Independent research laboratory
This laboratory pursues genuine scientific contributions from open space data — NASA/ESA archives, raw stellar time series, and signals from space. Its organising principle is that classification of known objects is saturated, whereas discovery — finding the rare, the anomalous, and the overlooked in under-mined or freshly released datasets — remains open to a motivated individual working on modest compute. Each project is reproducible, ties its output to a recognised publication pathway, and runs on a laptop or on Google Colab's free tier.
Full technical report: PDF (LaTeX, 33 pp) · Source code: github.com/en970/space-ml-lab
Pick a large, public, under-mined archive → define a crisp needle-in-a-haystack task where the standard pipeline is weak → apply an accessible model (train on synthetic data if labels are scarce) → tie the result to a real community and publication pathway (a Research Note of the AAS, a Zenodo DOI, and an official registry such as VSX / TNS / ExoFOP).
This is the recipe behind recent recognised work by young researchers (e.g. M. Paz, F. Wang, W. Cukier). The supporting survey and the publication pathways are documented in the repository under docs/.
A convolutional detector for solar radio bursts in e-CALLISTO dynamic spectra, labelled automatically from the network's burst lists, with a planned cross-modal extension to space-weather prediction. Trains locally on an Apple M2 in ~10 s.
Unsupervised discovery of open clusters via HDBSCAN + UMAP in Gaia's 5-D phase space. On real Gaia DR3 data it recovers NGC 2516 (1,618 members, 411.6 pc), validated against the Hunt & Reffert (2023) catalogue to $\sim$1′ and a few $\mu$as.
A self-supervised autoencoder + isolation-forest pipeline to surface unusual spectra, designed for SPHEREx NIR data and demonstrated on 4,990 real SDSS DR17 spectra (SPHEREx catalogues are not yet public — see report). The latent space self-organises by class, and the top-100 anomalies are $3.1\times$ enriched in extreme-emission-line objects.
A 1-D convolutional autoencoder plus Isolation Forest ranks the most unusual light curves among 700 real TESS (QLP) targets in Sector 67; the ranking surfaces high-amplitude variables, deep dippers, and high-scatter stars.
A 1-D CNN trained by synthetic batman transit injection detects single-transit (monotransit) events that periodic search misses; held-out ROC-AUC 0.998 with a full injection-recovery analysis.
| Path | Contents |
|---|---|
| docs/ | Research survey, the winning recipe & publication pathways, data-source catalog |
| projects/pN-.../ | Per-project code, README, results, and website |