An Optically Motivated Gamma-ray Study of Fermi-LAT Novae

The study investigates the relationship between optical and gamma-ray emission from nova eruptions, motivated by the theory that gamma-rays from these systems are generated in collisionless non-relativistic shocks, and that a portion of the optical luminosity is reprocessed shock power. The authors analyze all 26 novae detected by Fermi-LAT from August 2008 through June 2024, using a standard maximum likelihood analysis with the fermipy package. They explore a range of time bin sizes (tγ) for each source, defining t∗γ as the time bin that maximizes the detection significance (Test Statistic, TS). They find that "across our population, a nova’s t3 (the time it takes the nova V band brightness to decay 3 magnitudes) appears to be the favored analysis bin that optimizes the Fermi-LAT detection significance, although there is significant spread. The distribution of optical magnitude drops corresponding to t∗γ peaks for a brightness drop of ∼ 3 magnitudes, but the spread is substantial," with first and third quartiles of N ≈ 2 and 4, respectively.

The authors developed an open-source machine learning-based tool called nova-times, in collaboration with the LSST-DA Project Dovetail, to measure tN decay times directly from AAVSO optical lightcurves using a gradient boosted machine. They report that t3 and t4 have similar root mean square distances between their respective maxima and the 1:1 line, while the t2 results are farther from the 1:1 line when comparing tγ against measured optical decay times.

For the 11 novae within 0.5◦ of an unidentified 4FGL point source, they performed off-peak analyses to determine whether these sources should be associated with the novae. They report that three sources (V1324 Sco, V5855 Sgr, and V549 Vel) should be associated with a coincident 4FGL source based on the lack of significance outside the eruption window, and recommend removing these 4FGL sources from the model during the eruption window.

The authors also report that "V679 Car, a source previously noted as a marginally detected γ-ray nova, as a > 5σ detection in this work." Four other marginally detected sources (V6598 Sgr, V407 Lup, V1535 Sco, and V745 Sco) have gamma-ray significance at the 2σ level.

They performed cross-correlation analysis of gamma-ray and optical lightcurves for 18 novae with TS > 30, using three methods (lightcurves, first-order finite differences, and second-order finite differences). They recover the anticipated lag of 0 days for most novae, albeit with considerable uncertainty. However, for RS Oph, V1723 Sco, V357 Mus, and V5856 Sgr, they measure lags of ∼ 1 day, which might deviate from zero-lag due to fundamental differences in system makeup, such as RS Oph being a symbiotic nova with a red giant companion where ejecta interact with a more extended circumstellar environment. For V1674 Her, the first-order finite difference cross-correlation test yields an optical lag of 4.56 hours, but the other two tests show no lag.

The authors conclude that the optimized window to observe a Fermi-LAT nova eruption (t∗γ) varies widely across the population, and they propose that the optimal analysis window can be based on the optical lightcurve, supported by tight correlations we measure between the optical and γ-ray lightcurves.

Improvements to AI Systems Based on This Paper:

1. Adaptive Time-Window Selection for Transient Detection

• Improvement: Train an AI model to predict the optimal Fermi-LAT analysis bin (t*γ) for a nova using optical lightcurve features (e.g., decay rate, peak brightness, t2/t3/t4 values) instead of fixed or arbitrary bins.

• Capability: The AI can automatically select the best temporal bin for gamma-ray detection in real-time or archival data, maximizing signal-to-noise for new transients without manual tuning.

2. Multiwavelength Cross-Correlation with Uncertainty-Aware Lag Estimation

• Improvement: Implement a deep learning model that jointly processes optical and gamma-ray lightcurves, using Bayesian or Monte Carlo dropout methods to output lag distributions (e.g., 0-day lag vs. 1-day lag) with credible intervals, rather than point estimates.

• Capability: The AI can robustly identify physical lags (e.g., 1 day for symbiotic novae) and flag systems with unusual circumstellar environments, even with sparse or noisy data.

3. Automated Source Association and Contamination Removal

• Improvement: Build a classifier that uses off-peak gamma-ray significance, spatial coincidence, and optical eruption timing to decide whether a 4FGL source should be associated with a nova or removed from the model during the eruption window.

• Capability: The AI can autonomously clean gamma-ray models for upcoming surveys (e.g., LSST + Fermi synergies), preventing false associations and improving flux measurements for both the nova and background sources.

4. Machine-Learning-Based Lightcurve Decay-Time Estimator (Nova-Times Upgrade)

• Improvement: Extend the gradient-boosted machine to predict tN (N=2,3,4) with uncertainty bounds, and integrate it into a pipeline that feeds directly into the optimal bin selection (improvement #1).

• Capability: The AI can provide real-time decay-time estimates from streaming optical data (e.g., from LSST alerts), enabling immediate gamma-ray follow-up scheduling and prioritization.

5. Population-Level Outlier Detection for Shock Physics

• Improvement: Train an unsupervised anomaly detector on the joint distribution of optical decay times, gamma-ray significance, and cross-correlation lags to identify novae with atypical shock behavior (e.g., V1674 Her’s 4.56-hour lag).

• Capability: The AI can flag rare subclasses of novae for detailed theoretical modeling, improving our understanding of particle acceleration and shock physics in different progenitor systems.

6. Predictive Model for Gamma-Ray Detectability

• Improvement: Use the measured t*γ distribution (peaking at 3 mag drop, with quartiles 2–4) to train a regression model that predicts the probability of a nova being detected by Fermi-LAT (>5σ) given its early optical lightcurve.

• Capability: The AI can prioritize targets for gamma-ray observatories (e.g., CTA, Swift) by estimating detectability within the first days of eruption, optimizing telescope time allocation.

7. Time-Series Feature Fusion for Multi-Messenger Alerts

• Improvement: Develop a transformer-based model that fuses optical, gamma-ray, and (future) neutrino data, using attention to learn cross-band temporal relationships and automatically output a lag spectrum across multiple timescales.

• Capability: The AI can generate real-time multi-messenger alerts for novae, distinguishing between prompt (0-day) and delayed (1-day) emission mechanisms, and trigger follow-up observations across observatories.

High-energy (GeV) gamma-ray emission from nova eruptions was a surprise detection by the Fermi-LAT in 2010. Since then, it has been suggested that the gamma-rays from these systems are generated in collisionless non-relativistic shocks. In this theory, the observed correlation between optical and gamma-ray nova lightcurves naturally arises if a portion of the optical luminosity is reprocessed shock power. In this work, we investigate this scenario by analyzing Fermi-LAT-detected novae in varied time bins and then correlating the bin size that maximizes the nova's significance (tgam) with the nova's optical eruption data. Furthermore, we investigate whether there is a common optical decay across the sources' measured tgam values in the population. We find that across our population, a nova's t 3 (the time it takes the nova V band brightness to decay 3 magnitudes) appears to be the favored analysis bin that optimizes the Fermi-LAT detection significance, although there is significant spread. Additionally, we report V679 Car, a source previously noted as a marginally detected gamma-ray nova, as a >5 sigma detection in this work. We perform cross-correlation analysis of gamma-ray and optical lightcurves.