Fisika & Matematika editorial
Open AccessOA2022
A Deep Domain Decomposition Method Based on Fourier Features
The paper introduces F-D3M, a Fourier feature-based deep domain decomposition method for PDEs. It uses overlapping domain decomposition and multi Fourier feature networks to reduce high-frequency modes and improve accuracy and efficiency.
Sen Li; Ying-Zhi Xia; Yu Liu; Qi-Feng Liaoยท Journal of Computational and Applied Mathematicsยท 2022ยท DOI 10.48550/arXiv.2205.01884
The core problem
Deep neural network (DNN) based methods for solving partial differential equations (PDEs) have gained popularity, but their efficiency can degrade for problems with high-frequency modes. The authors propose a Fourier feature based deep domain decomposition method (F-D3M) to address this issue. The key idea is to combine overlapping domain decomposition with multi Fourier feature networks (MFFNets) to reduce high-frequency modes to relatively low-frequency ones in each subdomain, thereby improving the performance of DNNs.
Innovation
The F-D3M strategy involves partitioning the spatial domain into overlapping subdomains. In each subdomain, a multi Fourier feature network (MFFNet) is constructed. The loss function for each subdomain includes boundary and interface treatments to ensure continuity and accuracy. The mathematical framework of F-D3M is presented, detailing the construction of MFFNets and the treatment of boundary and interface conditions. The method leverages Fourier features to map inputs to a higher-dimensional space, enabling better approximation of high-frequency functions.
Introduction
Deep neural network (DNN) based methods for solving partial differential equations (PDEs) have gained popularity, but their efficiency can degrade for problems with high-frequency modes. The authors propose a Fourier feature based deep domain decomposition method (F-D3M) to address this issue. The key idea is to combine overlapping domain decomposition with multi Fourier feature networks (MFFNets) to reduce high-frequency modes to relatively low-frequency ones in each subdomain, thereby improving the performance of DNNs.
Why it matters
The paper discusses the advantages of F-D3M, particularly its ability to handle high-frequency problems by decomposing the domain and using Fourier features. The authors also analyze the computational efficiency and potential limitations. The general mathematical framework allows for extension to various types of PDEs and boundary conditions.
Who should read this
Pembaca Research Digest dan praktisi rumpun ini.
Opening member contentโฆ