.png){kind=icon}
.png){kind=icon}
Return
DOMAIN DECOMPOSITION LEARNING METHODS FOR SOLVING ELLIPTIC PROBLEMS
.png){kind=icon}
OA
AI
DOI：10.1137/22M1515392.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
With the aid of hardware and software developments, there has been a surge of interest in solving PDEs by deep learning techniques, and the integration with domain decomposition strategies has recently attracted considerable attention due to its enhanced representation and parallelization capacity of the network solution. While there are already several works that substitute the numerical solver of overlapping Schwarz methods with the deep learning approach, the nonoverlapping counterpart has not been thoroughly studied yet because of the inevitable interface overfitting problem that would propagate the errors to neighboring sub domains and eventually hamper the convergence of outer iteration. In this work, a novel learning approach, i.e., the compensated deep Ritz method using neural network extension operators, is proposed to enable the flux transmission across subregion interfaces with guaranteed accuracy, thereby allowing us to construct effective learning algorithms for realizing the more general nonoverlapping domain decomposition methods in the presence of overfitted interface conditions. Numerical experiments on a series of elliptic boundary value problems, including the regular and irregular interfaces, low and high dimensions, and smooth and high-contrast coefficients on multidomains, are carried out to validate the effectiveness of our proposed domain decomposition learning algorithms.
Keywords:
neural networks
domain decomposition methods
elliptic PDEs
compensated deep Ritz method
Journal
IF:
2.6
Papers:
5.1K
Citations:
1.8W