[Paper Review] Identifying microlensing events using neural networks

P. Mróz|CaltechAUTHORS (California Institute of Technology)|Aug 27, 2020

Stellar, planetary, and galactic studies3 references10 citations

TL;DR

This paper proposes two deep learning classifiers—based on fully connected neural networks and convolutional neural networks (CNNs)—to automatically detect single- and binary-lens microlensing events in photometric light curves. Trained on OGLE-III and OGLE-IV data, the models achieve ~98% recall for single-lens events and 80–85% for binary-lens events, demonstrating robust performance on ZTF data and strong potential for future space-based surveys like the Roman Space Telescope.

ABSTRACT

Current gravitational microlensing surveys are observing hundreds of millions of stars in the Galactic bulge - which makes finding rare microlensing events a challenging tasks. In almost all previous works, microlensing events have been detected either by applying very strict selection cuts or manually inspecting tens of thousands of light curves. However, the number of microlensing events expected in the future space-based microlensing experiments forces us to consider fully-automated approaches. They are especially important for selecting binary-lens events that often exhibit complex light curve morphologies and are otherwise difficult to find. There are no dedicated selection algorithms for binary-lens events in the literature, which hampers their statistical studies. Here, we present two simple neural-network-based classifiers for detecting single and binary microlensing events. We demonstrate their robustness using OGLE-III and OGLE-IV data sets and show they perform well on microlensing events detected in data from the Zwicky Transient Facility (ZTF). Classifiers are able to correctly recognize ~98% of single-lens events and 80-85% of binary-lens events.

Motivation & Objective

• To address the challenge of identifying rare microlensing events in large-scale photometric surveys, where manual inspection is infeasible.
• To develop automated, scalable classifiers for detecting both single-lens and complex binary-lens microlensing events, which are difficult to identify with traditional selection cuts.
• To create a robust, transferable machine learning framework that generalizes across different survey cadences and noise characteristics, including ground-based (OGLE, ZTF) and future space-based (Roman) data.
• To enable unbiased, high-completeness samples of microlensing events for statistical studies, especially for binary-lens systems that lack dedicated detection algorithms.
• To demonstrate that deep learning can outperform or match traditional methods in detecting microlensing events while reducing human bias and labor.

Proposed method

• A fully connected neural network is trained on 15 hand-crafted light curve features (e.g., peak magnitude, duration, χ² per degree of freedom) to classify single-lens microlensing events.
• A convolutional neural network (CNN) is applied to 2D representations of light curves, constructed by stacking phased light curves over time, to detect both single- and binary-lens events.
• The 2D light curve representation is designed to preserve temporal structure and is robust to gaps and variable cadence, though performance is optimized under uniform cadence, such as expected from the Roman Space Telescope.
• The models are trained on real OGLE-III and OGLE-IV light curves, including genuine microlensing events and non-microlensing artifacts, to improve generalization and avoid overfitting to synthetic data.
• Training uses a balanced dataset with real microlensing events and non-microlensing light curves to ensure high recall and robustness to noise and observational systematics.
• Performance is evaluated on independent test sets from OGLE-IV and ZTF, with metrics including recall, precision, and F1-score to assess detection efficiency and purity.

Experimental results

Research questions

• RQ1 Can deep learning classifiers detect single-lens microlensing events with higher completeness than traditional selection cuts, while maintaining acceptable purity?
• RQ2 Can a single neural network model detect both single-lens and binary-lens microlensing events with comparable performance, despite the morphological complexity of binary events?
• RQ3 How well do these classifiers generalize across different photometric surveys with varying cadence, noise, and data quality (e.g., OGLE vs. ZTF)?
• RQ4 Can a 2D CNN representation of light curves effectively capture microlensing signatures, especially in the presence of gaps and non-uniform sampling?
• RQ5 To what extent can these models outperform human inspection or classical statistical filters in identifying rare microlensing events, particularly binary-lens systems?

Key findings

• The fully connected neural network classifier achieves a recall of 97.5–98.0% for single-lens microlensing events on independent test sets, outperforming traditional methods in completeness.
• The binary-lens classifier achieves a recall of 80–85% on test sets, demonstrating the feasibility of automated detection for complex light curve morphologies.
• The CNN-based classifier achieves a recall of 96.7% for single-lens events and 84.1% for binary-lens events on the same test set, showing strong generalization across datasets.
• The models generalize well to ZTF data, correctly identifying microlensing events despite differences in cadence and noise structure compared to OGLE data.
• The classifiers maintain high performance even when trained on real data including artifacts, outperforming models trained on synthetic light curves due to preserved noise characteristics.
• The study demonstrates that deep learning can achieve performance comparable to or better than human experts in identifying microlensing events, especially for binary-lens systems where traditional methods fail.