A Hybrid Framework for Kilonova Anomaly Detection using Single-Epoch SEDs from the 7-Dimensional Telescope

1. Problem Statement

The advent of the Vera C. Rubin Observatory's Legacy Survey of Space and Time (LSST) will generate $\sim 10^7$ transient alerts per night, creating a critical bottleneck in identifying rare, scientifically valuable events like Kilonovae (KNe). KNe are electromagnetic counterparts to gravitational wave (GW) events (e.g., binary neutron star mergers) but are intrinsically faint, rapidly fading, and difficult to distinguish from a vast background of "impostors" such as Supernovae (SNe), Active Galactic Nuclei (AGN), and stellar variables.

Current classification methods face significant limitations:

• Spectroscopy: While reliable, it requires high signal-to-noise ratios and is often impractical for faint, fast-evolving sources due to telescope time constraints.
• Multi-epoch Photometry: Relies on light-curve evolution, which introduces latency (days to weeks) before a classification can be made, missing the critical early phase of KNe.
• Single-Epoch Limitations: Existing low-resolution spectro-photometric approaches often lack the ability to handle diverse transient types simultaneously or detect KNe without direct training data (due to the scarcity of observed KNe spectra).

The paper addresses the need for a rapid, single-epoch classification framework capable of distinguishing KNe from common transients using medium-band spectral energy distributions (SEDs) to facilitate immediate follow-up for GW events.

2. Methodology

The authors propose a Hybrid Classification Framework specifically designed for the 7-Dimensional Telescope (7DT), a unique array of 20 simultaneously operating 0.5-m telescopes in Chile.

A. Data Generation and Simulation

• 7DT Simulator: A dedicated tool (7DT-Simulator) was developed to generate realistic synthetic photometry. It convolves input spectra with the telescope's optical throughput, atmospheric transmission (Cerro Paranal model), and the specific response of 20 or 40 medium-band filters (spanning 400–887 nm).
• Training Data:
  • Non-KN Classes: Archival spectra from the Open Supernova Catalog (OSC) and WISeREP were used for eight classes: Type Ia, Ibc, and II SNe, Superluminous SNe (SLSNe), Tidal Disruption Events (TDEs), AGN, Stellar Variables (SVs), and Asteroids.
  • KNe (Anomalies): No KNe were used for training. Instead, a grid of simulated KNe based on the R. T. Wollaeger et al. (2021) radiative-transfer models (varying ejecta mass, velocity, and viewing angle) was used for testing.
  • Real Test Data: Observed 7DT data for SN 2025fvw (Type Ia), SN 2024diq (Type II), and AT 2024ett (CV) were used for validation.
• Feature Engineering: Instead of raw fluxes, the model uses color indices (magnitude differences between all filter pairs) as input features. This approach captures spectral slopes and diagnostic features while being robust to non-detections (masked as NaN).

B. The Hybrid Framework

The framework consists of two parallel components:

1. Supervised Multi-Class Classifier (XGBoost):
   • Trained to classify the eight common non-KN transient types.
   • Uses XGBoost (selected over LightGBM and CatBoost for its robustness and SHAP compatibility).
   • Outputs a maximum class probability ($P_{max}$) and a classification uncertainty metric ($1 - P_{max}$).
2. Unsupervised Anomaly Detector (Isolation Forest):
   • Trained exclusively on non-KN data.
   • Identifies KNe as "out-of-distribution" anomalies without needing labeled KNe examples.
   • Outputs an anomaly score ($P_{ano}$).

C. Decision Logic

A joint decision rule combines the two scores:

• An event is flagged as a KNe candidate only if it exceeds thresholds for both the anomaly score ($P_{ano}$) and the classification uncertainty ($1 - P_{max}$).
• Thresholds are optimized using Youden's J statistic to maximize the trade-off between recall and false positive rate.

3. Key Contributions

• First Single-Epoch KNe Detection: Demonstrates that KNe can be reliably identified as anomalies from a single-epoch medium-band SED, eliminating the latency of multi-epoch light-curve monitoring.
• Hybrid Architecture: Successfully integrates supervised learning (for common transients) with unsupervised anomaly detection (for rare, under-sampled classes like KNe), mitigating the risk of overfitting to scarce KNe data.
• Filter Optimization via SHAP: Utilized SHapley Additive exPlanations (SHAP) to quantify filter importance, revealing that ~40–50% of the most informative filters (primarily those covering H$\alpha$ and Si II features) are sufficient to maintain near-baseline performance.
• Synergy with LSST: Showed that combining a subset of 7DT medium-band filters with a single LSST broad-band filter (e.g., $u$ or $y$) reproduces full-model accuracy, offering a flexible strategy for resource-constrained follow-up.

4. Key Results

• Classification Performance:
  • The supervised XGBoost classifier achieved a macro F1-score of ~0.80 (20 filters) and ~0.82 (40 filters) across eight transient classes.
  • Type Ia SNe were classified with high robustness (F1 $\approx$ 0.87–0.90).
  • Confusion was observed between Type Ibc/II SNe and between SLSNe and Stellar Variables, typical for low-resolution ($R \sim 30-70$) data.
• Anomaly Detection (KNe Recovery):
  • The hybrid framework recovered >90% of simulated and observed (AT 2017gfo) optically detectable KNe.
  • The false positive rate (contamination from non-KN transients) was reduced to ~10% (down from ~50% if using anomaly detection alone).
  • AT 2017gfo Validation: Most epochs were correctly flagged as anomalies. The earliest epoch (+1.43 days) was misclassified as an asteroid due to its featureless, blue continuum, but this was identified as a known limitation of early-phase SEDs.
• Filter Importance:
  • Red-end filters (>700 nm) contributed little to classification due to lower instrumental sensitivity.
  • Removing the bottom 50% of filters (retaining the top 10–20) resulted in negligible performance loss, suggesting operational flexibility.
• Real-World Validation:
  • Applied to real 7DT observations of SN 2025fvw, SN 2024diq, and AT 2024ett, the framework correctly identified them as non-anomalous (normal transients), preventing false alarms.

5. Significance and Implications

• GW Follow-up: This framework provides a scalable, rapid triage tool for the Rubin era, capable of filtering millions of alerts to isolate KNe candidates within minutes of detection, crucial for multi-messenger astronomy.
• Operational Efficiency: The finding that a reduced filter set (or a hybrid 7DT+LSST setup) maintains high accuracy allows for optimized observing strategies, such as increasing survey area or exposure depth when full spectral coverage is not strictly necessary.
• Interpretability: The use of SHAP values ensures the model's decisions are physically interpretable, linking classification weights to known diagnostic features (e.g., H$\alpha$, Si II) rather than acting as a "black box."
• Limitations & Future Work:
  • The model is currently biased toward low-redshift ($z < 0.15$) and bright events; performance at high redshift or for faint, lanthanide-rich KNe requires further study.
  • Early-phase KNe (<1.5 days) remain challenging to distinguish from hot, featureless impostors (e.g., early SNe, asteroids), necessitating contextual checks (e.g., motion tracking, host galaxy association).
  • Future iterations should incorporate temporal evolution (multi-epoch data) and host-galaxy priors to further reduce degeneracies.

In conclusion, the paper establishes that medium-band, single-epoch SEDs are a powerful tool for transient classification, and the proposed hybrid supervised-unsupervised framework offers a robust, interpretable, and operationally viable solution for the upcoming flood of time-domain data.