Microlensify: a Transformer Based Machine Learning Classifier for Microlensing Events Trained on TESS Light Curves
Authors: Atousa Kalantari, Somayeh Khakpash, Sedighe Sajadian, Hosein Haghi, Willow Fox Fortino, Rosanne Di Stefano
Organizations: Department of Physics, Institute for Advanced Studies in Basic Sciences (IASBS), Zanjan 45137-66731, Iran · Department of Physics, Lehigh University, 16 Memorial Drive East, Bethlehem, PA 18015, USA · Department of Astronomy, University of Virginia, 530 McCormick Road, Charlottesville, VA 22904, USA · Department of Physics, Isfahan University of Technology, Isfahan 84156-83111, Iran · University of Delaware, Department of Physics and Astronomy, 217 Sharp Lab, Newark, DE 19716 USA · Harvard-Smithsonian Center for Astrophysics, Cambridge, MA, 02138, USA
Microlensing can reveal populations of faint compact objects that are otherwise difficult to detect. Depending on their design, all-sky surveys have the potential to search for these objects across the sky. The Transiting Exoplanet Survey Satellite (TESS), primarily designed to detect transiting exoplanets, also provides near all-sky coverage with high cadence. In this work, we use TESS data to search for microlensing candidates using both traditional and machine-learning methods and to identify associated false positives in high-cadence surveys. Microlensify is a physics-informed, transformer-based variational autoencoder trained on simulated single-lens microlensing light curves and real TESS Sector 12 data. The model classifies events, reconstructs light curves, and estimates microlensing event durations. Applied to ∼5.6 million TESS light curves, it identified between 0.036% and 1.89% as microlensing candidates across different TESS pipelines. After applying microlensing detection metrics and cross-matching with SIMBAD, we obtained a final list of candidates and identified false positives including long-period variables, Mira variables, cataclysmic variables, red giants, and transients. We also found Gaussian-like peaks caused by asteroid crossings, a potential source of false positives in high-cadence microlensing surveys. The model also predicts event duration with an accuracy of R2=0.97. The model was further tested on published events from different ground-based microlensing surveys, confirming 92.7% as microlensing, demonstrating its applicability across surveys with different cadences.
Figures & tables
Figure 1: Distribution of microlensing events from KMTNet, MOA, and OGLE, overlaid with the TESS observing fields in its first two years, plotted in ecliptic coordinates.
Figure 2: Relationship between log10(err/flux) and log10(flux) for all data points from the 12,000 selected light curves extracted with the Eleanor pipeline, shown as green dots. The pink dashed curve shows the fitted two-dimensional polynomial function.
Figure 3: Simulated TESS microlensing light curves with the corresponding best-fit KMTNet PSPL models overplotted as pink dashed lines. Panels (a)–(c) are rejected by our quality cuts: (a) due to insufficient baseline coverage on one side of the event and low SNR; (b) and (c) due to low SNR, low ΔχKMTNet2 . Panels (d)–(f) satisfy all selection criteria.
Figure 4: Parameter distributions of the full microlensing simulation sample (blue) and the K-means-selected training subset of 17,000 light curves (pink). Panels show blending ratio, t0 (0-–27.4 days as TESS observation time), tE (days), u0 , log10ρ , and baseline magnitude of the flat light curve. The clustered subset preserves the original distributions.
Figure 5: Architecture of Microlensify, a physics-informed transformer-based conditional variational autoencoder trained on ∼ 34,500 TESS (Eleanor, Sector 12) light curves. The model takes a normalized flux time series (length 940) and five statistical features as input, simultaneously classifies microlensing vs. non-microlensing candidates, reconstructs the input light curve, and predicts event duration via a physics-informed latent dimension. The example light curve shown in green is the known microlensing event Gaia20fnr (found in TESS with TIC ID 123793270); the model reconstruction is in pink produced by the decoder. For details on this event see subsection VI.2 .
Figure 6: Three true-positive microlensing light curves from the test set. Green points show the original simulation light curve, and the dashed pink line shows Microlensify’s reconstruction. For each event, the true duration and the duration predicted by Microlensify, in days, are indicated.
Figure 7: Predicted versus true microlensing event duration for all true-positive events in the test set. Microlensify achieves R2=0.97 , MSE = 7.84, and MAE = 1.22. The teal dashed line represents the 1:1 relation, and the gray dash-dot lines indicate ±5% deviation.
Eleanor Sector12
SPOC Sector12
QLP Sector12
QLP Sector33
Light curves fed to model
1,500,000
133,329
2,857,972
1,130,307
Predicted with probability=1.0
545
108
54,091
11,770
Remaining after threshold cleaning
363
86
24,907
2,383
Remaining after visual cleaning
328
69
3,554
1,532
SIMBAD Object Types in Final Clean Sample
Star
6
19
317
238
Table 1: Classification summary for TESS pipelines (Eleanor Sector 12, TESS-SPOC Sector 12, QLP Sector 12, and QLP Sector 33). Shown are the initial number of light curves processed, the number of sources assigned classification probability P=1.0 by the Microlensify model, and the final counts after quality cleaning ( ΔχKMTNet2>400 , peak SNR >30 , and position inside the skewness-–Von Neumann boundary with 0.1 tolerance). The lower panel lists SIMBAD object types found in the cleaned high-confidence samples. Note that the object types are not deterministic, and the model prediction is based on microlensing-like peaks, which may be caused by microlensing false positives, blending with nearby sources, systematics, or asteroid crossings.
Figure 8: Example light curves from the final high-confidence sample in each pipeline: Eleanor (Sector 12), SPOC (Sector 12), QLP (Sector 12), and QLP (Sector 33). Each panel is labeled with its TIC ID. Plotted flux is raw flux for Eleanor and SPOC pipelines, and normalized raw flux for QLP pipeline. The x-axis shows time in days (BJD − 2457000). Corresponding sky coordinates (R.A., Dec), best-fit KMTNet parameters, SNR, model-predicted 4FWHM values, and SIMBAD object types are reported in Table 2 .
TIC ID
Mission
RA [ ∘ ]
DEC [ ∘ ]
f0
f1
teff [days]
t0 (days)
Δχ2
SNR
4FWHMpred [days]
Obj Type
17115525
Eleanor 12
247.8453
-53.7273
22837.4
4288.4
14.25
1639.22
990.9
103.4
95.8
LPV* Cand
13594596
Eleanor 12
247.4081
-53.4260
6283.6
4149.7
14.25
1639.22
1055.9
38.5
52.8
LPV* Cand
10954658
Eleanor 12
247.2055
-52.9932
7591.2
1817.9
6.012
1639.001
1053.5
38.0
30.5
LPV* Cand
10949358
Eleanor 12
247.1984
-52.2926
4892.5
1160.3
14.250
1639.224
975.6
79.2
86.1
Mira
1311842998
Eleanor 12
259.6461
-63.9124
18875.4
15232.7
14.250
1639.223
1025.8
35.5
67.4
Not Found
125958042
SPOC 12
256.8225
-46.7657
43881.7
19999.7
25.333
1645.488
912.5
65.7
77.0
C*
Table 2: Parameters for the example light curves shown in Figure 8 . Listed are sky coordinates (R.A., Dec), best-fit KMTNet model parameters ( f0 , f1 , teff , t0 , ΔχKMTNet2 ), SNR, model-predicted 4FWHM values, and SIMBAD object types including long-period variables (LPV* and candidates), Mira variables, carbon stars (C*), EclBin (Eclipsing Binary), cataclysmic variables (CV*), asymptotic giant branch stars (AGB*), spectroscopic binaries (SB*), and regular stars; a fraction of sources have no counterpart in SIMBAD ("Not Found"). The flux parameters f0 and f1 are given in units of e−s−1 for Eleanor and SPOC light curves and are dimensionless for QLP light curves.
Figure 9: Light curves of three microlensing candidates (all extracted from the QLP pipeline). (a) TIC 123793270, cross-matched with Gaia20fnr, observed in TESS Sector 33; Microlensify predicts probability 1.0. (b) TIC 378615385, cross-matched with Gaia23bra, observed in Sector 64; a clear binary microlensing event well extracted by QLP but not classified as microlensing by our model. (c) Free-floating planet candidate reported by Kunimoto et al. [32] in Sector 61; Microlensify successfully recovers it as a microlensing candidate.
Figure 10: Three examples from TESS Sector 12 processed by all three light curve extraction pipelines (Eleanor, QLP, and SPOC). TIC 35302284 (SIMBAD type: AGB star): Microlensify probability 1.0 in QLP and SPOC, poor-quality Eleanor light curve. TIC 30716342 (SIMBAD type: star): clear normalization offset across the sector gap in QLP, cleanly extracted by SPOC, poor Eleanor extraction; not selected as a microlensing candidate in any pipeline. TIC 25133286 (SIMBAD type: cataclysmic variable): good-quality extraction and probability 1.0 in all three pipelines.
The discovery of exoplanets at scale has become one of the defining data science challenges in modern astrophysics. NASA's Transiting Exoplanet Survey Satellite (TESS) had catalogued over 7,800 planet candidates by early 2026, yet confirmation stands at fewer than 720. This paper introduces ExoNet, a multimodal deep learning framework that jointly processes phase-folded global and local light curve views alongside stellar parameter features through a calibrated late-fusion architecture combining 1D Convolutional Neural Networks, 8-head Multi-Head Attention over temporal feature maps, and a residual fusion head with post-hoc Temperature Scaling calibration. Trained on 7,585 labeled Kepler Objects of Interest, ExoNet achieves Test AUC = 0.9549 and 86.3% accuracy. Applied to 4,720 verified unconfirmed TESS Planet Candidates with TOI-TIC cross-identification verified against the NASA Exoplanet Archive, the model yields 1,754 high-confidence signals, 52 habitable-zone candidates, and six Earth-sized habitable-zone targets below 1.6 Earth radii. TOI-5728.01 and TOI-6716.01 emerge as the most Earth-like unconfirmed candidates. Full ablation confirms each modality improves AUC. Code and catalog are openly released.
Md. Rashadul Islam
Dept. of Computer Science and Engineering Daffodil International University Dhaka, Bangladesh
I present EXOVEIL, a transit detection system that learns what a star's brightness should look like and flags when reality disagrees. Unlike existing systems that require phase-folded input, EXOVEIL operates on raw flux time series and can detect planets that transit only once.A Transformer world model, trained on 16,499 Kepler light curves with transit-masked self-supervised learning, predicts expected stellar flux. A matched-filter detector with variance weighting extracts transit signals from the prediction residuals. A learned classifier (XGBoost) separates planets from false positives, achieving AUC 0.938 on Kepler DR25. Applied to single-transit injection-recovery, EXOVEIL recovers 32% of transits at 1000 ppm depth a task where all classification-based systems score 0% by construction. A blind search of 3,737 Kepler stars yields 179 new transit-like signals not present in the DR25 TCE catalogue, including 46 monotransit candidates. Applied withoutretraining to 47 confirmed TESS planets in the PLATO LOPS2 field, EXOVEIL achieves 100% recovery, demonstrating zero-shot cross-mission transfer. At PLATO's 25-second cadence, detection reaches 100 ppm -- approaching the Earth-analog regime. I provide the first application of conformal prediction to transit detection (95.9% empirical coverage) and release the system as pip install exoveil with pretrained weights and a candidate catalogue.
The Nancy Grace Roman Space Telescope (Roman), set for launch as early as September 2026, will conduct wide-field infrared imaging surveys with unprecedented spatial resolution and cadence, enabling the discovery of millions of astronomical transients. Hence, it is necessary to have automated pipelines for generating alerts in place so that the telescope can begin discovering reliable transients and variable objects soon after it is launched. However, no real Roman data currently exist, making the development of such pipelines difficult. In this work, we present a machine learning model RuBR and a general methodology for distinguishing genuine transient and variable detections from spurious (bogus) detections within the RAPID pipeline. In particular, we present three models using this methodology: RuBRcomb trained and tested on combined locally injected and OpenUniverse2024 transients, RuBRloc trained on locally injected transients and tested on OpenUniverse2024 transients, and RuBRDA that combines locally injected transients with a fraction of OpenUniverse2024 transients in domain-adaptation mode for training. This paves the way for strategies to adapt the RuBRcomb model to real observations in the absence of any ground-truth labels during the early phases of the Roman mission. While the image differencing pipeline continues to be improved, our experimental results demonstrate the effectiveness of the proposed approach and its promise for robust real-bogus classification in the Roman era.
Karan Gandhi, Ashish A. Mahabal, Jacob E. Jencson +6
Department of Computer Science and Engineering, Indian Institute of Technology, Gandhinagar, India · Division of Physics, Mathematics, and Astronomy, California Institute of Technology, Pasadena, CA 91125, USA · Center for Data Driven Discovery, California Institute of Technology, Pasadena, CA 91125, USA +2