Swapping tanh for a Biot–Savart-shaped activation cuts BS-PINN velocity error by about a third across 21 flow comparisons
Synopsis
Physics-informed neural networks for incompressible flow have long relied on a generic tanh activation, leaving the network to learn spatial flow structure through optimization alone. BS-PINN instead takes an odd rational profile from a regularized Biot–Savart kernel as the hidden-layer activation, keeping the usual velocity–pressure outputs and equation-residual training and offering fixed-scale and learnable-scale versions. Across 21 benchmark–architecture comparisons spanning Kovasznay flow, cylinder wake, three-dimensional Beltrami flow, Taylor–Green vortex and double shear layer, both versions beat the tanh baseline on combined velocity error, with mean relative reductions of 31.7% and 33.2%.
Interpretation
For incompressible-flow PINNs, the analytical structure of the flow can be written directly into the hidden-layer activation rather than pursued only through loss functions and sampling strategies. Earlier PINNs used generic activations such as tanh, so the network had to learn the solution's spatial features entirely through optimization; prior activation work includes adaptive activations, Fourier features, periodic activations and trainable sinusoidal activations. Across five benchmarks and three network architectures, 21 comparisons in all, both the fixed-scale and learnable-scale versions had lower combined velocity error than the tanh baseline, with mean relative reductions of 31.7% and 33.2%.
The construction turns a scalar profile of the Biot–Savart kernel into the hidden-layer nonlinearity while leaving the rest of the network unchanged. The Biot–Savart kernel maps vorticity to velocity, and its spatial form is fixed by the equations rather than chosen as an arbitrary basis; here a one-dimensional slice of the regularized two-dimensional kernel is normalized and given a small stabilization term, yielding a smooth, odd, bounded rational activation with algebraic tails. Both versions keep the conventional velocity–pressure outputs and Navier–Stokes residual objective, and within each benchmark and architecture the sampling points, loss weights and optimization schedule are identical, so the difference is the activation alone, with the learnable version additionally optimizing the kernel scale.
The error reduction appears not only on smooth analytical solutions but also in flows with localized shear and vortical structures. Multiple spatial scales, strong velocity gradients and evolving vortical structures have been the hard cases for PINN accuracy, with errors well above those on smooth steady solutions. On the Taylor–Green vortex the learnable version reduced mean combined velocity error by 69.1% and pressure error by 68.7%; on the double shear layer the velocity-error reduction reached 70.1% for a tested steepness parameter and architecture, with vorticity error falling as well.
Perspective
The findings rest on five incompressible-flow benchmarks covering steady and unsteady, two- and three-dimensional, periodic-boundary and localized-shear cases, all trained and evaluated without external forcing in whole-space or periodic settings; the long-time double-shear-layer roll-up comparison is qualitative imagery with no corresponding DNS error estimate.
Learning the kernel scale is not consistently better than fixing it: the fixed version has the lowest mean velocity error for all three network sizes on the Beltrami flow and also leads on several Taylor–Green and pressure metrics. The activation itself guarantees neither incompressibility nor boundary conditions, which are still imposed through equation, divergence and boundary losses; the authors also note it is not a direct implementation of the stochastic integral formula or the Biot–Savart integral, and its approximation properties, parameter sensitivity and long-time accuracy remain to be characterized quantitatively.
