🔭 astrophysics

SNID-SAGE: A Modern Framework for Interactive Supernova Classification and Spectral Analysis

The paper introduces SNID-SAGE, a modern framework combining an interactive graphical interface and a scriptable pipeline to perform automated supernova spectral classification and redshift estimation using FFT-based cross-correlation against a curated library of approximately 6,000 templates, which has been validated through cross-validation and large-scale testing on nearly 46,000 WISeREP spectra.

Original authors: Fiorenzo Stoppa, Stephen J. Smartt

Published 2026-03-31

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Fiorenzo Stoppa, Stephen J. Smartt

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the night sky as a massive, chaotic library where new "books" (exploding stars called supernovae) are being discovered every single night. The problem? We have too many books and not enough librarians to read them all. We need a way to quickly figure out what kind of book each new discovery is, who wrote it (which galaxy), and when it was published (how long ago the explosion happened).

Enter SNID-SAGE, a new digital tool created by astronomers at the University of Oxford to solve this problem. Think of it as a super-smart, automated librarian that can read a blurry, messy page from a new book and instantly tell you exactly what it is.

Here is how it works, broken down into simple concepts:

1. The "Fingerprint" Match (Cross-Correlation)

Every type of exploding star leaves a unique "fingerprint" on the light it emits. This fingerprint looks like a series of peaks and valleys on a graph.

The Old Way: Imagine trying to match a new fingerprint against a wall of 6,000 old fingerprints by squinting and guessing. It's slow and prone to human error.
The SNID-SAGE Way: This tool uses a high-speed "sliding puzzle" technique. It takes the new, messy fingerprint and slides it back and forth against its massive library of 6,000 known fingerprints. It doesn't just look for a perfect match; it calculates how well the lines line up, even if the new one is a bit blurry or shifted in time.

2. The "Quality Score" (The New Metric)

In the past, these tools would sometimes get tricked. If a fingerprint had a few sharp lines that happened to match by accident, the tool might say, "This is a perfect match!" even if the rest of the picture was wrong.

The Upgrade: SNID-SAGE adds a "skeptical detective" step. It asks two new questions:
1. How sharp is the match? (Is the alignment tight, or is it sloppy?)
2. Do the heights of the lines match? (If the new fingerprint is much brighter or dimmer than the template, the tool flags it as a "maybe," not a "definitely.")
The Result: It gives every match a score. If the score is low, it says, "I'm not sure about this." If it's high, it says, "I'm confident."

3. The "Group Hug" (Clustering)

Sometimes, a new star might look a little like a Type Ia supernova and a little like a Type Ib. The old tools would pick just one and ignore the rest.

The New Approach: SNID-SAGE gathers all the "maybe" matches and puts them in groups. It looks for clusters. If 50 different templates say "This is a Type Ia" and only 2 say "This is a Type Ib," the tool trusts the group. It combines the information from the whole crowd to give a final, more reliable answer, rather than relying on a single opinion.

4. The Two Faces of the Tool

The researchers built this tool to work in two different ways, depending on who is using it:

The Interactive Dashboard (The GUI): For human astronomers who want to look closely. It's like a high-tech microscope. You can zoom in, highlight specific lines, and even ask an AI assistant to write a plain-English summary of what the tool found (e.g., "This looks like a Type Ia supernova from 10 days after the explosion").
The Assembly Line (The Command Line): For massive surveys. Imagine a factory conveyor belt where thousands of spectra (light readings) are fed in, and the tool spits out classifications for all of them automatically, without a human needing to touch a single button.

5. Did It Work? (The Test Drive)

The team put SNID-SAGE through two tough tests:

The "Hide and Seek" Test: They took a known fingerprint, hid it in the library, and asked the tool to find it again. It got the type right 99% of the time for common stars and did a great job guessing the distance (redshift) and timing (phase).
The "Real World" Test: They fed it 46,000 real spectra from a public database (WISeREP). They compared the tool's distance estimates against the known distances of the galaxies hosting the stars. The results were incredibly accurate, proving that the tool works even on messy, real-world data.

Why Does This Matter?

Astronomy is moving into an era where we will discover millions of transient events. We can't have humans reading every single one. SNID-SAGE is the bridge that allows us to turn a flood of raw data into organized knowledge. It ensures that when a new, rare, or weird explosion happens, we can identify it quickly, understand what it is, and know exactly where to look for it in the universe.

In short: SNID-SAGE is the ultimate "spell-checker" for the universe's most dramatic explosions, helping us read the story of the cosmos faster and more accurately than ever before.

1. Problem Statement

The discovery rate of astronomical transients (supernovae, tidal disruption events, etc.) is accelerating due to wide-field time-domain surveys (e.g., LSST, ZTF, Rubin Observatory). This surge creates a bottleneck in spectroscopic follow-up and classification.

Scale and Heterogeneity: Current and future spectroscopic programs generate massive volumes of data with varying resolutions, sensitivities, and cadences.
Limitations of Existing Tools: Traditional tools like SNID (SuperNova IDentification) and machine learning classifiers (e.g., DASH) face challenges. SNID relies on a single best-matching template and uses the Tonry–Davis $R$ statistic, which can be unreliable for noise estimation and insensitive to flux amplitude mismatches. Machine learning methods often require retraining for new or rare transient types.
Need: A robust, scalable, and reproducible pipeline that can handle diverse transient classes, integrate new templates without retraining, and provide reliable confidence estimates for large-scale survey workflows.

2. Methodology: The SNID-SAGE Framework

SNID-SAGE (SuperNova IDentification—Spectral Analysis and Guided Exploration) is a Python-based framework offering both an interactive graphical interface (GUI) and a scriptable command-line pipeline. It retains the core FFT-based cross-correlation of SNID but introduces significant methodological upgrades.

A. Preprocessing

The pipeline ingests wavelength-calibrated 1D spectra and applies a deterministic sequence:

Quality Control: Automatic masking of cosmic rays and outliers; optional exclusion of telluric bands.
Rebinning: Interpolation onto a common logarithmic wavelength grid ( $\ln \lambda$ ) spanning 2500–10000 Å (1024 bins). This ensures cosmological redshift corresponds to an additive shift, enabling efficient FFT matching.
Continuum Removal: Fitting and dividing by a low-order spline to isolate line structures, followed by tapering with a cosine window to suppress Fourier artifacts.

B. Enhanced Cross-Correlation and Similarity Metrics

Instead of relying solely on the classical Tonry–Davis $R$ statistic, SNID-SAGE introduces a composite quality metric ( $H_{\sigma LAP-CCC}$ ) to address two specific limitations of previous methods:

Noise Estimation: The antisymmetric component of the correlation function (used in $R$ ) is not a reliable noise estimator. SNID-SAGE replaces this with a local scale parameter $\sigma$ derived from the FWHM of the correlation peak and the standard deviation of the residual spectrum (observed minus template).
Amplitude/Offset Sensitivity: Cross-correlation ignores flux scaling. SNID-SAGE incorporates the Lin Concordance Correlation Coefficient (CCC) to measure agreement in both correlation and flux scaling.

The final ranking metric is:
$H_{\sigma LAP-CCC} = \frac{h \cdot LAP \cdot CCC}{\sqrt{\sigma}}$
Where $h$ is peak height, $LAP$ is fractional wavelength overlap, and $\sigma$ is the sharpness-mismatch scale. Matches with scores $\leq 1.5$ are discarded.

C. Clustering-Based Inference

Rather than selecting a single "best" template, SNID-SAGE aggregates information from multiple matches:

Gaussian Mixture Modeling (GMM): For each transient type (e.g., Ia, II, Ib), redshift estimates from retained templates are clustered using a 1D GMM. The number of components is selected via the Bayesian Information Criterion (BIC).
Cluster Quality Score ( $Q$ ): A cluster's quality is defined by the mean score of its top 5 members, penalized if fewer than 5 templates support it.
Final Inference: The cluster with the highest $Q$ determines the primary type and subtype. Redshift ( $\hat{z}$ ) and phase ( $\hat{t}$ ) are calculated as weighted means of the contributing templates, where weights are exponential functions of the match scores.
Uncertainty Estimation: Internal scatter is calculated using an unbiased weighted dispersion, which is inflated if the solution relies on too few templates (low effective sample size).

D. Additional Features

Host Redshift Constraints: Users can fix or constrain the search range using known host-galaxy redshifts.
Line Identification: Interactive tools for measuring line velocities (e.g., H $\alpha$ , [O III]).
AI Summaries: An optional module using an LLM to generate natural-language reports of the classification results (strictly descriptive, not inferential).

3. Key Contributions

New Quality Metric: The introduction of $H_{\sigma LAP-CCC}$ significantly improves robustness against noise and flux mismatches compared to the traditional $R_{LAP}$ metric.
Clustering Architecture: Moving from a "single best match" to a "GMM clustering of multiple matches" provides more stable redshift and phase estimates and naturally handles degenerate solutions.
Dual Interface: A unified framework serving both human analysts (via GUI with line identification and AI summaries) and automated survey pipelines (via CLI).
Expanded Template Library: A curated library of ~6,000 spectra (698 templates) covering 15 transient types, including rare classes like TDEs, LFBOTs, and Kilonovae.

4. Results and Validation

The authors validated SNID-SAGE using two complementary approaches:

A. Leave-One-Out Cross-Validation (Template Library)

Type Recovery:
- Type Ia: Near-perfect Top-1 recovery (99.3%).
- Type II: Strong Top-1 recovery (~89%).
- Stripped-Envelope (Ib/Ic): Moderate Top-1 (78–80%) but very high Top-2/Top-3 recovery (>94%), indicating the clustering correctly identifies the type even if the exact subtype is ambiguous.
- Rare Classes: Lower Top-1 rates for SLSNe, TDEs, and Icn due to intrinsic heterogeneity and sparse templates, but often recovered in Top-2/3.
Redshift & Phase: Median absolute redshift errors are small ( $\sim 10^{-3}$ ). Phase errors are generally modest but larger for heterogeneous classes (e.g., TDEs).
Uncertainty Calibration: Redshift uncertainties are well-calibrated (residuals follow $N(0,1)$ ). Phase uncertainties are slightly conservative (pipeline overestimates error).

B. Large-Scale Application (WISeREP Database)

Dataset: Applied to ~46,000 spectra from WISeREP.
Validation Subset: 24,223 spectra with secure host associations and consistent redshifts (compared against the REGALADE catalogue).
Redshift Accuracy:
- High/Medium Quality Matches: Robust, nearly unbiased redshift recovery with minimal scatter.
- Low/Very Low Quality: Higher incidence of outliers, often due to low signal-to-noise or poor wavelength coverage.
- Instrument Bias: Spectra from lower-resolution instruments (e.g., Palomar P60/SEDM) showed slightly larger scatter and bias, particularly at $z \approx 0.06-0.08$ .
Conclusion: The pipeline performs robustly across diverse survey data, with the quality metric effectively flagging unreliable solutions.

5. Significance

SNID-SAGE represents a critical evolution in supernova spectral analysis, transitioning from manual or single-template matching to a statistically rigorous, clustering-based framework.

Scalability: It is designed to handle the data deluge from upcoming surveys (e.g., LSST), shifting the bottleneck from discovery to reproducible classification.
Reliability: By explicitly modeling uncertainty and using multiple templates, it reduces the risk of misclassification for rare or evolving transients.
Community Resource: The framework is open-source, includes a growing template library, and provides tools for both automated processing and interactive human-in-the-loop analysis.
Future-Proofing: The modular design allows for the easy addition of new transient types and spectral ranges (e.g., near-infrared) without retraining machine learning models.

In summary, SNID-SAGE provides a modern, robust, and flexible solution for the classification of astronomical transients, balancing the speed required for large surveys with the accuracy needed for scientific discovery.