Open-source BOS tomography dataset of high-speed flow over a flight body

This paper introduces an open-source, 70-view background-oriented schlieren dataset of high-speed flow over a flight body and demonstrates its utility through neural-implicit reconstruction and data assimilation techniques that achieve high-fidelity shock resolution, 3D state estimation, and efficient uncertainty quantification.

The Gist

Imagine you are trying to figure out what a speeding car looks like while it's driving through a thick fog. You can't see the car directly, but you can see how the fog swirls and bends around it. If you take photos of that swirling fog from just one angle, you only get a flat, blurry picture. But if you have a whole team of photographers standing in a circle, snapping pictures from every side, you can piece together a perfect 3D hologram of the car, even though you never actually saw the car itself.

That is essentially what this paper is about, but instead of a car in fog, scientists are looking at airplanes (or models of them) flying at supersonic speeds through invisible air currents.

Here is the breakdown of their work using simple analogies:

1. The Problem: Seeing the Invisible

When an object flies faster than sound, it creates invisible shockwaves (like the sonic boom you hear, but visible as distortions in the air). Scientists use a technique called BOS (Background-Oriented Schlieren).

• The Analogy: Imagine holding a piece of paper with a wavy pattern on it behind a glass of water. If you drop a rock in the water, the water bends the light, making the wavy pattern look distorted. By studying how the pattern is distorted, you can figure out exactly what the water (or air) is doing, even though you can't see the water itself.

2. The Challenge: Too Few Angles

Usually, scientists only have a few cameras (maybe 5 to 15) looking at the object. It's like trying to solve a 3D puzzle with only a few pieces. You can guess the shape, but the edges get blurry, and you might see "ghosts" or fake shapes that aren't really there.

3. The Solution: A Massive Data Set

The authors created a public treasure chest of data.

• The Setup: They built a model of a futuristic flight vehicle (a weird, pyramid-shaped object) and put it in a wind tunnel.
• The Magic: Instead of just a few cameras, they used 70 different views. They did this by taking a single row of 7 cameras and rotating the model 10 times.
• The Result: They now have a massive library of 70 angles of the air bending around the model. This is like having a 360-degree, high-definition video of the invisible air, which is incredibly rare and valuable.

4. The Brain: AI and Math (NIRT)

Having 70 photos is great, but how do you turn them into a 3D movie? The authors used a special computer program called NIRT (Neural-Implicit Reconstruction Technique).

• The Analogy: Think of NIRT as a super-smart detective.
  • The Clues: The detective looks at the 70 photos of the distorted background pattern.
  • The Rules: The detective knows the laws of physics (like how air pressure and speed work).
  • The Deduction: The detective doesn't just guess; it uses math to fill in the blanks. It asks, "If the air is doing this here, and that there, what must the air be doing in the middle?"
  • The Bonus: They also used a technique called Data Assimilation. This is like the detective not just guessing the shape of the car, but also guessing how fast the wind is blowing and how hot the air is, even though the cameras can't measure temperature directly. They used the laws of physics to "fill in the missing pieces" of the puzzle.

5. The Results: Crystal Clear 3D

Because they had so many angles and a smart AI detective:

• Sharp Edges: They could see the sharp "shockwaves" (the sonic booms) very clearly, without the blurry smears you usually get.
• No Ghosts: They didn't see fake shapes (artifacts) that looked like the air was expanding where it shouldn't.
• New Discoveries: For the first time, they were able to estimate the temperature and speed of the air just by looking at the distorted background pattern. It's like being able to tell how hot a stove is just by looking at the heat waves rising from it, without touching it.

Why Does This Matter?

This paper is a gift to the scientific community.

1. Open Source: They are giving away all their photos, the model designs, and the computer code for free. Anyone can download it and try to build better tools.
2. Better Design: Engineers can use this data to design better, faster, and more efficient airplanes and rockets.
3. Testing Ground: It's a "benchmark." If a new AI algorithm claims it can see 3D air flows, scientists can test it against this perfect dataset to see if it actually works.

In short: They took a picture of invisible air from 70 different angles, used a smart AI to turn those 2D pictures into a perfect 3D movie, and shared the whole kit with the world so everyone can learn how to see the invisible.

Technical Summary

1. Problem Statement

Background-Oriented Schlieren (BOS) tomography is a powerful, non-intrusive optical diagnostic for measuring refractive index (and thus density) fields in compressible flows. However, reconstructing fully 3D, asymmetric flows from BOS data faces significant challenges:

• Limited View Counts: Time-resolved measurements often restrict the number of available camera perspectives (typically 5–15 views), leading to ill-posed inverse problems.
• Artifacts: Traditional reconstruction methods (e.g., filtered backprojection) often suffer from smearing, non-physical features (e.g., expansion waves upstream of shocks), and sensitivity to alignment errors.
• Lack of Standardized Data: There is a scarcity of open-source, high-fidelity datasets containing raw images, calibration data, and ground-truth-like reconstructions for benchmarking new algorithms in supersonic flight body flows.

2. Methodology

The authors addressed these challenges by conducting a comprehensive experiment and developing a novel open-source dataset and processing pipeline.

A. Experimental Setup

• Test Article: A generic, asymmetric square-pyramid flight body (semi-angle 20.1°, length 82 mm) mounted at a 10° angle of attack (AoA). The model was additively manufactured in polycarbonate to minimize deflection (estimated <1 mm).
• Facility: VKF Wind Tunnel D at AEDC, operating at Mach 4.8 with a stagnation pressure of 60 psia.
• Imaging System:
  • Camera Array: Seven monochromatic sCMOS cameras arranged on three rails (one upstream, one downstream, one "master" rail).
  • Rotation Strategy: The model was mounted on a rotating strut. Over 10 tests, the model was rotated in 10° increments from 0° to 90°, yielding 70 distinct angular views (10 positions × 7 cameras).
  • Illumination: A backlit BOS pattern (2D sinusoidal) illuminated by a pulsed Nd:YAG laser (10 ns pulses, 10 Hz) to freeze flow motion.
• Calibration: A rigorous two-stage calibration process using a two-level target plate and Perspective-n-Point (PnP) algorithms for background pose estimation, achieving sub-pixel reprojection errors (<0.5 px).

B. Data Processing & Reconstruction

The authors utilized a Neural-Implicit Reconstruction Technique (NIRT) framework:

1. Deflection Sensing: Horn–Schunck optical flow with local image normalization was used to extract 2D deflection fields from flow-off/flow-on image pairs. This outperformed cross-correlation and wavelet-based methods in preserving shock sharpness.
2. Tomographic Reconstruction:
   • A neural network (6 layers, 300 nodes) was trained to represent the continuous 3D density field.
   • Regularization: Total Variation (TV) regularization was applied to preserve sharp discontinuities (shocks).
   • Training: The network minimized a loss function combining data fidelity (matching observed deflections via a cone-ray operator) and TV regularization.
3. Data Assimilation:
   • The TV term was replaced by residuals from the 3D compressible Euler equations.
   • This allowed the recovery of unmeasured flow variables (velocity, temperature, pressure) directly from the density field derived from schlieren data, enforcing physical consistency.

3. Key Contributions

• Open-Source Dataset: The release of a comprehensive dataset including:
  • 70 views of high-speed flow (reference and deflected images).
  • Calibration data, masks, and deflection fields.
  • 3D reconstructed flow fields.
  • Full source code for the NIRT algorithm, deflection sensing, and camera calibration.
• First 3D State Estimation from Schlieren: Demonstrated the first successful application of data assimilation using Euler equations to recover full 3D state variables (density, velocity, temperature) directly from experimental BOS measurements.
• Robustness to Limited Data: Showed that high-fidelity 3D reconstructions of asymmetric shocks can be achieved using only 9 views (a subset of the total data), resolving sharp shock structures with minimal artifacts.
• Uncertainty Quantification: The NIRT framework enabled efficient uncertainty quantification (aleatoric and epistemic), identifying well-resolved flow features and guiding experimental design.

4. Results

• Reconstruction Quality:
  • The NIRT with TV regularization successfully resolved attached oblique shocks and a detached ellipsoidal shock wave at the top of the body.
  • It captured the expansion fan emanating from the tail and the turbulent boundary layer growth.
  • Reconstructions exhibited minimal artifacts (e.g., no non-physical upstream expansions), outperforming previous studies that used filtered backprojection.
• Validation:
  • Reconstructions based on 9 views were validated against held-out views (blue borders in Fig. 3), showing high fidelity in reproducing deflection fields.
  • The 3D density field matched physical expectations, with sharp compression regions and smooth expansion zones.
• Data Assimilation Outcomes:
  • Successfully recovered velocity and temperature fields.
  • The flow field showed expected deceleration over the body (compression) and acceleration in the wake (expansion).
  • Uncertainty maps correctly highlighted high-error regions (shock fronts and expansion fans where signal-to-noise drops) and low-error regions (freestream).

5. Significance

This work represents a major step forward in the field of experimental fluid dynamics and optical diagnostics:

• Benchmarking: It provides a "gold standard" dataset for the community to benchmark new tomographic algorithms, deflection sensing techniques, and data assimilation workflows.
• Algorithm Development: The open-source NIRT code facilitates the development of neural-network-based reconstruction methods tailored for compressible flows.
• Physical Insight: By demonstrating that 3D state estimation is possible directly from schlieren data, the study bridges the gap between optical measurements and full flow field characterization, reducing the need for intrusive probes in high-speed wind tunnels.
• Reproducibility: The inclusion of all raw data, CAD models, and code ensures that the research is fully reproducible, accelerating progress in supersonic flow diagnostics.

Data Availability: The full dataset and code are publicly available at doi: 10.26208/1VE2-5C19.