Core Concepts
The proposed hybrid method combines the expressive power of neural networks and the convergence of finite difference schemes to efficiently solve Stokes equations with singular forces on an embedded interface in regular domains.
Abstract
The paper presents a hybrid neural-network and finite-difference method for solving Stokes equations with singular forces on an interface in regular domains. The key idea is to decompose the solution into singular and regular parts.
The singular part solution is obtained using neural network learning machinery to capture the non-smooth behavior across the interface. The neural networks are trained to satisfy the jump conditions along the interface.
The regular part solution is then computed using a traditional MAC (Marker-And-Cell) scheme and Uzawa-type algorithm to solve the resulting Stokes-like system. The regular part solution is smooth across the interface.
The overall computational cost is efficient, as it leverages well-established fast Poisson solvers. The method is demonstrated through several numerical experiments in both 2D and 3D, showing second-order convergence for the velocity and first-order convergence for the pressure. The hybrid approach avoids the need for additional discretization efforts near the interface, unlike traditional sharp interface methods.
Stats
The velocity field u = (u1, u2) is given by:
u1(r, θ) = (1/8r^2 cos(2θ) + 1/16r^4 cos(4θ) - 1/4r^4 cos(2θ)) for r < 1, and (-1/8r^-2 cos(2θ) + 5/16r^-4 cos(4θ) - 1/4r^-2 cos(4θ)) for r ≥ 1.
u2(r, θ) = (-1/8r^2 sin(2θ) + 1/16r^4 sin(4θ) + 1/4r^4 sin(2θ)) for r < 1, and (1/8r^-2 sin(2θ) + 5/16r^-4 sin(4θ) - 1/4r^-2 sin(4θ)) for r ≥ 1.
The pressure p(x, y) is given by: x^3 + cos(πx)cos(πy) for r < 1, and cos(πx)cos(πy) for r ≥ 1.
The external force g = (g1, g2) is given by:
g1(x, y) = (-π sin(πx)cos(πy) + 6x^2 - 3y^2) for r < 1, and (-π sin(πx)cos(πy) - 3(x^4 - 6x^2y^2 + y^4)/(x^2 + y^2)^4) for r ≥ 1.
g2(x, y) = (-π sin(πx)cos(πy) - 6xy) for r < 1, and (-π sin(πx)cos(πy) - 12(x^3y - xy^3)/(x^2 + y^2)^4) for r ≥ 1.