Skip to content

Latest commit

 

History

History
200 lines (158 loc) · 9.39 KB

TODO.rst

File metadata and controls

200 lines (158 loc) · 9.39 KB

The Skyfield To-Do List

This list includes both a section on immediately-reachable goals that are appropriate for sprints and collaboration, and longer-term goals that the code base is not quite ready for yet but that we do not want to forget.

  • While is_sunlit() is easy to use with a bare Earth satellite because it expects to compute the positions of both Earth and Sun, the phase_angle() method instead needs the position to already know its barycentric center because it’s only willing to compute the position of the Sun. Should they share logic, that would use a barycentric coordinate if it already exists, but that otherwise would compute and cache it? Or, should users have to do the work themselves, but maybe we could make it easier?

  • Somewhere the Skyfield docs should mention this URL: https://celestrak.org/NORAD/elements/active.txt

  • If a user builds a wgs84.latlon(), and then calls its at() to get a position, and then tries to use it to observe(), then they get:

    AttributeError: 'Geocentric' object has no attribute 'observe'
    

    This might be what I’ve been looking for: a refutation of the idea that position classes should have different methods. My idea had been that if only the Barycentric class had observe(), that users couldn’t call it on the wrong sort of position.

    That plan might have worked if users all used IDEs that prevented them from typing .observe() on a class with no such method. But even then, it would offer no help in how to get hold of a class that did have the method.

    But when typing in an editor, nothing stops them from typing .observe(), and the error they get back gives zero information about how to fix the problem — generating a support load for the maintainer on GitHub.

    We should look at what it would be like for .observe() to be on all positions, and raise informative exceptions for positions to which it doesn’t apply.

  • Incorporate historical polar x and y (link shared in #372): https://datacenter.iers.org/data/latestVersion/186_EOP_C01_2000.1846_NOW_V2013_01186.txt

  • Switch to a cube-root falloff, rather than a discontinuity, for Earth deflection.

  • The “finals2000A.all” file expresses DUT1 like 0.8084178 with 7 digits after the decimal place. How accurate is that? It goes down to tenths of microseconds. Roughly how much arc does the Earth rotate through in a tenth of a microsecond?

    (1e-7 / 24 / 60 / 60) * 360 * 60 * 60 * 1e3 = 0.0015

    So we know the Earth’s orientation to within 0.0015 mas. How does that compare to the error of storing TT in a 64-bit float? Per the output of design/time_precision.py, using a single float for TT provides steps of around 4e-5 and 5e-5 from 1950 through 2050: about 500 times less precise than the DUT1 numbers! Therefore, we should design delta-T to use both TT floats.

  • Should the deflection routine really say this instead?:

    limb_angle, nadir_angle = compute_limb_angle(
        target_au + gcrs_position, gcrs_position)
    
  • Improve the situation around “observer data”.

    • We currently special-case two components of the full center-to-target vector: the barycenter-to-target vector for things like deflection, aberration, and like magnitude being affected by distance from the sun; and Earth-to-target, for Earth deflection. First, those could reasonably live on the position object itself, as they are not really specific to the observer; and, second, we could instead simply save the components we added together. But maybe that would take too much RAM? Let’s continue to special case those two sub-results only, to save RAM, but from now on keep them on the position itself.
    • It feels like further simplifications might happen in VectorSum but we’ll see. Maybe ._at() should finally become something more visible, like .vectors_at()? Its purpose is to return a position and velocity without building a whole position object that would just get thrown away in the case of a sum of two vectors, for which the user will only need a final position object for the combined sum.
  • Several interesting API questions arise because of this Stack Overflow question. Should I finally go through with renaming .position to .xyz? Should .from_altaz() retain observer data so that a follow-up .altaz() returns the same coordinates? Should positions fully support math, so the computation of the Topos position can only be computed once?

  • For #145, skip deflections of planets that can’t affect an observation.

  • For #145, create a good syntax for combining two ephemerides.

  • Expand a bit on the documentation for stars, now that they can have an epoch for their position.

  • When I wrote add_deflection() and needed to know whether Jupiter itself is available in an ephemeris, or whether the Jupiter Barycenter should be used in its place, I tried writing the test if name not in ephemeris:. But this sent the code into an infinite loop! Why does in not work on an ephemeris object? This should be fixed, and a test written to keep it fixed.

  • The deflection code should really use integer identifiers instead of using names, which are slower because they need decoding.

  • The deflection code should have a quick way to reach in and ask an ephemeris for a raw position in au, without having to spin up a body object and have it spin up a Distance object.

  • We currently download most SPICE kernels from NAIF, but have to use FTP for fetching DE422. Are the files from the two sites equivalent and do they have the same data? Should we prefer one or the other?

  • We should have an illustration of Earth satellite heights above the surface, plotted against a blue atmosphere fading out into the black of space as the plot goes upwards towards the top.

  • The Timescale object does not currently know where its data comes from, so its repr() is pretty uninformative. Should it someday be put in charge of loading its data from the data files specified, so that its repr() can print out which leap second file and delta T file it is using? Should it also display how up to date the files are, and what leap seconds it knows about? Also: it should be told the expiration date of all of its data, so that it can print it out as part of its repr().

  • In stars.rst, document the other alternatives for how to set the RA and dec of a new Star object.

  • Solar eclipses.

  • Some users need to add offsets to vectors like .position and .velocity so should there be some official way to do so? We should either document the maneuver .position = Distance(au=old_position.au + xyz) or (gulp!) make Distance objects susceptible of being modified in-place without their various units going out of sync.

  • The almanac routines sunrise_sunset() and dark_twilight_day() currently use an expensive step size — with 25 steps per day — to catch even brief days near the poles; see Issue 571 for a user who noticed when we tried adjusting to 6⅔ steps per day. But what if instead we used the very uniform rate of the Earth’s rotation to zero in first on noon and midnight? Both noon and midnight could surely be found within, say, one second, with a single round of search, right? Then the iterative search could look for the transition, with zero chance of missing a, say, one-minute arctic day or night.

  • After computing a satellite position, should the position’s .target be the satellite object itself instead of a big negative integer? It’s an incompatible change, but more in agreement with how other geocentric objects work. Was the old behavior documented? (Not that I can find!)

  • Should the .target string of _KeplerOrbit become .target_name instead?

Adding more smarts to ephemeris handling

  • When we add or subtract vectors, the new VectorSum needs to inherit an ephemeris from one of the segments being combined. Right now it just grabs the first one. What if, of the choices of ephemeris among the segments, grabbing the first one gives us an ephemeris that is missing several key large bodies, and so we run into an exception when .apparent() tries to compute gravitational deflection? Should we be more intelligent in our choice? Or should we even combine the various ephemerides our segments might offer? Or should we specifically go ahead and look for the deflectors we need and try to find a segment for them each?
  • And additionally: the error when not enough bodies are available for deflection maybe someday needs to be more helpful.
  • Bring support for Type 1 and Type 21 JPL ephemerides into Skyfield from the third-party libraries where it now lives. The routines will need to be vectorized and updated to use a Skyfield approach to vector and matrix operations. When complete and documented, make a comment at: skyfielders#350

For 2.0

  • Remove old deprecation warnings for pre-1.0 behaviors.
  • Remove support and tests for old ephemeris Python packages.