Decoupling numerical method based on deep neural network for nonlinear degenerate interface problems

Fan Chen

doi:10.1016/j.cpc.2024.109275

2024-10-01

Fan Chen

Muhammad Aamir Ali

Zhiyue Zhang *

DOI：10.1016/j.cpc.2024.109275

原文链接

查看原文

摘要

En 中文

Many practical problems, including modeling composite materials, nuclear waste disposal, oil reservoir simulations, and flows in porous medium, commonly involve interface problems. However, the solution to interface problems with discontinuous coefficients of PDEs using fully decoupled numerical methods is challenging. The main objective is to solve the interface problems with fully decoupled numerical methods. This paper proposes an efficient decoupled numerical method for solving degenerate interface problems with double singularities. First, we divide the whole domain into singular and regular subdomains. Then, we use the Deep Neural Network (DNN) to find the solution on the singular subdomain and approximate the solution on the regular subdomain using the finite difference method. The scheme combines the solutions of singular and regular subdomains, which is an exciting idea. The key to the new approach is to split nonlinear degenerate partial differential equations with an interface into two independent boundary value problems based on deep learning. In this way, the expansion of the solution on the singular domain does not contain undetermined parameters, and two independent boundary value problems can be solved with any well-known traditional numerical methods. The main advantage of the proposed scheme is that we not only get the order of convergence of the degenerate interface problems on the whole domain, but we also can calculate VERY BIG jump ratio (such as 10 12 : 1 or 1 : 10 12 ) for the interface problems including degenerate and non -degenerate cases. Finally, with examples, we demonstrate the efficiency and accuracy of methods for 1 and 2D problems. It is also interesting that the proposed method is valid for the interface problems with degenerate and non -degenerate cases, we show it with some examples.

Keyword:

Nonlinear degenerate interface problems

Deep neural network

Fully decoupled method

Very big jump ratio

Convergence order

Sharp edge interface