Eintrag in der Universitätsbibliographie der TU Chemnitz
Volltext zugänglich unter
URN: urn:nbn:de:bsz:ch1-qucosa2-1019236
Farahani, Aida
Hamker, Fred (Prof. Dr. ) ; Hasse, Alexander (Prof. Dr. sc. ETH) (Gutachter)
Exploring Deep Learning Approaches for 3D Deformation : Toward Finite Element Method Distillation
Kurzfassung in englisch
This thesis explores neural approaches for 3D deformation modeling with the objective of distilling FEM principles into efficient predictive frameworks. The study investigates both single-step and multi-step deformation modeling to enhance predictive accuracy and computational efficiency. In the single-step deformation approach, implicit representations and signed distance fields are employed as a meshless method to approximate FEM-based deformations on both synthetic data and industry datasets. This technique enables fast and efficient handling of high-resolution meshes while preserving predictive accuracy. Additionally, it significantly reduces processing time, making computation approximately 400 times faster than traditional FEM simulations for shell meshes. While the single-step approach focuses on immediate deformation effects, the multi-step deformation approach formulates deformation as a sequential decision-making process within a deep reinforcement learning framework. By encoding 3D shape variations into a latent space, it leverages two encoding strategies: a mesh encoder to capture geometric surface features and an image encoder to extract depth-based structural information. Optimization through PlaNet and CEM enhances accuracy and efficiency in predicting sequential deformations, improving both computational performance and predictive reliability. To address the lack of standard datasets in this domain, two custom datasets, DefBeam and DefCube, were created. DefBeam captures controlled, single-force deformations, providing insights into the immediate effects of applied forces. In contrast, DefCube records cumulative deformations under sequential forces, enabling the assessment of long-term predictive accuracy and generalization. The evaluation of these models demonstrates that deep learning and reinforcement learning effectively complement FEM simulations by improving prediction accuracy and computational efficiency. By integrating these AI-driven methods, the proposed framework assists experts in refining simulation workflows and advancing applications in material design, virtual prototyping, and industrial forming.
| Universität: | Technische Universität Chemnitz | |
| Institut: | Professur Künstliche Intelligenz | |
| Fakultät: | Fakultät für Informatik | |
| Dokumentart: | Dissertation | |
| Betreuer: | Hamker, Fred H. (Prof. Dr.) | |
| DOI: | doi:10.60687/2026-0029 | |
| SWD-Schlagwörter: | Finite-Elemente-Methode , Deep Learning , Bestärkendes Lernen |
|
| Freie Schlagwörter (Englisch): | Deformation Modeling , Finite Element Method , Deep Learning , Neural Representations , Implicit Representations , Reinforcement Learning , Sequential Deformation Prediction , Signed Distance Fields | |
| DDC-Sachgruppe: | 006.3 | |
| Sprache: | englisch | |
| Tag der mündlichen Prüfung | 19.01.2026 | |
| OA-Lizenz | CC BY 4.0 |
