Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AppleCiDEr II: SpectraNet -- A Deep Learning Network for Spectroscopic Data

Maojie Xu, Argyro Sasli, Alexandra Junell, Felipe Fontinele Nunes, Yu-Jing Qin, Christoffer Fremling, Sam Rose, Theophile Jegou Du Laz, Benny Border, Antoine Le Calloch, Sushant Sharma Chaudhary, Hailey Markoff, Avyukt Raghuvanshi, Nabeel Rehemtulla, Jesper Sollerman, Yashvi Sharma, Niharika Sravan, Judy Adler, Tracy X. Chen, Richard Dekany, Reed Riddle, Mansi M. Kasliwal, Matthew J. Graham, Michael W. Coughlin

TL;DR

SpectraNet addresses the need for robust spectral classification of time-domain transients by learning multiscale spectral representations through a CNN architecture integrated into the AppleCiDEr multimodal pipeline. It combines rest-frame spectral preprocessing, a redshift regression head, and a five-block multi-scale CNN with a 1×1 projection, optimized with class-balanced focal loss and training techniques like EMA and mixed precision. The approach yields state-of-the-art classification performance and precise redshift estimates, with strong generalization to narrower wavelength coverage as demonstrated on NGPS data. This work enables real-time, spectroscopically informed transient classification and paves the way for LSST-era deployment with domain adaptation and evolution-aware extensions.

Abstract

Time-domain surveys such as the Zwicky Transient Facility (ZTF) have opened a new frontier in the discovery and characterization of transients. While photometric light curves provide broad temporal coverage, spectroscopic observations remain crucial for physical interpretation and source classification. However, existing spectral analysis methods -- often reliant on template fitting or parametric models -- are limited in their ability to capture the complex and evolving spectra characteristic of such sources, which are sometimes only available at low resolution. In this work, we introduce SpectraNet, a deep convolutional neural network designed to learn robust representations of optical spectra from transients. Our model combines multi-scale convolution kernels and multi-scale pooling to extract features from preprocessed spectra in a hierarchical and interpretable manner. We train and validate SpectraNet on low-resolution time-series spectra obtained from the Spectral Energy Distribution Machine (SEDM) and other instruments, demonstrating state-of-the-art performance in classification. Furthermore, in redshift prediction tasks, SpectraNet achieves a root mean squared relative redshift error of 0.02, highlighting its effectiveness in precise regression tasks as well.

AppleCiDEr II: SpectraNet -- A Deep Learning Network for Spectroscopic Data

TL;DR

SpectraNet addresses the need for robust spectral classification of time-domain transients by learning multiscale spectral representations through a CNN architecture integrated into the AppleCiDEr multimodal pipeline. It combines rest-frame spectral preprocessing, a redshift regression head, and a five-block multi-scale CNN with a 1×1 projection, optimized with class-balanced focal loss and training techniques like EMA and mixed precision. The approach yields state-of-the-art classification performance and precise redshift estimates, with strong generalization to narrower wavelength coverage as demonstrated on NGPS data. This work enables real-time, spectroscopically informed transient classification and paves the way for LSST-era deployment with domain adaptation and evolution-aware extensions.

Abstract

Time-domain surveys such as the Zwicky Transient Facility (ZTF) have opened a new frontier in the discovery and characterization of transients. While photometric light curves provide broad temporal coverage, spectroscopic observations remain crucial for physical interpretation and source classification. However, existing spectral analysis methods -- often reliant on template fitting or parametric models -- are limited in their ability to capture the complex and evolving spectra characteristic of such sources, which are sometimes only available at low resolution. In this work, we introduce SpectraNet, a deep convolutional neural network designed to learn robust representations of optical spectra from transients. Our model combines multi-scale convolution kernels and multi-scale pooling to extract features from preprocessed spectra in a hierarchical and interpretable manner. We train and validate SpectraNet on low-resolution time-series spectra obtained from the Spectral Energy Distribution Machine (SEDM) and other instruments, demonstrating state-of-the-art performance in classification. Furthermore, in redshift prediction tasks, SpectraNet achieves a root mean squared relative redshift error of 0.02, highlighting its effectiveness in precise regression tasks as well.

Paper Structure

This paper contains 21 sections, 7 equations, 9 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Dataset Description
  3. Spectral Preprocessing and Calibration
  4. Data Splitting Strategy
  5. Data Augmentation
  6. Model Architecture and Training Strategy
  7. Process Explanation
  8. Training Configuration
  9. Ablation Study
  10. Design Choices: Kernel Size and Channel Expansion
  11. Kernel Size
  12. Channel Width
  13. Normalization
  14. Spectra close to the photometric peak Vs multiple epochs
  15. Comparison across other works
  16. ...and 6 more sections

Figures (9)

  • Figure 1: Redshift distribution of the transient objects in the final dataset. The histogram is truncated at the 99th percentile to improve visibility at the low-$z$ range.
  • Figure 2: An examples of spectra representation before (gray) and after (blue) preprocessing. This corresponds to "ZTF19abjpelp" case.
  • Figure 3: Schematic architecture of SpectraNet. The green block represents the SpectraBlock, and the red line in the last block indicates that no pooling and downsampling are implemented. The three kernel sizes correspond to small, medium and large scale CNN-1D, presented in Fig. \ref{['fig:block']}.
  • Figure 4: Schematic architecture of the SpectraBlock.
  • Figure 5: Comparison of AppleCiDEr performance (a) with and (b) without spectra information. The evaluated spectra are close to the photometric peak.
  • ...and 4 more figures