Generative AI for image reconstruction in Intensity Interferometry: a first attempt

Technical Summary: Generative AI for Image Reconstruction in Intensity Interferometry

Problem Statement
Intensity Interferometry (II) has recently regained prominence as a technique for achieving high-precision resolution of stellar objects at optical wavelengths, particularly using Imaging Atmospheric Cherenkov Telescopes (IACTs). However, a fundamental limitation of II is the loss of phase information; detectors measure only the intensity correlation (squared visibility) of photons, not the phase. This makes image reconstruction an ill-posed inverse problem. While traditional methods such as triple-intensity correlation, iterative phase retrieval (e.g., Fienup's algorithm), and Transport-of-Intensity Equation (TIE) approaches have been explored, they often struggle with convergence, noise sensitivity, or computational complexity, especially when dealing with sparsely sampled data from finite telescope arrays.

Methodology
The authors propose a novel approach using a Conditional Generative Adversarial Network (cGAN) to reconstruct images of fast-rotating stars (FRS) from simulated II observations. The methodology involves the following components:

1. Data Simulation:

   • Source Models: Synthetic "ground truth" images of fictitious fast-rotating stars were generated. These models account for oblateness and gravity darkening (brighter poles, darker equator), resulting in 31,460 unique training images.
   • II Observation Simulation: Two hypothetical ground-based facilities were modeled: one with $N_T=6$ IACTs and another with $N_T=9$ IACTs. The simulation incorporated Earth's rotation to generate baseline tracks in the $(u, v)$ Fourier plane.
   • Input Generation: The "ground truth" images were transformed into the Fourier domain, and their absolute values were sampled along the specific baseline tracks of the 6-telescope and 9-telescope arrays. This resulted in sparsely sampled, phase-free intensity maps representing the II observations. Salt-and-pepper noise was added to enhance model robustness.

2. cGAN Architecture:

   • Framework: The study utilizes a Pix2Pix cGAN architecture (Isola et al., 2017), which performs image-to-image translation.
   • Conditioning: The sparse, noisy II observation patterns serve as the conditioning variable ($y$), while the high-fidelity stellar surface images serve as the target "ground truth" ($x$).
   • Generator: Implemented as a U-Net convolutional network with an encoder-decoder structure and skip connections. This allows the network to capture both global context and local high-frequency details (e.g., limb darkening, oblateness).
   • Discriminator: A PatchGAN model was employed to evaluate the "realness" of local image patches rather than the entire image, improving texture recovery and reducing computational cost.
   • Loss Function: The total loss combines the adversarial loss (minimax game) with an L1 loss (Mean Absolute Error) to enforce pixel-wise fidelity between the generated image and the ground truth.

3. Training and Validation:

   • The dataset was split into 80% training, 10% validation, and 10% testing.
   • Hyperparameters were tuned, including a learning rate of $2 \times 10^{-4}$, a kernel size of 5, and a discriminator repetition ratio of 1 (equal training steps for both networks).
   • The model was trained to minimize the discrepancy between the generated image and the ground truth while satisfying the constraints imposed by the sparse II data.

Key Results
The trained cGAN model was evaluated on a testing set of unseen fast-rotating stars using both the 6-telescope and 9-telescope configurations:

• Visual Reconstruction: The model successfully reconstructed the shape, size, and brightness distribution of the stars. The predicted images visually matched the ground truth, capturing the oblate shape and gravity darkening effects despite the input being sparsely sampled data from only 15 (for $N_T=6$) or 36 (for $N_T=9$) baselines.
• Quantitative Evaluation (Image Moments): The authors assessed the reconstruction quality using statistical image moments (monopole, centroids, and central moments up to the third order).
  • The scatter metrics between the ground truth and predicted moments were generally small, indicating robust learning.
  • Interestingly, the 6-telescope configuration ($N_T=6$) yielded slightly lower scatter values than the 9-telescope configuration ($N_T=9$). The authors attribute this to the specific geometry of the baseline tracks, suggesting that optimized $(u, v)$-plane coverage is more critical for reconstruction quality than simply increasing the total number of baselines.
• Parameter Recovery: The model effectively recovered physical parameters such as orientation, semi-major/minor axes, and eccentricity derived from the second-order central moments.

Significance and Claims
The paper presents this work as a "first attempt" to apply cGANs to image reconstruction in Intensity Interferometry. The authors claim that:

1. Feasibility: It is possible to reconstruct high-fidelity images of complex stellar systems (specifically gravity-darkened fast rotators) from the sparse, phase-less data typical of II, provided a sufficient number of baselines are available.
2. Scalability: While the current example (fast rotators) could theoretically be addressed via parameter fitting, the authors suggest that machine learning techniques are better suited for reconstructing more complex systems with varied surface features as IACT arrays grow larger (e.g., future Cherenkov Telescope Array Observatory).
3. Methodological Advancement: The study demonstrates that cGANs can effectively bridge the gap between sparse interferometric measurements and full image reconstruction, offering a promising alternative to traditional phase retrieval algorithms that struggle with noise and convergence.

The authors maintain a modest tone, noting that the results are based on simulated data and that future work must address detector efficiencies, real-world noise characteristics, and the optimization of telescope array layouts to fully realize the potential of this approach.