Abstract
This study proposes a radically alternate approach for extracting quantitative information from schlieren images. The method uses a scaled, derivative-enhanced Gaussian process model to obtain true density estimates from two corresponding schlieren images with the knife-edge at horizontal and vertical orientations. We illustrate our approach on schlieren images taken from a wind tunnel sting model, a supersonic aircraft in flight, and a high-order numerical shock tube simulation.
Introduction: From Qualitative to Quantitative Visualization
Schlieren imaging has been a foundational tool in fluid dynamics for centuries, prized for its ability to visualize refractive index gradients in transparent media. By capturing how light deviates as it passes through varying fluid densities, it produces high-contrast images of shock waves, expansion fans, and turbulent plumes. However, despite its visual power, schlieren has historically remained a primarily qualitative technique. While it clearly identifies the location of flow features, extracting precise numerical density values—a process known as quantitative schlieren—has traditionally required complex calibration, specialized hardware, or highly controlled environments.
Existing quantitative methods, such as Background Oriented Schlieren (BOS), are effective but often "fussy," requiring specific background patterns and high-resolution imaging. This research addresses the challenge of "retrospective quantization," aiming to invert the physics of standard schlieren images to recover underlying density fields using a machine learning framework.
Methodology: Derivative-Enhanced Gaussian Processes
The proposed method bridges the gap between pixel intensity and physical density through a Bayesian nonparametric approach. The core of the technique relies on the Gladstone-Dale relationship, which establishes that the refractive index of a gas ($n$) is linearly proportional to its density ($\rho$): $n - 1 = \kappa \rho$.
Because schlieren pixel intensities are functions of the refractive index gradients (typically $\partial n/\partial x$ or $\partial n/\partial y$), the image serves as a direct map of density gradients. Rather than performing error-prone numerical integration, the researchers model the underlying density field as a Gaussian Process (GP). By treating the schlieren image as a linear operator acting on the GP, they can jointly model the latent density and the observed pixel intensities. This framework allows the model to infer the most likely density field given the observed gradients, provided a few "anchor points" (known density values at specific locations) are available to set the absolute scale.
Experimental Validation and Results
The model was validated across several challenging scenarios, ranging from synthetic benchmarks to real-world flight data.
One primary test case involved a wind tunnel sting model in supersonic flow ($M=2.0$). The GP reconstruction was compared against established BOS data, successfully capturing the sharp density discontinuities across oblique shocks and the smooth transitions within expansion fans.
Further validation included:
- T-38 Supersonic Aircraft: Analyzing NASA flight images to resolve density wakes behind the aircraft, demonstrating robustness against complex geometries and natural backgrounds.
- Numerical Shock Tube: Using high-order CFD simulations (the Sod problem) to test the model's handling of sharp discontinuities and turbulent features.
In these cases, the Bayesian approach provided not only density estimates but also uncertainty quantification maps, highlighting regions where reconstruction reliability might be lower due to image saturation or noise.
Discussion and Future Outlook
This machine learning approach offers a powerful tool for fluid dynamics, enabling researchers to extract hard data from legacy images or simple lab setups. While the method currently faces computational scaling challenges—Gaussian Processes typically scale cubically with the number of data points—the use of sparse GP approximations and downsampling makes it viable for practical research.
Future work aims to extend this framework to schlieren cinematography. By incorporating temporal kernels into the GP, the model could potentially learn from the movement of flow features over time, further improving the accuracy and robustness of density reconstructions in dynamic systems.
Summary updated