LensNet: Enhancing Real-time Microlensing Event Discovery with Recurrent Neural Networks in the Korea Microlensing Telescope Network

TL;DR

LensNet presents a branched Recurrent Neural Network pipeline tailored to the Korea Microlensing Telescope Network (KMTNet) for real-time microlensing event discovery. By processing time-series flux and contextual features from three telescopes and leveraging data augmentation, time-relativization, and careful preprocessing, LensNet achieves strong binary (≈87.5% test accuracy) and robust multi-class (≈78%) performance while tolerating partial alert visibility. The system operates downstream of AlertFinder and automated vetting to reduce manual review workloads, with threshold tuning enabling high-purity or higher-recall alerting. These results demonstrate a practical path toward real-time, scalable microlensing alerts that can accelerate exoplanet discoveries and inform deployment in future surveys, including space-based missions.

Abstract

Traditional microlensing event vetting methods require highly trained human experts, and the process is both complex and time-consuming. This reliance on manual inspection often leads to inefficiencies and constrains the ability to scale for widespread exoplanet detection, ultimately hindering discovery rates. To address the limits of traditional microlensing event vetting, we have developed LensNet, a machine learning pipeline specifically designed to distinguish legitimate microlensing events from false positives caused by instrumental artifacts, such as pixel bleed trails and diffraction spikes. Our system operates in conjunction with a preliminary algorithm that detects increasing trends in flux. These flagged instances are then passed to LensNet for further classification, allowing for timely alerts and follow-up observations. Tailored for the multi-observatory setup of the Korea Microlensing Telescope Network (KMTNet) and trained on a rich dataset of manually classified events, LensNet is optimized for early detection and warning of microlensing occurrences, enabling astronomers to organize follow-up observations promptly. The internal model of the pipeline employs a multi-branch Recurrent Neural Network (RNN) architecture that evaluates time-series flux data with contextual information, including sky background, the full width at half maximum of the target star, flux errors, PSF quality flags, and air mass for each observation. We demonstrate a classification accuracy above 87.5%, and anticipate further improvements as we expand our training set and continue to refine the algorithm.

Figures (12)

• Figure 1: Illustration of the gravitational microlensing phenomenon used to detect exoplanets. As a foreground lens star passes in front of a more distant background star, the gravitational field of the lens star warps spacetime, magnifying the light from the background star. The diagram shows three stages of this event, where the lens star and its exoplanet's relative positions cause distinct distortions in the background star's light, as depicted by the warped grid lines. The resulting light curve, displayed in the inset, features two peaks: a primary one due to the lens star and a smaller secondary peak caused by the exoplanet.
• Figure 2: Comparison of the current and future pipelines for the KMTNet. The current pipeline, which processes around 5,000 alerts per day, involves manual vetting of light curves and difference images by co-author KHH. This process reduces the number of alerts to around 20 per day. The future pipeline, incorporating LensNet and additional work from deBeurs_inprep, will process a higher volume of alerts—around 20,000 per day—while reducing the number of manually reviewed alerts.
• Figure 3: Map of the KMTNet BLG observing fields, showing the distribution of the four CCDs per field, with each field color-coded according to its observational cadence, ranging from 0.5 hours (purple) to 5.0 hours (yellow). The fields are plotted in galactic coordinates, longitude ($\ell$) and latitude ($b$). Credit: Matthew Penny.
• Figure 4: Demonstration of the process to identify potential microlensing events by analyzing the light curves from the three different telescopes: CTIO, SAAO, and SSO. The point $t_{\text{break}}$ represents the time at which the algorithm identifies the start of the rising event, while $t_{\text{cut}}$ is the moment at which the data was evaluated. The point $t_0$ corresponds to the peak of the microlensing event assuming it is a genuine event. In this particular case it was the CTIO telescope that triggered the alert, and thus the specific time points $t_{\text{break}}$, $t_{\text{cut}}$, and $t_0$ refer to the CTIO's time series data. The "unalerted" data represents the baseline flux before any indication of a microlensing event, while the "alerted" data shows a noticeable increase in flux. The fitted line over the alerted data highlights the trend in flux increase, which is used to detect the alert. The figure captures the progression of the event from its real beginning ($t_{\text{begin}}$) to its end ($t_{\text{end}}$). This figure also shows how the microlensing pattern can be "seen" in all the three telescopes, despite the fact that is only alerted from the data of one of them.
• Figure 5: Examples of a real microlensing event (top) compared to two common types of false positives: a bleed trail (center) and a diffraction spike (bottom). Left panels show the original images, and the right panels show the difference images. A magenta circle indicates the location of the catalog star on which the candidate was detected. White arrows indicate the relevant feature causing the false positive.