Skip to main content
Back to timeline
arXivSource publication:

Writing the vorticity layer into the trial function: sl-PINN keeps its error flat across six decades of viscosity

Synopsis

At small viscosity the two-dimensional incompressible Navier–Stokes equations form a vorticity layer at the wall whose width shrinks with viscosity, and a physics-informed network trained on uniformly sampled points struggles to resolve it. sl-PINN embeds the asymptotic corrector and the law fixing its amplitude directly in the trial function, with the amplitude tied to the network's own wall trace, so no prior knowledge of the limit solution is needed. On a flat channel with slip boundary conditions the vorticity error stays uniform over six decades of viscosity, whereas an enrichment that uses only the thin scale and leaves profile and amplitude free degrades as viscosity falls.

Source-provided article image: Singular layer neural networks for the Navier-Stokes equations at small viscosity
arXiv

Interpretation

At small viscosity the vorticity forms a thin structure near the wall, and a network that learns only from uniformly sampled points cannot reproduce that structure at the wall. In control cases whose exact solution is a fixed smooth Fourier mode with no thin layer, lowering the viscosity from a moderate value to a very small one changed the error by less than a factor of four; once the exact solution carries a vorticity layer whose width shrinks with viscosity, the same network, sampling and optimizer settings produce rapidly growing vorticity error. Across a viscosity sweep of eight decades the relative vorticity error grew by about two orders of magnitude and the maximum vorticity error rose from about 10^-3 to the order of the solution itself, while stream-function and velocity errors changed by less than a factor of four because the Poisson solve smooths the layer away.

Writing the boundary-layer corrector together with the law that fixes its amplitude into the trial function leaves the network only a smooth outer solution that does not vary with viscosity. Earlier methods of this kind used only the thin scale, in a stretched variable or a composite expansion, and left the profile and amplitude for the network to learn freely. On the flat channel with slip boundary conditions the corrector-informed sl-PINN kept the vorticity error essentially unchanged over six decades of viscosity, while the scale-only enrichment grew by about a factor of three and its maximum pointwise error inside the layer reached the amplitude of the layer.

Once the corrector is available in closed form, the Laplacian in the residual can be discretized by a finite-volume flux difference on a coarse grid that does not resolve the thin layer. Applying the same discretization directly to the whole field is not consistent: at viscosity 1e-6 the discrete Laplacian of the exact solution is in error by four orders of magnitude. After splitting the field into a smooth regular part and an analytic layer part and applying the flux difference only to the former, the sl-FVNN kept the same viscosity-uniform accuracy as the sl-PINN on both the Stokes and the Navier–Stokes problems, at about half the cost of a training step.

Perspective

All experiments are posed on a flat channel periodic in x with slip boundary conditions on both walls, and the slip speed of the limit flow is constant along the wall, so the corrector has a closed form; the conclusions concern the vorticity layer in this setting, not the no-slip Prandtl layer.

A curved wall, a slip speed that varies along the wall and the no-slip boundary condition each restore a term that the present construction removes, and none of the three is tested; with a varying slip speed the corrector becomes a Duhamel integral and the analytic profile is only a leading-order approximation with a remainder. The authors list varying slip speed and the no-slip case as future work.