What "modelled, not measured" means for a coolant lead

What a lead is, and what it is not

When a company says a coolant is its lead, a reader may picture a bottle on a bench. Ours is not that. Our lead is a result of a computer model. The list of known limits states the position in one plain sentence: no FluxZero fluid has been mixed or measured yet.

That is the whole point of this post. A modelled lead is a prioritised guess with error bars, useful for deciding what to test first. It is not a performance claim about a fluid.

How the number is made

The lead is a mixture of a small amount of methanol in a larger amount of triethylene glycol, with a trace of boron nitride. The model gives that mixture a heat-flux figure built from a textbook boiling limit for pure methanol, multiplied by assumed boosts. Its median answer is 357.6 W/cm².

The model does not know that figure exactly. It runs 20,000 random draws. In each draw it nudges the inputs within an assumed uncertainty, for example by 15% on surface tension. The spread of answers gives a range that holds nine draws in ten, from 203.7 to 556.1 W/cm².

What the interval does and does not cover

The interval covers the uncertainties we assumed. It does not cover errors we did not think of. In the companion post on checking simulations, our property estimates missed manufacturers’ datasheets more often than they hit. If the inputs are biased, a wide interval can still sit in the wrong place.

The heat-flux figure is also built from a textbook boiling limit for pure methanol, multiplied by assumed boosts. It is not a model of this blend boiling on its own terms. The published results file gives neither that bare limit as a number nor the size of each boost: it marks the bare limit as an estimate and the boosts as taken from published studies and not tested. Most of the figure may come from those assumed boosts.

Why the model’s order still helps

In the draws, this blend came out ahead of our other three in-house candidates in 78.6% of them. That share rests on the same assumed boosts, and we have not published whether the blend still comes out ahead with them set to zero. It suggests what to test first, if the boosts hold up. It does not tell us the blend will win a real test.

The reasoning goes like this:

• An order set before testing narrows a large field to a short list.
• The short list sets the order of experiments.
• The experiments then confirm or overturn that order.

The model gives an order to test in, and it tells us in advance which results would surprise us. Whether its order beats an expert’s guess is untested until the bench tests run.

The fire question is still open

A reader should ask whether the lead burns. Methanol is flammable. The lead’s flash point has not been measured, and the vapour it gives off while boiling would be mostly methanol. So the lead is not shown to be non-flammable. Its electrical behaviour also limits it. The permittivity is too high for direct immersion, so it fits cooling loops that run beside the chip, not liquid in contact with the electronics.

We repeat these limits beside the research results because they are the questions a buyer asks first.

Why publish a modelled lead at all

It is fair to ask why a company would show a result it has not measured. We have three reasons.

• A modelled lead is a commitment. It names the blend, the method and the uncertainty, so anyone can open the results file and check that it holds the figures quoted here.
• It shows how we think. A buyer or a lab can judge the reasoning before spending anything.
• It is the cheapest honest way to say what we plan to test first.

The cost of publishing it is that some readers will treat a model as a result. That is why every page that mentions it also says that nothing has been mixed or measured.

What would turn it into a measured result

The bench steps

A measured result needs a bench test: mix the blend, check that the two liquids stay together, measure the flash point, and run a boiling test on a heated surface. The order matters, because each step can end the project cheaply. A blend that separates, or that ignites easily, never reaches the boiling rig.

Until those tests are run and published, the lead stays a model.

What a buyer can ask for

A buyer or a partner lab can ask for four things before taking a modelled lead seriously:

• the exact composition, including the trace additive;
• the method and the seed behind the draws, so the run can be repeated;
• the list of inputs that were assumed and not measured;
• the order of bench tests and the result that would stop the project.

The first can be answered from the published files today, except the basis of its percentages (by weight, mole or volume), which the file does not state. The results file names the method’s settings and the seed, but not the size of each assumed boost, and the code is not public, so the run cannot yet be repeated from outside; the list of assumed inputs is published only in part. The fourth is described above, under the bench steps; the plan document behind it is not published, because it still quotes figures we no longer use.