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

Abstract

We develop a hybrid framework to identify kilonovae (KNe), using single-epoch, medium-band spectral energy distributions from the 7-Dimensional Telescope (7DT). The framework integrates an unsupervised anomaly classifier (\texttt{Isolation Forest}) to flag unusual events with a supervised multi-class classifier (\texttt{XGBoost}) that characterizes eight common transient types. Trained on realistically simulated 7DT photometry accounting for per-filter sensitivity, the classifier achieves macro $F_{1}\sim0.80$ ($\sim0.82$) with 20 (40) filters across eight classes, Type~Ia/Ibc/II SNe, SLSNe, TDEs, AGN, stellar variables, and asteroids. Without direct training, the anomaly detector recovers $>$90\% of simulated and observed optically detectable KNe (AT~2017gfo) with a low contamination fraction, with a caveat of limitations of the training sample such as limited redshift range of SNe ($z < 0.15$), and a relatively small number of early non-KNe spectra. A SHAP-based feature analysis reveals that only $\sim$40--50\% of the most informative filters are sufficient to retain near-baseline performance, while red-end filters contribute little. Combining the top-ranked half of the 40 7DT filters with a single LSST band reproduces the full-model accuracy within 1--2\%, suggesting practical follow-up strategies. These results demonstrate that 7DT's medium-band system enables rapid, interpretable classifications and reliable anomaly alerts from single-epoch data -- promising for gravitational-wave follow-up, Rubin alert stream filtering, and serendipitous transient discovery in the 7DT survey.

Figures (24)

• Figure 1: System response and $5\sigma$ point-source depth for the 7DT medium-band set, computed with 7DT-Simulator. Top: Light-colored curves show the nominal filter transmission profiles (top-hat-shaped). The gray, blue, and purple curves denote the detector QE (CMOS), atmospheric transmission, and optics throughput, respectively. Rainbow-colored curves are the final system responses (filter $\times$ optics $\times$ atmosphere $\times$ QE). Bottom: Colored triangles mark the $5\sigma$ depths (AB mag) for individual medium bands.
• Figure 2: Examples of synthetic photometry (markers with connecting lines) and corresponding spectra generated by the 7DT-Simulator using the 40 medium-band filter set of 7DT. Each spectrum is normalized to zero magnitude at 550 nm and vertically offset and denoted their name, type, approximate phase after discovery date. The KN spectrum is computed from the ENGRAVE observation of AT 2017gfo 2022MNRAS.515..631G2024MNRAS.529.2918G, while the asteroid spectrum corresponds to a C-type object based on the Bus--DeMeo taxonomy template 2009Icar..202..160D. All other transients are derived from spectral templates obtained via the OSC or WISeREP. For each transient type, one representative spectrum was randomly selected from the available templates to visually illustrate its characteristic spectral features.
• Figure 3: Redshift distributions for SN subclasses (Type II, Type Ibc, Type Ia, and SLSN) with valid redshift metadata. Left: Class-wise histograms of $\log_{10}(z)$; the dashed gray line shows the distribution for the combined sample. Only objects with usable redshift entries were included in each histogram. Right: Empirical cumulative distribution of $\log_{10}(z)$ for the same sample. The red star marks the 90th percentile ($z_{90}\approx0.1467$), indicating that $\sim$90% of the usable sample lies at $z<0.15$.
• Figure 4: Detection efficiency of simulated KNe with the 7DT as a function of model parameters. Bars show the number of simulated samples per parameter bin (left axis), while black squares and lines indicate the corresponding detection fractions (right axis). A KN is considered detected if at least one medium-band filter achieves a $5\sigma$ detection in the simulated 7DT observation. The parameters shown are the dynamical ejecta mass ($m_{d}$), dynamical ejecta velocity ($v_{d}$), wind ejecta mass ($m_{w}$), wind ejecta velocity ($v_{w}$), viewing angle ($\theta$), and phase after merger. Detection fractions are computed only for simulated KN that are detectable under the assumed 7DT sensitivity.
• Figure 5: Transient SEDs observed by 7DT in 20 filter set. The stacked curves at the bottom trace the temporal evolution of the Type Ia SN, SN 2025fvw over $\sim$3–42 days post-discovery (2025 March 26 19:09:59; 2025TNSTR1151....1I); points are shown with dot markers connected by solid lines, and magnitudes are vertically offset for clarity (AB system; the $x$-axis is the filter central wavelength in nm). The three curves at the top are single-epoch SEDs for real test samples: the Type II SN, SN 2024diq at 2.0 days post-discovery (2024 February 28 08:09:46.368; 2024TNSCR.615....1C2024TNSCR.614....1D), plotted with square markers connected by dashed lines; and the CV, AT 2024ett at 2.5 and 4.6 days post-discovery (2024 March 19 11:40:04.224; 2024TNSCR.896....1D), plotted with square (2.5 d) and diamond (4.6 d) markers connected by dotted lines. The corresponding photometry is summarized in Table \ref{['tab:7dt_val']}.