Enregistré dans:
Détails bibliographiques
Auteurs principaux: Rahman, Md Ashiqur, George, Robert Joseph, Elleithy, Mogab, Leibovici, Daniel, Li, Zongyi, Bonev, Boris, White, Colin, Berner, Julius, Yeh, Raymond A., Kossaifi, Jean, Azizzadenesheli, Kamyar, Anandkumar, Anima
Format: Preprint
Publié: 2024
Sujets:
Accès en ligne:https://arxiv.org/abs/2403.12553
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866912098922004480
author Rahman, Md Ashiqur
George, Robert Joseph
Elleithy, Mogab
Leibovici, Daniel
Li, Zongyi
Bonev, Boris
White, Colin
Berner, Julius
Yeh, Raymond A.
Kossaifi, Jean
Azizzadenesheli, Kamyar
Anandkumar, Anima
author_facet Rahman, Md Ashiqur
George, Robert Joseph
Elleithy, Mogab
Leibovici, Daniel
Li, Zongyi
Bonev, Boris
White, Colin
Berner, Julius
Yeh, Raymond A.
Kossaifi, Jean
Azizzadenesheli, Kamyar
Anandkumar, Anima
contents Existing neural operator architectures face challenges when solving multiphysics problems with coupled partial differential equations (PDEs) due to complex geometries, interactions between physical variables, and the limited amounts of high-resolution training data. To address these issues, we propose Codomain Attention Neural Operator (CoDA-NO), which tokenizes functions along the codomain or channel space, enabling self-supervised learning or pretraining of multiple PDE systems. Specifically, we extend positional encoding, self-attention, and normalization layers to function spaces. CoDA-NO can learn representations of different PDE systems with a single model. We evaluate CoDA-NO's potential as a backbone for learning multiphysics PDEs over multiple systems by considering few-shot learning settings. On complex downstream tasks with limited data, such as fluid flow simulations, fluid-structure interactions, and Rayleigh-Bénard convection, we found CoDA-NO to outperform existing methods by over 36%.
format Preprint
id arxiv_https___arxiv_org_abs_2403_12553
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Pretraining Codomain Attention Neural Operators for Solving Multiphysics PDEs
Rahman, Md Ashiqur
George, Robert Joseph
Elleithy, Mogab
Leibovici, Daniel
Li, Zongyi
Bonev, Boris
White, Colin
Berner, Julius
Yeh, Raymond A.
Kossaifi, Jean
Azizzadenesheli, Kamyar
Anandkumar, Anima
Machine Learning
Existing neural operator architectures face challenges when solving multiphysics problems with coupled partial differential equations (PDEs) due to complex geometries, interactions between physical variables, and the limited amounts of high-resolution training data. To address these issues, we propose Codomain Attention Neural Operator (CoDA-NO), which tokenizes functions along the codomain or channel space, enabling self-supervised learning or pretraining of multiple PDE systems. Specifically, we extend positional encoding, self-attention, and normalization layers to function spaces. CoDA-NO can learn representations of different PDE systems with a single model. We evaluate CoDA-NO's potential as a backbone for learning multiphysics PDEs over multiple systems by considering few-shot learning settings. On complex downstream tasks with limited data, such as fluid flow simulations, fluid-structure interactions, and Rayleigh-Bénard convection, we found CoDA-NO to outperform existing methods by over 36%.
title Pretraining Codomain Attention Neural Operators for Solving Multiphysics PDEs
topic Machine Learning
url https://arxiv.org/abs/2403.12553