Marc Hon

Marc Hon

Postdoctoral Researcher

Research Projects

Discovery of a Planet Escaping Engulfment

We reveal the first discovery of a planet escaping engulfment from its host star up the red giant branch. The planet Halla (8 UMi b) orbits its host star Baekdu (8 UMi) perilously close at a distance of only 0.5 AU. We discover that Baekdu is a red clump star, implying that it should have once been sufficiently large to engulf Halla. In this study published in Nature, we explore the unexpected survival of Halla, which suggests a unique survival pathway involving a binary star history for Baekdu.

An artist's visualization of the merger between an expanding red giant and a white dwarf. Planet Halla, in the foreground, survives engulfment in this scenario as the merger prevents the red giant from expanding sufficiently large to reach the planet. Adam Makarenko/Keck Observatory.

The Henry Draper-TESS (HD-TESS) Catalog

I have published a catalog of asteroseismic masses, radii, and evolutionary states of > 1,500 bright giants with high SNR oscillation spectra in TESS’s Continuous Viewing Zones. The catalog is called HD-TESS, representing the some of the best and the brightest observed by the TESS space mission.

A starmap of bright red giants observed by TESS near the Continuous Viewing Zones cataloged in HD-TESS. Bordering the starmap are the oscillation spectra of several notable, bright red giants. M. Hon et al., 2022.

All-Sky Asteroseismology with TESS

I am interested in mapping the asteroseismology of stellar populations around the entire sky with NASA’s Transiting Exoplanet Survey Satellite. Curiously, we have yet to listen to the pulsations of much of the night sky around us, and my work with TESS seeks to change that.

The view of NASA's Transiting Exoplanet Survey Satellite across the night sky. The image dissolves into the heatmap of stellar masses inferred using the sounds of stars. Credit: NASA/MIT/TESS and Ethan Kruse (USRA), M. Hon et al., 2021.

Computer Vision in Asteroseismology (SLOSH)

In my paper I introduce the Solar-like Oscillations Shape Hunter (SLOSH), which is a 2D convolutional neural network that identifies the presence of red giant oscillations from images of frequency power spectra of stars. Red giant power spectra have very characteristic frequency-power profiles that show rich levels of complexity; this makes it non-trivial for asteroseismic data-fitting pipelines to automatically extract seismic information from a large sample of stars.

Power density of an oscillating red giant on log-log space. The power excess is where the oscillation modes that we use to probe the stellar interior are located in frequency. All red giants have this unique frequency-power morphology: a 'hump' sitting on a sloping background.

However, the morphology of these power spectra can be easily recognized visually by a human expert. Since convolutional neural networks are powerful deep learning tools that can mimic human visual recognition, SLOSH acts as a ‘machine expert’ that can visually identify oscillations within thousands of stars within seconds.

The lack of any explicit fitting of a mathematical model to power spectra makes SLOSH incredibly flexible. It has been used to find red giants in all Kepler data, and also for asteroseismic target selection in the K2 Galactic Archaeology Program. It is now being implemented for stellar classification on a grand scale with TESS through the TESS Asteroseismic Science Consortium, with an integrated version available here.

What makes SLOSH unique is that it uses a 2D representation - an image - of 1D data (power density) as input. There are a few advantages to such an approach:

  • The image representation is analagous to how human experts assess the presence of red giant oscillations - by viewing the plot of the power spectrum in log-log space (see this paper as an example).
  • It is easy to handle observations of different length. 4 year Kepler observations have high frequency resolution (high number of frequency bins), while 1 month TESS observations have low frequency resolution (fewer frequency bins). If a 1D neural network is used, the power spectrum must be manually binned for different observation lengths (neural networks typically have a fixed size of its input), and extra care must be taken such that the data is not over-smoothed or too sparse. Using an image representation, however, mitigates this requirement, allowing the network to learn on its own how a power spectrum with more/fewer frequency bins would appear.
AI Heatmap
The deep learning seismic detector determines the presence of red giant oscillations and precisely measures the characteristic oscillation frequency, $\nu_{\mathrm{max}}$, from the grayscale image (right panel) of a power spectrum plot (left panel). The red-purple regions on the right image are parts of the power spectrum that are given high `attention' by the detector -- these correspond to the very same features a human visual expert uses for identifying oscillations.

The use of 2D representations for 1D data is not a new concept, and methods such as Recurrence Plots and Gramian Angular Fields have been used in literature for 1D time series analyses. SLOSH shows that a suitable choice of data representation goes a long way in training a machine learning model.

Despite its name, SLOSH can also be applied to classify stars showing other forms of variability! The interative plot below shows how SLOSH can separate different stellar types fairly well just by looking at their frequencies. The distances between clusters represent how alike or different the power spectra are.

Red - Delta Scuti/Beta Cepheid; Gray - RR Lyrae/Cepheid; Orange - Constant (No Variability); Green - Contact Binary/Rotation; Pink - Eclipsing Binary; Blue - Aperiodic Variability; Purple - Solarlike Oscillator; Brown - Gamma Doradus/SPB.

Deep Learning Evolutionary State Classifier

Red giants are incredibly useful for a variety of Galactic studies because they are luminous, long-lived, and exist just about anywhere in the Milky Way. As these stars first evolve up the red giant branch (RGB), they burn hydrogen from a shell within the stellar interior. Eventually, they start burning helium in their core (HeB) and become clump stars. Measuring the ages of red giants requires us to know what evolutionary state red giants are in (RGB or HeB); the problem is both types occupy more or less the same temperature and luminosity ranges, making them look very similar from the outside!

While similar on the outside, the interiors of both types of giants are very different! Asteroseismology is the definitive way to differentiate them apart by measuring the period spacing of core-sensitive oscillation modes, which identifies clear differences between the internal structures of both types of stars. However, period spacing measurements require a star’s frequency spectrum to be finely resolved and have high S/N: requirements not often met by stars with shorter observations such as those from K2 or TESS.

The left column shows the oscillation modes of RGB and HeB stars. The former typically have 'neater' spectra as opposed to the latter. Folding the spectra by their large frequency separation (\Delta\nu) as showin in the right column further accentuates the orderliness of the frequency pattern. The evolutionary state classifier uses folded spectra as its input.

In my work, I train a 1D convolutional neural network to classify the arrangement of core-sensitive oscillation modes – these are very different for more evolved HeB stars as compared to RGB stars. The neural network classifies very well on 4-year Kepler data (99%), and it importantly allows the evolutionary states of vast numbers of red giants to be determined very quickly. In my follow-up work, I apply this to data of low frequency resolution (as low as 1 month’s worth of observations), and find that accuracies over 90% can still be recovered.

This method will soon be updated again to be even more robust to noisy data. It is currently used in the K2 Galactic Archaeology Program, and will soon be applied to TESS data!