Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Telluric Line Hack Week Wrap-Up

David W Hogg
February 28, 2019

Telluric Line Hack Week Wrap-Up

The slides presented at the end of the Telluric Line Hack Week held at the Flatiron Institute in 2019 February.

David W Hogg

February 28, 2019
Tweet

More Decks by David W Hogg

Other Decks in Science

Transcript

  1. Telluric Line Hack Week
    Wrap-up slides

    View Slide

  2. David W. Hogg (NYU) (MPIA) (Flatiron)
    What I did:
    ● I worked out (thanks Zechmeister) that the wobble post-processing is a
    self-calibration; this has data-combining value beyond wobble itself
    ● I completed my proof that binary convolution + Gaussian fit is equivalent to
    x-corr with a synthetic template.
    What I learned:
    ● There are hybrid methods (physics + data-driven) that would be easy to
    implement now.
    ● The implicit “divide by tellurics” in wobble is technically slightly wrong (thanks
    Sharon X Wang)!
    What I want to know more about:
    ● Are there observing strategy modifications that could improve our robustness
    to tellurics? Are these in conflict with other desiderata?

    View Slide

  3. Lee Rosenthal (Caltech)
    What I did:
    ● Ran wobble on Automated Planet Finder (APF) data, used variable RV option
    to extract good-looking telluric model
    What I learned
    ● How wobble works
    ● A ton about the current understanding of tellurics: how they are modeled, how
    individual instruments merit different approaches, etc.
    What I still want to learn:
    ● How telluric contamination affects measurement of H-alpha. Can hopefully
    quantify this with wobble-modeled APF data

    View Slide

  4. BJ Fulton (Caltech/IPAC/NExScI)
    Network of Robotic Echelle Spectrographs (NRES)
    Experiments with wobble
    Tau Ceti
    RMS = 7.1 m/s
    RMS = 17.5 m/s
    HD 19916

    View Slide

  5. Megan Bedell (Flatiron)
    What I did:
    ● (Hopefully) helped people get started with wobble
    ● Made various bug fixes to wobble, especially to the RV uncertainties (although
    these may still need work)
    What I learned:
    ● Every instrument has its own data challenges! (I keep relearning this...)
    ● Even if we have perfect knowledge of the tellurics, removing them is a
    mathematical challenge (cf. Sharon Wang’s work)
    What I want to know more about:
    ● Which external measurements (local environment monitoring, national-level
    weather service predictions, etc) are most feasible & most useful for predicting
    telluric variability behavior?

    View Slide

  6. Arpita Roy
    What I still want to do:
    What I did:
    What I learned:
    What I still want to learn:
    - Felt less alone in the battle against tellurics
    - Gathered strategies from other instrument
    teams for pipeline dev
    - Spend more time working with Chad on
    TERRASPEC
    - Implement automated (believable) telluric
    correction into HPF, NEID, KPF pipelines
    - Variety of mathematical tools we might still
    explore
    - What are going to be our common metrics
    to compare success at telluric correction
    -
    -

    View Slide

  7. Neil Cook (Universté de Montréal)
    What I did:
    ● To pull out the code from the SPIRou pipeline and try it on other data sets (as a proof of concept)
    ● Tested with SPIRou and CARMENES data
    ● Progress on github: https://github.com/njcuk9999/tellu_pca
    What I learned:
    ● Many having the same problems and wanting to use data-drive methods to correct tellurics (i.e. WOBBLE)
    ● Removing OH lines is important for work with Helium, HITRAN is used everywhere, dividing by tellurics is bad but
    everyone does it anyway, technical details about CARMENES, HPF etc
    Result:
    ● Have the code working but some problems → will need more work → extend to more instruments
    Worked with: Evangelos Nagel, Solène Ulmer-Moll
    Work this week
    (with spirou data)
    SPIRou pipeline
    (with same spirou data)
    Test on
    CARMENES

    View Slide

  8. Greg Mace (University of Texas at Austin)
    ● Planetary Spectrum Generator (PSG) https://psg.gsfc.nasa.gov
    ● “Earth Transmittance” - Load - set location, target, and instrument information - Generate spectra
    ● Command line API available with curl interface

    View Slide

  9. Brian Thorsbro (Lund)
    What I did:
    ● Played with the PSG
    What I learned
    ● Got a much deeper understanding
    of the issues around dealing with
    telluric line removals
    What I want to know more about
    ● How to combine stellar models and
    telluric models for grid searching
    HIP 89584 - O6
    Arcturus - K2

    View Slide

  10. What I did:
    ● Flattened IGRINS A0 spectra.
    ● Used Planetary Spectrum Generator to compute
    telluric model.
    ● Fit the telluric model, component by component, as
    a linear model to IGRINS observations

    What’s next:
    ● Improve IGRINS spectrum flattening.
    ● Regularize the wavelength solution.
    ● Solve instrument line profile to replace Gaussian.
    ● Test robustness of using a single model.
    What I learned:
    ● Telluric modeling has the potential to negate the
    need for A0s(!)
    Joe Llama (Lowell)
    --- IGRINS data --- PSG telluric model
    --- Optimized telluric fit
    With help from Greg Mace, Brian Thorsbro,
    Dan Foreman-Mackey

    View Slide

  11. Joe P. Ninan (PennState)
    What I did:
    ● Worked on developing non-linear dimension reduction of Telluric line
    variations using Diffusion maps.
    ● Worked on running HPF data in wobble.
    What I learned
    ● How close forward modelling techniques are to fitting Telluric lines as well as
    sky emission lines.
    ● How various groups are doing Telluric correction.
    ● The inverse transform from non-linear space of Diffusion maps is not trivial.
    What I want to know more about
    ● Computational methods to do inverse transform of Diffusion maps.

    View Slide

  12. Evangelos Nagel (University of Hamburg)
    What I did:
    ● Talking
    ● Apply PCA approach to CARMENES NIR data (Neil) is not trivial
    What I learned
    ● How other teams deal with telluric lines (especially Spirou => PCA approach
    developed by Etienne A. & Neil)
    ● That molecfit works with the outdated HITRAN version of 2008 => use 2016
    to solve badly corrected oxygen bands
    ● How Kyle models the OH sky emission lines (forward model with three
    parameters)
    ● Many things about TAPAS & HITRAN
    ● Using a hybrid method is the way to go in the future
    ● Many technicals details about Spirou & HPF

    View Slide

  13. Adrian Kaminski (Landessternwarte Königstuhl, University of Heidelberg)
    What I did:
    ● Comparing Wobble vs. Serval RVs
    ● Optimizing, how Wobble handles CARMENES spectra
    What I learned and would like to follow up on:
    ● The RVs are comparable in general
    ● Orders behave differently (d_RV(order); some fail)
    -> a) robust way of combining them for final RVs
    b) finding the reason for that
    i) quality of spectra? (but no dependency on SNR or airmass)
    ii) amount of tellurics and position wrt. the stellar spectrum (d_RVs show clear systematics
    with BERVs)

    View Slide

  14. Sharon Xuesong Wang (Carnegie DTM)
    What I did:
    ● Got wobble working on simulated data - not giving better RVs than CCF or
    forward modeling... still looking into it
    ● Improved forward modeling algorithm for fitting tellurics
    What I learned
    ● Make sure to use HITRAN 2016, which would make a difference for especially
    water and oxygen lines
    ● I’m feeling that the most optimal way forward is to combine ground-up telluric
    modeling with the data driven method
    What I want to know more about
    ● Can we pin point to places, one by one, where the synthetic model couldn’t
    match the observations, and why.

    View Slide

  15. What I did:
    ● Computed RV precision masking
    ● and after telluric correction (Y, J, H)
    With CARMENES spectrum corrected with Molecfit
    > pb: synthetic data scaled to SNR 100 in J band
    https://github.com/jason-neal/eniric
    (Neal 2018, Figueira et al. 2016)
    ● Ran wobble on HARPS data of HD41248
    > need to find best parameters for the
    regularisation
    What I learned:
    ● New techniques to correct tellurics:
    PCA with Neil Cook
    Wobble with Megan Bedell
    ● HITRAN 2016 should improve water &
    oxygen lines modelling
    What I want to know more about:
    ● Quantify the gain of correcting for the
    tellurics
    ● Test out hitran 2016
    ● Impact of the wind on the telluric lines
    Solène Ulmer-Moll
    IA - Porto

    View Slide

  16. Caleb Cañas (PSU)
    What I did:
    ● Worked with wobble - plan to see how it behaves with HPF data
    ● Using different tools (TelFit/Molecfit/terraspec) on HPF data
    What I learned
    ● The latest HITRAN is preferred (but AER 3.6 ≃ HITRAN2016)
    ● About different data driven approaches to telluric correction
    What I want to know more about
    ● How do data driven models compare with forward modelling
    ● Metric to compare these methods of telluric correction

    View Slide

  17. Mathias Zechmeister (U Göttingen)
    ● run SERVAL on HPF (GJ 436)
    ● noted that SERVAL is already used for
    HPF (Joe)
    ● interpolation methods (linear, cubic,
    band-limited)
    ● got a nice introduction for Spirou (Neil)
    ● wobble
    ○ segmentation fault (opensuse?)

    View Slide

  18. Sam Halverson (MIT)
    What I did:
    ● Simulated effects of differential illumination variations for fiber-fed spectrometer --
    gauge feasibility of ‘vanilla’ sky subtraction in the NIR.
    ● Thought experiments for estimating the current limits of micro-telluric
    contamination on optical PRV measurements.
    What I learned:
    ● Framework of wobble
    ● HITRAN 2016
    ● Active interest in full, 2D RVs from multiple directions.
    What I want to know more about:
    ● How can we robustly estimate the current limits to ground-based radial velocity
    measurements, even in the most ‘clean’ spectral windows? -- tie into PRV white
    paper?
    ● Can we reduce observing overheads by in-situ telluric correction (reduce frequency
    of hot star observations as telluric standards?)

    View Slide

  19. [email protected]
    •what I learned:
    •night glow of O2 at HPF with Kyle Kaplan
    clarify some spectroscopy questions with Iouli
    Gordon (HITRAN)
    (not relevant to RV exoplanets, but relevant to
    Climate CO2 monitoring from space)
    The usefullness of a co-ordinated network of
    Hi-res spectometers HZ exoplanets hunters
    •make exhaustive inventory of EPHZ around stars nearby sun
    (increasing distance, to prepare space observations)
    • Earth longitude distribution to measure host star oscillation regime,
    to determine age of system
    •distribution of target stars between various observatories !
    otherwise, everybody look at the same star…
    •each star to be monitored by at least 2 or 3 spectrometers.
    •a bulletin or Newsletter: fast exchange of informations ?
    •The network needs not to be formalized by international
    agreements, or bindings agreements
    •Examples: TCCON network, NDACC network, etc…
    a volunteer to initiate and
    manage this network?

    View Slide

  20. Dino Chih-Chun Hsu UC San Diego
    ● MCMC forward-modeling A0V K Keck/NIRSPEC data +airmass and pwv
    ● Fit better around 23220 Angstrom
    ● Close to the value with airmass (1.18 vs. 1.25)
    ● LSF slightly improves (Gaussian profile from 5.03 to 5.0 km/s)
    Before After

    View Slide

  21. Ashley Baker UPenn
    What I did: Started getting set up with Terraspec. Looked into the differences b/n
    atmospheric models isolating changes in HITRAN parameters & different atmospheric
    models. Saw that these can produce errors in telluric modeling on a similar order of
    magnitude (at least b/n HIT2008 & AER 3.6 which is ~HIT2016 for water)
    What I learned: About the many great tools being developed for telluric modeling/removal
    & what RV folks worry about in removing
    tellurics (e.g. mismatch in line shape)
    Remaining questions: Wobble sees
    success w/ 3 principle components - can do
    PCA on telluric models for diff atm models to
    physically inform these components + PCA
    on extracted telluric data from solar data
    (right).

    View Slide

  22. Andreas Quirrenbach
    What I still want to do:
    What I did:
    What I learned:
    What I still want to learn:
    - Listened to presentations
    - Discussed with many different people
    - Thought about applications to CARMENES
    data
    - Test modeling of OH lines, in particular near
    He 101830 lines
    - Data-driven approaches should work well
    for CARMENES data
    - PCA should be ok although problem is
    non-linear
    - OH airglow spectrum can be described by a
    rather small number of parameters
    - Whether site (altitude, PWV) has a strong
    influence on RV precision
    - Whether there is a rationale to go to space
    for RV measurements
    -
    -

    View Slide

  23. What we did:
    ● Attempted to implement the
    Bolton & Schlegel
    “spectro-perfectionism”
    algorithm on simple fake data.
    ● Worked on code to
    characterize PSF over PaRVI
    array from LFC exposures
    What we learned:
    ● Proof of concept
    ● Constructing the
    calibration matrix A will be
    the most difficult part.
    ● Should run this on SIG
    data, not yet lab data
    Rose Gibson / Ricky Nilsson
    from AMNH+Columbia U / Caltech
    What we want to know more about:
    ● Hope to continue conversations
    about this extraction method
    with others
    [email protected]
    [email protected]

    View Slide

  24. Rose Gibson / Ricky Nilsson
    from AMNH+Columbia U / Caltech
    These plots are much better than the
    original ones on here...

    View Slide

  25. Jerome de Leon
    UTokyo
    What I still want to do:
    What I did:
    What I learned:
    What I still want to learn:
    - Ran wobble on sample datasets
    - Packaging my data into proper format: took
    time! Should have done it earlier if possible!
    - Worked on group project: (non-)linearity of
    underlying telluric manifold [unfinished!]
    - Running wobble on actual IRD data
    - Using the wobble as a black-box tool is
    easy; understanding its output is hard!
    - A lot of existing tools!
    - More intuitive feel on data-driven paradigm
    on modelling tellurics
    -
    -

    View Slide

  26. Michael Zhang
    What I learned/did:
    ● How wobble works and how to use it
    ● How to do non-linear dimension reduction
    ● How to implement a variational autoencoder (VAE) with PyTorch to subtract
    tellurics (thanks to Miles Cranmer!)
    What I still want to know:
    ● How to prevent overfitting (aka subtracting out stellar spectrum) by dimension
    reduction and/or VAE techniques
    ● How well wobble works for high resolution cross correlation spectroscopy

    View Slide

  27. Kyle Kaplan (Univ. of Arizona)
    ● Added ability to forward model O2 sky emission along with OH sky emission
    ○ Specifically the 1.27 micron O2 band
    ○ Molecular data from HITRAN2016
    ○ O2 level populations described by single temperature (Boltzmann distribution)
    ● To do…
    ○ Combine forward modeled sky emission with telluric absorption model

    View Slide

  28. Christopher Leet
    What I learned/did:
    ● Non-linear dimensionality reduction techniques on telluric data (ISOMAP,
    diffusion maps, variational autoencoder)
    ● How to use Wobble
    ● How better use MolecFit and TAPAS.
    What I still want to know:
    ● Whether non-linear dimensionality reduction can usefully fit telluric residuals.
    ● What the shape of the telluric manifold looks like.
    ● How microtellurics (which are often highly transient) can be accurately
    modelled.
    Yale University

    View Slide

  29. Chad Bender
    What I did: What I learned:
    - Updated TERRASPEC to run latest LBLRTM
    code (HITRAN 2012 with backports of H2O, O2,
    (possibly more) from 2016.
    - Nearly all O2 A lines shifted ~50 MHz after
    lab measurement with LFC
    -
    -
    -

    View Slide