Neural networks sort supernovae and estimate their distances from brightness alone

The work

Technical · from the record

This thesis develops machine learning tools for supernova cosmology from photometry alone. SCONE, a convolutional neural network applied to Gaussian-process-interpolated wavelength-time heatmaps of supernova lightcurves, reached 99.73±0.26% test accuracy separating simulated type Ia from non-Ia supernovae and 75% accuracy on 6-way typing at the date of trigger with redshift information. Photo-zSNthesis, a related CNN, predicts full redshift probability distributions from lightcurves and was compared against LCFIT+Z on simulated LSST and SDSS data and on 489 observed SDSS supernovae. A study of directional-light-radius host galaxy matching on DES-SN5YR-like simulations found 1.7% of simulated supernovae matched to the wrong host and a shift in the dark energy equation of state parameter of Δw=0.0013±0.0026 with a CMB prior, and the Connect Later framework fine-tunes pretrained models with targeted augmentations on astronomical, wildlife and histopathology benchmarks.

How AI was used

Simulated supernova lightcurves were generated with SNANA following PLAsTiCC, SDSS and DES survey models, passed through quality and SALT lightcurve-fit selection cuts, and class-balanced into training, validation and test splits. Each lightcurve was encoded by fitting a two-dimensional Gaussian process in wavelength and time with a Matérn 3/2 kernel at a fixed 6000 Å wavelength length scale, fitting the time length scale, and sampling the fitted model on a 32×180 grid to produce stacked flux and uncertainty heatmaps normalised to [0,1]. Convolutional networks with full-height kernels were trained from scratch on these heatmaps: SCONE with binary cross-entropy for Ia versus non-Ia and sparse categorical cross-entropy for 6-way typing, with a variant concatenating redshift and redshift error into the fully connected classifier; Photo-zSNthesis with residual blocks and a softmax layer over discretised redshift bins, calibrated afterwards by temperature scaling. Trained models were run over held-out splits, truncated early-time lightcurves, bright subsets, out-of-distribution spectroscopic-like samples and observed SDSS and DES lightcurves. In the host-mismatch study, a self-organizing map trained on griz fluxes assigned photometric redshifts to catalogue galaxies, the directional light radius method matched supernovae to hosts in both data and simulations, and SuperNNova and SCONE supplied Ia probabilities that weighted the BEAMS likelihood before cosmology fitting with wfit. Connect Later pretrained an Informer encoder by masked autoencoding of lightcurve observations and SwAV-pretrained ResNet-50 and DenseNet121 image models, then fine-tuned with linear probing followed by fine-tuning using targeted augmentations that resample each object to a new redshift or randomise image background and stain colour.