises out of the necessity to convey an intuitive understanding of geometrical patterns that cannot be mapped into a classical rule-based computer algorithm. Within the last couple of years deep learning has succeeded rule-based approaches in computer vision and natural language applications. While deep learning in the context of geometry understanding is still an area of emerging research, it has already successfully been applied as a process automation tool in Computer Aided Engineering.

In this paper we focus on a specific application of deep learning in the engineering domain: The detection, classification and modeling of complex 3D features. Every CAD model consists of geometrical features which vary deeply in complexity. Simple features such as holes, chamfers and fillets can be robustly recognized by every CAD software from both native and non-native CAD formats. In contrast, more complex features (e.g. ribs, undercuts, 6 sided pockets, connection elements) can typically not be reliable detected and measured. However, there are several applications where this type of feature modeling is required:

Modeling for simulation: It is often desirable to replace small parts with complex physical behavior such as connection elements with numerically efficient surrogate models. If these features are not included in the CAD model, the corresponding geometry has to be identified and modeled manually.

Manufacturability analysis: Regardless of the specific manufacturing process used it is important to secure the manufacturability of a part during the design stage. This task can be solved by detecting relevant geometrical features automatically.

Cost estimation: The possibility to determine the manufacturing costs of a part early in the design process (design to cost) is an important competitive advantage. This requires an automated detection and classification of cost relevant geometric structures.

Deep learning offers the possibility to create robust feature detection algorithms by learning from past examples. In this review paper we detail some applications as well as the underlying technology and highlight differences to traditional methods.

Document Details

AuthorSuwelack. S
Date 18th June 2019
OrganisationRenumics GmbH


Back to Search Results