Alex Baker · Research Documentation
Three open-source research projects at the intersection of stellar spectroscopy, galaxy morphology, and Bayesian cosmology.
* AstroSpectro: XGBoost v6 (349,998 spectra) includes 4 Gaia/photometric features, so its metrics are not purely spectroscopic; see the Data Leakage Audit. Interpretable machine-learning pipeline for stellar spectral classification using LAMOST DR5 × Gaia DR3. v6 · includes 4 Gaia/photometric features — see Data Leakage Audit Late-fusion multimodal classifier combining DINOv2 visual features, morphometrics, and SDSS photometry. Best image-only model. Late fusion reaches 86.3% on a different split — not a modality gain. Reproducible Bayesian MCMC pipeline comparing ΛCDM, CPL, and Ξosc dark-energy models. ⭐ Key Scientific Finding · AstroSpectro · SHAP Interpretability · v6 · G2 · 349,998 spectra Caveat: v6 includes 4 Gaia/photometric features (M_BP, M_G, M_RP, ag_gspphot), which hold 3.2% of the total importance — see the Data Leakage Audit.Three research projects
AstroSpectro
AstroVision
Dark Energy · BAO MCMC
"SHAP analysis of the XGBoost classifier (350k spectra, ROC-AUC 0.986) reveals that Mg b equivalent width ranks first globally, Ca II K fourth — metallicity-sensitive features dominate over classical Balmer temperature indicators in the learned classification structure. 96.8% of feature importance derives from physically interpretable spectroscopic descriptors."