🔭 astrophysics

Exploring Long-period Architectures: Four New Planet Candidates from Kepler with Periods >342 days

This paper presents a new single-transit detection pipeline utilizing convolutional neural networks and spacecraft diagnostics to identify four long-period planet candidates (with periods exceeding 342 days) in Kepler systems, all of which host inner planets exhibiting transit timing variations.

Original authors: Matthew T. Hansen, Jason A. Dittmann

Published 2026-08-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Matthew T. Hansen, Jason A. Dittmann

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

For decades, astronomers have been busy counting the worlds orbiting other stars, but their view has been skewed by the tools they use. The most successful method for finding these planets involves watching a star for a tiny, periodic dip in its brightness, which occurs when a planet passes directly in front of it. This technique is incredibly effective at finding planets that hug their stars closely, because they cross the star's face frequently, offering many chances to be seen within a few years of observation. However, this same method struggles with planets that take a long time to complete an orbit. If a planet takes several years to circle its star, it might only pass in front of it once or twice during the entire lifespan of a telescope's mission. These distant worlds remain hidden in the data, leaving a significant gap in our understanding of how planetary systems are built. We know little about the architecture of systems that might resemble our own solar system, where giant planets orbit far from the sun.

A team of researchers has now developed a new way to hunt for these elusive, long-period worlds using data from the Kepler space telescope, which observed the same patch of sky for four years. Instead of relying on the traditional method of looking for repeating patterns, which fails when there are very few repeats, the team built a machine learning system designed to spot a single transit event. This system was trained not just on the brightness of the stars, but also on the internal health data of the spacecraft itself, such as the temperature of its instruments and the performance of its stabilizing wheels. By combining these two streams of information, the computer learned to distinguish between a genuine planet crossing a star and the subtle, systematic errors that often mimic them. The researchers applied this new pipeline to thousands of star systems that were already known to host at least one planet, specifically looking for additional, longer-period companions that might have been missed.

The search yielded four new planet candidates, all found in systems where the known inner planets were already behaving strangely. These inner planets were exhibiting transit timing variations, meaning they were crossing their stars at slightly different times than a perfect clockwork orbit would predict. This irregularity is a strong hint that another, unseen body is tugging on them gravitationally. The new candidates represent the potential culprits for these gravitational nudges, though the team found that simply adding these new planets to the models did not fully explain the observed wobbles without assuming complex, unstable orbital configurations. Two of the new candidates, Kepler 1752.02 and Kepler 199.03, were detected crossing their stars twice, allowing the team to calculate their orbital periods with high precision. Kepler 1752.02 takes approximately 778 days to orbit its star, while Kepler 199.03 takes about 505 days. These are massive worlds, roughly 3.5 and 2.7 times the size of Earth, respectively. The other two candidates, Kepler 1897.02 and Kepler 1811.02, were seen only once, making their exact orbital periods impossible to pin down, though they are known to take at least 342 and 544 days to complete a single lap.

Despite finding these promising signals, the researchers faced a significant hurdle: the new planets they found could not fully explain the gravitational disturbances seen in the inner systems if they were moving in simple, circular orbits. The team ran detailed computer simulations to test if these new planets could be the source of the timing variations. They found that while the new planets existed, they were too far away and too massive to cause the specific wobbles observed in the inner planets without creating a dynamically unstable system that would likely result in a collision or ejection of a planet. This suggests that the true cause of the timing variations might be a different, unseen planet located somewhere between the inner and outer worlds, one that is too small or dim to have been detected by the telescope. For the systems where the team could not find a simple explanation, they explored the possibility that the hidden perturber is not even aligned with the plane of the other planets, making it invisible to the transit method entirely.

The discovery of these four candidates adds to a very small but growing list of long-period exoplanets found by the Kepler mission. Currently, only a tiny fraction of the thousands of planet candidates identified by Kepler have orbits longer than a year. These new findings help fill in the map of our galaxy's planetary diversity, showing that systems with widely spaced worlds are not just theoretical possibilities but real, observable structures. However, confirming these planets is a formidable challenge. Because they take so long to orbit, waiting for the next transit event could mean waiting years, and the transits themselves are long, lasting more than ten hours, which makes them difficult to catch with ground-based telescopes. The researchers suggest that future space missions, such as the upcoming PLATO telescope, may be the only instruments capable of catching these rare, slow-moving worlds in the act. Until then, these candidates remain intriguing possibilities, hinting at complex, multi-layered solar systems that are still waiting to be fully understood.

Technical Summary: Exploring Long-period Architectures

Problem Statement
The Kepler detection pipeline and the transit method inherently possess a bias toward shorter orbital periods, resulting in a significant dearth of detections for long-period exoplanets. This lack of data leaves the architectures of exoplanet systems in the long-period regime incomplete. Furthermore, the determination of $\eta$ -Earth (the frequency of Earth-like planets) is currently dominated by systematic corrections and extrapolations due to the falloff in detections at long periods. While visual inspection of lightcurves has historically been used to find long-period candidates, it is tedious and often misses signals that automated detrending techniques might remove if the transit duration exceeds the frequency threshold of the detrending algorithm. Additionally, many known systems exhibit Transit Timing Variations (TTVs) in their inner planets, suggesting the presence of perturbing bodies that have not yet been detected, particularly in the long-period regime.

Methodology
The authors developed and applied an updated single-transit detection pipeline to address these gaps. The methodology involves the following key components:

Neural Network Architecture: The pipeline utilizes a binary classification Convolutional Neural Network (CNN) combined with a Fully Connected Neural Network (FCNN). The input consists of 128 long-cadence flux segments from stellar lightcurves, concatenated with nine onboard spacecraft engineering attributes (diagnostics such as reaction wheel states and attitude pointing errors) processed through an FCNN block. This "flux-engineering" approach aims to improve classification accuracy by leveraging spacecraft systematics that correlate with false positives.
Training Data: The network was trained on a dataset derived from the NASA Exoplanet Archive, filtered for orbital periods >12.5 days and transit depths <350 ppm to focus on small planet signals. The dataset was split into training, validation, and testing sets based on Kepler Input Catalog (KIC) numbers to prevent data leakage.
Detection Strategy: The pipeline applies a sliding window approach across the lightcurves of 1,605 systems known to host at least one planet with a period >6 days. It calculates a "fractional classification rate" for each time stamp, representing the frequency with which the network flags a window as containing a transit.
Candidate Identification: Threshold Crossing Events (TCEs) are identified where the fractional classification rate exceeds 50%. The pipeline filters out TCEs associated with known planets, systematic artifacts (identified via time-bin histograms and visual inspection of common false positives), and poor detrending.
Vetting and Modeling: Remaining TCEs undergo rigorous manual vetting, including checks for centroid offsets, blending from background stars, and quality flags. For candidates with multiple transits, periods are derived via Lomb–Scargle periodograms. For single-transit candidates, allowed periods are constrained by analyzing data gaps within the Kepler dataset.
Dynamical Analysis: The authors employ N-body simulations (REBOUND) and TTV fitting codes (TTVFast) to test if the new candidates can explain the observed TTVs of inner planets. They also perform injection-recovery tests using Box Least Squares (BLS) to determine the detection thresholds for undetected perturbers in the "in-between" regions of these systems.

Key Results
The application of this pipeline to the Kepler field yielded four new long-period planet candidates, all located in systems where inner planets exhibit TTVs:

Kepler 1752.02: A candidate with two visible transits, consistent with an orbital period of $777.78 \pm 0.02$ days and a radius of $3.55 \pm 0.15 R_{\oplus}$ . This would place it among the top 20 longest-period transiting planets known.
Kepler 199.03: A candidate with two visible transits separated by 1010 days. Due to a large data gap between them, a 2:1 alias period of $505.495 \pm 0.004$ days is also permitted. The candidate has a radius of $2.74 \pm 0.05 R_{\oplus}$ .
Kepler 1897.02: A single-transit candidate with a radius of $4.81 \pm 0.19 R_{\oplus}$ . The minimum allowed period, constrained by data gaps, is 342 days, though a range of periods up to ~700 days is possible.
Kepler 1811.02: A single-transit candidate with a radius of $3.25 \pm 0.30 R_{\oplus}$ . The minimum allowed period is 544 days.

Dynamical Constraints and TTVs:

The new candidates themselves, assuming circular orbits, are incapable of reproducing the observed TTV signals in the inner systems of their respective hosts.
The authors investigated the possibility of an undetected "in-between" perturber causing the TTVs. For Kepler 1752 and Kepler 1897, they found that a coplanar, transiting perturber (small enough to evade BLS detection) could theoretically explain the TTVs while maintaining dynamical stability.
For Kepler 199 and Kepler 1811, no stable, coplanar, transiting solution was found that could explain the TTVs. For Kepler 1811, a solution was found when allowing for a non-coplanar (mutually inclined) perturber; however, the predicted stability (0.707) fell below the authors' chosen threshold of 0.80, indicating a precarious or less robust stable configuration compared to other systems.

Significance and Claims
The paper claims to expand the known population of long-period transiting planets, addressing the observational bias against long-period detections. By utilizing engineering data alongside photometry, the authors demonstrate a method to detect single-transit events that might otherwise be missed or discarded by standard pipelines.

The authors modestly state that while these candidates are significant additions to the long-period census, the current Kepler dataset is insufficient to fully constrain their orbital parameters (specifically eccentricity and precise periods for single-transit candidates) or to definitively identify the perturbers causing the TTVs. They emphasize that follow-up observations are necessary to confirm these candidates and to dynamically model the systems. The paper suggests that future missions like PLATO, with longer continuous observation baselines, are essential for confirming these long-period candidates and bridging the gap between Kepler's photometric detections and the astrometric yields expected from Gaia and Roman. The work highlights the complex architectures of multi-planet systems, where inner and outer planets may be dynamically decoupled, and underscores the need for improved detection techniques to fully map the exoplanet population.

More like this