Two DVMs

Is the Fluke meter not quite showing yet another digit in the photo? Might that be an 8 or 9 as the .00000X digit?

If so, then indeed, it isn't doing badly!

John :-#)#

Maybe it's the Keysight that's off by ~10 uV :-)

Ed

I think there was one more digit that the phone camera missed.

But they agree to 1 PPM!

I have a request to design a board that will be a couple-of-PPM accurate programmable voltage source. That's scary.

I'm thinking of using several (or many) ADR420-series parts, temperature controlled. Or a few LM199s, except it's obsolete.

ADR1000 is still available.

REF-01 is remarkably stable.

I wonder what those DVMs use.

The Fluke uses a "reference transistor", which is an NPN with a zener in the emitter.

At a SWAG, three LTZ1000s and some software.

Cheers

Phil Hobbs

Great longterm stability. Maybe they've been aged.

$93 from Digikey.

It's just a transistor and a zener.

The trick with reference transistors was to select a resistor to tweak the operating current to get the tempco to zero. That's a nuisance in production.

If I ovenize my references, the tweak could be automated, whatever part I use.

Could be. They're probably a bit cheaper if you buy lots at a time, but there's obviously a lot of TLC involved.

Stabilizing the temperature adequately will require some care, for sure. I sometimes put voltage references on paddles routed out of the PCB. This is mostly to avoid shifts due to mechanical stress on the package, but could come in very handy to (more or less) eliminate temperature gradients.

A nice analytically-calculable geometry, e.g. a paddle within a paddle, joined with two pairs of thin bars, would let one put two nested temperature control zones in a small space.

Some nice squeaky styrofoam, and maybe a bit of wraparound metal here and there to get rid of incidental vertical gradients, would probably do a good job.

Cheers

Phil Hobbs

I have a Philips PM2518 which lost its fine calibration when the memory battery ran out. It pre-dates the takeover by Fluke of Philips Instruments and this one had stickers showing it was last calibrated by the Luftwaffe in 2003.

Earlier this year I sent it off to Instrotech of Watford, who took a bit of persuading that it was worth recalibrating. They succeeded, despite a lack of information, and returned it to me promptly at a sensible price with a proper Calibration Certificate.

Disclaimer: I have no connection with Instrotech other than as a satisfied customer.

I'm thinking about a baby board with some mosfet heaters in the corners and a clever arrangement of references and thermistors towards the middle. It could go on my main board on spacers, with a cover. That could attenuate external temp changes by 50:1 maybe.

Four refs would cut the noise in half.

I could mux a nanovolt-resolution differential delta-sigma ADC between multiple references and see if any is an outlier.

The potential customer has been talking about this for over a year. In typical form, when they finally place an order they will want it in three weeks.

The key thing about “analytically calculable” is that you can vary all those sliders you keep wishing for, using only a plotting program.

In the present case, I expect that you’d find that you’d be better off optimizing for maximum closed-loop bandwidth, and letting the geometry do most of the work of smoothing out the gradients. (This is based on having actually done something similar for a stabilized laser for a downhole application.)

That means one sensor right up close to each heater, with a local feedback loop wrapped around each pair.

The math isn’t complicated. It’s mostly the same as you’d need for your spice model—conductivities and thermal masses and such—but with some first-order estimates of diffusion delays based on the homogenous-material model I linked to upthread.

Really repays the effort!

Cheers

Phil Hobbs

Here there be dragons - and madness:

.

Joe Gwinn

Or keep my PCB pretty isothermal, measure temperature, and correct the gain and offset errors with a pair of tweak dacs per channel. Or a 2d polynomial!

Maybe bolt the critical PCB to a sheet of aluminum with a gap-pad between, to keep the board pretty isothermal.

A heater would just be used at factory cal time to generate the correction tables. Runtime, no heat would be applied, so gradients would be minimal. I can delegate the coding!

Yes, we'd need one of those crazy expensive system DVMs to test and calibrate units.

I could imagine an automated test/cal cycle taking a full day.

None of that helps the bandwidth, which is what controls the rejection of external forcing. That’s where the math comes in.

Cheers

Phil Hobbs

They don't use this.

There's always the South Korean room temperature superconductor if it proves to be real.

I would use a cheaper reference of known characteristics and use software correction. Anyone can throw $100 ltz1000 at the problem, but there must be a smarter way...

Fluke used the reference amplifer, transistor / zener combination in a whole range of test gear. Original part was in the 1969 General Electric Transistor manual, so idea had a very long life. Looking at the the Fluke 731B voltage standard, doesn't look anything special in the circuitry, but the reference zener / transistor devices came as a selected pair, device and resistor. Bet Fluke also designed so that the various tempcos of the resistors all cancelled out. Fluke were masters of analog design, but takes a lot of work to get that sort of performance from such innocent looking circuitry. Special sauce, indeed...

And if you want long term stability, you stabilise the temperature at the lowest leven you can get away with. Peltier junctions are better than resistive heaters for that.

The more rabid temperature stabilisation freaks stabilise an external block at the desired temperature just to minimise temperature gradients in the central core.

It's in the literature. It didn't sweep the world.

If you knew exactly what you were doing.

The LTZ1000 data sheet talks about thermocouple voltages from kovar-to-copper connections. The cure might be worse than the disease.

Analog Device wouldn't have bothered to keep on offering the LTZ1000 if there was.

If you want to use software correction you have to characterise your cheaper part very accurately, which takes time and costs money. there is no free lunch.