Research results

The check that held: the open-source boiling solver we use landed within 1.43% of its authors’ exact answer to a one-dimensional evaporation problem, after a short start-up period the comparison leaves out, against a pass bar of 5% written down before the comparison was computed.

What FluxZero has found so far, one result per card, each with its limit beside it. Everything here was computed: no FluxZero fluid has been mixed or measured. We compared three sets of our numbers with outside references, an exact textbook answer and makers’ published values, and we publish the misses as well as the match.

One page per result

Each of these results has its own page: what it shows in plain words, why it matters, who should care, its limits quoted word for word, the published work it builds on, and available source files. Where records or checking code are not public, the page states that limit.

  1. Checked against an exact answer

    Our boiling simulation against an exact textbook answer

    The phase-change solver behind our own blend simulations was run on a standard problem with a known exact answer. Its error stayed well inside a pass bar we wrote down before the comparison was computed.

    Limit It tests the solver, not a FluxZero fluid, and the runs use a constant surface tension.

  2. Checked against published values

    Surface tension of three commercial fluids

    We simulated the surface tension of three fluids sold for electronics cooling under a pass band set in advance. All three came out high, and our control run missed as well.

    Limit The likely cause, a simulated liquid far less dense than the real one, was found after the runs and is not yet fixed.

  3. Graded against datasheets

    Our earlier predictions against the makers’ datasheets

    We graded our earlier property predictions against the makers’ published datasheets. Most missed; the one that held was a dielectric constant.

    Limit The grades were computed after the predictions were made, so they measure how far our methods sit from real data rather than test them in advance.

  4. Modelled

    The modelled order of four candidate blends

    A seeded Monte-Carlo model over published properties ranks a methanol blend first by modelled heat flux among the four candidates we modelled.

    Limit It is a model, not a measurement: the figure is a textbook boiling limit multiplied by assumed factors, and the blend has not been mixed.

  5. A finding of the screen

    Blends rated non-flammable under the GWP ceiling exist in our screen, and all of them contain fluorine

    Across every two-part blend in our catalogue, the ones rated non-flammable that also clear the climate ceiling all use a fluorinated fuel (recounted from the blend table, which is not published). Our lead blends contain no fluorine, and neither is shown to be non-flammable: the pentane lead is flammable.

    Limit It is a finding about our catalogue of molecules, not proof that no non-flammable coolant without fluorine can exist.

  6. Counted from the screen

    The screen and its bounds

    Every two-part blend we can build from catalogue molecules, scored on heat flux, permittivity, safety, climate impact and boiling window, with the bounds the screen enforces.

    Limit The scores come from estimates and simulations; several of the estimators missed published values, as the comparisons below show.

How our computed numbers compare with outside references

Here are three comparisons of our computed numbers with outside references, an exact textbook answer and makers’ published values, and we show the misses as plainly as the match. No FluxZero fluid has been mixed or measured yet: everything below is our computed numbers graded against other people’s work. One comparison held; the other two missed, one on every fluid and the other on eight of nine predictions.

  1. Held

    1.43%pass bar 5%

    The orange bar is the largest gap between the solver’s interface position and the exact answer after start-up; the tick is the typical gap; the line is the pass bar, written down before the comparison was computed.

    A boiling simulation against an exact textbook answer

    The boiling solver we use was run on a standard problem that has an exact answer, and the position of its liquid-vapour interface was compared with that answer at every recorded time after a short start-up period. The worst difference was 1.43%, against a pass bar of 5% that was written down before the comparison was computed. Across 6,111 comparison points the typical (root-mean-square) difference was 1.08%.

    The problem and its exact answer come from the solver’s authors, so this tests the solver, not our fluids. Our three blend runs use the same one-dimensional setup with a constant surface tension, so they contain no surface-tension-driven flow and show nothing about it.

    Open the file this is read from

  2. Missed

    Each row is one sold fluid. The shaded band is the pass band around the maker’s published value (the white line); the hollow dots are our three runs and the orange dot their mean. All three means land above their bands.

    Surface tension of three commercial fluids with published datasheets

    We simulated the surface tension of three commercial fluids and compared each result with the maker’s published value. The pass band, within ±15%, was written down before the runs. All three simulations came out too high.

    • Novec 7100

      Maker’s published value
      13.6 mN/m
      Our simulation
      18.5 mN/m
      Higher by
      36%

    • Fluorinert FC-72 (as n-perfluorohexane)

      Maker’s published value
      10.0 mN/m
      Our simulation
      13.3 mN/m
      Higher by
      33%

    • Chemours Opteon SF33 (HFO-1336mzz-Z)

      Maker’s published value
      13.0 mN/m
      Our simulation
      18.4 mN/m
      Higher by
      41%

    A control run on pentane, which should reproduce our own earlier simulated value, also missed its bar under the averaging method chosen in advance, so the method is in doubt, not only these fluids. A second averaging method recorded in the same file passes that control but still leaves all three fluids outside the band. After the runs we also found that our simulated liquids came out 26% to 59% less dense than published: a likely cause of the high surface tension, not a proven one.

    Open the file this is read from

  3. Mostly missed

    1 of 9

    Each row is one earlier prediction. The shaded band is its pass bar around the maker’s value (the centre line), on a scale that treats too high and too low alike; the dot is our estimate. One dot lands inside its band.

    Our earlier property predictions against the makers’ datasheets

    We also graded 9 of our earlier predictions, for fluids we do not own, against the makers’ published datasheet values. 1 fell within its bar. The one that held was the dielectric constant of FC-72: we estimated 2 and the maker publishes 1.75. The clearest miss was the fuel HFO-1336mzz-Z: we estimated a dielectric constant of 2; the maker publishes 32.

    These grades were computed after the predictions were made, and the bars were chosen after the published values were read. They show how far our methods sit from manufacturer data; they are not tests written down in advance.

    Open the file this is read from

How the lab checks its own work

What the lab can show so far, each count read from a published file:

AI agents do our research and engineering. These checks rerun on the lab's own machine; the code is not public, and none of them is a measurement of a FluxZero fluid. The counts and the checks behind them

Published Data Files

29 files, served from this site with no key: the published data the screen ran on, its stage-by-stage counts, a ten-blend score table and the uncertainty behind them; the per-blend score table is not published.

Operational Envelope & Disclosures

  • No FluxZero fluid has been mixed or measured yet: every surface tension, permittivity and heat-flux figure here is computed or taken from the literature.
  • Heat-flux figures are Monte-Carlo model outputs over literature properties with assumed enhancement multipliers, reported with 90% intervals; they are not bench measurements.
  • The pentane with dimethyl carbonate lead is flammable and needs a closed, nitrogen-blanketed tank, and its permittivity is only an estimate. It is meant for closed-tank direct immersion, if its permittivity measures low enough and its dielectric strength high enough; neither is measured yet. The methanol lead is for cooling loops only: its permittivity is too high for direct immersion, and its flash point has not been measured, so it is not shown to be non-flammable. Of our four in-house candidates, none that is rated non-flammable has a permittivity below 6, the cut this screen uses for a fluid in direct contact with the electronics.
  • For four of the direct-immersion pairs, Hansen solubility predicts the components will separate while simulation shows one phase; the shake-flask test decides.
  • Permittivities in the screen are group-contribution and mixing-rule estimates. Graded against its maker’s datasheet, the estimate for the fluorinated fuel Opteon SF33 was off by an order of magnitude (see the datasheet grades above), so no fluorinated blend is quoted here.
  • In one pair, our fitted slope of surface tension with temperature cannot be told apart from the range our file records for similar liquids, whose source it does not name (−0.177 against −0.05 to −0.15 mN/m per kelvin): four temperatures are too few; the sweep is of Pentane+DMF 80:20.
  • Graded after the fact against manufacturers' datasheets, 1 of 9 of our property predictions falls within its bar. Our simulated surface tensions run high on the fluorinated fluids we checked, and one fluorinated fuel's published dielectric constant is 32, where we had estimated 2.
  • Under bars written down before the runs (a ±15% band around each manufacturer's published value), 0 of 3 of our simulated surface tensions for commercial fluids landed inside the band; all three came out high, and the pipeline's own control run failed too, so the simulation, not the fluids, is in question.
  • The largest single-run surface-tension drive among the ten blends scored against Novec 7100, pentane with γ-butyrolactone at 27.6 mN/m between pump and blend, is not separated from the runner-up. It compares a published value for the pump with a simulated value for the blend; our simulations came out too high on three fluorinated fluids (see the comparisons above) and too low on pentane, methanol and triethylene glycol, so the size and even the direction of its error are unknown: treat it as a model output.
  • None has yet been shown to replace a fluorinated immersion fluid. No lab has been engaged yet.
  • The bench plan is a nine-sample panel with one sample expected to fail and Novec 7100 as the reference.
  • Every blend our screen rates non-flammable that also clears the climate ceiling has a fluorinated fuel (recounted from our blend table, which is not published).
  • FluxZero has no public code repository yet. Available data files are served under Published files; some records, raw runs and checking code are not public, as their pages disclose. A served file and its hash support checking those bytes, not independent reproduction or the accuracy of the model.
How the numbers on this page are checked
Each underlined number is read from a published file when the site is built, and the build stops if the page and the file disagree. A number would be marked if
  • its file were still loading in your browser,
  • the question behind it had not been looked into,
  • it had been looked for and not found,
  • its file held no value for it,
  • its file were missing or had changed,
  • two files disagreed about what it describes,
  • its sample were too small for the claim,
  • two files gave it different values,
  • its file could not be published,
  • the file it is read from had last changed more than thirty days before this build,
  • the question did not apply to this page or
  • the program that makes it had stopped with an error; no underlined number on this page is marked.
A number in plain text, without the underline, is not checked this way and carries no mark; the date each file last changed is in the list of published files.