The TESS Ten Thousand Catalog: 10,001 Uniformly Vetted and Validated Eclipsing Binary Stars Detected in Full-frame Image Data by Machine Learning and Analyzed by Citizen Scientists

Concurrently with the development of the neural network, we also needed to refine our method of preprocessing the lightcurves, which, as in many ML applications, was critical to the performance of the neural network. ML methods require data to be of the same shape. This of course presents a problem with TESS FFI lightcurves, especially after the per-time-step quality flags are masked from the lightcurve, resulting in a wide variety of 1D array sizes. The temporal discontinuities caused by the data downlink gap and the quality mask also create a temporally irregular data set. We chose to ignore the temporal component and treat the lightcurve as a 1D shape rather than a time-dependent signal. This approach is consistent with our selection of a CNN rather than a recurrent neural network (D. E. Rumelhart et al. 1986) or another time-dependent methodology such as long short-term memory (LSTM; S. Hochreiter & J. Schmidhuber 1997), convolutional LSTM (ConvLSTM; X. Shi et al. 2015), or temporal convolutional networks (C. Lea et al. 2017), among others.

To create homogeneous shapes from the irregular lightcurves, our options were to either truncate longer lightcurves or pad shorter lightcurves. Truncation, of course, risks missing a lightcurve where a single eclipse occurs in the truncated section. Padding provides its own risks in providing artificial information in the discontinuity of the data shape. We decided to pad the lightcurves to a maximum length of 1400 elements, with the padding containing a mirror of the data. We emphasize again that the neural network has no time dependency, and therefore no understanding of periodicity, and it was determined that the neural network could learn to ignore the discontinuity in the padding in the same manner as it would the discontinuity in the collapsed time gaps. We also note that we developed the neural network during Year 2 of the TESS mission, when the 30 minute cadence data provided relatively short lightcurves. With the continued shortening of the TESS cadence over subsequent years, we have not retrained the network. Rather, we downsampled longer lightcurves to fit our required data shape.

TESS systematics presented a different set of challenges. In many TESS FFI lightcurves, indeed, the dominant signal is scattered light systematics, which often produce features in the lightcurves that resemble eclipses. In Years 1 and 2 of the TESS data, this problem is particularly pronounced in Sectors 1, 12, 13, 14, 15, 23, 24, and 26. In neural networks, and ML in general, the magnitude of a value is a representation of its importance. As such, strong eclipse-like systematic signals have the potential to dominate an ML method if not properly diminished in importance, hence the need for data scaling. We specifically selected the quantile transform as our method of scaling, using the built-in scikit-learn (F. Pedregosa et al. 2011) package functionality. With this method, scattered light systematics are reduced to effectively the same size as the eclipses, forcing the neural network to learn the shape of the eclipse signal in contrast to the systematics. Figure 2 demonstrates the outcome of this scaling on a lightcurve dominated by a scattered light feature. The top panel shows the unscaled lightcurve, while the bottom panel shows the quantile-scaled lightcurve processed as input to the neural network. Although it is clearly more difficult for the human eye to distinguish eclipses in the scaled form, this method proved to be superior to the neural network for understanding subtle differences between eclipses and eclipse-shaped noise or systematics. This is not to say that we were able to avoid outcomes where the neural network classifies such features as eclipses entirely, but this method did substantially diminish the problem.

Zoom In Zoom Out Reset image size

While quantile scaling underemphasizes large features, it also has the effect of overemphasizing small features. To our benefit, this helped in the identification of shallow eclipses. However, we also found that our network will identify planet transits as well as small eclipse-like shapes in noise patterns, which became a substantial source of error (discussed further in Section 3.3). Thus, we effectively made the decision to trade large noise effects for small noise effects. We make no assertion that this trade was ideal, nor our method of classifying the eclipse shape versus identifying an EB directly.

A particular challenge of this effort was the collection and augmentation of the training data set. Generally, the performance of a classifier will track directly with the quantity of training data samples to an asymptotic limit (e.g., C. Sun et al. 2017). Additional difficulties arise for our particular application in that eclipses can be vastly different in appearance, thus requiring a substantial amount of training data for a neural network to effectively generalize the features of an eclipse.

At the time of development of our neural network (TESS Year 2), we had tens of millions of lightcurves, but only a handful of manually selected EB lightcurves. To gather a sufficiently sized data set to effectively train our neural network by manually sorting through individual lightcurves would have been an intractable task. As such, we progressively augmented our data set by iteratively using a weakly trained neural network to find new training samples among the full data set of lightcurves. After each iteration of training and inference, we would select (i) lightcurves given a score near unity that clearly did not show an eclipse, or (ii) lightcurves given a score near zero that clearly showed an eclipse. The former represented false positives and the latter false negatives. We used these as properly labeled training samples in the next iteration of training, effectively filling gaps in the understanding of the neural network. With each of these iterations, the neural network became progressively more capable. By the time we were satisfied with the performance of the neural network, we had built a training set of ∼40,000 samples.

We emphasize that our neural network was not trained to find EBs. It has no concept of repeated features or periodicity. Rather, it was trained to find eclipses, or, more broadly, features resembling eclipses. We show how the neural network activates on the shape of the eclipse in the saliency map of the activation weights of the penultimate layer in Figure 3. Note that the neural network will emphasize a single eclipse in determining the output score of the lightcurve.

Zoom In Zoom Out Reset image size

Knowing that the neural network would be providing candidates for manual review rather than directly populating a list of near-certain EBs, we wanted to allow for interesting results that would not fit the conventional shape of an eclipse, e.g., a complex syzygy or lopsided eclipse. This process has allowed us to find multiple star systems with complex outer orbital eclipses, among other interesting phenomena, the body of work on which has established the effectiveness of our methods given that this neural network has contributed to many discoveries (e.g., V. B. Kostov et al. 2021a,2021b, 2022b, 2023, 2024a, 2024b; B. P. Powell et al. 2021a, 2021b, 2022b, 2023, 2025; T. Borkovits et al. 2022; B. K. Capistrant et al. 2022; S. A. Rappaport et al. 2022,2023, 2024; R. Jayaraman et al. 2024; T. Mitnyan et al. 2024a; K. Oláh et al. 2025).

Our results have been qualitative, with a manual review of outputs in multiple stages (described further in Section 4). As such, an assessment of performance of the neural network against a data set of known EBs is hardly direct. However, we provide such a comparison here to provide the reader with the context of our process as well as the contents of our catalog.

An evaluation of the model would be most complete with a section of the sky where we could consider all EBs within TESS’s limiting magnitude to be known. As such, the Kepler (W. J. Borucki et al. 2010) field provided an ideal testing ground, with the full data set having been thoroughly evaluated for the presence of EBs, resulting in the production of an EB catalog (A. Prša et al. 2011; R. W. Slawson et al. 2011; G. Matijevič et al. 2012; K. E. Conroy et al. 2014a, 2014b; D. M. LaCourse et al. 2015; M. Abdul-Masih et al. 2016; B. Kirk et al. 2016), hereafter referred to as the “Kepler EB catalog.” By comparing the number and characteristics of lightcurves identified in our catalog against those of the 2920 EBs of the Kepler EB catalog, we could make an estimate of the performance of our catalog.

We also considered the catalog of 4584 TESS EBs from short-cadence data in Sectors 1–26 (A. Prša et al. 2022), hereafter referred to as the “TESS EB catalog.” Although less comprehensive in terms of a full survey of a section of the sky, a direct comparison to TESS EBs rather than Kepler allows for fewer independent sources of error such as different noise amplitudes or photometric capabilities.

We cross-referenced our catalog with the Kepler field boundaries,²² finding each object from our catalog that would have been observed by the original Kepler mission, which resulted in 9768 unique TIC IDs. We then reduced the Kepler EB catalog to those objects with T_mag < 15, as this was the limit of our lightcurve construction, resulting in 2458 EBs out of the original 2920. Our catalog contains 1371 of these 2458 EBs, or ∼55.8%. To compare our catalog to the TESS EB catalog, we cross-referenced our catalog with lists of the 2 minute targets for TESS Sectors 1–26,²³ from which the TESS EB catalog was derived. In total, there were 507,898 unique 2 minute targets in these sectors, 8910 of which were identified by our neural network. Comparing this sample directly to the TESS EB catalog revealed that our neural network found 3884 of the 4584 EBs therein, or ∼84.7%. In this comparison of true positives, it is clear that our neural network performed far better than the TESS EB catalog, which we assess to be likely due to the systematic differences between the TESS and Kepler data.

In Figure 4, we show a scatter plot of the morphology parameter versus T_mag of the Kepler EBs found (blue) and not found (red) by our neural network. The morphology parameter (described further in A. Prša et al. 2011) is a measure of the EB type, with values less than 0.5 corresponding to detached EBs, values in the range 0.5–0.7 corresponding to semidetached EBs, values in the range 0.7–0.8 corresponding to overcontact EBs, and values greater than 0.8 corresponding to ellipsoidal or unknown classifications. We exclude analysis where the morphology parameter was given a value of −1, indicating the lightcurve was unclassifiable by the methods of A. Prša et al. (2011). Furthermore, we make the same comparison using the EB period versus T_mag in Figure 5 and the EB period versus the morphology parameter in Figure 6. We can make two conclusions from these figures:

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

(i) The T_mag histograms in the left panels (Kepler comparisons) of Figures 4 and 5 show a clear decrease in performance with decreasing T_mag, while we do not see the same decrease in the right panels (TESS comparisons) of the same figures. Although there is a somewhat artificial lower limit on magnitude in the TESS 2 minute cadence targets due to the selection bias for bright stars, we can still see a generally uniform trend of the fraction identified in the T_mag histograms of the TESS comparisons. However, these targets are still somewhat idealized in comparison to a full sample over a section of the sky. As such, we consider our neural network’s performance against the TESS EB catalog 2 minute cadence targets to be an upper limit, while the performance against the Kepler EB catalog should be considered a lower limit.

(ii) For the Kepler (left) panels in Figures 4 and 6, the morphology histograms show a clear weakness of our neural network for the extremes of the morphology range. Aside from showing the general relationship between EB period and morphology, Figure 6 demonstrates a clear trend of weakness for our neural network in identifying EBs at the extremes of the period range for any given morphology range in both the Kepler (left) and TESS (right) comparisons. That is, the red points seem to dominate the blue both to the far left and to the far right of the general trend line. We assess that both of these trends demonstrate a weakness in generalization of the neural network to less common types of eclipse patterns.

Having examined the nature of our true positives and false negatives, we turned to the much larger set of false positives. As previously discussed, 1371 of the 9768 unique TIC IDs from the Kepler field and 3884 of the 8910 in the TESS Sector 1–26 short-cadence targets found by our neural network were true positives, leaving the remaining 8397 (∼86%) of the Kepler sample and 5026 (∼56%) of the TESS sample as false positives. Naturally, the questions arise as to what these false positives are and why they are so numerous. To determine their nature, we manually examined a subset of bright false positives in the Kepler field with T_mag < 10, totaling 126 unique TIC IDs.

Of these lightcurves, we found that 16 (∼13%) showed clear EBs. Since these were not in the Kepler EB catalog, we assumed that most of these are likely the result of blending. That is, the 21″ pixels of TESS will frequently cause the lightcurves of bright stars to show in the lightcurves of dimmer close neighbors. To demonstrate this effect, we show the Kepler field with true positives, false positives, and false negatives in Figure 7, where the clustering of several large groups of false positives can be seen, likely as a result of blending. While this was only a problem in ∼13% of our false positives with T_mag < 10, it can be reasonably expected that lightcurves of dimmer stars will show this type of contamination more frequently. We examined each of the 16 lightcurves showing clear EBs and confirmed that contamination from nearby brighter stars was indeed the source of the signal. However, we assessed two of the 16 lightcurves showing clear eclipses as the true source of an EB, TIC 26542657 (KIC 12013550) and TIC 63454475 (KIC 10342012). We confirmed that neither of these targets are present in the Kepler EB catalog. Furthermore, we found that there was no Kepler data available for TIC 63454475, while the Kepler lightcurves for TIC 26542657 (first identified as an EB in the TESS EB catalog by A. Prša et al. 2022) indeed showed no eclipses, confirming that neither of the targets were missed accidentally in the creation of the Kepler EB catalog. Given that TIC 26542657 has Kepler lightcurves without eclipses, we assess that it must be a higher-order system, likely a triple, with so-called “disappearing eclipses.” We will discuss this particular system further (also confirming it as a triple) as well as identify other examples of this type of system in Section 5.4. Although it is beyond the scope of this effort to analyze these systems, we note briefly that the changes in binary inclination causing periods of eclipsing and noneclipsing behavior are driven by interaction with the outer body, and we refer the reader to T. Borkovits (2022) for a detailed discussion of the nature of this type of triple system, among others. Given we found one such system out of only 126 in our crossmatch with the Kepler EB catalog with T_mag < 10, we expect there to be several more such systems in our 8397 false positives from the Kepler field, which may merit an investigation in its own right.

Zoom In Zoom Out Reset image size

Returning to our false positives from the T_mag < 10 sample, oscillations resembling eclipses comprised 16 (∼13%) of the false positives, while 21 (∼17%) showed scattered light systematics resembling eclipses, as in Figure 2. The latter were particularly prominent in TESS Sector 14, which overlapped with the Kepler field. The bulk of the false positives, 74 (∼59%), contained noise patterns that resembled eclipses. We provide examples of these false positives in Figure 8.

Zoom In Zoom Out Reset image size

Another type of scientifically valuable false positive found by our neural network was planet transits. We compared our catalog to the NASA Exoplanet Archive Kepler catalog and found an overlap of 1502 out of the 8397 TIC IDs. Of these, 125 were determined to be genuine exoplanet candidates. The overlap with the NASA Exoplanet Archive TESS exoplanet candidate catalog²⁴ is even more pronounced. Of the 7576 TESS exoplanet candidates, our catalog contains 2445 (∼32%). Given how we scale the lightcurves (see Figure 2), the neural network is generally not able to discern a difference between an eclipse and a clean transit signal, as it was not trained to do so. As such, we expect that our catalog is also rich with planet candidates. Of particular interest is the fact that, since our neural network has no periodicity requirement, we expect there to be many single transit events that evade discovery through periodicity-based analyses.

We emphasize, again, that our neural network was trained to find features resembling eclipses, not EBs. Our results were qualitative and manually reviewed. As such, the comparisons provided in the previous section should not be considered a fully accurate measure of the performance of the neural network as much as context for the reader to understand our process and the contents of our catalog. Our method should be considered as a means of reducing an extremely large data set (hundreds of millions of lightcurves) to a much smaller, manageable data set with a high concentration of scientific value.

Our full catalog consists of 1,223,603 unique TIC IDs with lightcurves that our neural network gave a score of ≥0.9. The distribution of these, in terms of TESS magnitude and ecliptic coordinates, is shown in Figure 9. Most of the targets are on the faint side, with a median TESS magnitude of ≈14, and the majority are near the Galactic plane. It is from these candidates that we distilled the much smaller catalog of vetted EBs to be discussed in the remainder of this paper. This catalog could be employed by interested researchers with the caveats of the analysis in the preceding section. We use the Kepler EB catalog comparison as the lower end of our estimate and the TESS EB catalog comparison as the upper end of our estimate in the following summary of caveats and contents:

• 1.  
  This is a catalog of lightcurves with eclipse-like features, not an EB catalog.
• 2.  
  A total of 14%–44% of the catalog should be expected to be EBs.
• 3.  
  It follows that 56%–86% of the catalog should be expected to not be EBs. These will consist of contaminated lightcurves, transiting exoplanets, dippers, systematics, and other sharp ellipsoidal features.
• 4.  
  The catalog should be expected to contain 55%–84% of all EBs in the TESS lightcurves for Sectors 1–82 with T_mag < 15. The fraction of completeness will decrease with brightness.
• 5.  
  The full catalog is entirely unvetted and we offer no guarantee as to its contents. We have, however, found the neural network outputs to be scientifically valuable, so we offer it to the community in its entirety for their own purposes.

Zoom In Zoom Out Reset image size

With these caveats in mind, we transition in the next section to explaining how we analyzed a subset of these, producing 10,001 genuine EBs for inclusion in a well-vetted catalog.

A particularly powerful method to constrain the pixel position of the source of detected eclipses (or transits) is the photocenter-based analysis that is routinely used in surveys aimed at finding transiting exoplanets (e.g., J. L. Coughlin et al. 2014; S. E. Thompson et al. 2015, 2018; J. D. Twicken et al. 2018; V. B. Kostov et al. 2019; D. J. Armstrong et al. 2021; H. Valizadegan et al. 2025; and references therein). Briefly, the method uses the target pixel files to measure the in-eclipse center-of-light (“photocenter”) pixel position for each eclipse detected in the difference image,²⁵ and compare it to the out-of-eclipse photocenter and/or to the catalog pixel position of the target. If there is no statistical difference between these and the in-eclipse photocenter, the eclipses are considered to be “on-target”; otherwise the eclipses are “off-target,” likely coming from a nearby field EB, and the candidate is marked as a false positive.

Obtaining robust photocenter measurements depends on multiple factors such as the signal-to-noise ratio (SNR) of the target pixel files, the depth of the detected eclipses, the presence of nearby comparably bright field stars (worst-case scenario: ones that are much brighter and variable on timescales comparable to the duration of the eclipses), etc. In theory, the difference images used to measure the per-eclipse photocenters resemble a well-defined, bright pixelated spot superimposed on an otherwise dark background (Figure 10, first three columns from the left). In practice, the difference images are often distorted due to various astrophysical, systematic, or instrumental effects (Figure 10, rightmost column), making the corresponding photocenter measurements unreliable (e.g., L. Cacciapuoti et al. 2022; V. B. Kostov et al. 2022b, 2024a; C. Magliano et al. 2023).

Zoom In Zoom Out Reset image size

Overall, based on our experience with TESS data—and depending on the peculiarities of the specific target—measurements of genuine photocenter offsets of ≳0.2–0.3 pixels (i.e., ≳4″–6″) are often trustworthy (V. B. Kostov et al. 2022b, 2024a; C. Magliano et al. 2023). However, considering measured offsets of ≲0.1–0.2 pixels as significant can be extremely challenging or potentially even impossible. Thus, throughout this work we adopt a photocenter offset threshold of 0.2 pixels (≈4″) such that cases below are considered as likely “on-target” and those above “off-target.”

Ideally, genuine photocenter offsets 3–5 times larger than this threshold (i.e., ∼12″–21″) should be relatively easy to measure (and accept as reliable). Thus, to account for potential false positives due to known EBs, we first evaluated whether the TESS EB candidates our neural network identified are within a 21″ (1 TESS pixel) sky-projected separation of EBs listed in various catalogs. In particular, we queried EB catalogs from ASAS-SN (∼150,000 EBs), ATLAS (∼30,000 EBs), Gaia (∼2,200,000 EBs), OGLE (∼430,000 EBs), Simbad (∼2,400,000 EBs, ∼200,000 spectroscopic binaries), TESS (∼50,000 EBs, ∼10,000 planet candidates from ExoFOP-TESS²⁶ ), the International Variable Star Index (VSX; ∼900,000 EBs), and WISE (∼50,000 EBs), taking into account the respective overlaps between the different data sets. Unsurprisingly, given the pixel size of TESS and the corresponding crowding, about one-quarter of our ∼1.2 million candidates fulfill the above criteria (see Table 1). We note that this consideration does not immediately rule these out as false positives, but it marks them as likely suspects. Conversely, those that are not within 1 pixel of known EBs (∼900,000 TIC IDs) can potentially be confirmed as bona fide new EBs through careful photocenter analysis. For the benefit of the community, we provide these targets as an auxiliary data set (see Table 2).

To investigate this matter further, we conducted a deep dive into a subset of ∼60,000 (hereafter 60K) targets, representing ∼5% of the likely known and potentially new EB candidates from our preliminary list. The targets were randomly selected and evenly split on either side of the 21″ demarcation line, and are representative of the TESS magnitude and sky position distributions shown in Figure 9. These 60K targets were subjected to comprehensive ephemeris determination and photocenter measurements, and analyzed in depth via the two-step process outlined below.

During the first step, we developed and utilized an automated pipeline to calculate ephemerides and measure photocenter offsets. Specifically, we applied the box least-squares algorithm (BLS; G. Kovács et al. 2002) to the available eleanor FFI lightcurves to measure periods and conjunction times, limiting the minimum/maximum period searched for to 0.5/40 days, respectively. The BLS results were further improved by fitting a generalized Gaussian model to each detected eclipse adopting the methodology of V. B. Kostov et al. (2022b), and testing for period deviations from the linear ephemeris. The latter helps take into account potential eclipse timing variations (ETVs) that may decrease the precision of the BLS measurements, and also provides robust measurements for the eclipse depths and durations. Next, we used the refined ephemerides and eclipse durations to construct the appropriate difference images for each EB candidate, following the prescription of V. B. Kostov et al. (2019). Finally, we obtained photocenter measurements by fitting to each difference image the TESS pixel response function and a Gaussian point-spread function, and adopting the average of the two as the corresponding photocenter of the image.

Preliminary results from the first step are highlighted in Figure 11, showing the distributions of the 60K sample in terms of TESS magnitude, measured period, and photocenter offset. At this stage, the relevant measurements have not yet undergone the rigorous vetting and validation analysis required for promoting a target as a genuine EB, and are thus likely affected by various systematics. For example, the period distribution of the new EBs seen in Figure 11 shows a local maximum near 14 days. This is close to half the duration of a TESS sector, which makes the potential periods suspicious.

Zoom In Zoom Out Reset image size

The second step of the process addresses the issue of potentially incorrect ephemeris measurements and, consequently, incorrect photocenter measurements produced by the automated pipeline outlined above. This can occur when the period search is misled by, for example, strong systematic effects such as prominent out-of-eclipse lightcurve variations (due to, e.g., starspots) that can dominate or even completely overwhelm the eclipse signal (see Figures 12 and 13, upper panels). Additionally, EBs where the primary and secondary eclipses are similar in depth, duration, and shape can lead to situations where the automatically measured period is an integer ratio of the true period (see Figure 13, lower panels). While we tried to minimize the impact of such issues on the ephemeris and photocenter pipeline as much as possible, it was challenging to account for all possible complications without negatively affecting the signals of interest. For example, we only used data points with good quality flags as provided by eleanor; removed those that were near known issues, such as momentum dumps (TESS Instrument Handbook²⁷ ), or that we identified as potentially suspect sections of the lightcurve; and, where appropriate, utilized low-order polynomial detrending.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

To account for these and other challenges, we manually inspected the products of the automated pipeline for each of the 60K targets as outlined below.

Analyzing vast amounts of data using automated methods remains a highly nontrivial process as various sources of “noise”—be it astrophysical, instrumental, or systematic—can introduce subtleties that are challenging to automatically account for. Additionally, while autonomous methods excel at, for example, finding a periodic signal in time series such as the transits of an exoplanet (e.g., using BLS), they often lack the insight to discover unique and unexpected features, or various unusual and uncommon astrophysical objects or phenomena. For example, all known transiting circumbinary planets have been identified by visual inspection of Kepler and TESS lightcurves (V. B. Kostov 2023). Notably, while human intervention can help fill in the voids left by algorithms, given the complexity and scope of astronomical data sets it is quite challenging for a traditional research group to manually inspect hundreds of thousands of targets.

Thankfully, to the rescue come volunteers from all walks of life that boost the capacity of bandwidth-limited professional astronomers manyfold and help tackle the ever-increasing volume of publicly available astronomical data. This so-called citizen science approach is not a new concept—professional and amateur astronomers have a fantastic and strongly intertwined history. One famous example is the “Harvard Computers” project, where some of the participants started with no formal astronomy training yet helped revolutionize astronomy and became some of the most successful professional astronomers (e.g., S. Nelson 2008). Among the more recent examples, many transiting planets with orbital periods longer than 1 yr have been discovered by citizen scientists (e.g., J. Wang et al. 2015), and hundreds of eclipsing triple- and quadruple-star systems have been spotted by eagle-eyed volunteers (e.g., T. Borkovits 2022 and references therein). Citizen scientists have been responsible for many “firsts,” e.g., (i) the unusual Boyajian’s Star (T. S. Boyajian et al. 2016), (ii) an exocomet transiting its host star (S. Rappaport et al. 2018), and (iii) a newly discovered class of objects called “tidally tilted pulsators” (e.g., G. Handler et al. 2020). Time and again, volunteers have demonstrated they can extract interesting signals from noise in numerous cases.

The state of citizen science is strong, with multiple projects tackling various astronomical data sets and making important scientific contributions on a regular basis thanks to, e.g., Planet Hunters and Planet Hunters TESS,²⁸ Exoplanet Explorers,²⁹ Citizen ASAS-SN,³⁰ SuperWASP Variable Stars,³¹ Planet Patrol,³² Eclipsing Binary Patrol (EBP),³³ Exoplanet Watch,³⁴ UNITE: Unistellar Network Investigating TESS Exoplanets,³⁵ and the Visual Survey Group (VSG; M. H. K. Kristiansen et al. 2022). The power and dedication of citizen scientists is truly incredible—for example, members of the VSG have visually inspected tens of millions of lightcurves from Kepler and TESS (M. H. K. Kristiansen et al. 2022), and helped make important new discoveries in multiple branches of astrophysics.

4.3.1. Exogram

The manual inspection of the 60K targets proceeded as follows. First, we adapted our online vetting portal Exogram³⁶ for rapid visual scrutiny of a randomly drawn subset of about 10,000 targets out of the 60K set by our core science team composed of professional astronomers and highly knowledgeable citizen scientist “superusers.” The team formed during the Planet Patrol project, and we have been working together ever since. The custom interface breaks down the vetting process into three main questions: “Is this an EB?,” “Is the measured period correct?,” and a space for additional comments (both predefined and free text). An example screenshot from the Exogram EB vetting portal is shown in Figure 14.

Zoom In Zoom Out Reset image size

Briefly, the user first evaluates the data for clear evidence of eclipse-like features, paying close attention to the two panels on the right, and to the lower left panel. Next, Exogram automatically proceeds to the second question, where the user scrutinizes the phase-folded plot (middle left panel) and decides whether the measured period is correct. Finally, the portal takes the user to the last question, which provides the opportunity to mark the target as particularly noteworthy. Several predefined options are provided, corresponding to the most commonly observed characteristics.

Importantly, Exogram is designed to enable fast image classification through the use of keyboard shortcuts. We found this to be a critical advantage as it allows an expert vetter to classify images with a typical “cruising speed” on the order of seconds, especially when the data clearly indicates a typical EB system.³⁷ Naturally, more interesting cases such as those exhibiting additional eclipse-like features take longer to inspect, as do targets dominated by systematics such as momentum dumps, but even for these the access to keyboard shortcuts significantly decreases the response time.

The portal also provides links to external tools such as the Fast Lightcurve Inspector (FLI³⁸ ) and LATTE (N. Eisner 2022) that enable deeper investigation of potentially interesting or particularly challenging targets. Both tools allow researchers to interactively examine the entire lightcurve, making it easier, for example, to distinguish between genuine eclipses and momentum dumps or other sources of interference. It is worth noting that FLI was created by one of our superusers (J. S. de Lambilly) and is designed as a free, online, user-friendly, interactive tool for visual inspection of TESS data, including BLS analysis and phase folding. FLI uses the Lightkurve package (Lightkurve Collaboration et al. 2018) to query MAST for all available GSFC-ELEANOR-LITE, QLP, SPOC, and TESS-SPOC lightcurves, along with corresponding diagnostics (e.g., background flux and centroid measurements), and presents them in a Bokeh/Plotly environment.

Exogram was developed using SvelteKit, a modern web application framework. The database and authentication are handled by Google’s Firebase platform, and the website itself is hosted by Vercel. We store the lightcurve images on Google Drive. Behind the scenes, the Exogram server tallies the number of responses, labels targets as fully classified if three users have already commented, and removes them from the pool of images shown. Additionally, the Exogram platform integrates social media–like features to encourage collaboration between users. For example, users can “star” a target to save it for later inspection. Starred EBs are public, and users can see which targets were saved by others. This makes it easy for vetters to find targets that others deemed interesting or unusual. Users can also share targets with each other, and the integrated notification feature alerts users when something has been shared with them. Finally, the platform also shows a vetting leaderboard to encourage friendly competition among the users.

4.3.2. EBP

To further capitalize on the power of citizen science, and inspired by the success of the Planet Patrol project (V. B. Kostov et al. 2022a), we proceeded with the investigation of all 60K targets by developing the EBP project.³⁹ EBP is hosted on Zooniverse and provides a streamlined, interactive, and user-friendly platform to visually inspect a summary of the results produced by the automated pipeline described above. EBP was launched on 2024 September 3, and was completed on 2025 March 26, during which period ∼1800 participants produced ∼320,000 classifications.

The EBP workflow consists of four questions aimed at evaluating whether the target is indeed an EB candidate, whether the calculated period is correct, and whether the photocenter measurements are reliable. An example screenshot of the classification scheme is shown in Figure 15. The volunteers inspect the original and phase-folded lightcurves, and decide whether they see periodic eclipse-like signals, check if the period is correct, and scrutinize the lightcurve for secondary eclipses. Additionally, they evaluate the quality of the difference image and classify it as either appropriate for photocenter measurements—i.e., the image shows a well-defined bright spot on an otherwise dark background—or otherwise.

Zoom In Zoom Out Reset image size

The EBP portal provides extensive background information on the science of EBs and their astrophysical importance, a comprehensive tutorial with a step-by-step demonstration of the workflow, a field guide presenting relevant examples, edge cases, etc. The portal also includes guidelines on how to interpret and classify the images, as well as an active talkboard where volunteers can discuss targets of interest and ask for help from the science team. Each image presented to the volunteers also contains additional auxiliary information enabling more detailed investigation of the inspected target, particularly with the help of FLI. We note that classifications on EBP are not strictly blind. Volunteers could freely look up outside information, based on the provided TIC ID, which could potentially affect their evaluation. Finally, volunteers interested in contributing to the vetting process beyond the Zooniverse project are invited to join the science team.

For completeness, we would like to briefly share the experience gained and lessons learned from EBP. First and foremost, frequent interactions between the science team and volunteers on the talkboards, especially during the initial stages, were critical for the success of the project. These interactions ensured timely resolution of technical issues, addressed vetting and scientific questions, and enabled live updates aimed at improving the overall workflow. For example, prompted by feedback from volunteers we quickly refined the FAQ and tutorial by adding new examples, clarifying existing instructions, etc. Finally, consistent communication, including social media posts highlighting interesting targets and celebrating milestones, helped retain user engagement throughout the duration of the project, averaging about 1000 classifications per day, even months after launch.

The workflow of EBP is designed such that each target is considered as fully classified when at least five different volunteers have inspected the corresponding image and answered the provided questions. It is worth pointing out that at the launch of the project, the image “retirement” limit was set to nine. However, that proved to be too high as the rate of completed classifications was rather slow. Thus, in order to speed up the vetting and complete the 60K sample in a timely manner, three weeks after the project was launched we reduced the limit to seven, and shortly after down to five.

To adopt an aggregate response to each question, we tested three options: (i) a simple majority, i.e., at least three out of five volunteers select the same answer; (ii) at least four out of five; and (iii) five out of five. To evaluate the reliability of these aggregates, we checked the corresponding classifications for a random sample of 1000 targets where the measured period was classified as correct. Overall, we found that about 75%, 85%, and 90% of the responses are correct for options (i), (ii), and (iii), respectively. In order to increase the fidelity and maximize the reliability of the classifications, members of the science team performed complementary visual inspection of all potentially new EB candidates for which at least three out of five volunteers indicate that the period is correct.

Altogether, 7936 targets passed the following vetting and validation tests: (1) the eleanor lightcurve shows clear eclipses; (2) the measured period is correct; (3) the difference images used for photocenter analysis are of sufficiently high quality for reliable measurements; (4) the measured photocenter offsets are smaller than 0.2 pixels; and (5) no field stars from the Gaia catalog and TIC are within 0.2 pixels of the target, and bright enough to produce the detected eclipses as contamination. An example of a target that passes the first four tests but fails the last is TIC 187172446, shown in Figure 16. Here, there is a nearby field star, TIC 510123334, that is about 1 TESS magnitude fainter and at a projected separation of 0$\mathop{.}\limits^{\unicode{x02033}}$17. Thus, while the measured photocenter offset is about 0.14 pixels, it is impossible to tell from the TESS data which of these two stars is producing the detected eclipses.

Zoom In Zoom Out Reset image size

To measure the eclipse times, depths, and durations, we adopted the methodology of V. B. Kostov et al. (2022b) and, for each sector, fit each eclipse with its generalized Gaussian model:

$$\begin{eqnarray}&&F(t)=A-B{e}^{-{\left(\tfrac{| t-{t}_{{\rm{o}}}| }{\omega }\right)}^{\beta }}+C(t-{t}_{{\rm{o}}}).\end{eqnarray} \tag{ 1 }$$

For illustrative purposes, Figure 18 shows the model fit to all phase-folded primary eclipses of TIC 470715046, as well as a Gaussian and a trapezoid fit for comparison. As seen from the figure, the generalized Gaussian model provides an excellent fit to the data—certainly better than both the trapezoid and the narrower Gaussian model—and we used it for measuring the eclipse depths and durations.

Zoom In Zoom Out Reset image size

Table 3 provides the median depths and durations for the new EBs presented here. The corresponding distributions are shown in Figure 19. The primary depth distribution has mean/median/95th percentile values of 91/62/276 ppt, respectively; the mean/median/95th percentile values for the primary duration distribution are 4.2/3.9/7.4 hr, respectively. Roughly half of the targets exhibit secondary eclipses. As highlighted in Figure 20, most of these occur near an orbital phase of 0.5, and about 95%/99% of them reside within the phase range of ∼0.43–0.58/∼0.3–0.72, respectively. The two most extreme secondary phases in our catalog are for TIC 149673382, with a secondary phase of ≈0.87, and TIC 337097515, with a secondary phase of ≈0.11, both exhibiting a pronounced heartbeat “bump” between the primary and secondary eclipses.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

We note that several targets in our catalog have primary depths larger than 0.5 according to the current version of eleanor data, likely due to systematics. The three most extreme cases are TIC 42066695 (average depth of ≈848 ppt), TIC 446208053 (depth of ≈804 ppt), and TIC 192305147 (depth of ≈751 ppt); the phase-folded lightcurves for the first two are shown in Figure 21. The average primary depth for 42066695 is much smaller, ≈434 ppt; for the other two targets there is no publicly available QLP data at the time of writing.

Zoom In Zoom Out Reset image size

It is important to note that the observed eclipse depths often vary from one sector to the next. With a handful of exceptions, mentioned below, these depth variations are due to systematic effects inherent to the lightcurve extraction process. An example is shown in Figure 22 for the case of TIC 5232381, where the eclipses in Sector 9 (left panel) are about half as deep as those in Sector 62 (right panel). This is likely due to the sector-specific background subtraction being affected by TIC 5232374, ≈13″ away and about 2 mag fainter.

Zoom In Zoom Out Reset image size

TIC 5232381 is neither an isolated occurrence nor an outlier. Sometimes, eclipse depths can fluctuate even within a single sector, showing differences before and after TESS data downlink gaps. In the most extreme cases, the eclipses can be virtually undetectable in certain sectors, as highlighted in Figure 23 for TIC 77392704 (also, e.g., for TIC 63165670). Thus, in these cases it is preferable to exclude such sectors when phase-folding the lightcurve, which we did by visual inspection on a target-by-target and sector-by-sector basis.

Zoom In Zoom Out Reset image size

These complications are often further exacerbated when the number of detected eclipses is small due to relatively long orbital periods, data gaps, and systematic effects. Sometimes, even the addition of new sectors does not help improve the measurements. An example of this is shown in Figure 24 for the case of TIC 173706211, where the Sector 84 lightcurve is completely dominated by systematics and there is a single useful eclipse near the end of the sector. As a result, while obtaining reliable eclipse depths and durations for individual sectors is, in general, relatively straightforward, extending these measurements across multiple sectors is challenging and sector-averaged depths and durations can be misleading.

Zoom In Zoom Out Reset image size

Collectively, these factors underscore the complexity of obtaining accurate and precise measurements for many TESS FFI EBs. Thus, it is important to emphasize that even after thorough scrutiny it is still possible that the ephemerides provided in this catalog are slightly off, especially when the number of TESS observations is small, the eclipses are few and shallow, the SNR is low, and the lightcurve is dominated by systematics. Unfortunately, resolving these issues by, e.g., cross-checking EBs between different lightcurve pipelines is far from straightforward. For instance, it is not uncommon to see depth differences when comparing eleanor data to QLP or TESS-SPOC data, likely due to different treatments of crowding correction.⁴⁰

Finally, it is worth pointing out two further considerations regarding eclipse depth variations in particular. First, these do not substantially affect the ephemeris measurements as these are done on an eclipse-by-eclipse basis—each eclipse is independently modeled with a generalized Gaussian that fits for the depth by design. Neither do the depth variations dramatically impact our photocenter-based vetting as it mostly depends on the eclipse durations.⁴¹

Multiple stellar systems are not uncommon. About one in 10 binary stars reside in hierarchical (2+1) triples, and thousands of even higher order systems have already been discovered (e.g., D. Raghavan et al. 2010; S. Tremaine 2020; and references therein). The higher the multiplicity of the system, the higher its complexity in terms of orbital and physical parameters, formation and evolution pathways, and long-term dynamical stability (e.g., M. Moe & R. Di Stefano 2017; S. Tremaine 2020; A. Tokovinin 2021; and references therein). In general, wide multiple systems are prime targets for long-term astrometric monitoring, while compact multiples are ideally suited for observing short-term dynamical interactions between their components such as ETVs.

Crossmatching our catalog with the Gaia DR3 astrometric measurements, we extracted the available astrometric_excess_noise (AEN), astrometric_excess_noise_sig (AENS), and renormalized unit weight error (RUWE). These can be used to test for unseen companions (e.g., V. Belokurov et al. 2020; Z. Penoyre et al. 2020; K. G. Stassun & G. Torres 2021; P. Gandhi et al. 2022; S. R. Majewski et al. 2025, in preparation; and references therein), which, if indeed present, would potentially mark the EBs as components in systems of three (or more) stars. The corresponding distributions are shown in Figure 25, highlighting several interesting features. In particular, the AEN is greater than 10 mas for hundreds of targets, and the AENS is greater than 3 for ∼40% of the EBs, reaching values of tens to even hundreds of thousands for dozens of targets. Similarly, the RUWE is greater than 1.4—suggesting unresolved companions (K. G. Stassun & G. Torres 2021)—for about one in every four targets. Altogether, these considerations indicate that a potentially large fraction of the 7936 new EBs presented here may reside in multiple stellar systems.

Zoom In Zoom Out Reset image size

Another option for finding multiple stellar systems is through the presence of extra events in the lightcurves of EBs. Indeed, TESS has already enabled the detection of thousands of such events, practically revolutionizing the field by discovering hundreds of new 2+1 triply eclipsing triple systems (e.g., T. Borkovits 2022; S. A. Rappaport et al. 2022, 2024; V. B. Kostov et al. 2024b; and references therein) and 2+2 eclipsing quadruple systems (V. B. Kostov et al. 2022b, 2024a; P. Zasche et al. 2024; B. P. Powell et al. 2025), as well as unusual (2+1)+1 eclipsing quadruples (e.g., B. P. Powell et al. 2022b), several (2+1)+2 quintuple systems (V. B. Kostov et al. 2022b, 2024a), the first two (2+2)+2 eclipsing sextuple systems (B. P. Powell et al. 2021a; P. Zasche et al. 2023), and even two transiting circumbinary planets (V. B. Kostov et al. 2020, 2021a). Volunteers at EBP have independently rediscovered many of these and, naturally, a significant number of false positives that mimic 2+2 eclipsing quadruples due to blended light from two unrelated EBs,⁴² and also identified several new eclipsing triple and quadruple candidates (V. B. Kostov 2025, in preparation).

Given our ML search was effectively blind, it was inevitable that it picked up a large number of known EBs. And indeed, as discussed above, about one in four of the identified candidates are within 1 pixel of known EBs. Thus, in order to verify the efficiency and reliability of our automated ephemeris and vetting pipeline, we applied it to ≈30,000 such targets and tracked its performance. Interestingly, during the early stages of the EBP project, the volunteers noticed that the correctly measured periods from TESS were sometimes different from the literature values. Altogether, we marked 2065 such cases. As an example, the distributions of the period ratios between TESS on the one hand and Gaia, ASAS-SN, ATLAS, and VSX on the other are highlighted in Figure 26.

Zoom In Zoom Out Reset image size

As seen from the figure, most of the Gaia, ASAS-SN, ATLAS, and VSX periods are close to an integer fraction of the true period, where for simplicity “close” is defined as within 10% of integer fractions of 2 (from 1/2 to 10/2) and 3 (from 1/3 to 20/3).⁴³ This is perhaps not too surprising given the much longer continuous baseline coverage and higher cadence of TESS observations compared to other surveys. As an example, Figure 27 shows TIC 2597145, where the correct period measured from TESS is 1.4143 days, twice the period listed in Gaia (0.7072 days). For this target, the periods listed in ASAS-SN and VSX are correct. Another example is TIC 9473243, where the correct period measured from TESS is 2.2669 days, whereas Gaia gives a period of 9.0680 days (4 times as long), ASAS-SN gives a period of 4.5338 days (2 times as long), and WISE gives a period of 1.1335 days (only half as long). Additionally, TESS excels at enabling the detection of shallow secondary eclipses. An example of this is shown in Figure 28 for TIC 403072759, highlighting the shallow but clear secondary eclipse near the phase of 0.5. Here, the true period measured from TESS is 1.3029 days whereas Gaia gives a period of 0.52 days, i.e., a 2/5 fraction of the true period.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Interestingly, about 20% of the Gaia EB periods seem to be unrelated to the TESS periods at all. These cases are represented in Figure 26 by the peak at zero. One example is TIC 2239760, where the correct TESS period is 5.8855 days, while the Gaia period is 3.2370 days, a ratio of ≈0.55 (Figure 29, left panel). Another is TIC 143060048, where the TESS period is 4.2852 days and the Gaia period is 30.4715 days (ratio of ≈7.11; Figure 29, right panel). Some of the most extreme discrepancies are for TIC 443450339, 152328270, 353628656, 138032974, and 34853800, for which the TESS periods are 2.9032, 9.0706, 2.5816, 1.8713, and 3.2028 days, respectively, whereas the corresponding Gaia periods are orders of magnitude longer, i.e., 406.0884, 381.0741, 355.6257, 189.4914, and 164.4351 days.

Zoom In Zoom Out Reset image size

Table 4 highlights 10 random rows of the catalog of 2065 known EBs with updated ephemerides produced as part of this work.

Here, we highlight some of the more interesting targets independently identified as part of this effort, split into the following categories:

• 1.  
  Additional eclipses: Targets exhibiting extra events not associated with the EB signal, such as a second set of eclipses following a different period (representing a 2+2 quadruple system consisting of two EBs) or complex tertiary events (representing triply eclipsing 2+1 triple systems or eclipsing (2+1)+1 quadruple systems). Figure 30 highlights two such examples.
• 2.  
  ETVs: Targets where the eclipse times deviate from the linear ephemeris, suggesting potential dynamical interactions with additional bodies. An example is shown in the upper panel of Figure 31 for the case of the 2+2 quadruple system TIC 219006972, where the two EB subsystems are dynamically interactive on observable timescales (V. B. Kostov et al. 2023). Another example is the known EB TIC 26542657 (A. Prša et al. 2022), which, through our comparison with the Kepler catalog, we determined does not show the ∼11 day eclipses in Kepler and must therefore be a higher-order system. It is separated by only $\sim 1{\rm{^{\prime} }}{\rm{^{\prime} }}$ from TIC 1882992210, which is only ∼0.1 mag fainter in TESS, making photocenter confirmation of the eclipse source effectively impossible from TESS. However, the target exhibits clear primary and secondary ETVs (see Figure 31, lower panel), a tertiary eclipse in Sector 81, and prominent changes in the shape of both the primary and secondary eclipses between Sectors 14/15 (narrow, sharper primary, more rounded secondary) and later sectors (more rounded primary, flat secondary; see Figure 32). Taken together, these provide strong evidence that either TIC 26542657 or TIC 1882992210 is a dynamically interacting, triply eclipsing triple system with an outer period of about 300 days. Additionally, as seen from Figure 33, the Kepler lightcurve of the target shows one tertiary eclipse suggesting that the system was out of the eclipsing window for the ∼11 day EB during the Kepler era due to orbital precession.⁴⁴
• 3.  
  Disappearing eclipses: Targets where the detected eclipses exhibit prominent depth variations due to precession of the EB orbital plane, to the point of eventually ceasing altogether. Figure 34 shows the TESS lightcurves of TIC 236774836 (T. Mitnyan et al. 2024b) and TIC 220410224, indicating dynamical interaction with unseen companions.
• 4.  
  “Switching” eclipses: Similar to the previous example, but here the depth ratio between the primary and secondary eclipses changes between sectors. Figure 35 shows an example of this effect for the case of the known EB TIC 234229841.
• 5.  
  Apsidal motion: Targets exhibiting pronounced “smear” of the secondary eclipses in orbital phase, indicating apsidal motion. Figure 36 highlights two such targets, TIC 189281140 and TIC 470715046.
• 6.  
  Stellar variability: Targets exhibiting prominent lightcurve modulations due to, e.g., rotational variability (spotted stars), pulsating components, heartbeat patterns, etc. Figures 37 and 38 show examples of each category, represented by TIC 21159577, TIC 22621932, and TIC 336538437.
• 7.  
  Transiting planets: It is only logical that a search for stellar eclipses will result in finding planetary transits as well. Indeed, our ML pipeline picked up the confirmed planet TIC 408310006 (WASP-166 b). Interestingly, an eagle-eyed volunteer on EBP (D. Iannone) noticed an additional transit-like event in the lightcurve of the target in Sector 62.⁴⁵ Further investigation showed another event in Sector 89 (see Figure 39), suggesting the potential presence of a second transiting planet in the system.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size