Can PINNs Reconstruct Velocity and Pressure Fields from Passive Scalar Visualizations?

Background and Motivation

The authors review earlier PINNs work and introduce the Science paper that pushes the inverse problem further: given only scattered concentration measurements of a passive scalar (e.g., smoke, dye, MRI contrast agent), can the hidden velocity and pressure fields of an incompressible flow be recovered without any boundary‑condition or geometry information?

Core Problem Formulation

The inverse problem is severely under‑determined because the sole observation is the scalar concentration c(t,x,y). The unknowns are the three velocity components and pressure, totaling four fields. By embedding the Navier‑Stokes equations and the scalar transport equation as hard constraints, the concentration data become a conduit that unlocks the hidden fields.

Mathematical Framework

The coupled PDE system consists of the incompressible Navier‑Stokes momentum equations, the continuity equation, and the advection‑diffusion equation for the passive scalar. The authors emphasize that without the transport equation the scalar data would provide no information about the velocity, but the coupling forces the velocity to be consistent with the observed concentration evolution.

Network Architecture

HFM employs a two‑stage network. The first stage is a physics‑uninformed fully‑connected network that maps spatio‑temporal coordinates (t,x,y,z) to five outputs (c,u,v,w,p). The second stage computes the five PDE residuals via automatic differentiation. The network has ten hidden layers with 250 neurons per layer (effectively 5 × 50 neurons per output variable). Swish (SiLU) activation is used for smooth higher‑order derivatives, and weight normalization replaces batch normalization to keep residuals independent of batch composition.

Loss Function

The total loss is the sum of a data term (mean‑squared error between predicted and measured concentration) and five physics terms (MSE of each PDE residual). All terms are equally weighted in the original work, which the authors note can lead to gradient‑flow imbalance.

Training Details

Optimizer: Adam with learning rate ≈ 1e‑3.

Iterations: 10⁵–10⁶ steps (≈ 42 h for 2‑D cases, 70 h for 3‑D on an NVIDIA Titan V).

Mini‑batch size: 10 000 points.

Training loss shows a rapid drop followed by a slower convergence; the transport‑equation residual converges first, while the continuity residual is the slowest, reflecting the gradient‑pathology discussed in earlier PINNs studies.

Benchmark Cases and Results

Key experiments include:

Da Vinci vortex reconstruction : Using scattered concentration points, the network recovers velocity streamlines and pressure that visually match DNS reference.

2‑D cylinder wake (Re=100) : Training on a flower‑shaped domain with no explicit velocity/pressure boundaries yields velocity errors of 2–5 % and pressure errors of 5–8 % across a wide range of spatial (down to 250 points) and temporal (as few as 5 frames per vortex cycle) resolutions. Adding Gaussian noise up to 160 % of the signal increases error by at most a factor of three, demonstrating strong denoising from the physics regularizer.

Parameter identification : When Reynolds and Péclet numbers are treated as unknowns, the network recovers them with ≈ 6–7 % relative error, slightly higher than the <1 % error achieved when only velocity data are used.

3‑D cylinder wake : Extending to three dimensions raises errors to 5–10 % (velocity) and 8–15 % (pressure), confirming the method’s scalability.

Narrow channel wall‑shear stress : From concentration data alone, wall shear stress is computed via automatic differentiation, matching high‑order spectral element reference solutions without numerical differentiation artifacts.

3‑D intracranial aneurysm : Using synthetic contrast‑agent data confined to the aneurysm sac, the network predicts velocity, pressure, and streamlines, illustrating potential clinical relevance. The authors note that these results rely on synthetic data generated with the Nektar++ spectral element solver.

Comparison with Earlier PINNs Work

A side‑by‑side table (converted to prose) highlights three major advances over the 2019 JCP PINNs paper: (1) the observation switches from velocity to passive scalar concentration, (2) boundary‑condition dependence is largely removed when concentration gradients are sufficient, and (3) the method handles arbitrary geometries and both 2‑D and 3‑D domains.

Impact and Significance

Publishing in Science marks a milestone for physics‑informed machine learning, demonstrating a concrete application that turns qualitative flow visualizations into quantitative measurements. The work bridges the gap between method development and real‑world utility, inspiring extensions to MRI‑based blood‑flow analysis, atmospheric remote sensing, and other fields where only scalar visualizations are available.

Limitations and Future Directions

All benchmarks use synthetic data; real experimental images require preprocessing to extract concentration fields, introducing additional uncertainty.

The method assumes incompressible flow and sufficient scalar gradients; regions with uniform concentration lead to non‑unique velocity reconstructions.

Training is computationally intensive, limiting rapid clinical deployment; transfer learning and neural operators are suggested pathways for acceleration.

Extension to turbulent, compressible, or non‑Newtonian flows remains an open research challenge.

Overall, the article provides a thorough step‑by‑step walkthrough of the Hidden Fluid Mechanics (HFM) framework, from problem definition through implementation, evaluation, and critical discussion of its scope.