TY  - JOUR
AB  - Computer-aided design (CAD) plays a critical role in the development of innovative 3D models and tackles many different industrial branches ranging from small enterprises to large corporations. In recent years, growing interest has emerged in automating CAD workflows, using deep learning, to reduce repetitive manual tasks and allow engineers to focus on higher-level design and innovation. However,despite major advances in deep learning, especially in Natural Language Processing (NLP), applying these methods to CAD data remains challenging. One major challenge is an efficient representation of CAD data. Traditional methods such as voxel-based and point-cloud representations suffer from limitations, such as high memory consumption and loss of geometric detail. On the other side, standard tokenization techniques for NLP such as WordPiece, originally developed for text, may not handle the discrete and structured nature of CAD data. We focus on advanced text-based representations for CAD, building upon the DeepCAD framework, which is well suited for transformer-based architectures. We compare three representation strategies: a common method from NLP, a method inspired by DeepCAD,and a hybrid approach that combines elements of both. Our study aims at evaluating their effectiveness and at identifying pathways toward more robust and scalable CAD representations for design automation.
AU  - Habibi, Sayeda Hadisa
AU  - Bergelt, Julia
AU  - Teichmann, Michael
AU  - Hamker, Fred H.
DO  - 10.17619/UNIPB/1-2654
PB  - Universitätsbibliothek
DP  - Universität Paderborn
LA  - eng
PY  - 2026
SP  - 1 Online-Ressource (Seite 139-148) : Diagramme
T2  - 1st International Symposium: March 24 – 26, 2026, Heinz Nixdorf Institute, Paderborn University
TI  - Influence of tokenization strategies on the prediction of CAD model descriptions
UR  - https://nbn-resolving.org/urn:nbn:de:hbz:466:2-58914
Y2  - 2026-10-04T19:11:50
ER  -