1. Forum
  2. Usenet
  3. REC.ARTS.COMICS.STRIPS

  • Re: xkcd: Tukey

    From Thomas Koenig@21:1/5 to Lynn McGuire on Sat Jun 21 08:51:37 2025
    XPost: rec.arts.sf.written

    Lynn McGuire <[email protected]> schrieb:
    xkcd: Tukey
    https://www.xkcd.com/3104/

    That one is really good, and reminds me of Mark Twain's
    quote about the length of the Missisippi.

    So true, so true. I can always tell who is a new user of simulation software, they expect to get 9 (ppb, parts per billion) or 12 (ppt,
    parts per trillion) digits of precision out of our software. I will go through my explanation of how simulation software is based on
    experimental data of 2 or 3 digits of precision and watch their faces
    change when they start to understand.

    You could also try explaining about the analytics. If they hand
    you an analysis which is accurate to 9 ppb in one of the main
    components (not a trace component where 9 ppb which, for xome
    reason, can be found at that level and where the 9 ppb is a large
    fraction of what is in there), and prove that it's accurate even
    when two different labs analyze it, several times, and the labs
    don't know they are analyzing the same sample,

    Or the number of theoretical stages in a column - that is a
    model based on a false assumption, but (AFAIK) everybody
    lumps mass transfer and equilibrium together; the HTU/NTU
    method that I also learned at university is not really used,
    and you cannot use a fractional number of theoretical stages.

    But if they are asking for REPRODUCIBILITY, that is a very
    much different matter, and can be quite justified. (The value
    converged to should be the same up to a certain accuracy for
    different reasonable starting conditions). They might want to use
    a gradient-based optimization, which requires numerical derivatives,
    for example.

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From Paul S Person@21:1/5 to [email protected] on Sat Jun 21 08:53:00 2025
    XPost: rec.arts.sf.written

    On Fri, 20 Jun 2025 17:56:59 -0500, Lynn McGuire
    <[email protected]> wrote:

    xkcd: Tukey
    https://www.xkcd.com/3104/

    So true, so true. I can always tell who is a new user of simulation >software, they expect to get 9 (ppb, parts per billion) or 12 (ppt,
    parts per trillion) digits of precision out of our software. I will go >through my explanation of how simulation software is based on
    experimental data of 2 or 3 digits of precision and watch their faces
    change when they start to understand.

    Explained at:
    https://www.explainxkcd.com/wiki/index.php/3104:_Tukey

    Of course they can get 9ppb out. Computers have no problem computing
    on and on if you let them. It's just that everything after the second
    decimal digit (or whatever the value for the answer is) will be
    garbage.

    These are the people the <https://en.wikipedia.org/wiki/Blinkenlights>
    posters were meant to address.
    --
    "Here lies the Tuscan poet Aretino,
    Who evil spoke of everyone but God,
    Giving as his excuse, 'I never knew him.'"

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From Thomas Koenig@21:1/5 to Lynn McGuire on Sun Jun 22 07:50:19 2025
    XPost: rec.arts.sf.written

    Lynn McGuire <[email protected]> schrieb:

    The worst thing is getting the young inexperienced engineers to
    understand that even though we are first principles simulation software,
    they think that any simulation is good for making billion dollar
    decisions on. They need to validate that simulation with a pilot plant
    and extreme laboratory data first. Few do nowadays.

    Judging the quality particular model is definitely part of the
    art of engineering; judging if errors are on the safe side or
    not also plays a large role.

    A colleague of mine once stated, ironically, "Convergence means
    correct", which has become a favorite quip in our group.

    If I were to write a new simulation program from scratch, (which
    I'm not), I would probably include sensitivity analysis into the
    model right from the start, so people can now (if they care to know)
    how a difference in composition, temperatre, pressure or material
    properties will affect their results. This could also provide a
    guideline to those young engineers where the problems are.

    Plus (as much as this pains me to say, as you know I'm a Fortran
    person) I would probably build this package on Julia, which can do autodifferentiation and analytic Jacobians right out of the box,
    has cool ODE solvers and is reasonably fast because functions
    are compiled.

    Maybe a little anecdote: Once upon a time, some people wanted to
    build a distillation column. The did the calculation using the
    material properties provided by a well-known simulation package.
    Somebody noted that things were a little too close for comfort
    to an azeotrope, and asked the thermodynamics people to do some
    measurements to confirm the design. The thermodynamics people
    measured and found that the azeotrope was indeed much closer than
    previously calculated, and that the column actually needed a factor
    of four more theoretical stages than been originally simulated.
    The column started up on time and delivered in-spec product.

    They caught this in time, but less experienced engineers might
    not have, and sensitivity analysis could have pointed a less
    experienced engineer into the right direction.

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From Thomas Koenig@21:1/5 to Lynn McGuire on Sun Jun 22 08:50:46 2025
    XPost: rec.arts.sf.written

    Lynn McGuire <[email protected]> schrieb:
    On 6/21/2025 3:51 AM, Thomas Koenig wrote:

    But if they are asking for REPRODUCIBILITY, that is a very
    much different matter, and can be quite justified. (The value
    converged to should be the same up to a certain accuracy for
    different reasonable starting conditions). They might want to use
    a gradient-based optimization, which requires numerical derivatives,
    for example.

    [...]

    Several of the theoretical stage columns use numerical derivatives. One method that my fellow engineer created while he was teaching at Rice.

    What I meant was the the results were reproduceble (and smooth)
    enough so they could be used in an external gradient-based method
    of optimization.

    And a couple of methods based on the Boston Inside-Out method.

    That must be a reference to Henry Morgan's "Boston Inside Out! Sins
    of a Great City!", right? :-) (Actually, I know what it is).

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)