On the joint estimation of flow fields and particle properties from Lagrangian data

This paper presents a data assimilation framework that successfully demonstrates the joint estimation of carrier flow fields and unknown particle properties (such as position, size, and density) from sparse and noisy Lagrangian particle tracking data across turbulent, inertial, and compressible supersonic flow regimes.

The Gist

Imagine you are trying to figure out the wind patterns inside a giant, invisible tornado. You can't see the air itself, so you throw thousands of tiny, colorful balloons into the storm and film them with a high-speed camera. This is essentially what scientists do in Lagrangian Particle Tracking (LPT): they track particles to understand how the fluid (like air or water) is moving.

However, there are two big problems with this method:

1. The balloons aren't perfect: Sometimes the camera is blurry, or the balloons are heavy and get blown off course by their own momentum (inertia), so they don't follow the wind exactly.
2. The data is messy: You only see the balloons at specific points in time, and your camera might make small mistakes about exactly where they are.

Traditionally, scientists would try to guess the wind speed by just connecting the dots between where the balloons were. But if the balloons are heavy or the camera is shaky, this guess is often wrong.

The New Idea: "The Detective and the Physics"

This paper introduces a new method called NIPA (Neural-Implicit Particle Advection). Think of it as a super-smart detective who doesn't just look at the clues (the balloon positions) but also knows the laws of physics by heart.

Instead of just drawing lines between the dots, NIPA runs a simulation in its head. It asks: "If the wind were moving like this, and these balloons were this heavy, would they end up where the camera saw them?"

It does this by juggling two things at the same time:

1. The Flow Field: The invisible map of wind speed, pressure, and density.
2. The Particle Properties: How heavy the balloons are, how big they are, and how they react to the wind.

The system tweaks both the wind map and the balloon properties until they perfectly match the camera footage and obey the laws of physics.

The Three Test Cases (The "Exams")

The authors tested their detective on three different scenarios to see how well it works:

1. The "Ghost" Balloons (Ideal Tracers)

• The Scenario: Tiny, feather-light balloons that follow the wind perfectly, but the camera is shaky and blurry.
• The Result: The detective was amazing. Even with a blurry camera, it could figure out exactly where the balloons really were and reconstruct the wind map perfectly. It was like looking at a blurry photo of a car and perfectly reconstructing the road it drove on.

2. The "Heavy" Balloons (Inertial Particles)

• The Scenario: Heavier balloons that don't follow the wind perfectly; they lag behind or drift. Plus, the researcher doesn't know how heavy they are!
• The Result: This is the magic trick. The detective didn't just guess the wind; it figured out the weight and size of the balloons just by watching how they moved. It realized, "Ah, these balloons are lagging behind the wind, so they must be heavy," and used that clue to fix the wind map. It's like guessing a person's weight just by watching how fast they run on a treadmill.

3. The "Supersonic" Balloons (Shock Waves)

• The Scenario: A jet flying faster than sound, creating shock waves. The balloons are tiny specks of dust (Titanium Dioxide) that get smashed and pushed by the shock.
• The Result: This is the hardest test. The wind changes instantly, and the balloons struggle to keep up. The detective successfully reconstructed the shock waves and the air density, even though the balloons were behaving erratically. It proved that even in extreme conditions, you can learn about the air and the dust simultaneously.

Why Does This Matter?

Imagine you are trying to fix a broken engine.

• Old Way: You look at the smoke coming out (the particles) and guess how the engine is running. If the smoke is thick or the camera is bad, your guess is wrong.
• New Way (NIPA): You use a computer model that knows how engines work. You feed it the smoke data, and it says, "Based on this smoke pattern and the laws of thermodynamics, the engine is running at X speed, and the fuel mixture is Y."

The Key Takeaways

• Don't just look at the dots: By combining the raw data with the laws of physics, you can get a much clearer picture than by looking at the data alone.
• You can learn two things at once: You don't need to know the particle size beforehand. The system can figure out the wind and the particle size simultaneously.
• More data is better, but physics helps: If you have very few balloons, the system gets confused. But even with fewer balloons, the "physics knowledge" helps the system guess better than before.
• Noise is manageable: Even if your camera is shaky, the laws of physics act as a safety net, preventing the system from making wild, impossible guesses.

In short, this paper teaches computers to be better detectives. Instead of just staring at a messy crime scene, they use their knowledge of how the world works to reconstruct the truth, even when the evidence is imperfect.

Technical Summary

1. Problem Statement

Lagrangian Particle Tracking (LPT) is a powerful diagnostic tool for measuring volumetric, time-resolved flow fields. However, extracting accurate Eulerian flow fields (velocity, pressure, density) from LPT data faces two primary challenges:

1. Measurement Noise and Sparsity: Experimental tracks are spatially sparse and subject to localization errors (especially in single-camera systems), which degrade velocity estimates derived via finite differencing or filtering.
2. Inertial Particle Transport: In many practical flows, particles have finite response times (Stokes number $St > 0$), causing them to deviate from the carrier fluid velocity (slip). Traditional reconstruction methods often assume ideal tracers ($St \to 0$), leading to significant errors when applied to inertial particles. Furthermore, particle properties (size, density) are often unknown, making it impossible to correct for slip without prior knowledge.

The central question addressed is: Can we jointly estimate the carrier flow field and unknown particle properties (position, size, density) directly from noisy, sparse Lagrangian tracks by enforcing the governing equations of disperse multiphase flow?

2. Methodology: Neural-Implicit Particle Advection (NIPA)

The authors propose a novel data assimilation (DA) framework called NIPA, which couples an Eulerian flow model with Lagrangian particle models using Physics-Informed Neural Networks (PINNs).

Core Architecture

• Eulerian Flow Model: The flow field (velocity $\mathbf{u}$, pressure $p$, density $\rho$, etc.) is represented by coordinate neural networks. These networks take space-time coordinates $(\mathbf{x}, t)$ as input and output flow variables. They are trained to minimize residuals of the Navier-Stokes equations (carrier phase).
• Lagrangian Particle Models (KCT): Each particle $k$ is modeled by a dedicated Kinematics-Constrained Track (KCT) model.
  • Instead of treating velocity as the primary variable, the model treats the particle's position trajectory as a continuous function of time.
  • The advection equation ($\mathbf{x}(t) = \int \mathbf{v}(t) dt$) is embedded as a hard constraint within the model structure. This ensures that the estimated velocity and acceleration always integrate back to the particle positions, avoiding the noise amplification inherent in differentiating raw tracks.
  • Particle properties (e.g., diameter $d_p$, density $\rho_p$) are treated as trainable parameters within the particle model.
• Physics Constraints:
  • Carrier Phase: Residuals of the Navier-Stokes equations (incompressible or compressible).
  • Disperse Phase: Residuals of the Maxey-Riley equation (extended for compressible flows). For inertial particles, the slip velocity depends on $d_p$ and $\rho_p$.
• Objective Function: The total loss function ($J_{total}$) combines four terms:
  1. Data Fidelity ($J_{data}$): Penalizes the discrepancy between measured and estimated particle positions, weighted by localization uncertainty (Mahalanobis distance).
  2. Flow Physics ($J_{flow}$): Navier-Stokes residuals.
  3. Particle Physics ($J_{part}$): Maxey-Riley residuals.
  4. Boundary Conditions ($J_{bound}$): Enforces conditions like no-slip at walls.

By minimizing this composite loss, NIPA simultaneously optimizes the flow field, the particle trajectories, and the unknown particle properties.

3. Key Contributions

1. Joint Estimation Framework: Development of NIPA, the first framework to simultaneously infer carrier flow states and unknown particle properties (size, density) from Lagrangian data without assuming ideal tracer behavior.
2. Kinematics-Constrained Tracks (KCT): Introduction of a track model that embeds the advection equation as a hard constraint, allowing for the direct optimization of particle positions rather than velocities, thereby bypassing noise amplification.
3. Implicit Particle Characterization: Demonstration that particle properties can be "learned" from the flow dynamics and track data alone, without prior calibration or image-based sizing.
4. Physics-Informed Tracking: Showing that governing equations can correct noisy track geometries, effectively recovering true particle positions even when localization errors are significant.

4. Results

The method was validated across three distinct flow regimes using synthetic Direct Numerical Simulation (DNS) and CFD data:

Case 1: Turbulent Boundary Layer (TBL) with Noisy Tracers ($St \to 0$)

• Goal: Recover true particle positions and flow fields from noisy tracks.
• Findings: Joint estimation significantly outperformed traditional filtering (B-splines, TrackFit) and finite differencing.
  • At high noise levels, joint estimation reduced position errors by up to 60% compared to filtering.
  • It successfully recovered the underlying flow structures and true trajectories, demonstrating that physics constraints can "denoise" the tracks.

Case 2: Homogeneous Isotropic Turbulence (HIT) with Bidisperse Inertial Particles ($St \sim 1-5$)

• Goal: Recover flow fields and particle diameters for particles with unknown sizes.
• Findings:
  • Flow Recovery: Joint estimation accurately recovered turbulent kinetic energy (TKE) spectra and coherent structures, whereas flow-only methods (assuming $St=0$) failed to capture high-frequency dynamics and produced spurious structures.
  • Property Recovery: The model successfully inferred particle diameters, separating the bidisperse distribution with a Pearson correlation of 0.95 between true and estimated sizes.
  • Sensitivity: Reconstruction accuracy was highly dependent on seeding density. Dense seeding allowed for super-resolution (recovering scales beyond the Nyquist wavenumber), while sparse seeding led to ill-posed inversions and filtered flow fields.

Case 3: Supersonic Cone-Cylinder Flow ($M=2$)

• Goal: Jointly reconstruct compressible flow (velocity, pressure, density, temperature) and inertial particle properties in the presence of shock waves.
• Findings:
  • Joint estimation resolved shock structures and expansion fans with high fidelity, whereas flow-only methods smeared shocks and introduced artifacts.
  • NRMSEs: Joint estimation achieved low errors (e.g., 1% for axial velocity, 3% for density) compared to flow-only (e.g., 25% for radial velocity).
  • Particle Properties: While diameter ($d_p$) was inferred with high accuracy (correlation 0.8), density ($\rho_p$) was harder to constrain due to the quadratic dependence of response time on diameter. However, fixing $\rho_p$ improved $d_p$ inference further, highlighting the complementary nature of the two phases.

Sensitivity Analysis

• Seeding Density: Critical for well-posedness. Sparse seeding leads to ill-conditioned problems where particle properties cannot be uniquely identified.
• Noise & Stokes Number: Accuracy degrades with increasing noise and $St$. High $St$ particles amplify acceleration errors. However, joint estimation remains robust up to moderate $St$ levels where flow-only methods fail completely.

5. Significance and Implications

• Overcoming Experimental Limitations: NIPA enables the use of LPT in regimes where ideal tracers are unavailable (e.g., high-speed aerodynamics, sediment transport, atmospheric studies) by mathematically correcting for inertial effects.
• Data Efficiency: The framework can recover flow information at scales finer than the particle sampling limit (super-resolution) by leveraging physical constraints, reducing the need for extremely dense seeding.
• Unified Inverse Problem: It transforms the problem from a two-step process (track smoothing $\to$ flow reconstruction) into a single, coupled inverse problem, ensuring physical consistency between the measured data and the reconstructed fields.
• Future Applications: This approach paves the way for "smart" LPT experiments where particle properties are not just known inputs but are simultaneously inferred, allowing for more accurate characterization of complex multiphase flows in real-world applications.

In conclusion, the paper demonstrates that by treating Lagrangian data and governing physics as a unified system, it is possible to overcome the limitations of noise and particle inertia, achieving high-fidelity reconstruction of both flow fields and particle characteristics.