🔢 mathematics

Physics-Informed Neural Networks for the High-Resolution Reconstruction of Flow Measurement Indicators in Fluid Dynamics

This paper proposes a physics-informed neural network (PINN) framework that integrates the incompressible Navier-Stokes equations with sparse experimental flow data to accurately reconstruct high-resolution velocity fields and estimate critical hemodynamic indicators like wall shear stress and pressure, demonstrating superior performance over standard methods in both FDA nozzle and aneurysm models.

Original authors: Irena Radišić, Raffaele Tirotta, Alberto Zingaro, Stefano Pagani, Luca Dede'

Published 2026-07-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Irena Radišić, Raffaele Tirotta, Alberto Zingaro, Stefano Pagani, Luca Dede'

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a giant, invisible puzzle of how blood flows through your body. You have a camera (called 4D flow MRI) that takes pictures of the blood moving, but the pictures are a bit blurry, the edges are fuzzy, and some pieces are just missing. It's like trying to guess the shape of a river by looking at a few scattered pebbles on the bank.

This is the problem the authors of this paper are tackling. They want to fix those blurry, incomplete pictures of blood flow so doctors can see exactly how the blood is moving, even in tricky spots like the walls of a blood vessel or inside a bulging aneurysm.

The Magic Trick: Teaching Computers the Rules of Physics

The authors didn't just try to make the pictures sharper using standard photo-editing tricks. Instead, they taught a special kind of computer brain (a "Physics-Informed Neural Network," or PINN) the actual laws of physics that govern how fluids move. Specifically, they fed the computer the Navier–Stokes equations. Think of these equations as the "rulebook" for how water and blood behave—how they swirl, how they speed up, and how they push against walls.

The computer's job was to look at the blurry, incomplete data from the MRI scans and say, "Okay, I see these few points, but I also know the rulebook. If the blood is moving this way here, it must be moving that way there to obey the rules." By mixing the real (but messy) data with the perfect rules of physics, the computer could fill in the missing pieces and create a high-definition, 3D map of the blood flow.

The Test Drives: A Nozzle and a Bulging Balloon

To see if their idea actually worked, the team ran two different tests.

First, they used a FDA nozzle benchmark. This is a standard, tube-shaped model used by scientists to test flow. They had two types of data:

Perfect Simulations: They created a perfect, computer-generated flow (like a video game) and pretended some parts were missing.
Real Experiments: They used actual measurements taken with a technique called PIV (Particle Image Velocimetry), which is like taking photos of tiny particles floating in the water.

In these tests, the PINN was able to reconstruct the full flow field. When they compared the PINN's guess to the "perfect" simulation, it was much closer than just using the raw data alone. Even better, when they used the real, noisy experimental data, the PINN managed to smooth out the errors and predict things the camera couldn't see, like the pressure inside the tube and the wall shear stress (how hard the blood is rubbing against the tube walls).

Second, they tried a more complex scenario: a model of a brain aneurysm (a bulging weak spot in a blood vessel). This time, they used real 4D flow MRI data from a lab experiment. The data was sparse and noisy, especially near the edges. The PINN successfully reconstructed the velocity (speed and direction) and the pressure. It even managed to spot swirling patterns (vortices) that were hidden in the original blurry data.

What They Found (and What They Didn't)

The main finding is that this "physics-plus-data" approach works really well. It turns low-quality, sparse measurements into a high-resolution, 3D story of how blood flows.

Pressure: The paper shows that the method can estimate pressure, which is usually impossible to measure directly with these MRI scans. The reconstructed pressure maps looked very similar to the "ground truth" from the simulations.
Wall Shear Stress: This is a tricky one. The paper found that while the method improved the results, predicting the exact force of the blood rubbing against the wall is still difficult, especially if the data is very noisy or if the boundary conditions (like the exact shape of the vessel) aren't perfectly known.
The "No-Go" Zone: The paper explicitly rules out the idea that you can just use a standard computer simulation (CFD) or a simple data-driven AI without physics. Standard simulations are slow and hard to customize for every patient, while pure AI without physics rules tends to get confused by noisy data and produces "non-physical" results (like blood flowing in impossible ways). The authors argue that you need the physics rules to make the AI work correctly.

How Sure Are They?

The authors are very confident in their results, but they are careful about what they claim.

Simulations: For the FDA nozzle and the aneurysm, they proved the method works by comparing it to "ground truth" data from high-fidelity computer simulations. In these controlled, simulated environments, the method was highly accurate.
Experiments: When they used real experimental data (PIV and MRI), the results were promising and matched the physical reality better than standard methods. However, for the aneurysm case, they note that the Wall Shear Stress results were "poor and difficult to assess" because the data was so dispersed and noisy. They don't claim to have solved the problem of measuring wall stress perfectly yet; they just showed that their method is a better starting point than the alternatives.
Limitations: The paper admits that their current tests were mostly on laminar (smooth) flow or simplified models. They didn't test it on fully turbulent, chaotic blood flow, which happens in real human bodies. They also noted that their method relies on knowing the boundary conditions (like the exact speed of blood entering the vessel), which is often hard to know for real patients.

The Bottom Line

This paper suggests that by teaching computers the "rules of the game" (physics) alongside the "scoreboard" (data), we can turn blurry, incomplete medical scans into clear, detailed maps of blood flow. It's not a magic cure-all yet—especially for the trickiest measurements like wall stress—but it's a powerful new tool that makes the blurry pictures much sharper and reveals hidden details that were previously invisible.

Technical Summary: Physics-Informed Neural Networks for High-Resolution Reconstruction of Flow Measurement Indicators

1. Problem Statement

Accurate, spatially resolved flow field measurements are critical for assessing hemodynamic quantities in cardiovascular research and clinical practice. However, experimental techniques such as 4D flow MRI, Particle Image Velocimetry (PIV), and Doppler ultrasound often yield data that are sparse, noisy, or spatially under-resolved. These limitations are particularly pronounced near vessel walls and in regions of complex flow, such as recirculation zones or transitions to turbulence. Consequently, the fidelity of derived hemodynamic indicators—specifically Wall Shear Stress (WSS) and pressure distributions—is compromised, reducing their clinical utility for diagnosing pathologies like stenosis, aneurysms, and assessing rupture risks.

While Computational Fluid Dynamics (CFD) can provide high-resolution data, standard simulation pipelines (involving segmentation, meshing, and parameter tuning) are computationally expensive and laborious to personalize. Furthermore, pure data-driven approaches often fail to generalize well when data is sparse or noisy. The authors propose a solution that bridges this gap by integrating physical laws with experimental data to reconstruct high-fidelity flow fields.

2. Methodology

The study employs Physics-Informed Neural Networks (PINNs), a class of artificial neural networks trained to simultaneously leverage data and physical knowledge encoded via Partial Differential Equations (PDEs).

2.1 Mathematical Formulation

The framework assumes blood behaves as an incompressible Newtonian fluid governed by the dimensionless steady-state Navier–Stokes equations. The network approximates the velocity field $u(x,t)$ and pressure field $p(x,t)$ by minimizing a total loss function $\mathcal{L}$ composed of three main terms:
$\mathcal{L}(\tilde{\theta}) = \mathcal{L}_{PDE}(\tilde{\theta}) + \mathcal{L}_{BC}(\tilde{\theta}) + \mathcal{L}_{train}(\tilde{\theta})$

Physics-Based Loss ( $\mathcal{L}_{PDE}$ and $\mathcal{L}_{BC}$ ): These terms enforce the momentum balance, mass conservation (continuity), and boundary conditions (inlet, no-slip wall, and outflow) by computing residuals at a set of collocation points distributed throughout the domain and on the boundaries.
- To handle residuals of varying magnitudes, the authors incorporate a Mean Squared Logarithmic Error (MSLE) term alongside the standard Mean Squared Error (MSE). This allows for accurate approximation in regions where physics residuals are small.
- A zero-mean pressure constraint ( $\mathcal{L}_{\bar{p}}$ ) is introduced to ensure pressure uniqueness in the absence of specific pressure boundary conditions.
Data-Driven Loss ( $\mathcal{L}_{train}$ ): This term minimizes the discrepancy between the network output and available experimental measurements (velocity).
- The loss includes terms for velocity magnitude and, crucially, a term for the velocity direction unit vector ( $\hat{u}$ ). This enhances the reconstruction of flow direction even when velocity magnitudes are low, avoiding issues with zero-magnitude vectors.
- Notably, the network is trained without pressure data; pressure is reconstructed solely through the momentum balance equation.

2.2 Network Architecture and Training

Architecture: Fully connected neural networks with three hidden layers, hyperbolic tangent activation functions, and batch normalization to mitigate vanishing/exploding gradients.
Optimization: A two-stage training process is used: initial optimization with the ADAM optimizer followed by fine-tuning with the L-BFGS-B quasi-Newton algorithm.
Input/Output: Inputs are spatial coordinates ( $x, y, z$ ); outputs are velocity components and pressure.

3. Key Contributions and Experimental Validation

The authors validate their framework on two distinct test cases: the FDA nozzle benchmark and a patient-derived aneurysm model.

3.1 FDA Nozzle Benchmark

This case utilizes both in silico (CFD/FEM) data and experimental PIV data.

In Silico Verification: The authors performed a sensitivity analysis on seven different PINN configurations (FDA-CFD-1 to 7) to optimize hyperparameters.
- Findings: Physically consistent scaling of loss weights (ensuring residuals are of similar magnitude) yielded more robust results. Including MSLE improved pressure gradient reconstruction in low-gradient regions. Increasing network size (neurons) improved accuracy but required careful balancing to avoid overfitting numerical noise.
- Result: The PINN successfully reconstructed velocity and pressure fields from sparse cross-sectional data, outperforming standard CFD in matching specific experimental constraints when trained on PIV data.
Experimental Validation: Using PIV data from the FDA benchmark, the PINN was trained on a subset of cross-sections and tested on removed sections.
- Findings: The PINN compensated for data sparsity and asymmetry in the experimental measurements. It achieved lower errors compared to pure CFD simulations when validated against the experimental dataset, demonstrating the benefit of physics-informed regularization in noisy, sparse environments.

3.2 Aneurysm Model

This case focuses on blood flow in a realistic, patient-derived aneurysm geometry using in vitro 4D flow MRI data.

Setup: The model uses pulsatile flow data. The PINN was trained on low-resolution (1.5 mm³) MRI data to reconstruct high-resolution fields.
Findings:
- Velocity and Vorticity: The PINN successfully recovered main flow features and identified vorticity paths not visible in the raw, low-resolution training data, showing better agreement with high-resolution (1.0 mm³) MRI data than pure data-driven approaches.
- Pressure: The framework enabled non-invasive estimation of the pressure distribution, a quantity not directly measured in 4D flow MRI.
- Wall Shear Stress (WSS): While the physics-based regularization improved overall fidelity, WSS estimation remained challenging due to high sensitivity to boundary conditions and data dispersion. However, the PINN provided a more coherent distribution than raw data interpolation.
- Integral Quantities: The method successfully estimated kinetic energy and enstrophy over a heartbeat, with specific configurations (e.g., AN-4DMRI-2) providing a good compromise between data fidelity and physical consistency.

4. Significance and Claims

The paper claims that the proposed PINN framework offers a computationally efficient and accurate approach to enhancing 4D flow MRI post-processing. Its primary significance lies in:

Data Fusion: Successfully integrating sparse, noisy, or under-resolved experimental data with the incompressible Navier–Stokes equations.
Superior Reconstruction: Achieving spatially resolved solutions that align more closely with ground truth (in silico) and experimental observations than standard CFD or pure data-driven methods, particularly in regions with complex flow or near boundaries.
Derivation of Unmeasured Quantities: Enabling the estimation of clinically relevant but difficult-to-measure quantities, specifically pressure and Wall Shear Stress, without requiring invasive measurements.
Mesh-Free Advantage: Avoiding the laborious steps of geometry segmentation and mesh generation required by traditional CFD, as PINNs utilize a mesh-free formulation.

The authors conclude that while challenges remain regarding hyperparameter selection and the handling of highly turbulent or transient flows, PINNs represent a valuable tool for improving the fidelity of hemodynamic indicators derived from emerging imaging technologies like 4D flow MRI.