Inference of Black Hole Fluid-Dynamics from Sparse Interferometric Measurements

Aviad Levis, Daeyoung Lee, Joel A. Tropp, Charles F. Gammie, Katherine L. Bouman

Presented at CVPR 2021, this research establishes a sophisticated bridge between computer vision and astrophysics. It introduces a method to "see" the turbulent motion of plasma around a black hole using the same kind of sparse, noisy data that produced the famous first image of M87*. By leveraging physics-motivated surrogate models, the authors demonstrate that it is possible to extract meaningful scientific information from data too sparse for traditional video reconstruction.

The Challenge of Dynamic Black Hole Imaging

Imaging black holes is an exercise in extreme "sparse recovery." Radio telescope arrays, such as the Event Horizon Telescope (EHT), collect data as "visibilities"—individual points in the Fourier domain. While the M87 black hole was massive enough to appear static during an evening's observation, Sgr A (at the center of our galaxy) is smaller and evolves on timescales of minutes.

Traditional imaging methods face a fundamental dilemma:

  1. Static Imaging: Assumes the source is stationary, which results in motion blur for dynamic targets like Sgr A*.
  2. Video Reconstruction: Attempts to solve for every pixel in every frame. Due to the extreme sparsity of the data, this requires heavy regularization (smoothing) that often wipes out the very dynamics scientists wish to study.

The core question of this paper is: Instead of trying to reconstruct a high-fidelity movie, can we directly infer the physical parameters governing the motion?

A Physics-Motivated Surrogate Model

The authors develop a surrogate model using Stochastic Partial Differential Equations (SPDEs) to infer fluid dynamics. Rather than defining a massive covariance matrix for every pixel across time—a computationally prohibitive task—they model the evolving plasma as a solution to an anisotropic diffusion equation. This allows the system to describe complex, multi-scale correlations using only a few local interaction coefficients, such as velocity and correlation length.

Overview of the SPDE-based inference pipeline, mapping noise samples through physical parameters ($\\Theta$) to generate Gaussian Random Field (GRF) samples for comparison with telescope data.

By treating the plasma as a Spatio-temporal Gaussian Random Field (GRF), the method focuses on recovering the statistical parameters ($\\Theta$) of the flow. This approach bypasses the need for individual pixel accuracy, focusing instead on the underlying physical process.

Modeling Local Interactions and Dimensionality Reduction

To make the inference tractable, the model uses Randomized Subspace Iteration (RSI). This matrix-free dimensionality reduction technique allows the researchers to identify the most significant "modes" of the dynamic system without explicitly constructing the gargantuan covariance matrix.

Visualization of the SPDE components, including correlation lengths $\\lambda(r)$, velocity fields $v(x)$, and the resulting time-evolving snapshots of the plasma flow.

The inference pipeline alternates between estimating a static "envelope" (the time-averaged ring shape) and the dynamic parameters. The "best" physical parameters are identified by minimizing the distance between the projected model and the actual sparse telescope measurements.

Key Results and Scientific Interpretability

The approach was evaluated using synthetic SPDE videos and high-fidelity General Relativistic Magneto-Hydrodynamics (GRMHD) simulations. These simulations, which require tens of thousands of core hours to produce, serve as the ground truth for realistic black hole physics.

  • Superior Dynamic Recovery: In comparisons with StarWarps (a state-of-the-art dynamic imaging tool), the proposed method successfully identified rotation direction and spiral structure where traditional tools produced blurry averages.
  • Robustness to Sparsity: The method recovered flow parameters even using the highly limited EHT 2017 telescope configuration, demonstrating that a model-based approach can effectively "fill in the gaps" of missing data.
  • Interpretability: Unlike "black box" neural networks, the recovered parameters (such as flow angles and correlation scales) are directly interpretable by physicists, providing immediate scientific value.

Limitations and Future Directions

While promising, the current model assumes the statistics of the flow are stationary over the observation window. Sudden events, such as massive flares, might challenge this assumption. Additionally, the current study focused on "face-on" accretion disks; future work will need to account for high-inclination disks and relativistic effects like Doppler beaming. Finally, incorporating "closure phases" into the loss function will be necessary to handle complex atmospheric phase errors in real-world radio astronomy data.

Summary updated

Abstract

We develop an approach to recover the underlying properties of fluid-dynamical processes from sparse measurements. We are motivated by the task of imaging the stochastically evolving environment surrounding black holes, and demonstrate how flow parameters can be estimated from sparse interferometric measurements used in radio astronomical imaging. To model the stochastic flow we use spatio-temporal Gaussian Random Fields (GRFs). The high dimensionality of the underlying source video makes direct representation via a GRF’s full covariance matrix intractable. In contrast, stochastic partial differential equations are able to capture correlations at multiple scales by specifying only local interaction coefficients. Our approach estimates the coefficients of a space-time diffusion equation that dictates the stationary statistics of the dynamical process. We analyze our approach on realistic simulations of black hole evolution and demonstrate its advantage over state-of-the-art dynamic black hole imaging techniques.
BibTeX
@inproceedings{levis2021inference,
  author = {Aviad Levis and Daeyoung Lee and Joel A. Tropp and Charles F. Gammie and Katherine L. Bouman},
  title = {{Inference of Black Hole Fluid-Dynamics from Sparse Interferometric Measurements}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2021},
  doi = {10.1109/iccv48922.2021.00234},
}

Loading related content...

Figure