Flow field tomography with uncertainty quantification using a Bayesian physics-informed neural network

This paper presents a Bayesian physics-informed neural network approach for flow field tomography that leverages governing fluid dynamics equations to reconstruct 2D flows from sparse line-of-sight measurements while providing comprehensive uncertainty quantification to address semi-convergence issues caused by noise.

The Gist

Imagine you are trying to figure out what a complex, swirling storm looks like inside a sealed, opaque box. You can't open the box, but you have a flashlight. You shine the light through the box from many different angles, and on the other side, you see a shadow or a blur of light intensity.

This is the challenge of Flow Field Tomography. Scientists want to see the invisible currents of air, fire, or liquid inside engines, furnaces, or the atmosphere. They take "Line-of-Sight" measurements (like your flashlight beams), but these measurements are just 1D slices. Reconstructing the full 3D picture from these slices is like trying to guess the shape of a hidden object just by looking at its shadows from a few angles. It's a math puzzle with infinite wrong answers.

Here is how this paper solves that puzzle, explained simply:

1. The Old Way: Guessing and Smoothing

Traditionally, scientists used computer algorithms to guess the shape of the flow.

• The Problem: Because there are so many possible shapes that could create the same shadows, the computers would get confused. They would either produce a blurry mess or start "hallucinating" details that weren't there, especially if the data was a little noisy (like a shaky flashlight).
• The Fix (Old): They would add "rules" to force the image to look smooth. But these rules were often too generic, like saying "everything must be round," which doesn't help if the flow is actually jagged or turbulent.

2. The New Tool: The "Physics-Savvy" AI

The authors introduce a new type of Artificial Intelligence called a Physics-Informed Neural Network (PINN).

Think of a standard AI as a student who has only seen pictures of cats and dogs. If you show it a picture of a tiger, it might guess "cat" because it looks similar.

• The PINN Student: This student has also studied the laws of nature. They know that air can't just teleport from one side of the room to the other, and that heat moves in specific ways.
• How it works: Instead of just trying to match the shadows (the data), this AI is forced to obey the Navier-Stokes equations (the fundamental laws of how fluids move). It's like telling the AI: "You can guess the shape, but if your guess breaks the laws of physics, you get a failing grade."

3. The Big Leap: Direct vs. Post-Processing

Before this paper, people used AI to fix blurry images made by old methods.

• The Analogy: Imagine a photographer takes a blurry photo of a car. They use an AI app to sharpen it. The AI is limited by how bad the original photo was.
• This Paper's Method: The authors built the camera inside the AI. They didn't take a blurry photo first and then fix it. Instead, they taught the AI to look at the raw shadows and immediately build the 3D car, ensuring the car obeys the laws of aerodynamics as it's being built.
• The Result: This "Direct Reconstruction" was vastly superior. Even with very few flashlight beams (sparse data) or shaky hands (noisy data), the AI produced a crystal-clear picture of the flow.

4. The Trap: The "Semi-Convergence" Problem

Here is a tricky part. When the data is noisy (shaky flashlight), the AI starts to learn the noise instead of the real flow.

• The Analogy: Imagine you are trying to learn a song, but there is a lot of static on the radio. At first, you learn the melody correctly. But if you keep listening too long, you start memorizing the static as part of the song. The more you listen, the worse the song sounds.
• The Solution: The authors realized the AI goes through specific "phases" of learning. They found a precise moment to stop the training right before the AI starts memorizing the static. They created a "stop sign" based on how well the AI is obeying the laws of physics. If the physics starts to break, you stop immediately.

5. The Bayesian Twist: Knowing What You Don't Know

The final and most brilliant part of the paper is the Bayesian approach.

• The Problem: Standard AI gives you one answer. "The wind speed is 50 mph." But what if the AI is wrong? How sure is it?
• The Bayesian Solution: Instead of giving one answer, the Bayesian AI (B-PINN) gives you a cloud of possibilities. It says, "I think the wind is 50 mph, but it could be anywhere between 48 and 52, and here is a map showing exactly where I am most unsure."
• Why it matters: In science, knowing how uncertain you are is just as important as the answer itself. This allows scientists to say, "We are 95% confident in this part of the flow, but we need more data for that corner."

Summary

This paper is about teaching a super-smart AI to solve a 3D puzzle using two superpowers:

1. Physics: It refuses to guess anything that breaks the laws of nature.
2. Uncertainty: Instead of just giving a single answer, it tells you how confident it is in that answer.

By combining these, the authors can see inside complex flows (like jet engines or flames) with much higher accuracy and reliability than ever before, even when the data they have is messy or incomplete. It's like upgrading from a blurry, guess-and-check flashlight to a high-tech, self-correcting X-ray vision that knows exactly where it's guessing.

Technical Summary

1. Problem Statement

Flow field tomography aims to reconstruct 2D or 3D fluid flow structures (velocity, pressure, concentration) from a limited set of line-of-sight (LoS) integrated measurements (projections). This is an ill-posed inverse problem characterized by:

• Rank Deficiency: There are fewer measurements than unknowns, leading to an infinite set of solutions that satisfy the projection data.
• Noise Amplification: Standard reconstruction algorithms (e.g., ART, SIRT, Tikhonov) amplify measurement noise and discretization errors.
• Semi-Convergence: Iterative algorithms often exhibit "semi-convergence," where early iterations capture low-frequency, robust features, but later iterations fit high-frequency noise, degrading the solution.
• Limitations of Current Methods:
  • Regularization: Traditional methods use penalty terms (e.g., smoothness) that are often incompatible with actual flow physics.
  • Deep Learning Post-Processing: Existing Physics-Informed Neural Network (PINN) approaches often rely on "post-processing" error-laden reconstructions from conventional algorithms. This limits the PINN's ability to correct fundamental reconstruction errors because it treats the flawed initial estimate as ground truth.
  • Lack of Uncertainty Quantification (UQ): Most methods provide a single point estimate without quantifying the confidence or uncertainty associated with the reconstruction, which is critical for scientific validation.

2. Methodology

The authors propose a framework using Physics-Informed Neural Networks (PINNs) and Bayesian PINNs (B-PINNs) to directly reconstruct flow fields from raw projection data.

A. Direct Reconstruction Architecture

Instead of post-processing, the PINN is trained to map spatio-temporal coordinates $(x, y, t)$ directly to flow variables $(c, u, v, p)$ (concentration, velocity components, pressure).

• Loss Function: The total loss is a weighted sum of a measurement loss and a physics loss.
  • Measurement Loss ($L_{meas}$): Instead of comparing the network output to a pre-reconstructed field, the network output is projected through the forward model (matrix $A$) and compared directly to the raw experimental projections ($B$).
    $$L_{meas} = \|AC_{out} - B\|_F^2$$
  • Physics Loss ($L_{phys}$): The network outputs are differentiated (via Automatic Differentiation) and plugged into the governing equations (incompressible Navier-Stokes and advection-diffusion equations). The residuals are minimized.
    $$L_{phys} = \|E\|_F^2$$
• Training Strategy: The network minimizes $L_{total} = \gamma L_{meas} + L_{phys}$. The authors identify three distinct training phases:
  1. Measurement Dominated: The network fits the projection data quickly.
  2. Hybrid: Physics residuals stabilize while velocity fields begin to improve.
  3. Fully-Coupled: All fields and losses are simultaneously minimized.

B. Handling Semi-Convergence (Noise)

When training on noisy data, the PINN exhibits semi-convergence similar to iterative solvers. To address this without heuristic stopping:

• Stopping Criterion: The authors utilize the transition between training phases. They stop training when the continuity residual (mass conservation error) reaches 95% of its maximum value. This point correlates with the minimum reconstruction error before the network begins overfitting to noise.

C. Bayesian Inference (B-PINN)

To address uncertainty and the limitations of point estimates:

• Framework: The network parameters $\theta$ are treated as random variables. The goal is to sample the posterior distribution $f(\theta|B)$ rather than finding a single Maximum A Posteriori (MAP) estimate.
• Likelihood & Prior:
  • Likelihood: Modeled based on Gaussian measurement noise.
  • Prior: Encodes the physics residuals (Navier-Stokes) as a Gaussian distribution.
• Sampling: The Hamiltonian Monte Carlo (HMC) method is used to sample the posterior distribution of the network parameters. This generates a distribution of flow fields rather than a single deterministic output.
• Output: The result is a Conditional Mean (CM) estimate and Credible Intervals (CI) (e.g., 95% CI) that quantify uncertainty.

3. Key Contributions

1. Direct Reconstruction via PINN: Demonstrated that embedding the projection model directly into the PINN loss function yields significantly superior results compared to post-processing conventional reconstructions. This avoids propagating initial reconstruction errors.
2. Identification of Training Regimes: Characterized the specific phases of PINN training for tomography, revealing that noise causes semi-convergence at the transition from the "measurement-dominated" to the "fully-coupled" phase.
3. Robust Stopping Criterion: Developed a physics-based stopping criterion (monitoring continuity residuals) that effectively regularizes noisy data without manual tuning.
4. Bayesian UQ Implementation: Successfully applied B-PINNs to flow tomography, enabling the quantification of uncertainty in reconstructed fields.
5. Superiority of CM over MAP: Showed that for noisy data, the Bayesian Conditional Mean (CM) estimate is more robust and physically consistent than the MAP estimate (standard PINN output), allowing for the use of stricter physics priors that would otherwise cause standard PINNs to fail.

4. Results

The methods were tested on a 2D flow field (cylinder wake, $Re=100, Pe=100$) using synthetic data from Direct Numerical Simulation (DNS).

• Accuracy:
  • Direct vs. Post-Processing: Direct reconstruction reduced concentration errors by ~99% compared to post-processing conventional Bayesian reconstructions.
  • Sparse Data: Direct reconstruction using only 20 beams (noisy) produced better results than post-processing using 100 beams (clean).
  • Noise Robustness: The proposed stopping criterion successfully halted training at the point of minimum error, preventing overfitting to 5% Gaussian noise.
• Uncertainty Quantification:
  • B-PINN reconstructions provided credible intervals that correctly highlighted regions of high uncertainty (e.g., gaps between beam paths and domain edges).
  • In noisy scenarios, the B-PINN Conditional Mean (CM) remained close to the ground truth, whereas the standard PINN (MAP) and conventional methods produced significant artifacts.
• Grid Resolution: Increasing grid resolution significantly improved direct reconstruction accuracy in noise-free cases but had diminishing returns in high-noise scenarios, confirming that noise intensity, not grid density, is the limiting factor for high-noise data.

5. Significance

This work represents a paradigm shift in flow field tomography:

• From Post-Processing to Direct Inference: It proves that PINNs can solve the inverse problem directly from raw measurements, bypassing the error-prone intermediate steps of traditional algorithms.
• Physics-Driven Regularization: By using the governing equations as the primary regularizer, the method produces physically plausible estimates even with sparse and noisy data.
• Scientific Rigor: The integration of Bayesian inference provides a necessary layer of Uncertainty Quantification, allowing researchers to distinguish between reliable features and artifacts in reconstructed flow fields. This is crucial for validating computational fluid dynamics (CFD) codes and understanding complex fluid phenomena where data is inherently limited.
• Scalability: The approach is applicable to various tomographic modalities (PIV, absorption spectroscopy, etc.) and offers a robust framework for handling the ill-posed nature of inverse problems in fluid diagnostics.