pith. sign in

arxiv: 2508.03398 · v3 · submitted 2025-08-05 · 🌌 astro-ph.IM

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

Km Nitu Rai , Yuri van der Burg , Soumen Basak , Prasenjit Saha , Subrata Sarangi This is my paper

Pith reviewed 2026-05-19 00:31 UTC · model grok-4.3

classification 🌌 astro-ph.IM
keywords intensity interferometryimage reconstructiongenerative adversarial networksstellar imagingCherenkov telescopesmachine learningrotating starsastronomical interferometry
0
0 comments X

The pith

A conditional generative adversarial network reconstructs images of fast-rotating stars from intensity interferometry data.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper shows that a conditional Generative Adversarial Network can recover the shape, size, and surface brightness distribution of simulated fast-rotating stars when given only the sparsely sampled spatial power spectra that intensity interferometry supplies. Intensity interferometry records photon correlations rather than phases, so conventional imaging methods struggle; the network learns the mapping from those correlations to full images. The demonstration uses two hypothetical ground-based arrays of six and nine Cherenkov telescopes. The result indicates that machine-learning methods may open intensity interferometry to targets too complex for simple parameter fitting once larger arrays become available.

Core claim

The paper claims that a conditional Generative Adversarial Network successfully reconstructs the shape, size, and brightness distribution of simulated fast-rotating stars from sparsely sampled spatial power spectra obtained with two hypothetical ground-based intensity interferometry facilities composed of six and nine Imaging Atmospheric Cherenkov Telescopes respectively. Although parameter fitting could address this particular simple case, the results indicate that machine-learning techniques applied to intensity interferometry data could reconstruct much more complex systems with varied surface features when larger arrays are used.

What carries the argument

A conditional Generative Adversarial Network trained to map sparsely sampled spatial power spectra from intensity interferometry observations into reconstructed stellar images.

If this is right

  • Larger arrays of Imaging Atmospheric Cherenkov Telescopes would allow machine-learning reconstruction of complex stellar systems that exhibit varied surface features.
  • Machine-learning methods applied to intensity interferometry can address image reconstruction problems beyond the reach of parameter fitting for simple cases.
  • The approach provides an alternative route to overcoming phase-retrieval difficulties that arise in intensity interferometry because only power spectra are measured.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same network architecture could be retrained on data from existing Cherenkov telescope arrays to test performance on actual rather than simulated observations.
  • Extending the method to binary stars or stars with spots would show whether the network can handle targets whose complexity exceeds the current training set.
  • Combining the reconstructed images with simultaneous data from other wavelengths might constrain stellar atmosphere models more tightly than either technique alone.

Load-bearing premise

The simulated intensity interferometry measurements and the training distribution of fast-rotating stars are representative enough that the trained network will generalize to real observations or to more complex targets.

What would settle it

Applying the trained network to actual intensity interferometry observations of a real fast-rotating star and finding that the output shape, size, or brightness distribution disagrees with independent measurements from spectroscopy or other techniques.

Figures

Figures reproduced from arXiv: 2508.03398 by Km Nitu Rai, Prasenjit Saha, Soumen Basak, Subrata Sarangi, Yuri van der Burg.

Figure 1
Figure 1. Figure 1: — [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: — The tracks of the baselines provided by the four tele￾scopes arranged in fig. 1 for one night of observation [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: — This figure shows the simulated fast rotating star. The brightness is highest at the poles and there is gravitational dark￾ening along the equator. This section presents a brief conceptual overview of how an array of telescopes is used to perform II observa￾tions, and explains the Signal-to-Noise Ratio (SNR) from these measurements. 2.1. The signal for II As a simple example, let us consider a pair of IA… view at source ↗
Figure 4
Figure 4. Figure 4: — Absolute value of the two-dimensional Fast Fourier Transform of the source depicted in [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: — Absolute value of the two-dimensional Fast Fourier Transform of the source depicted in [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: — Merged image, which includes the original and the sparsely sampled Fourier plane. It is exactly what the GAN re￾ceives. The grey scale of the figure is normalized to the brightest pixel in the image. a limited observation schedule. We have simulated the II observation of the fictitious star by four telescopes (cor￾related with baselines as seen in [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: — Discriminator and Generator losses for three different learning rates. The left panel and the right panel show the total discriminator loss and the total generator loss (Eq. 17) respectively. There is no significant difference, but these figures indicate that higher learning rates might render the training prone to outliers. 0 10 20 30 40 50 60 70 80 Steps 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 dis… view at source ↗
Figure 8
Figure 8. Figure 8: — Discriminator and Generator losses (left and right panels respectively) for three different kernel sizes in the convolutional layers. Here, the smallest kernel size has many outliers, while the largest kernel size seems to be the most stable. 0 10 20 30 40 50 60 Steps 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 disc_loss alpha / beta 1 0.05 50 0 10 20 30 40 50 60 Steps 0 5 10 15 20 gen_total_loss alpha / beta 1 0.05… view at source ↗
Figure 9
Figure 9. Figure 9: — Effect of the Salt (alpha) and Pepper (beta) noise (explained in the text) introduced into the images. There is no significant effect of the alpha/beta ratio. These results are from training on 64 × 64-pixel images [PITH_FULL_IMAGE:figures/full_fig_p006_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: — Loss functions for two different batch sizes. Large batch sizes seem to make the training more robust, but also increase training time significantly. 0 10 20 30 40 50 Steps 0.0 0.5 1.0 1.5 2.0 disc_loss Disc. rep. 1 3 6 0 10 20 30 40 50 Steps 2 4 6 8 10 12 14 gen_total_loss Disc. rep. 1 3 6 [PITH_FULL_IMAGE:figures/full_fig_p007_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: — The number of episodes of Discriminator training per every episode of Generator training (termed Discriminator repetition or Disc. rep.), understandably, has a higher impact on the Discriminator loss than on the Generator loss. The left and the right panel show this effect on the Discriminator loss and the Generator loss respectively. 0 10 20 30 40 50 60 70 80 Steps 0.5 1.0 1.5 2.0 disc_loss Telescopes … view at source ↗
Figure 12
Figure 12. Figure 12: — Discriminator loss (left panel) and Generator loss (right panel) for different numbers of telescopes. The number of telescopes has a very significant impact on the model performance. If there are only two telescopes (1 baseline), both Discriminator and Generator are not trained smoothly: One gains a big advantage over the other. The result of four telescopes (6 baselines) is a lot better because the los… view at source ↗
Figure 13
Figure 13. Figure 13: — Example results of image reconstruction using the GAN model along with the II observations simulated in this work. Each row in this figure represents the results for a hypothetical fast-rotating star. Going from left to right in each row, the first panel represents the simulated (u, v) plane II signals obtained using the six baselines (of Fig.5). This image is presented, as the input, to the trained GAN… view at source ↗
Figure 14
Figure 14. Figure 14: — This set of figures shows the comparison of monopole, x-centroid, and y-centroid for ground truth and predicted images generated by trained GAN [PITH_FULL_IMAGE:figures/full_fig_p012_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: — The second-order central moments provide information about the size and shape of stellar objects. Shown here are all the second-order central moments for ground truth and predicted images generated by the trained GAN. From left to right these are µ11, µ20, µ02. this end, the central moment of an image is calculated according to: µpq = 1 M00 X x X y (x − xc) p (y − yc) q I(x, y). (22) The sum of p and q … view at source ↗
Figure 16
Figure 16. Figure 16: — Shown here are all the third-order central moments for ground truth and predicted images generated by the trained GAN. They represent the skewness of the brightness distributions. The panels in reading order show µ30, µ03, µ21, µ12. and δ = p 4µ 2 11 + (µ20 − µ02) 2 2 . (26) Using the calculated axis values, the eccentricity of the fast-rotating star is determined as: e = p 1 − a/b. (27) Eqs. 23-27 desc… view at source ↗
read the original abstract

In the last few years, Intensity Interferometry (II) has made significant strides in achieving high-precision resolution of stellar objects at optical wavelengths. Despite these advancements, phase retrieval remains a major challenge due to the nature of photon correlation. This paper explores the application of a conditional Generative Adversarial Network (cGAN) to tackle the problem of image reconstruction in II. This method successfully reconstructs the shape, size, and brightness distribution of simulated, fast-rotating stars based on sparsely sampled spatial power spectra obtained by using two different hypothetical ground-based II facilities composed of six and nine Imaging Atmospheric Cherenkov Telescopes (IACTs), respectively. Although this particular example could also be addressed using parameter fitting, our results suggest that with larger arrays of IACTs much more complex systems with varied surface features could be reconstructed by applying machine-learning techniques to II. Hence this approach merits closer examination.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript presents a first attempt to apply a conditional Generative Adversarial Network (cGAN) to image reconstruction in Intensity Interferometry. It reports that the network reconstructs the shape, size, and brightness distribution of simulated fast-rotating stars from sparsely sampled spatial power spectra generated for two hypothetical ground-based arrays consisting of six and nine Imaging Atmospheric Cherenkov Telescopes, respectively. The authors note that parameter fitting could also solve the specific case examined but argue that the ML approach may scale to more complex stellar surface features with larger arrays.

Significance. If the central claim is substantiated, the work constitutes a useful proof-of-concept demonstration that generative models can address the phase-retrieval problem in intensity interferometry. This is a timely direction given recent advances in II instrumentation, and the choice of simulated fast-rotating stars as a test case is reasonable. However, the absence of quantitative performance metrics, baseline comparisons, or robustness tests against realistic instrumental effects currently limits the assessed impact to that of an exploratory study rather than a definitive methodological advance.

major comments (2)
  1. [Abstract] Abstract and results description: the claim that the cGAN 'successfully reconstructs' shape, size, and brightness distribution is presented without any quantitative metrics (e.g., mean squared error, structural similarity index, or reconstruction fidelity scores), error bars, or comparison to traditional phase-retrieval or fitting baselines. This omission makes it impossible to evaluate the strength of support for the central claim from the information provided.
  2. [Methods] Methods and simulation description: the forward model used to generate both training and test power spectra is not shown to include or vary key real-world effects such as photon noise statistics, baseline-dependent visibility losses, telescope optical transfer functions, or residual atmospheric scintillation. Without such variation or hold-out tests, the reported reconstructions on simulated data do not yet demonstrate robustness to the distribution shift expected in actual telescope observations.
minor comments (2)
  1. [Methods] The manuscript would benefit from a clearer statement of the cGAN architecture details, training/validation split ratios, and hyperparameter choices, even if relegated to an appendix.
  2. [Figures] Figure captions should explicitly state the input power-spectrum sampling density and the exact array configurations (6 vs. 9 telescopes) used for each reconstruction example.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive feedback on our manuscript. We address each of the major comments in detail below and indicate the revisions we plan to make.

read point-by-point responses

  1. Referee: [Abstract] Abstract and results description: the claim that the cGAN 'successfully reconstructs' shape, size, and brightness distribution is presented without any quantitative metrics (e.g., mean squared error, structural similarity index, or reconstruction fidelity scores), error bars, or comparison to traditional phase-retrieval or fitting baselines. This omission makes it impossible to evaluate the strength of support for the central claim from the information provided.

    Authors: We acknowledge that quantitative performance metrics are important for substantiating the claims. Although the manuscript emphasizes the qualitative success in reconstructing the stellar images as a proof-of-concept, we agree that this could be strengthened. In the revised manuscript, we will add quantitative metrics including mean squared error and structural similarity index (SSIM) for the reconstructions. We will also include a comparison to a parameter-fitting approach for this specific case of fast-rotating stars, as noted in the abstract. This will allow readers to better evaluate the method's performance. revision: yes

  2. Referee: [Methods] Methods and simulation description: the forward model used to generate both training and test power spectra is not shown to include or vary key real-world effects such as photon noise statistics, baseline-dependent visibility losses, telescope optical transfer functions, or residual atmospheric scintillation. Without such variation or hold-out tests, the reported reconstructions on simulated data do not yet demonstrate robustness to the distribution shift expected in actual telescope observations.

    Authors: The simulations in the current work are indeed idealized, without inclusion of photon noise, optical transfer functions, or atmospheric effects, as this represents a first attempt to apply cGAN to II image reconstruction. We recognize this as a limitation for demonstrating robustness to real data. In the revised manuscript, we will expand the methods section to clearly describe the forward model assumptions and add a dedicated discussion on the challenges of applying the method to real observations, including potential distribution shifts. While we cannot fully address robustness in this initial study without additional simulations, we believe the current results provide a valuable starting point for future investigations that incorporate these effects. revision: partial

Circularity Check

0 steps flagged

No significant circularity; standard supervised cGAN training on simulations

full rationale

The paper describes training a conditional GAN on forward-simulated intensity interferometry power spectra (from 6- and 9-telescope IACT arrays) to reconstruct images of fast-rotating stars, then evaluating on held-out simulations. This is a conventional supervised learning pipeline with no equations, parameter fits, or self-citations that reduce the reported reconstructions to the training inputs by construction. The central claim is feasibility on synthetic data rather than a first-principles derivation, and the method remains self-contained without invoking uniqueness theorems or ansatzes from prior author work.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim depends on the realism of the simulated II data and the generalization properties of the trained cGAN; no new physical entities are introduced.

free parameters (1)
  • cGAN training hyperparameters and architecture choices
    Standard ML model parameters fitted during training on the simulated dataset.
axioms (1)
  • domain assumption Simulated power spectra from hypothetical 6- and 9-telescope IACT arrays accurately represent real intensity interferometry measurements
    Invoked when claiming the reconstructions demonstrate feasibility for actual observations.

pith-pipeline@v0.9.0 · 5697 in / 1348 out tokens · 44393 ms · 2026-05-19T00:31:42.425166+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. FluxFlow: Conservative Flow-Matching for Astronomical Image Super-Resolution

    cs.CV 2026-05 unverdicted novelty 7.0

    FluxFlow is a conservative pixel-space flow-matching framework for astronomical super-resolution that incorporates real atmospheric uncertainty and a training-free Wiener correction, outperforming baselines on a new 1...

  2. FluxFlow: Conservative Flow-Matching for Astronomical Image Super-Resolution

    cs.CV 2026-05 unverdicted novelty 5.0

    FluxFlow uses conservative pixel-space flow-matching with uncertainty weights and Wiener test-time correction to outperform baselines on photometric and scientific accuracy for ground-to-space super-resolution, valida...

Reference graph

Works this paper leans on

15 extracted references · 15 canonical work pages · cited by 1 Pith paper · 2 internal anchors

  1. [1]

    TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems

    Abadi et al., M. 2016, arXiv preprint arXiv:1603.04467 Abe et al., S. 2024, MNRAS, 529, 4387 Acciari et al., V. A. 2020, MNRAS, 491, 1540 Acharyya et al., A. 2024, The Astrophysical Journal, 966, 28 Aleksi´ c et al., J. 2016, Astroparticle Physics, 72, 76 Baumgartner et al., S. 2020, MNRAS, 498, 4577 Coccomini et al., D. 2021, arXiv preprint arXiv:2122.11...

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  2. [2]

    1982, Applied Optics, 21, 2758

    Fienup, J. 1982, Applied Optics, 21, 2758

    work page 1982

  3. [3]

    1963, Journal of Applied Physics, 34, 875

    Gamo, H. 1963, Journal of Applied Physics, 34, 875

    work page 1963

  4. [4]

    Gerchberg, R. W. 1972, Optik, 35, 237

    work page 1972

  5. [5]

    Glauber, R. J. 1963, Physical Review, 130, 2529 Goldberger et al., M. L. 1963, Physical Review, 132, 2764 Goodfellow et al., I. 2014, Advances in neural information processing systems, 27 Hanbury Brown, R. & Twiss, R. Q. 1956, Nature, 178, 1046 Hanbury Brown et al., R. 1954, The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, ...

    work page 1963

  6. [6]

    2002, Optics (Addison Wesley) Holmes et al., R

    Hecht, E. 2002, Optics (Addison Wesley) Holmes et al., R. 2010, in Adaptive Coded Aperture Imaging, Non-Imaging, and Unconventional Imaging Sensor Systems II, Vol. 7818, SPIE, 175–185

    work page 2002

  7. [7]

    1962, IRE transactions on information theory, 8, 179

    Hu, M.-K. 1962, IRE transactions on information theory, 8, 179

    work page 1962

  8. [8]

    Hunter, J. D. 2007, Computing in Science & Engineering, 9, 90 Isola et al., P. 2017, in Proceedings of the IEEE conference on computer vision and pattern recognition, 1125–1134 Kirisits et al., C. 2024, arXiv preprint arXiv:2406.14143v2 Le Bohec, S; Holder, J. 2006, The Astrophysical Journal, 649, 399 Leonard et al., M. 1995, Optical Coherence and Quantum...

    work page arXiv 2007

  9. [9]

    Lucy, L. B. 1967, Zeitschrift f¨ ur Astrophysik, Vol. 65, p. 89, 65, 89

    work page 1967

  10. [10]

    Conditional Generative Adversarial Nets

    Maeder, A. 1999, Astronomy & Astrophysics, 347, 185 McAlister et al., H. A. 2005, The Astrophysical Journal, 628, 439 Mirza et al., M. 2014, arXiv preprint arXiv:1411.1784

    work page internal anchor Pith review Pith/arXiv arXiv 1999

  11. [11]

    Murphy, K. P. 2022, Probabilistic machine learning: an introduction (MIT press) Mustafa et al., M. 2019, Computational Astrophysics and Cosmology, 6, 1 Nu˜ nez et al., P. D. 2010, in Optical and Infrared Interferometry

    work page 2022

  12. [12]

    Prince, S. J. 2023, Understanding deep learning (MIT press) Rai et al., K. N. 2021, Monthly Notices of the Royal Astronomical Society, 507, 2813 Rai et al., K. N. 2022, Monthly Notices of the Royal Astronomical Society, 516, 2864 Ronneberger et al., O. 2015, in Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conf...

    work page 2023

  13. [13]

    Teague, M. R. 1980, Josa, 70, 920

    work page 1980

  14. [14]

    Teague, M. R. 1983, Journal of Opticala Society of America, 73, 1434 Virtanen et al., P. 2020, Nature methods, 17, 261 Vogel et al., N. 2025, MNRAS, 537, 2334 Von Zeipel, H. 1924, Monthly Notices of the Royal Astronomical

    work page 1983

  15. [15]

    Society, Vol. 84, p. 665-683, 84, 665 Zhang et al., J. 2020, Opt. Lett., 45, 3649

    work page 2020