pith. sign in

arxiv: 2602.09223 · v1 · submitted 2026-02-09 · 🌌 astro-ph.IM

astromorph: Self-supervised machine learning pipeline for astronomical morphology analysis

Per Bjerkeli , Jouni Kainulainen , Maria Carmen Toribio , Leon Boschman , Otoniel Maya Lucas This is my paper

Pith reviewed 2026-05-16 05:01 UTC · model grok-4.3

classification 🌌 astro-ph.IM
keywords self-supervised learningastronomical morphologyimage embeddingsBYOLprotoplanetary disksinfrared dark cloudsFITS imagesspectral cubes
0
0 comments X

The pith

Astromorph adapts self-supervised BYOL learning to generate embeddings that capture morphological patterns in unlabeled astronomical images and cubes.

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

The paper introduces a pipeline called astromorph that adapts the BYOL self-supervised method to create image embeddings from large astronomical datasets. This label-free approach handles varying data formats such as single-channel FITS images and multi-channel spectral cubes, enabling tasks like clustering and similarity searches. Demonstrations on ALMA images of protoplanetary disks and Spitzer/Herschel observations of infrared dark clouds show that the embeddings group objects by morphological differences and similarities. The framework includes user-friendly scripts alongside customizable PyTorch classes for broader use across observational astronomy. If the embeddings prove robust, they reduce reliance on manual labeling for exploring complex structures in modern telescope data.

Core claim

Astromorph implements an adapted BYOL self-supervised framework that produces embeddings capturing morphological differences and similarities across large samples of astronomical data without labeled training examples. The pipeline accommodates data of varying dimensions and resolutions, including both single-channel FITS images and multi-channel spectral cubes, and is demonstrated on protoplanetary disks observed with ALMA and infrared dark clouds observed with Spitzer and Herschel. These embeddings support key tasks such as clustering, anomaly detection, and similarity-based exploration in a user-friendly manner with both streamlined scripts and deeper PyTorch customization options.

What carries the argument

The adapted BYOL (Bootstrap Your Own Latents) self-supervised training that learns representations from data augmentations to produce embeddings without any labels.

If this is right

  • Large unlabeled samples can be clustered by morphological similarity for statistical studies.
  • Anomaly detection becomes possible by identifying outliers in the embedding space.
  • Similarity searches can locate analogous objects across different surveys or wavelengths.
  • The same pipeline applies directly to both image and spectral cube data without retraining from scratch.
  • Broad applicability is suggested across imaging-rich contexts from varied telescopes.

Where Pith is reading between the lines

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

  • The embeddings could serve as input features for hybrid models that combine self-supervised pretraining with limited supervised fine-tuning on physical parameters.
  • Application to simulated datasets with injected known morphologies would provide a direct test of whether learned features align with physical structures.
  • Integration with upcoming large-scale surveys could lower the cost of initial data exploration before targeted follow-up observations.
  • Cross-domain transfer might allow embeddings trained on one telescope's data to initialize analysis for another instrument's images.

Load-bearing premise

The adapted BYOL training on astronomical images and cubes will produce embeddings that reflect physically relevant morphological properties rather than imaging artifacts or dataset-specific biases.

What would settle it

If expert visual classification of a held-out sample from the ALMA protoplanetary disk dataset shows no agreement with clusters formed from the learned embeddings, the claim of scientifically meaningful representations would be falsified.

Figures

Figures reproduced from arXiv: 2602.09223 by Jouni Kainulainen, Leon Boschman, Maria Carmen Toribio, Otoniel Maya Lucas, Per Bjerkeli.

Figure 1
Figure 1. Figure 1: Process diagram of the astromorph package. The pipeline_training.py script builds and trains a CNN model using the BYOL framework, with configuration defined in example_settings.toml. The BYOL architecture (upper right) operates on two augmented views on the input to learn meaningful representations without labels. The script pipeline_inference.py uses the trained model to compute and store embeddings that… view at source ↗
Figure 2
Figure 2. Figure 2: t-SNE projection of embedding vectors for ALMA continuum images selected from the ASA. Each point is represented by a thumbnail of the corresponding FITS image. For clarity, low-emission regions have been masked using an adaptive threshold derived from a Gaussian fit to the core of the distribution of normalized pixel values. Sources with stronger emission appear on top of weaker ones in the plot to enhanc… view at source ↗
Figure 3
Figure 3. Figure 3: ALMA continuum image of GG Tau (upper left) and the six most morphologically similar images, identified using cosine similarity. Each thumbnail shows the flux density in Jy beam−1 . Lower right panel presents a 2D PCA projection of embedding vectors, with the selected object and its closest matches highlighted. structure of the clouds as opposed to being dominated by the local maxima. We analysed the data … view at source ↗
Figure 4
Figure 4. Figure 4: Same as [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Principal component analysis of cloud morphologies. Upper panel: PC1 vs. PC2, where PC1 primarily seems to trace cloud size and PC2 reflects the distribution of the emission. Lower panel: PC2 vs. PC3 where PC2 again represent emission distribution, while PC3 seems to capture some measure of morphological complexity. 5. Conclusion We have presented astromorph, a Python package designed to apply the BYOL met… view at source ↗
read the original abstract

Modern telescopes generate increasingly large and diverse datasets, often consisting of complex and morphologically rich structures. To efficiently explore such data requires automated methods that can extract and organize physically meaningful information, ideally without the need for extensive manual interaction. We aim to provide a user-friendly implementation of a self-supervised machine learning framework to explore morphological properties of large datasets, based on the BYOL (Bootstrap Your Own Latents) method. By enabling the generation of meaningful image embeddings without manually labelled data, the framework will enable key tasks such as clustering, anomaly detection, and similarity based exploration. In contrast to existing BYOL implementations, astromorph accommodates data of varying dimensions and resolutions, including both single-channel FITS images and multi-channel spectral cubes. The package is built with usability in mind, offering streamlined pipeline scripts for ease of use as well as deeper customization options via PyTorch-based classes. To demonstrate the utility of astromorph, we apply it in two contrasting science cases representing different astronomical domains: images of protoplanetary disks observed with ALMA, and infrared dark clouds observed with Spitzer and Herschel. In both cases, we demonstrate how astromorph produces scientifically meaningful embeddings that capture morphological differences and similarities across large samples. astromorph enables users to apply a robust, label-free approach for uncovering morphological patterns in astronomical datasets. The successful application to two markedly different datasets suggest that the pipeline is broadly applicable across a wide range of imaging-rich astronomical context, providing a user friendly tool for advancing discovery in observational astronomy.

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

3 major / 3 minor

Summary. The paper presents astromorph, a PyTorch-based Python package implementing an adapted BYOL self-supervised learning pipeline for generating image embeddings from astronomical data. It accommodates single-channel FITS images and multi-channel spectral cubes of varying dimensions and resolutions, and demonstrates the pipeline on two datasets: ALMA observations of protoplanetary disks and Spitzer/Herschel infrared dark clouds. The central claim is that the resulting embeddings are scientifically meaningful, capturing morphological differences and similarities in a label-free manner to support tasks such as clustering and anomaly detection.

Significance. If the embeddings prove robustly meaningful, the package would offer a practical, user-friendly tool for label-free morphological exploration of large astronomical imaging datasets across modalities, filling a gap between generic self-supervised methods and domain-specific needs in observational astronomy.

major comments (3)
  1. [§4] §4 (ALMA protoplanetary disks application): The claim that embeddings 'capture morphological differences and similarities' is supported only by qualitative description; no quantitative metrics (e.g., correlation with catalogued disk parameters such as radius or inclination, silhouette scores, or comparison against supervised baselines) are reported, leaving the scientific meaningfulness unverified.
  2. [§5] §5 (infrared dark clouds application): Similarly, the demonstration on Spitzer/Herschel data asserts meaningful capture of cloud morphology (e.g., filamentarity) without ablation studies on the variable-resolution handling, controls for SNR or beam effects, or any embedding quality metrics, making it impossible to rule out that clusters arise from imaging covariates rather than astrophysical morphology.
  3. [Methods] Methods section (BYOL adaptation): The modifications for handling variable-resolution and multi-channel data are described at a high level but lack explicit equations or pseudocode for the loss function, augmentation strategy, or resolution-normalization step, preventing assessment of whether the adaptation preserves the original BYOL invariance properties.
minor comments (3)
  1. The abstract and introduction would benefit from explicit statements of dataset sizes (number of images/cubes) and hyperparameter choices used in the demonstrations.
  2. Figure captions for embedding visualizations should include details on the dimensionality reduction method (e.g., t-SNE perplexity) and any color-coding scheme.
  3. The package documentation and installation instructions could be expanded with a minimal reproducible example script.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed review of our manuscript. We address each major comment point by point below, providing clarifications and indicating where revisions have been made to strengthen the paper.

read point-by-point responses

  1. Referee: [§4] §4 (ALMA protoplanetary disks application): The claim that embeddings 'capture morphological differences and similarities' is supported only by qualitative description; no quantitative metrics (e.g., correlation with catalogued disk parameters such as radius or inclination, silhouette scores, or comparison against supervised baselines) are reported, leaving the scientific meaningfulness unverified.

    Authors: We agree that quantitative support would strengthen the demonstration in §4. In the revised manuscript we have added silhouette scores for the reported clusters, Pearson correlations between embedding dimensions and catalogued parameters (radius, inclination, and total flux), and a brief comparison of clustering purity against a supervised baseline trained on the available morphological labels. These results are now included in the text of §4 and in an updated Figure 4. revision: yes

  2. Referee: [§5] §5 (infrared dark clouds application): Similarly, the demonstration on Spitzer/Herschel data asserts meaningful capture of cloud morphology (e.g., filamentarity) without ablation studies on the variable-resolution handling, controls for SNR or beam effects, or any embedding quality metrics, making it impossible to rule out that clusters arise from imaging covariates rather than astrophysical morphology.

    Authors: We have added ablation experiments that isolate the effect of the resolution-normalization step and report quantitative embedding quality metrics (mean intra- versus inter-cluster Euclidean distances). However, exhaustive controls for SNR and beam convolution effects would require a dedicated suite of synthetic observations, which lies beyond the scope of the present pipeline-focused paper. We have expanded the discussion in §5 to explicitly acknowledge this limitation and to describe the steps taken to mitigate obvious imaging covariates. revision: partial

  3. Referee: Methods section (BYOL adaptation): The modifications for handling variable-resolution and multi-channel data are described at a high level but lack explicit equations or pseudocode for the loss function, augmentation strategy, or resolution-normalization step, preventing assessment of whether the adaptation preserves the original BYOL invariance properties.

    Authors: We have revised the Methods section to include (i) the explicit loss-function equation after the resolution-normalization and channel-handling modifications, (ii) pseudocode for the full augmentation pipeline, and (iii) a step-by-step derivation of the resolution-normalization operator. These additions allow readers to verify that the core BYOL invariance properties are retained. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain; software pipeline with qualitative demonstration

full rationale

The paper describes a user-friendly software implementation adapting the established BYOL self-supervised method for astronomical images of varying sizes and channels. No mathematical derivation, parameter fitting, or prediction step is present that could reduce to its own inputs by construction. Claims of 'scientifically meaningful embeddings' rest on application to two example datasets rather than any self-referential logic or self-citation chain. The methodology cites the original BYOL work as an external foundation and focuses on usability and adaptation for FITS data, with no load-bearing uniqueness theorems or ansatzes imported from the authors' prior work. This is a standard tool-description paper whose central assertions are not forced by internal definitions or fitted inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard assumptions from self-supervised learning rather than new physical postulates or fitted parameters.

axioms (1)
  • domain assumption Self-supervised contrastive learning via BYOL can extract morphologically meaningful features from unlabeled astronomical images
    This is the core premise enabling the entire label-free pipeline.

pith-pipeline@v0.9.0 · 5587 in / 1234 out tokens · 65165 ms · 2026-05-16T05:01:18.545562+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

14 extracted references · 14 canonical work pages

  1. [1]

    2015, TensorFlow: Large- Scale Machine Learning on Heterogeneous Systems, software avail- able from tensorflow.org

    Abadi, M., Agarwal, A., Barham, P., et al. 2015, TensorFlow: Large- Scale Machine Learning on Heterogeneous Systems, software avail- able from tensorflow.org

    work page 2015

  2. [2]

    1994, A&A, 286, 149

    Dutrey, A., Guilloteau, S., & Simon, M. 1994, A&A, 286, 149

    work page 1994

  3. [3]

    2020, Bootstrap your own latent: A new approach to self-supervised Learning

    Grill, J.-B., Strub, F., Altché, F., et al. 2020, Bootstrap your own latent: A new approach to self-supervised Learning

    work page 2020

  4. [4]

    E., Heitsch, F., et al

    Hacar, A., Clark, S. E., Heitsch, F., et al. 2023, in Astronomical So- ciety of the Pacific Conference Series, Vol. 534, Protostars and Planets VII, ed. S. Inutsuka, Y. Aikawa, T. Muto, K. Tomida, & M. Tamura, 153

    work page 2023

  5. [5]

    2016, in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR, 1

    He, K., Zhang, X., Ren, S., & Sun, J. 2016, in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR, 1

    work page 2016

  6. [6]

    2009, A&A, 508, L35

    Kainulainen, J., Beuther, H., Henning, T., & Plume, R. 2009, A&A, 508, L35

    work page 2009

  7. [7]

    W., Rosolowsky, E

    Koch, E. W., Rosolowsky, E. W., Boyden, R. D., et al. 2019, AJ, 158, 1

    work page 2019

  8. [8]

    Larson, R. B. 1981, MNRAS, 194, 809

    work page 1981

  9. [9]

    C., Sandstrom, K

    Lee, J. C., Sandstrom, K. M., Leroy, A. K., et al. 2023, ApJ, 944, L17 Lenkić, L., Nally, C., Jones, O. C., et al. 2024, ApJ, 967, 110 maintainers, T. & contributors. 2016, TorchVision: PyTorch’s Com- puter Vision library,https://github.com/pytorch/vision

    work page 2023

  10. [10]

    M., Barrientes, J., Blome, C., et al

    Pontoppidan, K. M., Barrientes, J., Blome, C., et al. 2022, ApJ, 936, L14

    work page 2022

  11. [11]

    H., Tam, A., Tang, Y., et al

    Richemond, P. H., Tam, A., Tang, Y., et al. 2023, in Proceedings of the 40th International Conference on Machine Learning, ICML’23 (JMLR.org)

    work page 2023

  12. [12]

    2011, A&A, 529, A1 van der Maaten, L

    Schneider, N., Bontemps, S., Simon, R., et al. 2011, A&A, 529, A1 van der Maaten, L. & Hinton, G. 2008, Journal of Machine Learning Research, 9, 2579

    work page 2011

  13. [13]

    N., Galvin, T

    Vantyghem, A. N., Galvin, T. J., Sebastian, B., et al. 2024, Astronomy and Computing, 47, 100824

    work page 2024

  14. [14]

    G., Lee, J

    Williams, T. G., Lee, J. C., Larson, K. L., et al. 2024, ApJS, 273, 13 Article number, page 12 of 12

    work page 2024