Skip to content

en970/space-ml-lab

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

6 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

space-ml-lab

Live reports (GitHub Pages): https://en970.github.io/space-ml-lab/ Full technical report (LaTeX, PDF, 33 pp, all five projects): reports/space-ml-lab-reports.pdf

An independent research repository applying machine learning to open space-science data (ESA/NASA archives, stellar catalogues, and radio observations), with an emphasis on discovery rather than the reproduction of established classification benchmarks. Every project is reproducible, connects its output to a recognised publication pathway, and runs on modest hardware (a laptop, or Google Colab's free tier).

Guiding principle

The reproduction of known classifications is a largely saturated activity: a supervised classifier can, by construction, only recover the classes it was trained on. The scientific opportunity lies in discovery — surfacing rare, anomalous, or overlooked structure in under-mined or newly released datasets. The methodology followed throughout is:

Select a large, public, under-mined archive; define a well-posed detection task in which the standard pipeline is weak; apply an accessible model (training on synthetic data where labels are scarce); and connect the result to a recognised community and publication pathway — a Research Note of the AAS, a Zenodo DOI, and an official registry such as VSX, TNS, or ExoFOP.

The supporting survey and the publication pathways are documented in docs/01-winning-recipe-and-publishing.md.

Projects

# Project Data Novelty Status
P3 Solar radio burst detection Dynamic spectra (image) High Working — CNN, validation AUC approx. 0.83 (local)
P1 Star clusters in Gaia DR3 Astrometric catalogue Medium Working — recovers NGC 2516 (1,618 members)
P2 Anomaly detection in spectra Optical/NIR spectra High Working — 3.1x emission-line enrichment (SDSS)
P4 TESS anomalous light-curve search Light curves High Working — 129 anomalies flagged in 700 TESS QLP light curves
P5 TESS single-transit search Light curves Medium-high Working — synthetic-injection CNN, held-out ROC-AUC 0.998

Each project directory contains a README.md describing how to build and run it from scratch, and a site/index.html report. Projects marked Working contain runnable code and real results; projects marked Design contain a complete plan only.

Documents

Repository layout

space-ml-lab/
  README.md
  index.html                 landing page (links to every project report)
  reports/                   combined LaTeX technical report (PDF)
  docs/                      research survey and reference material
  projects/
    p3-solar-radio-bursts/   working: CNN on e-CALLISTO dynamic spectra
    p1-gaia-star-clusters/   working: HDBSCAN + UMAP on Gaia DR3
    p2-spherex-anomaly/      working: autoencoder + Isolation Forest on spectra
    p4-tess-weird-stars/     working: convolutional autoencoder + Isolation Forest on TESS
    p5-tess-monotransit/     working: 1-D CNN + synthetic transit injection on TESS

Each working project directory holds README.md, src/ (runnable code), outputs/ (result figures and catalogues), site/ (the HTML report), and, where applicable, notebooks/.

Technology

Python 3 with astropy, astroquery, lightkurve, numpy, scikit-learn, PyTorch (Metal/MPS on Apple silicon), UMAP, and HDBSCAN. Data download and training run on the local machine or on Google Colab; the pipelines are deliberately modest in their compute requirements.

Data and licence

The code and written reports are released under the MIT Licence (see LICENSE). The analysed data are the property of their respective providers (ESA Gaia, SDSS, the e-CALLISTO network, STScI/DSS, NASA/SDO) and are used under their open-data terms; the original sources are cited in each project's report.

Note on results

All results reported here are candidate-level and method-validating; none constitute a confirmed scientific discovery. Data provenance, quality cuts, and failure modes are stated explicitly so that every claim can be independently reproduced and scrutinised.

About

Discovery-oriented machine learning on open space data: solar radio bursts (e-CALLISTO), star clusters in Gaia DR3, and spectral anomaly detection. Reproducible pipelines with formal reports.

Topics

Resources

License

Stars

0 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors