⚡ electrical engineering

FPINN: A proper orthogonal decomposition-assisted fourier-gradient physics-informed neural network for navier-stokes equations and flow reconstruction

This paper proposes a novel POD-assisted Fourier-gradient Physics-Informed Neural Network (FPINN) framework that effectively mitigates spectral bias, mass leakage, and high computational costs to achieve accurate and efficient reconstruction of Navier-Stokes flow fields across various Reynolds numbers.

Original authors: Qiancheng Tan, Xiaodi Wu, Yi Zhang, Yonghui Qin

Published 2026-07-28

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Qiancheng Tan, Xiaodi Wu, Yi Zhang, Yonghui Qin

Original paper licensed under CC BY 4.0 (https://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 trying to predict how a river flows around a rock, or how wind swirls around a skyscraper. For decades, scientists have used powerful computers to solve the "Navier-Stokes equations," a set of complex math rules that describe how fluids like water and air move. Think of these equations as the ultimate rulebook for fluid motion. However, solving them is like trying to count every single grain of sand on a beach while the tide is coming in; it's incredibly slow, expensive, and requires massive amounts of computer power.

Enter a newer, flashier tool called Physics-Informed Neural Networks, or PINNs. If traditional computers are like a meticulous accountant checking every single number, PINNs are like a super-smart student who learns the rules of the game so well that they can guess the answer without checking every single step. But even this smart student has trouble. They sometimes get "spectral bias," which is a fancy way of saying they are great at seeing the big, slow waves but terrible at spotting the tiny, fast ripples and sharp turns in the flow. They also sometimes "leak" mass, meaning they might predict that water appears out of nowhere or disappears, which breaks the laws of physics.

This paper introduces a new, upgraded student named FPINN. It's a "Fourier-gradient Physics-Informed Neural Network" that has been given a special pair of glasses and a memory aid. The glasses are "Fourier features," which help the student see those tiny, fast ripples clearly. The memory aid is "Proper Orthogonal Decomposition" (POD), a technique that helps the student ignore the boring, repetitive parts of the flow and focus only on the most important energy patterns. The researchers tested this new method on some classic fluid puzzles, like water flowing past a spinning cylinder and air swirling in a box with a moving lid. They found that FPINN is much better at predicting these flows, especially when the water is moving fast and chaotically, keeping the physics rules tight and the predictions accurate.

The Problem: The "Blurry Glasses" and the "Leaky Bucket"

For a long time, scientists have relied on traditional methods to simulate fluid flow. These methods are like building a giant grid of tiny boxes over a river and calculating the water's speed in every single box. While accurate, it takes forever and requires supercomputers. In recent years, a new approach called Physics-Informed Neural Networks (PINNs) arrived. Instead of a grid, PINNs use a neural network—a type of artificial intelligence—to learn the flow. It's like teaching a robot to understand the river by showing it the rules of physics rather than forcing it to count every drop.

However, standard PINNs have two big flaws. First, they suffer from "spectral bias." Imagine trying to draw a picture of a storm. A standard PINN might draw the big, dark clouds perfectly but miss the tiny, fast raindrops and the sharp edges of the lightning. It struggles with high-frequency details. Second, they sometimes violate the law of conservation of mass. In a real river, water doesn't just vanish or appear; it flows from one place to another. Standard PINNs sometimes "leak," predicting that water disappears in one spot and reappears in another, which makes the simulation physically impossible.

The Solution: A Team of Specialists

The authors of this paper, Qiancheng Tan and their team, decided to build a better version of this AI. They created a framework they call FPINN (Fourier-gradient Physics-Informed Neural Network), but they didn't stop there. They gave it a powerful assistant called POD (Proper Orthogonal Decomposition).

Think of the fluid flow as a chaotic dance party. There are thousands of people (fluid particles) moving in complex patterns.

POD is like a dance instructor who watches the party and says, "Forget about the individual twirls of every single person. Let's just focus on the three main moves that everyone is doing." It strips away the noise and keeps only the "dominant energy modes"—the most important patterns of movement. This makes the problem much simpler for the AI to solve.
Fourier Feature Mapping is like giving the AI "high-definition glasses." Standard AI often sees the world in low resolution, missing the sharp details. By feeding the AI data through a special mathematical filter (Fourier features), it can suddenly see the tiny, fast ripples and sharp gradients that it used to miss.
Gradient Penalties act as a strict referee. If the AI starts to predict that water is leaking or appearing out of thin air, the referee blows the whistle and forces the AI to correct its mistake immediately, ensuring the laws of physics are never broken.

The Training: Learning from Mistakes

Training this new AI wasn't just about feeding it data; it was about teaching it how to learn. The researchers used a clever strategy called SoftAdapt. Imagine you are studying for a test. You have three subjects: Math, History, and Science. If you are getting terrible grades in Math but great grades in History, your teacher (the AI's training system) should focus more time on Math.

In the FPINN system, the "loss" is a score of how wrong the AI is. If the AI makes a big mistake in predicting the pressure, the SoftAdapt system notices that the "Math score" is low. It automatically increases the weight of the pressure rules, telling the AI, "Hey, pay more attention to this part!" This dynamic balancing act prevents the AI from getting stuck or ignoring difficult parts of the problem.

They also used a sampling strategy called Latin Hypercube Sampling with a twist. Instead of just picking random points in the fluid to check, they made sure to pick extra points right near the walls and boundaries where the flow changes most drastically. It's like a detective who knows that the most important clues are usually found right next to the crime scene, so they check that area extra carefully.

The Results: From Calm Ponds to Turbulent Storms

The team tested their new FPINN on three famous fluid problems to see if it actually worked.

1. The Kovasznay Flow (The Calm Test)
First, they tested a flow with a known, exact solution (like a math problem with an answer key). They ran simulations at different speeds (Reynolds numbers of 20, 40, and 100). The results were impressive. The FPINN model predicted the flow with errors so small they were almost invisible (around $10^{-7}$ ). When compared to the old standard PINN, the new model was vastly superior, especially as the flow got faster. It proved that the "high-definition glasses" (Fourier features) helped it see the details clearly.

2. Flow Past a Cylinder (The Vortex Test)
Next, they simulated water flowing past a spinning cylinder. This creates a "wake" behind the cylinder, where the water swirls into complex vortices.

At low speed (Re = 100): The FPINN was incredibly accurate, matching the reference data almost perfectly over time.
At high speed (Re = 3900): This is where things get tricky. The flow becomes turbulent and chaotic. The paper notes that while the FPINN started with high accuracy, the errors began to grow as time went on, especially in the wake region where the vortices are strongest. The error for pressure, for instance, jumped from a tiny number ( $4.848 \times 10^{-2}$ ) to a much larger one ( $1.314 \times 10^{0}$ ) after a few seconds. This suggests that while the model is a huge improvement, predicting highly chaotic, high-speed flows for long periods is still a challenge. However, it still performed better than the traditional PINN.

3. The Lid-Driven Cavity (The Box Test)
Finally, they looked at a box of fluid where the top lid slides back and forth, dragging the fluid with it. They tested this at speeds ranging from a slow crawl (Re = 100) to a very fast slide (Re = 5000).

The FPINN successfully predicted the flow patterns at all these speeds.
Even at the highest speed (Re = 5000), where the flow is very turbulent, the FPINN kept the errors much lower than the standard PINN. For example, at Re = 5000, the error for the vertical velocity was $1.018 \times 10^{-2}$ for FPINN, compared to $1.883 \times 10^{-1}$ for the old PINN. This shows the model can handle high-speed, complex flows much better than before.

The Verdict: A Strong Step Forward

The paper concludes that combining POD (to simplify the problem) with FPINN (to see the details and obey physics) creates a powerful tool for fluid dynamics. It is faster to train, more accurate, and physically consistent.

However, the authors are honest about the limits. While the model is excellent, it isn't perfect yet. In the most chaotic, high-speed scenarios (like the cylinder wake at Re = 3900), small errors still tend to pile up over time. The model also relies heavily on the quality of the initial data used to create the POD "dance moves." If the initial snapshot is bad, the whole prediction suffers.

Despite these minor limitations, the FPINN framework represents a significant leap forward. It offers a way to simulate complex fluid flows without needing massive grids or supercomputers, making it a promising tool for future engineering applications, from designing better airplanes to understanding ocean currents. The researchers suggest that this approach could be the key to unlocking efficient, high-speed fluid simulations that were previously too difficult to compute.

Technical Summary: FPINN for Navier-Stokes Equations and Flow Reconstruction

Problem Statement
While Physics-Informed Neural Networks (PINNs) offer a mesh-free approach for solving the Navier-Stokes (N-S) equations, traditional models face significant hurdles in practical fluid dynamics applications. These include "spectral bias" (the inability to capture high-frequency, multi-scale structures), mass leakage (violation of the continuity equation), and prohibitive computational costs, particularly for high-Reynolds-number (Re) flows and inverse problems. Furthermore, integrating sparse data into existing algorithms and handling complex boundary conditions often require computationally expensive mesh generation or entirely new formulations.

Methodology
The paper proposes a novel framework termed FPINN (Fourier-gradient Physics-Informed Neural Network), which integrates Proper Orthogonal Decomposition (POD) with advanced neural network techniques to address the aforementioned limitations. The methodology consists of three core components:

POD-Assisted Dimensionality Reduction:
- The framework employs POD to extract dominant flow structures (core energy modes) from high-dimensional flow features.
- Specifically, Singular Value Decomposition (SVD) is applied to the initial state (t=0) to filter out small-scale spatial fluctuations. The reconstructed initial field serves as the target for the initial condition loss, ensuring a smooth starting point without sacrificing the network's capacity to learn complex dynamics over time.
- This reduces computational complexity by focusing the network on essential flow modes.
Fourier-Gradient Feature Mapping:
- To overcome spectral bias, the input spatiotemporal coordinates are transformed via Fourier feature mapping. This maps normalized coordinates into a higher-dimensional space using a fixed Gaussian random matrix, enabling the Multi-Layer Perceptron (MLP) to capture high-frequency, multi-scale flow dynamics.
- The network architecture utilizes a composite loss function that includes terms for the N-S equations, boundary conditions (BC), initial conditions (IC), and specific gradient penalties.
- Mass Conservation: Selective spatial gradient penalties are enforced to strictly ensure mass conservation ( $\nabla \cdot u = 0$ ).
Advanced Optimization Strategies:
- Boundary-Enriched Sampling: The training set is initialized using a boundary-enriched Latin Hypercube Sampling (LHS) strategy, allocating additional points to narrow boundary layers to capture steep gradients.
- Residual-Based Dynamic Adaptive Sampling: A discovery phase periodically evaluates PDE residuals on a candidate pool, dynamically adding collocation points to high-error regions to refine the solution.
- SoftAdapt Dynamic Weighting: A dynamic weighting mechanism (SoftAdapt) monitors the decay rates of individual loss components. It adaptively increases the weights of terms that decay slowly (or "spike"), preventing training stagnation and ensuring balanced convergence across multi-physics constraints.

Key Contributions

Physics-Informed Modeling: The integration of POD for dominant structure extraction, Fourier mapping for spectral bias mitigation, and strict spatial gradient constraints for mass conservation.
Optimization Robustness: The combination of boundary-enriched sampling and SoftAdapt dynamic loss weighting to handle complex multi-physics scenarios and sharp gradients effectively.
Engineering Application: A reduced-order framework capable of efficient, high-Reynolds-number unsteady flow predictions, bridging machine learning innovations with industrial fluid dynamics challenges.

Experimental Results
The framework was validated on three benchmark problems using data generated by COMSOL 6.4 or analytical solutions:

Kovasznay Flow:
- Tested at Re = 20, 40, and 100.
- FPINN achieved significantly lower errors compared to standard PINN and Fourier-feature PINN. For instance, at Re=40, the maximum pointwise absolute errors for velocity and pressure were tightly bounded within $10^{-3}$ .
- The model demonstrated superior stability even as Re scaled up to 100.
Flow Past a Circular Cylinder:
- Low Re (Re=100): FPINN exhibited high accuracy and long-term stability from t=0 to t=9.9, significantly outperforming traditional PINN in error metrics ( $\bar{E}_2$ , $R$ , $\bar{E}$ ).
- High Re (Re=3900): The model showed extremely high accuracy at the initial moment (t=0). However, as time progressed (t=2.5 to 4.95), the complex vortex structures in the wake region triggered significant error accumulation, with pressure errors increasing by orders of magnitude. This highlights a limitation in resolving highly nonlinear, high-Re wake dynamics over extended timeframes.
Lid-Driven Cavity Flow:
- Tested at Re = 100, 400, 1000, and 5000.
- FPINN consistently outperformed standard PINN across all Reynolds numbers in both steady-state and unsteady scenarios (t=1 to t=6).
- At Re=5000, FPINN maintained lower relative L2 errors compared to PINN, demonstrating robustness in high-Re regimes where standard models struggle.

Significance and Claims
The paper claims that the proposed POD-assisted FPINN framework establishes a highly flexible, mesh-free, reduced-order solver with exceptional robustness and computational efficiency. By coupling transfer learning across Reynolds numbers and enforcing physical consistency, the approach provides an efficient tool for flow field reconstruction.

However, the authors modestly acknowledge specific limitations:

Error Accumulation: Minor error accumulation is still observed when resolving highly nonlinear, high-Re wake dynamics over time.
Data Dependency: The performance of the reduced-order model inherently depends on the quality of the initial snapshot data used for POD basis extraction.

Overall, the work advances the implementation of deep learning in fluid dynamics by offering a physically consistent framework that minimizes prediction errors and achieves high accuracy in velocity and pressure field reconstructions, particularly in wake regions, while maintaining strict adherence to mass conservation laws.