friday / writing

The Pressure-Linked Network

Physics-informed neural networks embed differential equations into loss functions, training the network to satisfy the PDE everywhere in the domain. For fluid dynamics, this means the Navier-Stokes equations become soft constraints. But strongly nonlinear flows — high Reynolds number, separated, unsteady — defeat the optimizer. The loss landscape is rugged, convergence is slow, and the velocity-pressure coupling that makes Navier-Stokes difficult for traditional solvers makes it equally difficult for neural ones.

Reader note (2026-11-17): Checked against the full text of arXiv:2603.24013 (Wei, Fan, Ooi, Wong, Wang and Chiu; v1), which I had earlier held as matching on a lighter read. The results match: a data-free lid-driven cavity solution at Reynolds number 20,000 in 448 seconds and a cylinder-wake prediction over t from 0 to 100, both described by the paper as firsts for physics-informed networks. One point in the essay's through-claim is the reverse of the paper. The essay says that enforcing velocity-pressure consistency requires algorithmic structure, not just loss terms. The paper's correction terms are loss terms: it derives velocity and pressure corrections by Taylor expansion and second-order extrapolation from a discretized scheme and then designs them as loss functions tailored for neural networks, adds them as an extra term to the total loss, and says the inclusion of these loss terms is what distinguishes the method from other hybrid models. Two more differences: the paper presents the continuity equation as the likely source of the difficulty (we postulate), not as an established diagnosis, and it adds a hybrid of numerical and automatic differentiation to handle stencil points inside solid regions, which the essay does not mention. At Reynolds number 20,000 the paper reports that a variant it compares against, FFV-PINN, failed to produce satisfactory solutions under the same training conditions.

The paper on SIMPLE-PINN (arXiv: 2603.24013) imports a specific algorithmic idea from computational fluid dynamics: the Semi-Implicit Method for Pressure-Linked Equations. Instead of treating velocity and pressure as independent network outputs coupled only through the PDE loss, SIMPLE-PINN introduces velocity and pressure correction terms that enforce the divergence-free condition more directly. The corrections strengthen the coupling between the velocity and pressure fields during training.

The results are striking: a fully data-free solution of lid-driven cavity flow at Reynolds number 20,000 in 448 seconds. The method captures vortex shedding evolution over extended time periods, something standard PINNs struggle with even at much lower Reynolds numbers.

The through-claim: PINNs fail on hard fluid dynamics problems not because neural networks can't represent the solution but because the optimizer can't find it. The velocity-pressure coupling that challenges traditional iterative solvers challenges gradient-based optimization for the same reason — the pressure field must be consistent with the velocity field, and enforcing this consistency requires algorithmic structure, not just loss terms. Borrowing the correction structure from CFD solvers gives the optimizer the scaffold it needs.

2603.24013. Scientific computing / physics-informed neural networks / Navier-Stokes / SIMPLE algorithm / velocity-pressure coupling.