Neural inference of fluid-structure interactions from sparse off-body measurements

This paper presents a novel physics-informed neural framework that accurately reconstructs unsteady fluid-structure interactions and structural deformations from sparse, off-body flow measurements without requiring a solid constitutive model or direct surface position data.

The Gist

Imagine you are trying to figure out how a flexible underwater robot is swimming, but you can't see the robot itself. You can only see a few tiny, glowing specks floating in the water around it. The robot is moving, bending, and twisting, pushing the water around, but the camera is too far away or the water is too murky to see the robot's body directly.

This paper presents a clever new "detective tool" that solves this mystery. It uses Artificial Intelligence (AI) to look at those few floating specks and mathematically "fill in the blanks" to reconstruct exactly what the robot is doing and how the water is flowing, even though the data is sparse and noisy.

Here is how it works, broken down into simple concepts:

1. The Problem: The "Blind Spot"

In the real world, studying how fluids (like water or air) interact with moving objects (like fish, wind turbine blades, or heart valves) is hard.

• The Simulation Problem: Computer simulations are great, but they need perfect information about the object's material and the starting conditions. If you guess wrong about the material, the whole simulation is wrong.
• The Experiment Problem: Real experiments give you real data, but you can't measure everything. You might only see the water moving (using particle tracking), but you can't see the object's surface or know exactly how stiff it is.
• The Gap: We have a "rich" simulation that might be wrong, and "sparse" real data that is incomplete. We need a way to combine them.

2. The Solution: The "Physics-Savvy Detective"

The authors created a system called a Physics-Informed Neural Network (PINN). Think of this AI not as a magic black box, but as a detective who knows the laws of physics by heart.

• The Detective's Notebook (The Neural Network): The AI has two main "notebooks." One describes the water (velocity and pressure), and the other describes the object (how it bends and moves).
• The Rules (Physics Laws): The detective is strictly forbidden from writing down anything that breaks the laws of physics. For example, water can't just disappear, and the object can't pass through the water. The AI is trained to minimize "mistakes" in these laws.
• The Clues (Particle Tracks): The only hard evidence the detective has are the paths of the floating specks (particles). The AI uses these paths to figure out where the water is going.

3. How It Solves the Mystery

The AI works like a puzzle solver that adjusts two things simultaneously until everything fits:

1. The Water Flow: It guesses the water speed and pressure everywhere.
2. The Object's Shape: It guesses how the object is bending.

It checks its work against the "Rules" (Physics) and the "Clues" (Particle Tracks).

• If the AI guesses the object is bending one way, but the water particles are moving in a way that physics says is impossible for that bend, the AI changes its guess.
• If the AI guesses the water flow is smooth, but the particles show a sudden jump, the AI adjusts the flow.

Eventually, the AI finds the only combination of water flow and object movement that satisfies the laws of physics and matches the few particle tracks we actually saw.

4. The "Flexible Shape" Trick

One of the hardest parts is that the object changes shape. To handle this, the AI doesn't try to guess every single pixel of the object's surface (which would be too hard). Instead, it uses a "Lego Block" approach.

Imagine the object is made of a few standard "bending modes" (like a flexible ruler that can only bend in specific ways). The AI just has to figure out how much to bend each of those modes. This makes the problem much easier to solve, even with very little data.

5. Real-World Tests

The authors tested their detective on three scenarios:

• A Flapping Plate: A flexible sheet attached to a cylinder in a wind tunnel. The AI successfully figured out how the sheet was flapping just by watching the wind swirl around it.
• A Flexible Pipe: A pipe that expands and contracts like a blood vessel. The AI reconstructed the pulse wave moving through the pipe and the pipe's expansion, even though the data near the walls was very sparse.
• A Swimming Fish: A 3D fish swimming through water. The AI reconstructed the fish's undulating tail and the complex wake (swirls) it left behind, despite the fish's body being invisible to the "camera."

6. Why This Matters

This is a game-changer because:

• It works with bad data: Even if the particle tracks are noisy or missing pieces, the AI uses physics to "clean up" the data and fill in the gaps.
• It doesn't need to know the material: You don't need to know exactly how stiff the fish or pipe is beforehand. The AI figures out the motion directly from the water's reaction.
• It saves money: You don't need expensive, complex cameras to see the object itself. You just need to track the fluid around it.

The Bottom Line

Think of this technology as Sherlock Holmes for fluid dynamics. You give it a few blurry footprints (particle tracks) and the rulebook of physics, and it deduces the entire story: exactly how the suspect (the structure) moved and how the environment (the fluid) reacted, even if you never saw the suspect directly.

Technical Summary

1. Problem Statement

Fluid–Structure Interactions (FSI) are critical in applications ranging from aerospace to cardiovascular systems. Traditional computational methods (e.g., ALE, Immersed Boundary) require precise boundary conditions and material models, which are often unknown in real-world scenarios. Conversely, experimental measurements are often sparse, noisy, and limited to a single phase (typically the fluid), making it difficult to reconstruct the coupled dynamics of both the flow and the deforming structure.

Existing data assimilation (DA) methods struggle with:

• Single-phase limitations: Most algorithms require simultaneous measurements of both fluid and solid phases, which is costly or infeasible (e.g., inside blood vessels).
• Inverse problem complexity: Reconstructing evolving interfaces from incomplete data is ill-posed.
• Computational cost: Grid-based inverse solvers (like ODIL) often struggle with non-convex shapes, unsteady 3D domains, and high computational costs.

The authors propose a solution to reconstruct unsteady FSI fields (velocity, pressure, and structural deformation) using only sparse, off-body Lagrangian Particle Tracking (LPT) data of the fluid, without requiring direct structural measurements or constitutive models for the solid.

2. Methodology

The authors introduce a Physics-Informed Neural Network (PINN) framework tailored for FSI data assimilation. The core innovation is the joint training of three distinct neural models constrained by physical laws and kinematic data.

A. Neural Architecture

The framework utilizes three primary components:

1. Fluid Model (Flow PINN): A coordinate neural network mapping space-time coordinates $(x, t)$ to fluid velocity $\mathbf{u}$ and pressure $p$. It uses Fourier feature encoding to mitigate spectral bias and separate networks for velocity and pressure to handle different spectral characteristics.
2. Structure Model (Solid PINN): A neural network mapping time $t$ to modal coefficients $\beta(t)$. The structural surface is represented as a linear combination of a base surface and a set of deformation modes (eigenmodes or POD modes):
   $$\mathbf{p}(t) \approx \mathbf{p}_{base} + \Phi \beta(t)$$
   This low-dimensional representation regularizes the inverse problem, avoiding the need to infer infinite-dimensional surface shapes directly.
3. Particle Track Model: A kinematics-constrained model that maps time to particle velocity $\mathbf{v}^{(k)}$. Crucially, this model embeds the advection constraint ($\frac{d\mathbf{x}}{dt} = \mathbf{v}$) as a hard constraint using the theory of functional connections. This ensures that the predicted particle trajectories are mathematically consistent with their velocities.

B. Loss Function and Constraints

The models are trained by minimizing a composite loss function $\mathcal{J}_{total}$:
$$\mathcal{J}_{total} = \gamma_1 \mathcal{J}_{flow} + \gamma_2 \mathcal{J}_{surf} + \gamma_3 \mathcal{J}_{part} + \gamma_4 \mathcal{J}_{data}$$

• $\mathcal{J}_{flow}$ (Physics Loss): Weakly enforces the incompressible Navier-Stokes equations (continuity and momentum residuals) over the fluid domain.
• $\mathcal{J}_{surf}$ (Boundary Loss): Enforces the no-slip moving-wall condition at the fluid-solid interface by minimizing the mismatch between fluid velocity and the structural surface velocity derived from the Solid PINN.
• $\mathcal{J}_{part}$ (Particle Physics Loss): Couples the particle model to the flow field by minimizing the difference between the particle velocity and the fluid velocity at the particle's location ($\mathbf{v}^{(k)} - \mathbf{u}(\mathbf{x}^{(k)}, t)$).
• $\mathcal{J}_{data}$ (Data Fidelity): Penalizes deviations between estimated and measured particle positions. To handle noise, particle positions are treated as trainable variables, allowing the framework to refine trajectories based on physical consistency.

C. Sampling Strategy

Since the fluid domain and boundary evolve over time, the authors employ an adaptive Monte Carlo sampling strategy. The domain is partitioned into fixed far-field regions and dynamic near-field regions (meshed with triangles/tetrahedra). Samples are drawn proportional to the volume/area of these subregions to ensure uniform coverage of the transient space-time domain, which is critical for stable gradient estimation.

3. Key Contributions

1. Single-Phase FSI Reconstruction: The method successfully infers both flow fields and structural dynamics using only fluid-phase particle tracks, eliminating the need for simultaneous solid-phase measurements or known material properties.
2. Kinematics-Constrained Particle Models: By embedding the advection constraint directly into the particle model structure, the framework enforces physical consistency on the trajectory data, allowing for robust refinement of noisy measurements.
3. Modal Regularization: The use of a modal basis (physics-based or data-driven) for the structure regularizes the ill-posed inverse problem, enabling the inference of complex deformations from sparse data.
4. Robustness to Over-Parameterization: The study demonstrates that the reconstruction remains accurate even when the number of structural modes exceeds the necessary minimum, a crucial feature for experimental settings where the optimal basis is unknown.
5. Noise Resilience: The framework effectively handles localization errors by jointly optimizing particle positions and flow/structure states, significantly reducing errors in both trajectory and field reconstruction compared to raw data processing.

4. Results

The framework was validated on three canonical benchmarks using synthetic data generated from high-fidelity FSI simulations:

• 2D Flapping Plate (Vortex-Induced Vibration):
  • Successfully reconstructed vortex shedding and pressure fields.
  • Achieved ~15% NRMSE for vorticity and ~8% for pressure.
  • Accurately recovered structural mode coefficients, with the first two modes capturing >99% of deformation energy.
• 3D Flexible Pipe (Pulse-Wave Propagation):
  • Captured axial pressure wave propagation and transverse vortical structures.
  • Achieved ~14% error in transverse velocity and ~9% in pressure.
  • Surface deformation was recovered with <9% relative error, accurately tracking the traveling pressure bulge.
• 3D Swimming Fish (Undulatory Wake):
  • Reconstructed complex wake dynamics and caudal fin motion.
  • Achieved ~13% error in transverse velocity and ~17% in pressure (higher near the thin fin due to steep gradients).
  • Demonstrated that adding higher-order modes did not degrade performance, confirming robustness to over-specification.

Sensitivity Analysis:

• Localization Noise: Even with high noise levels (up to 2 pixels), the optimized tracks and flow fields remained stable. The framework reduced track velocity errors significantly compared to raw finite-difference estimates.
• Sampling Variance: The study highlighted that low-variance sampling (balancing spatial and temporal samples) is critical for training stability. High-variance sampling led to oscillatory loss convergence and poor reconstruction quality.

5. Significance

This work represents a significant advancement in inverse problems for fluid dynamics and experimental FSI analysis.

• Practical Application: It provides a pathway for quantitative FSI analysis in scenarios where structural measurements are asynchronous, unavailable, or impossible (e.g., internal blood flow, biological swimmers, or high-speed aerodynamics).
• Methodological Shift: It moves beyond traditional grid-based inverse solvers by leveraging the flexibility of neural networks and the regularization of modal bases to handle complex, moving boundaries.
• Data Efficiency: By utilizing sparse, off-body data and refining it through physics constraints, the method maximizes the utility of experimental data that would otherwise be considered insufficient for full-field reconstruction.

In summary, the paper presents a robust, physics-informed neural framework that bridges the gap between sparse experimental observations and the rich, coupled dynamics of fluid-structure systems, offering a powerful tool for both analysis and design in engineering and biology.