Deep Learning Unresolved Lensed Lightcurves

TL;DR

This paper demonstrates that convolutional neural networks can accurately classify and estimate time delays in unresolved gravitational lens lightcurves, significantly improving speed and robustness over previous methods.

Contribution

The authors develop CNN-based methods for classifying unresolved lensed lightcurves and estimating time delays, achieving high accuracy and speed, especially for delays over 6 days.

Findings

100% accuracy in classifying image configurations for delays >6 days

Estimated time delays with ~1 day precision

Robust performance with 10% flux noise levels

Abstract

Gravitationally lensed sources may have unresolved or blended multiple images, and for time varying sources the lightcurves from individual images can overlap. We use convolutional neural nets to both classify the lightcurves as due to unlensed, double, or quad lensed sources and fit for the time delays. Focusing on lensed supernova systems with time delays $\Delta t\gtrsim6$ days, we achieve 100\% precision and recall in identifying the number of images and then estimating the time delays to $\sigma_{\Delta t}\approx1$ day, with a $1000\times$ speedup relative to our previous Monte Carlo technique. This also succeeds for flux noise levels $\sim10\%$. For $\Delta t\in[2,6]$ days we obtain 94--98\% accuracy, depending on image configuration. We also explore using partial lightcurves where observations only start near maximum light, without the rise time data, and quantify the success.