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).
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.
| # | 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.
docs/00-research-summary.md— a survey of what is saturated versus open in machine learning applied to space data.docs/01-winning-recipe-and-publishing.md— how to turn a project into a recognised contribution; publication pathways.docs/02-data-sources.md— a catalogue of open data sources with access methods.
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/.
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.
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.
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.
This laboratory is organised around the claim that discovery in open archives remains open to an individual working on modest compute. That claim is measured rather than assumed in a companion repository, archival-discovery-search, which evaluates four discovery routes with a kill criterion fixed before each measurement ran.
Three close. The AAVSO Variable Star Index already carries about 80% of Gaia DR3's variability catalogue, so what is missing from it is untranscribed published objects rather than discoveries. The 1,777,356 Gaia sources with epoch photometry but no variability classification show less scatter than ordinary stars, not more. The comet route could not be assessed at all, because its reporting channel was unreachable.
The fourth passes its test: known irregular satellites of Uranus and Neptune are recovered blind from public archival imagery with sub-arcsecond astrometry, nineteen detections across five nights and two telescopes. It is bounded all the same — injection and recovery place its 50% completeness at the same magnitude the surveys that took those frames already reached.
No object was discovered, and ten errors made during the work are recorded rather than removed.