≡ Menu

Dataset Mining Reveals New Planets

I’m always interested in hearing about new ways to mine our abundant datasets. Who knows how many planets may yet turn up in the original Kepler and K2 data, once we’ve applied different algorithms crafted to tease out their evanescent signatures. On the broader front, who knows how long we’ll be making new discoveries with the Cassini data, gathered in such spectacular fashion over its run of orbital operations around Saturn. And we can anticipate that, locked up in archival materials from our great observatories, various discoveries still lurk.

Assuming, of course, we know how to find them and, just as important, how to confirm that we’re not just looking at noise. What scientists at the Max Planck Institute for Solar System Research (MPS), the Georg August University of Göttingen, and the Sonneberg Observatory have come up with is 18 new planets roughly of Earth size that they’ve dug out of K2, looking at 517 stars that, on the basis of earlier analysis, had already been determined to host at least one planet.

The Kepler mission was jeopardized by the failure of critical system components in its reaction wheel system, out of which emerged K2, looking not at the original 156,000 target stars but at a series of fields and a total of more than 100,000 stars for which light curves could be observed. None of the 18 new worlds were detected in earlier analysis because the search algorithms in play were not sensitive enough. But René Heller (MPS) and Michael Hippke (Sonneberg Observatory), working with Kai Rodenbeck (University of Göttingen) believed the sensitivity of the transit method would be enhanced by modeling a more realistic light curve.

“Standard search algorithms attempt to identify sudden drops in brightness,” explains Heller. “In reality, however, a stellar disk appears slightly darker at the edge than in the center. When a planet moves in front of a star, it therefore initially blocks less starlight than at the mid-time of the transit. The maximum dimming of the star occurs in the center of the transit just before the star becomes gradually brighter again.”

Most of the confirmed planets and those yet to be confirmed from both the original Kepler mission and K2 have been examined with a transit search algorithm called BLS — box least squares — in which, like other algorithms in play, the software looks for box-like decreases in flux from the light curve of the star. Heller and Hippke developed TLS — transit least squares — specifically to look for smaller planets by modeling the stellar limb darkening Heller mentioned above. The method is optimized for the detection of shallow periodic transits. The scientists believe the false-positive rate is suppressed and signal detection enhanced by this method.

From the paper:

In contrast to BLS, the test function of TLS is not a box, but an analytical model of a transit light curve… As a consequence, the residuals between the TLS search function and the observed data are substantially smaller than the residuals obtained with BLS or similar box-like algorithms, resulting in an enhancement of the signal detection efficiency for TLS, in particular for weak signals.

Image: If the orbit of an extrasolar planet is aligned in such a way that it passes in front of its star when viewed from Earth, the planet blocks out a small fraction of the star light in a very characteristic way. This process, which typically lasts only a few hours, is called a transit. From the frequency of this periodic dimming event, astronomers directly measure the length of the year on the planet, and from the transit depth they estimate the size ratio between planet and star. The new algorithm from Heller, Rodenbeck, and Hippke does not search for abrupt drops in brightness like previous standard algorithms, but for the characteristic, gradual dimming and recovery. This makes the new transit search algorithm much more sensitive to small planets the size of the Earth. Credit: © NASA/SDO (Sun), MPS/René Heller.

What we find in this initial cut using TLS is 18 planets that have been overlooked in previous work. The new worlds turn out to be, in most of the systems studied, the smallest planets, and most of them orbit their star closer than the previously known planetary companions. So with one exception we’re talking about planets with temperatures between 100 and 1,000 degrees Celsius. The exception is an apparent super-Earth (EPIC 201238110.02) circling a red dwarf star possibly in its habitable zone.

A system that stands out is that around the star known as K2-32. Heller and Hippke are able to identify a fourth transiting world here to join the three Neptune-size planets already known, this one with a radius roughly the size of Earth’s. With four planets orbiting it, K2-32 joins a list of about a dozen K2 stars with four or more transiting planet candidates. As the paper notes, only K2-138 has been found to host five such planets, all in super-Earth to sub-Neptune range.

The authors discuss the K2-32 system in detail in the first of two new papers, with the second examining the 17 other planets their methods have uncovered. Here we find worlds ranging from 0.7 Earth radii to 2.2 R. That smallest world (EPIC 201497682.03) is the second smallest ever discovered with K2. Half of the planets are smaller than 1.2 Earth radius.

The success of the TLS methodology is provocative and leads to expectations of further discoveries. After all, beyond the K2 data, we have thousands of datasets for other stars. From the paper:

Based on our discovery rate of one new planet around about 3.5 % of all stars from K2 with previously known planets, we expect that TLS can find another 100 small planets around the thousands of stars with planets and candidates from the Kepler primary mission that have been missed in previous searches.

Image: Almost all known exoplanets are larger than Earth and typically as large as the gas planet Neptune. The 18 newly discovered planets (here in orange and green), for comparison, are much smaller than Neptune, three of them even smaller than Earth and two more as large as Earth. Planet EPIC 201238110.02 is the only one of the new planets cool enough to potentially host liquid water on its surface. Credit: © NASA/JPL (Neptune), NASA/NOAA/GSFC/Suomi NPP/VIIRS/Norman Kuring (Earth), MPS/René Heller

The papers are Heller, Hippke & Rodenbeck, “Transit least-squares survey. I. Discovery and validation of an Earth-sized planet in the four-planet system K2-32 near the 1:2:5:7 resonance,” Astronomy & Astrophysics, 625, A31 (abstract); and “Transit least-squares survey. II. Discovery and validation of 17 new sub- to super-Earth-sized planets in multi-planet systems from K2,” Astronomy & Astrophysics, 2019, in press (abstract).


Comments on this entry are closed.

  • Gary Wilson May 24, 2019, 15:25

    A brilliant re-analysis of part of the K2 dataset. We should see many more planets be found with this method. Of the 18 found so far, one is in it’s star’s habitable zone and possibly a rocky planet I believe (EPIC 201238110.02 I think).

  • Gary Wilson May 24, 2019, 16:31

    Sorry, it’s already in the article. I should read more carefully next time. The Kepler data sets are going to continue to be a treasure trove for a long time to come.

  • Alex Tolley May 24, 2019, 17:48

    I’m a little surprised that the TLS algorithm can do better simply by modeling the effect of limb darkening. This seems such a simple change.
    What is not clear is why, in such noisy data, this has that much extra power.

    I hope that we eventually get methods that will allow some sort of “ground truth” verification to ensure that these various algorithms have good discrimination and very low false positive and false negative rates.

    At some point, if the previous high definition astrometry method is operational, it should be compared to the future versions of the transit method on the same stars to provide orthogonal methods to determine the overlap of results.

    At some point in the future, we are going to be sending probes out to the most interesting nearer stars, and we will need the best information we have to ensure that the targets planets are really what we expect so that the cost and long mission times will not prove wasteful.

    I am always reminded by Von Braun’s logic about the conditions of Mars in his book “The Mars Project”. He made an interesting logical case based on some data that Mars would have a thin atmosphere, enough to avoid needed full pressure suits, but more importantly, enough density for his winged landers to use for gliding to a landing. Laughable in retrospect, but I wonder what our computer-aided logic failures are today for exoplanet conditions. They are a modern equivalent of mistaking a map for the territory, and so their results should be used with some caution, if not skepticism until we can validate them.

    • Alex Tolley May 25, 2019, 11:58

      The “box” model will show a flat–bottomed dip with steeply sloping edges as the plant partially, then fully enters the transit. Their “limb” model extends the edges some/all distance into teh transit, either making a non-flat signal, or one with a narrower flat bottom region. Given the large amount of uncertainty and noise in the signals, it surprises me that this modification can expose more planets that were previously “hidden”. I suspect it is rather like finding some statistical p values changing from just marginally too high to just marginally inside the needed threshold. I suspect it probably also invalidates some findings too, simply due to chance.

      • Alex Tolley May 25, 2019, 13:25

        I was wrong. Reading the Hippke paper on the method [ https://doi.org/10.1051/0004-6361/201834672 ] the TLS method does indeed have a better true positive rate of detection for the same false positive rate.

        Whilst they use averaged transit data to create the template for the method, the ability to feed in different templates offers great promise for finding other objects:

        It would also be possible to feed TLS with an analytic
        description of exocometary transits (Rappaport et al. 2018;
        7 https://keplergo.arc.nasa.gov/
        Kennedy et al. 2019) or disintegrating planets with comet-like
        tails (Bochinski et al. 2015; Garai 2018), atmospheric refraction
        (Dalba 2017, 2018), exoplanetary rings (Barnes & Fortney
        2004; Ohta et al. 2009; Tusnski & Valio 2011; Aizawa et al.
        2017; Hatchett et al. 2018) or artificial shapes such as rectangles
        (Arnold 2005) as well as starshades at the Lagrange points
        (Gaidos 2017; Moores & Welch 2018).

        This also strikes me as a gift for SETI searches of alien artifacts. Model examples of, e.g. Dyson swarms, and scan the complete Kepler set for periodic matches to that template.

        I note that there is some effort to reduce the extra compute effort of this technique, with mention of workstation compute times. Can an astronomer please confirm that this is no longer an issue given the availability of compute engines in the cloud. Parallel instances in the 100s or 1000s are easily achieved, allowed complete runs against 100,000s of stars practically overnight, rather than running serially on workstations or lashed-up Linux box server farms of the turn of the century vintage.

  • Enzo May 24, 2019, 23:41

    Does anyone know whether 201238110.02 is in the “optimistic” HZ (i.e. the one that nearly includes Venus) or the realistic one (i.e. 0.9-0.95 AU solar system equivalent) ?
    BTW, interesting upgrade of the detection method. The results are yet more confirmation of the abundance of “compact systems” (i.e. systems with multiple earth sized/ mini neptunes planets in very tight orbits very close to the star).

    • Alex T. May 26, 2019, 3:36

      From article text you can easily detect that HZ it is super-optimistic approximation, because the real distance is not know yet.

  • Bruce D. Mayfield May 25, 2019, 1:06

    If they expect to find ~ 100 additional small planets by re-examining the primary Kepler planetary systems with this improved algorithm then they should also be able to find even thousands more planets by using this improvement on the rest of Kepler’s non-detection stars as well.

    • Hary R Ray May 28, 2019, 9:56

      Or even MORE important, CONFIRM the existance of REALLY INTERESTING CANDIDATES like KOI 7923.01, KOI 8012.01 and KOI 8174.01!

  • NS May 25, 2019, 2:50

    Are all possible Kepler planet detections verifiable with other techniques available today, or are there some that cannot be accepted (or rejected) because (say) Kepler was so sensitive that there are cases where no other instrument can get comparable data?

    • Bruce D. Mayfield May 25, 2019, 10:13

      Yes, since any system in which there is a transit would also have good RV method alignment. But are there enough spectrographic telescopes to look at all those unconfirmed targets enough times to build up the needed confirmation data?

  • wdk May 25, 2019, 10:12

    Verification of Kepler transit with astrometric means does pose some interesting problems. If we regard the celestial sphere as a plane in the immediate area of a star and planet, the luminous primary ( star) revolves around the plane either clockwise or counter clockwise with a projection based on the orbital plane of the two bodies and the intrinsic conic ( ellipse with an eccentricity). If the orbital and celestial plane match, the tracking is simplified. But short period oscillations such as hot jupiters would be the hardest to track. E.g., with 10 observing nights over a year I could plot out a circle were it in orbit similar to ours; a planet in an orbit of less than 36 days, well…

    But transits are 90 degrees to the celestial plane.

    • Bruce D. Mayfield May 26, 2019, 11:25

      An excellent feature of the astrometric method is that it can work with ANY orientation! For transiting systems stars will have sine wave curves (instead of loops) superimposed upon their proper motion paths.

  • Gary Wilson May 26, 2019, 14:40

    So with the new method are the researchers actually looking at the slope of the downward and upward curves of dimming and recovery? And are they actually mainly dependent upon the number of data points in the those areas as opposed to the overall dip and the total number of points contributing to that dip?

  • Michael May 26, 2019, 17:18

    Star spots are a great source of noise, just wondering if the new algorithm takes the limb entry and exit into account, subtracting those signals which should be very similar should help.

  • Gary Wilson May 27, 2019, 16:00

    I’m very excited to see exactly how many new planets in total this new algorithm turns up when it has run through both Kepler data sets. It should be a fairly large number surely?

  • Harry R Ray May 28, 2019, 9:47

    Paul Gilster: Could you ask Drs Heller and Hippke whether TLS is applicable to FOLLOW-UP observations of specific stars, like LHS 1140 by the Hubble and Spitzer space telescopes with respect to a SPECIFIC search for as yet undiscovered sub-Mercurian ultra short period planets? If so, then could you also include searches for ultra short period DWARF PLANETS around very small stars like Proxima Centauri and TRAPPIST-1? Spitzer recently conducted a 48 hour survey of Proxima Centauri and found no evidence for ANY Proxima b transits, but this EXISTING data could reveal sub Moon-sized objects with orbital periods of less than 48 hours. Finally, I am convinced that there SHOULD BE objects orbiting BETWEEN TRAPPIST-1b and TRAPPIST-1’s roche limit. TLS may be able to PROVE this, especially with a wealth of Kepler, Spitzer, AND Hubble observations.

    • Paul Gilster May 29, 2019, 6:15

      Here is René Heller’s response. I’ll follow this with Michael Hippke’s.

      Dear Harry Ray,
      sure, there is no reason why TLS cannot be applied to follow-up observations. Based on the Spitzer data quality reported by Jenkins+ (2019, https://arxiv.org/abs/1905.01336) transits of a planet that is 0.4 Earth radii in size would have been found with their method. Although TLS might be able to push the radius limit by a small amount, I think the key hurdle here is not so much in the transit detection algorithm but in the stellar activity. The resulting time-correlated noise can mimic a transit (see their Fig. 2; and see the disputed candidate transit by Li+ 2017: https://arxiv.org/abs/1712.04483).
      About the TRAPPIST-1, this system is already very tight. It might be possible from a stability point of view to pack another planet into an orbit interior to TRAPPIST-1b, but this hypothetical planet would have an orbital period of ~1 day and it would likely need to be sub-Earth mass. Michael Hippke and I actually looked for additional transits in the Kepler data of this star (https://arxiv.org/abs/1901.02015) but did not find evidence for another planet. I am convinced there is no more Earth-like transiting planet around TRAPPIST-1. What is the reason for you to be “convinced that there SHOULD BE objects orbiting BETWEEN TRAPPIST-1b and TRAPPIST-1’s roche limit.”?
      Best regards, René Heller

      • Harry R Ray May 29, 2019, 10:06

        I stand corrected. My original assumption was that there were initially one or more planets that formed interior to TRAPPIST-1b and migrated to positio0ns interior to TRAPPIST-1’s roche limit and were subsequently ripped apart by TRAPPIST-1’s strong tides and then some of the debris was pulled outside the roche limit due to gravitational interactions with TRAPPIST-1b with the debris congealing into several Moon to Ceres sized dwarf planets. However, if the only stable orbital period interior to TRAPPIST-1b is ~1 day, then I now doubt that TRAPPIST-1b’s gravitational pull is strong enough to lift debris into a circular orbit that high. Speaking of TRAPPIST-1: ArXiv:1905.11419. “The Chaotic Nature of TRAPPIST-1 Planetary Spin States.” by Vinson A., Tamayo D. & Hansen B. KEY QUOTE: “We show that these spin states are likely to be unable to sustain long-term stability within any of our simulations, suggesting that the spin evolves under the influence of tidal synchronization forces into quasi-stable atractor states which las on timescales of thousands of years.” Could Westeros be located on TRAPPIST-1e(lol)? On a more serious note, this is the second TRAPPIST-1 paper I have reported on this website in the last few days. I expect this to be just the tip of the iceberg as the TRAPPIST-1 conference approaches in 3 weeks.

    • Paul Gilster May 29, 2019, 6:16

      From Michael Hippke (and great thanks to both Drs. Hippke and Heller for responding):

      Yes, TLS well suited for the suggested transit searched. We estimate the TLS sensitivity to be typically ~10% better than BLS, so that smaller planets can be detected at the same signal-to-noise ratio threshold. It would be worthwhile to search the mentioned light curves with TLS in addition to previous BLS-based searches. In our original TLS paper, we have in fact performed a search in the Kepler K2 EVEREST / TRAPPIST data (see Figure 8 in https://arxiv.org/pdf/1901.02015.pdf). We did not find evidence for an additional planet in the TRAPPIST system. We have begun searching existing datasets and expect to publish many more candidate and validated planets in the future! Of course our resources are finite, but luckily TLS is open source and we have tutorials online to help everybody get started in joining the planet hunt! See https://github.com/hippke/tls

  • Geoffrey Hillend May 28, 2019, 16:52

    Quote by Bruce D. Mayfield: “An excellent feature of the astrometric method is that it can work with ANY orientation! For transiting systems stars will have sine wave curves (instead of loops) superimposed upon their proper motion paths.”

    The sign waves motion would occur when we look at the planetary system parallel to the orbital plane. When our view is perpendicular to it we see it from above as a circle, but as we change the angle more towards the horizontal it turns into an ellipse, but when we view it exactly at the orbital plane, the whole system would only appear move back and forth on a flat line instead of circle so the sine wave proper motion makes sense.

  • wdk May 28, 2019, 19:15

    Still contemplating the problem of astrometric measurements backing up transiting stars. If we’ve got a transit event in one telescope, then another
    telescope registers a specific set of celestial coordinates. Ninety degrees of planetary rotation later, a doppler reading ( not astrometric) would be registered toward the viewer, and in another 180 degrees, the doppler would be receding velocity. But the astrometric offset would be greatest at 180 degrees of planetary revolution after the transit.
    The effectiveness of astrometric measurements are going to fall off with distance to the primary star relative to parsecs. Transits can be observed (demonstrably with Kepler) over thousands of parsecs. Doppler will depend on ability to resolve wavelenths or lines. Astrometric measurements: If Jupiter causes 5 thousands of an AU
    magnitude oscillation, terrestrial planet oscillations are not going to be visible on the celestial sphere from many parsecs away.

  • Geoffrey Hillend May 31, 2019, 19:47

    Astrometry depends of the proper motion of the star, but not the angle of the orbital plane of the planets in the direction of the observer. Assume that the star and planetary system as a whole is are not moving towards the viewer or away with radial velocity but only vertically in relation to the observer.

    The doppler shift only applies to the radial velocity method of exoplanet detection, but is not necessary for the astrometry. The star system has to be moving towards or away from the observer, so that the light of the star is doppler shifted by it’s slight movement radially caused by the motion of the barycenter of the mass of the planets and Star..

  • Geoffrey Hillend May 31, 2019, 19:52

    My point is the wavelength of light should always be doppler shifted no matter how far away it is with radial velocity, but with astrometry distance might be more of a problem.