Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Photometric classification of supernovae detected by the Zwicky Transient Facility using noise augmentation

A. Townsend, J. Nordin, M. Kowalski, S. Reusch, J. P. Anderson, E. C. Bellm, U. Burgaz, T. X. Chen, T. -W. Chen, G. Dimitriadis, L. Galbany, A. Goobar, M. J. Graham, M. Gromadzki, C. P. Gutiérrez, D. Hale, C. Inserra, M. Kasliwal, Y. -L. Kim, K. Maguire, F. J. Masci, T. E. Müller-Bravo, D. A. Perley, R. L. Riddle, M. Rigault, J. van Santen, S. Schulze, M. Smith, J. Sollerman, S. Yang

TL;DR

This work tackles the challenge of constructing a large photometric SN Ia sample from ZTF data by combining ParSNIP-based light-curve representations with a robust noise augmentation strategy that mimics fainter, higher-redshift observations. Through a two-stage pipeline—latent-feature extraction via ParSNIP and a gradient-boosted classifier—the authors achieve SN Ia recall exceeding 98% for favorable observing conditions and maintain high overall classification accuracy while controlling contamination. Real-time performance is demonstrated in NoiZTF, where the classifier correctly identified the majority of targets for spectroscopic follow-up, including rare SLSNe, despite modest detection counts. The approach is shown to generalize across brightness ranges, DR2 validation, and practical cosmology considerations, with clear pathways to LSST-era applications and further improvements via expanded peculiar-class training and survey simulations.

Abstract

Modern time-domain surveys, such as the Zwicky Transient Facility (ZTF), detect far more extragalactic transients than can be spectroscopically classified. Photometric classification offers a scalable alternative, enabling the identification of larger, fainter, and higher-redshift supernova samples suitable for applications such as Type Ia supernova (SN Ia) cosmology. We present a feature-based photometric classifier for SNe detected by ZTF, with the primary goal of constructing a photometric SN Ia sample for cosmological analyses. Our approach utilises the autoencoder architecture of ParSNIP (Boone 2021) to capture the intrinsic diversity of SN light curves. We trained the model on a spectroscopically classified ZTF SN sample, incorporating a realistic noise augmentation procedure that simulates the flux uncertainties of fainter sources. Light curve features were used to train a gradient-boosted decision tree classifier, implemented in both binary (SN Ia vs. non-Ia) and multi-class configurations. We validated our classifier on independent, fainter ZTF data with and without noise augmentation. To evaluate real-time performance, we also applied our classifier to live ZTF alerts and conducted a spectroscopic classification survey within the ePESSTO+ collaboration. We found that noise augmentation significantly improves classification performance, particularly for fainter sources. Our binary classifier achieves an SN Ia recall of (98.1 $\pm$ 0.4)%, averaged across five train-test splits. SN Ia recall exceeds 98% for events with a peak apparent magnitude up to 20 and more than 10 detections, and remains above 96% up to magnitude 20.5. Overall, 95% of sources were correctly classified in both binary and multi-class modes. Our classifier performs efficiently on real ZTF data and enables construction of a large photometric SN Ia sample for cosmology.

Photometric classification of supernovae detected by the Zwicky Transient Facility using noise augmentation

TL;DR

This work tackles the challenge of constructing a large photometric SN Ia sample from ZTF data by combining ParSNIP-based light-curve representations with a robust noise augmentation strategy that mimics fainter, higher-redshift observations. Through a two-stage pipeline—latent-feature extraction via ParSNIP and a gradient-boosted classifier—the authors achieve SN Ia recall exceeding 98% for favorable observing conditions and maintain high overall classification accuracy while controlling contamination. Real-time performance is demonstrated in NoiZTF, where the classifier correctly identified the majority of targets for spectroscopic follow-up, including rare SLSNe, despite modest detection counts. The approach is shown to generalize across brightness ranges, DR2 validation, and practical cosmology considerations, with clear pathways to LSST-era applications and further improvements via expanded peculiar-class training and survey simulations.

Abstract

Modern time-domain surveys, such as the Zwicky Transient Facility (ZTF), detect far more extragalactic transients than can be spectroscopically classified. Photometric classification offers a scalable alternative, enabling the identification of larger, fainter, and higher-redshift supernova samples suitable for applications such as Type Ia supernova (SN Ia) cosmology. We present a feature-based photometric classifier for SNe detected by ZTF, with the primary goal of constructing a photometric SN Ia sample for cosmological analyses. Our approach utilises the autoencoder architecture of ParSNIP (Boone 2021) to capture the intrinsic diversity of SN light curves. We trained the model on a spectroscopically classified ZTF SN sample, incorporating a realistic noise augmentation procedure that simulates the flux uncertainties of fainter sources. Light curve features were used to train a gradient-boosted decision tree classifier, implemented in both binary (SN Ia vs. non-Ia) and multi-class configurations. We validated our classifier on independent, fainter ZTF data with and without noise augmentation. To evaluate real-time performance, we also applied our classifier to live ZTF alerts and conducted a spectroscopic classification survey within the ePESSTO+ collaboration. We found that noise augmentation significantly improves classification performance, particularly for fainter sources. Our binary classifier achieves an SN Ia recall of (98.1 0.4)%, averaged across five train-test splits. SN Ia recall exceeds 98% for events with a peak apparent magnitude up to 20 and more than 10 detections, and remains above 96% up to magnitude 20.5. Overall, 95% of sources were correctly classified in both binary and multi-class modes. Our classifier performs efficiently on real ZTF data and enables construction of a large photometric SN Ia sample for cosmology.
Paper Structure (30 sections, 6 equations, 18 figures, 8 tables)

This paper contains 30 sections, 6 equations, 18 figures, 8 tables.

Table of Contents

  1. Introduction
  2. The application of photometric classifiers in supernova cosmology
  3. Data
  4. Class breakdown
  5. Training and test samples
  6. BrightZTF and FaintZTF
  7. RandomZTF
  8. Augmentation method
  9. Modelling the ZTF flux errors at higher redshifts
  10. K-correction
  11. Noisification
  12. Testing the noisification with SALT2 fits to SNe Ia
  13. Classification procedure
  14. Generating the training and test samples
  15. Training the ParSNIP model and classifier
  16. ...and 15 more sections

Figures (18)

  • Figure 1: The peak apparent magnitude distribution for our sample, divided into classes of SN Ia (blue), SN II (orange), SN Ib/c (green), SN IIn (red), SLSN (pink). The filled histograms represent the test sample, FaintZTF, and the outline represents the full dataset. The peak apparent magnitude corresponds to the brightest detection in any ZTF band ($g$, $r$ or $i$).
  • Figure 2: Comparison of the $\sigma_\mathrm{t}/f_\mathrm{t}$ vs. $f_\mathrm{t}$ distributions of the original ZTF light curves (upper left) and the ZTF light curves with flux errors simulated according to Equation \ref{['equation:errmodel']} (upper right). The corresponding residual distributions for $\epsilon$ are shown in the lower panels.
  • Figure 3: Flux evolution of the random exponential component to the noise, $\epsilon$. The lower panels show histograms of $\epsilon$ for flux bins of width 400. The upper panel shows the corresponding exponential scale factor, $\lambda_f$, for each flux bin, which is fit with a linear model, $\lambda_f = m \cdot f_\mathrm{t} + c$.
  • Figure 4: Comparison of the simulated and real $\sigma_\mathrm{t}/f_\mathrm{t}$ values at flux bins in the low ($f_\mathrm{t}<400$), medium ($800<f_\mathrm{t}<1200$), and high ($f_\mathrm{t}>1600$) flux regimes.
  • Figure 5: K-corrections at peak brightness for the three ZTF bands ($g$: left, $r$: centre, $i$: right), applied to a representative set of SN types. For each class, the true redshift $z_\mathrm{true}$ (indicated by a dot at $(z_\mathrm{true}, 0)$) is taken as the median redshift from the BTS catalogue. Corrections are shown over the range $z_\mathrm{sim} \in [z_\mathrm{true}, z_\mathrm{true}+0.1]$. Template corrections are shown for SNe Ia (red), II (blue), Ib (green), Ic (purple), and IIn (orange), alongside the constant approximation adopted for SLSNe (brown).
  • ...and 13 more figures