🔭 astrophysics

Microlensify: a Transformer Based Machine Learning Classifier for Microlensing Events Trained on TESS Light Curves

The paper introduces Microlensify, a physics-informed transformer-based variational autoencoder trained on TESS data that effectively identifies microlensing candidates, filters false positives, and accurately estimates event durations across diverse high-cadence surveys.

Original authors: Atousa Kalantari, Somayeh Khakpash, Sedighe Sajadian, Hosein Haghi, Willow Fox Fortino, Rosanne Di Stefano

Published 2026-08-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Atousa Kalantari, Somayeh Khakpash, Sedighe Sajadian, Hosein Haghi, Willow Fox Fortino, Rosanne Di Stefano

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the vast, crowded streets of our galaxy, stars are so numerous that they often line up perfectly from our perspective on Earth. When a foreground star passes directly in front of a more distant background star, its gravity acts like a natural lens, bending the light of the background star and making it appear temporarily brighter. This phenomenon, known as gravitational microlensing, is a powerful tool for astronomers because it reveals objects that are otherwise invisible, such as faint, isolated stars or even free-floating planets that do not orbit any sun. However, finding these events is like searching for a specific type of cloud in a stormy sky; the brightening is often subtle, short-lived, and easily confused with other types of stellar behavior, such as a star pulsating or a binary system eclipsing its partner. To find these rare signals, astronomers need to sift through millions of light curves—the graphs that track how a star's brightness changes over time—and distinguish the unique signature of a gravitational lens from the noise of the universe.

A team of researchers has developed a new, highly efficient method to solve this problem using a sophisticated computer program called Microlensify. Instead of relying on traditional, slow mathematical fitting methods that struggle with the sheer volume of data, this team trained an artificial intelligence system to recognize the specific shape of a microlensing event. The system is built on a "transformer" architecture, a type of machine learning model originally designed to understand language but here repurposed to understand the patterns of light over time. What makes this approach unique is that the scientists did not just feed it raw data; they built the laws of physics directly into the training process. This means the AI not only learns to spot the event but also understands the physical rules governing how the light should behave, allowing it to reconstruct the event's timeline and estimate how long the lensing lasted with remarkable accuracy.

The researchers tested this system on data from the Transiting Exoplanet Survey Satellite, or TESS, a space telescope that scans almost the entire sky. While TESS was primarily designed to find planets by watching them cross in front of stars, its high-speed observations make it an excellent, if underutilized, tool for catching these fleeting microlensing events. The team trained their model on a mix of simulated events, created to mimic what a microlensing signal looks like in TESS data, and real light curves from a specific sector of the sky that contains a high density of stars. They then applied the trained model to millions of light curves from different data processing pipelines. The results were striking: the system successfully identified thousands of potential candidates, filtering them down to a manageable list of high-confidence events. In one test, it found a known microlensing event that had been flagged by a different space mission, proving that the model could recognize the signal even when the data came from a different source or had different characteristics.

However, the power of this tool lies not just in what it finds, but in what it reveals about the challenges of modern astronomy. As the researchers applied the model to millions of stars, they uncovered a surprising source of false alarms. They discovered that small asteroids crossing in front of stars can create sharp, brief peaks in the light curves that look almost identical to microlensing events. This finding is crucial for future surveys, as it highlights a specific type of confusion that must be accounted for when hunting for these elusive objects. The model also proved effective at identifying other types of variable stars, such as long-period pulsating stars and cataclysmic binaries, which often masquerade as microlensing candidates. By cross-referencing its findings with astronomical databases, the team was able to categorize these impostors, refining the list of true candidates to include only those stars that have not yet been classified or are known to be stable.

The success of Microlensify extends beyond the TESS mission. The researchers tested the model on data from ground-based telescopes in Chile, South Africa, and New Zealand, which observe the sky with different speeds and sensitivities than the space telescope. Despite these differences, the model correctly identified the vast majority of known microlensing events from these ground-based surveys, demonstrating that it has learned the fundamental shape of the phenomenon rather than just memorizing the specific data it was trained on. This ability to generalize across different instruments suggests that the tool can serve as a universal first filter for future, even larger surveys, such as those planned for the Nancy Grace Roman Space Telescope. By rapidly screening millions of stars and flagging only the most promising candidates, Microlensify allows astronomers to focus their time and resources on the events that are most likely to reveal new secrets about the invisible population of objects in our galaxy.

The study confirms that while the search for microlensing events is complex, the combination of high-cadence space data and physics-informed machine learning offers a clear path forward. The model does not just classify events; it reconstructs the light curves and estimates the duration of the lensing with high precision, providing immediate physical insights. The researchers have made their software available to the broader scientific community, ensuring that this new way of seeing the universe can be used by others to continue the hunt for the faint, hidden objects that shape our understanding of the cosmos. Through this work, the team has not only found new candidates but has also mapped the landscape of false positives, turning a chaotic search into a more precise and manageable endeavor.

Technical Summary: Microlensify – A Transformer-Based Classifier for Microlensing Events in TESS Light Curves

Problem Statement
Gravitational microlensing is a critical method for detecting faint compact objects and exoplanets, yet identifying these events in high-cadence, all-sky surveys like the Transiting Exoplanet Survey Satellite (TESS) presents significant challenges. Microlensing events are rare, and distinguishing them from the vast array of astrophysical false positives (e.g., variable stars, transients) and instrumental systematics requires analyzing millions of light curves. Traditional detection methods, such as fitting point-source point-lens (PSPL) models or applying specific selection cuts (e.g., $\Delta\chi^2$ , Signal-to-Noise Ratio), are computationally expensive and often struggle to balance false positive rates with detection efficiency when scaled to millions of targets. Furthermore, existing machine learning approaches have often relied solely on simulated data or extracted features, potentially missing the complex noise characteristics and systematics inherent in real survey data.

Methodology
The authors introduce Microlensify, a Physics-Informed Transformer-Based Variational Autoencoder (PI-TVAE) designed to classify single-lens microlensing events, reconstruct light curves, and estimate event durations.

Data Preparation: The model was trained on a hybrid dataset comprising:
1. Simulated Data: $\sim$ 93,000 single-lens microlensing light curves generated via Monte Carlo simulations tailored to TESS Sector 12 characteristics (using the Besançon Galactic model for stellar parameters and VBBINARYLENSING for finite-source effects). These were filtered through rigorous quality cuts ( $\Delta\chi^2_{KMTNet} > 400$ , $SNR > 30$ , and skewness-von Neumann space constraints) to ensure high-confidence training samples.
2. Real Data: $\sim$ 17,500 non-microlensing light curves from TESS Sector 12, extracted via the GSFC-ELEANOR-LITE pipeline. This inclusion of real data allows the model to learn the specific systematics and noise profiles of TESS observations.
- Input Features: The model processes a fixed-length sequence of 940 normalized flux points. To preserve amplitude information lost during normalization, five statistical features (minimum, maximum, median, standard deviation, and the ratio of standard deviation to the range) are concatenated with the flux time series.
Architecture:
- Encoder: Utilizes parallel 1D convolutional layers (kernel sizes 3, 5, 7) to capture multi-scale patterns, followed by Transformer blocks with multi-head self-attention. The encoder outputs a 20-dimensional latent vector.
- Physics-Informed Latent Space: The latent space is constrained by a Variational Autoencoder framework. Crucially, one specific dimension is dedicated to encoding the event duration ( $4 \times \text{FWHM}$ ) via a physics-informed loss function that aligns the latent distribution with the true duration.
- Decoder: Reconstructs the light curve from the latent vector using Transformer blocks and transposed convolutions.
- Classification Head: A separate branch processes the latent features to output a binary probability for microlensing detection.
- Loss Function: A composite loss minimizes reconstruction error (BCE + MSE), Kullback-Leibler divergence (regularizing the latent space), a physics-informed duration loss, and binary cross-entropy for classification.
Preprocessing for Variable Cadence: To apply the model to data with different cadences (e.g., 10-minute or 200-second) or lengths, the authors implemented a chunking and downsampling strategy. Light curves are split into segments, downsampled to match the training input shape, and processed in parallel to ensure robustness across different survey configurations.

Key Contributions

Novel Architecture: The development of a Physics-Informed Transformer VAE that integrates physical constraints (event duration) directly into the latent space, improving the physical interpretability of the model's internal representations.
Real-World Training: Unlike many prior works relying solely on simulations, Microlensify is trained on real TESS light curves (Eleanor pipeline) to capture genuine false positives and systematics, enhancing its generalization capabilities.
Multi-Task Learning: The model simultaneously performs classification, light curve reconstruction, and event duration estimation, providing a comprehensive analysis tool rather than a simple binary classifier.
Cross-Survey Applicability: The framework is demonstrated to generalize effectively across different TESS pipelines (Eleanor, SPOC, QLP) and independent ground-based surveys (OGLE, KMTNet, MOA) with varying cadences and photometric characteristics.

Results

Performance on TESS Sector 12: Applied to $\sim$ 1.5 million Eleanor light curves, Microlensify identified 545 candidates (0.036%). After applying traditional detection metrics ( $\Delta\chi^2$ , SNR, skewness-von Neumann) and visual inspection, 328 high-confidence candidates remained. Similar analyses on SPOC (133k curves) and QLP (2.86M curves) sectors yielded 69 and 3,554 final candidates, respectively.
False Positive Identification: The study successfully identified various false positives, including long-period variables, Mira variables, cataclysmic variables, red giants, and a specific class of false positives caused by asteroid crossings (producing Gaussian-like peaks).
Duration Estimation: The model predicts event durations with high accuracy ( $R^2 = 0.97$ , MAE = 1.22 days) on the test set.
Cross-Survey Validation: When tested on 200 published events each from OGLE, KMTNet, and MOA, Microlensify achieved an average detection efficiency of 92.7% (probability $\ge 0.99$ ).
Specific Case Studies:
- Successfully recovered the known binary-lens event Gaia20fnr (observed as a single-lens-like wing in TESS) with a probability of 1.0.
- Detected a candidate free-floating planet event reported by Kunimoto et al. (2025) with probability 1.0.
- Failed to assign high probability to the binary-lens event Gaia23bra, which is consistent with the model's training on single-lens simulations, though the flux variation was clearly visible in the data.

Significance and Claims
The paper claims that Microlensify represents a significant advancement in the automated search for microlensing events in high-cadence surveys. By combining the pattern recognition capabilities of Transformers with physical constraints and real-data training, the model offers a computationally efficient alternative to traditional fitting pipelines (roughly 300 times faster on a single CPU core). The authors emphasize that the model's ability to generalize across different cadences and survey types (from space-based TESS to ground-based OGLE/KMTNet) makes it a versatile tool for time-domain astronomy. The identification of asteroid crossings as a source of false positives is highlighted as a critical finding for future high-cadence surveys, such as those planned for the Nancy Grace Roman Space Telescope. The authors conclude that Microlensify serves as an effective initial screening tool, reducing the volume of data requiring human inspection and enabling the discovery of rare microlensing candidates in massive datasets.

More like this