1 of 21

AI Clustering of ALMA ISM �& Star Formation Data

Yun-Chi Wang(NTU)

Supervisor : Dr. Tien-Hao Hsieh (TARA)

2 of 21

Background

The Atacama Large Millimeter/submillimeter Array (ALMA)

  •  astronomical interferometer of 66 radio telescopes
  • Total Data Volume: ~2.3 PB (growing by ~300–400 TB/year)

George Kosugi, Eisuke Morita, Hiroshi Yoshida, et al. "Implementation and cost estimation of ALMA archive using cloud technology", Proc. SPIE 14155, Software and Cyberinfrastructure for Astronomy IX, 1415544 (19 Aug 2026); https://doi.org/10.1117/12.3104417

3 of 21

Background

4 of 21

Methodology

http://iopscience.iop.org/article/10.1086/671105

VICReg

https://doi.org/10.48550/arXiv.2105.04906

Umap Cluster

 interpret clusters physically

5 of 21

Methodology-UMAP

Dimension /Criterion

PCA

t-SNE

UMAP

Astrophysical Impact for ALMA

Manifold Geometry

Linear only

Non-linear

Non-linear

Captures complex velocity fields and varied emission morphologies

Global Structure

Preserved

Lost (local only)

Preserved

Preserves evolutionary connections

New Data Projection

Yes

No (requires full rerun)

Yes

Instantly projects future ALMA cycles without retraining

Cluster Morphology

Heavy overlap

Scattered clumps

Distinct "Islands"

Isolates clean physical archetypes (Disks, Outflows, Filaments)

6 of 21

Verification Labled Sample

Keplerian disks

outflows

Moment 0

(Symmetric Intensity)

Moment 0

(Collimated Lobes)

Moment 1

(Butterfly Velocity Gradient)

Moment 1

(Red/Blue-shifted

High-Velocity Wings)

Curated benchmark of ~100 known astrophysical sources used to validate unsupervised clustering.

7 of 21

UMAP

Min_dist

0.1

0.5

0.9

n_neighbors

20

50

100

8 of 21

🔴 Infall

🟩 Keplerian Disk ➕ Outflow

✖️ Point Source 🔶 Filament

🟡 Jet

Issue with primary beam correction

Without primary beam correction

9 of 21

PCA : technique that rotates the high-dimensional space to find the orthogonal axes of maximum variance

Measures total signal power

​

PC1

PC2

10 of 21

Umap of right island

11 of 21

orthogonal projection

project our vector onto the subspace that is strictly orthogonal to the PC1.

12 of 21

Conclusions

13 of 21

Future work

  • Try to improving the model
  • Exploring our new embedding data

​

​

14 of 21

Back-up slides

15 of 21

morphology of disks

16 of 21

morphology of outflows

17 of 21

morphology of infall

18 of 21

morphology of jet

19 of 21

morphology of filaments

20 of 21

morphology of point sources

21 of 21

Domain Variance Removal via Orthogonal Projection

  •