Physical Constrained Deep Neural Network

Event details
Date | 23.06.2021 |
Hour | 11:15 › 13:15 |
Speaker | Zhen Wei |
Category | Conferences - Seminars |
EDIC candidacy exam
exam president: Prof. Lenka Zdeborova
thesis advisor: Prof. Pascal Fua
co-examiner: Prof. Jan Hesthaven
Abstract
Shape optimization is an important technique that has been widely applied in various engineering fields.
However, the high computational cost of simulation prohibits conventional
shape optimization from becoming more efficient and effective.
To this end, the problem of automatic shape optimization is proposed to achieve a faster design process and
significantly improved optimized shapes.
In this proposal, the technical background for automatic shape optimization is discussed first.
Then three papers relating to this project are reviewed.
Specifically, an attempt that tackles the shape optimization problem with deep learning models is
introduced, which serves as a basis for future studies in this project.
To solve the problems in the existing implementation and develop novel functionalities for shape optimization,
the latest techniques of the mesh adaptation for CFD numerical methods and the graph learning models are reviewed,
providing inspirations as well as potential solutions for future study. The major directions for further investigations are also
elaborated in the last section of this proposal.
Background papers
1. E. Remelli et al., “MeshSDF: Differentiable Iso-Surface Extraction,” in NeurIPS, 2020.
https://papers.nips.cc/paper/2020/hash/fe40fb944ee700392ed51bfe84dd4e3d-Abstract.html
2. F. Xia et al., “Graph Learning: A Survey,” IEEE Transactions on Artificial Intelligence, 2021.
https://ieeexplore.ieee.org/abstract/document/9416834
3. F. Alauzet and A. Loseille, “A decade of progress on anisotropic mesh adaptation for computational fluid dynamics,” Computer-Aided Design, 2016.
https://www.sciencedirect.com/science/article/pii/S0010448515001517
exam president: Prof. Lenka Zdeborova
thesis advisor: Prof. Pascal Fua
co-examiner: Prof. Jan Hesthaven
Abstract
Shape optimization is an important technique that has been widely applied in various engineering fields.
However, the high computational cost of simulation prohibits conventional
shape optimization from becoming more efficient and effective.
To this end, the problem of automatic shape optimization is proposed to achieve a faster design process and
significantly improved optimized shapes.
In this proposal, the technical background for automatic shape optimization is discussed first.
Then three papers relating to this project are reviewed.
Specifically, an attempt that tackles the shape optimization problem with deep learning models is
introduced, which serves as a basis for future studies in this project.
To solve the problems in the existing implementation and develop novel functionalities for shape optimization,
the latest techniques of the mesh adaptation for CFD numerical methods and the graph learning models are reviewed,
providing inspirations as well as potential solutions for future study. The major directions for further investigations are also
elaborated in the last section of this proposal.
Background papers
1. E. Remelli et al., “MeshSDF: Differentiable Iso-Surface Extraction,” in NeurIPS, 2020.
https://papers.nips.cc/paper/2020/hash/fe40fb944ee700392ed51bfe84dd4e3d-Abstract.html
2. F. Xia et al., “Graph Learning: A Survey,” IEEE Transactions on Artificial Intelligence, 2021.
https://ieeexplore.ieee.org/abstract/document/9416834
3. F. Alauzet and A. Loseille, “A decade of progress on anisotropic mesh adaptation for computational fluid dynamics,” Computer-Aided Design, 2016.
https://www.sciencedirect.com/science/article/pii/S0010448515001517
Practical information
- General public
- Free
Organizer
- EDIC
Contact
- edic@epfl.ch