Long Term LM399 Stability (how I Learned to Stop Worrying and Love the HP 3457a)

[CaptnYellowShirt]

Our good friend DilligentMinds was lucky enough to be granted an audience with the great and powerful Wizard of Linear, Bob Dobkin. If the title of this post grabbed your attention and don't know what I'm talking about, its worth a read: https://www.eevblog.com/forum/projects/ultra-precision-reference-ltz1000/msg404500/#msg404500

This post is related to note #6 on that post dealing with the LM399:

the long term stability of the array of LM399's will be directly related to the power required to run the heater-- and this can be minimized with insulation-- the more the better!

Since it just so happens that I've recently acquired a third HP 3457a to add to my collection (yes I need help), I thought I might hold this claim to an empirical test.

Nothing too technical here. I just cut a small piece of styrofoam to fit over the LM399 and secured it with a small piece of kapton tape.

I run my HP3457a's 24/7 mainly because I don't like waiting for them to 'warm up'. Environmental and temperature drift arguments aside, this should make for a good test for the insulation idea. All three will be powered 24/7. I took a set of three voltage measurements (NPLC=100, Navg=25), and will do so again in a month. If the idea works, I will expect to see the percent error *between* DMMs #1 and #2 to stay about the same, while the percent errors between DMMs #1/#2 and #3 to be a similar number but different between first and second tests. If all the percent errors are similar after a month, I think I can say that they've all drifted a similar amount. (check me on my reasoning here).

The reference source is an Fluke 5100b:

	1.0000 V	10.000 V	100.00 V
DMM #1	1.0000718 V	9.9999457 V	100.00034 V
DMM #2	1.0000567 V	9.9995784 V	99.997358 V
DMM #3	1.0000564 V	9.9995584 V	99.996624 V

[Andreas]

Hello,

just some questions:

- is the LM399 soldered directly on the board (without gap) or is there some distance. (short / long leads?)
- did you isolate only the top side or also the bottom side of the reference PCB?
- how much drifted the instrument before / after thermal isolation?
  (I have up to 4 ppm difference between isolated / non isolated device depending on orientation).
- how do you keep air currents away from the solder joints? (styrofoam is not soft enough?)
- do you have a room with constant temprature + humidity?

I do not think that you can measure any drift of the reference on a 6,5 digit instrument with pre-aged reference.
The ageing drift will probably be below the resolution of the instrument if the humidity of the room is constant.

with best regards

Andreas

[CaptnYellowShirt]

-- The LM399 is soldered with tiny legs (maybe 2mm) in length.
-- I only isolated the top. The bottom of the PCB has a gap of about 6-8mm down to the mainboard PCB.
-- I have this data, but I'm going to have to crunch it and post it later.
-- No hope on them air currents. This is just a test about insulating the LM399.
-- Negative, I work in (what is basically) a hole in the ground. But for whatever the temp/humity is doing to me, hopefully it will effect all the units the same (?).

One variable in this experiment that has not been accounted for is the aging of DMM #3. #1 and #2 have been ON for at least the past year. #3 is a new acquisition. I have no history for its cumulative ON time.

[CaptnYellowShirt]

Attached are the crunched numbers.

Here's how I did the analysis. I assumed that each DMM was measuring the same EMF. Because each DMM was acquiring data at the exact same time, the actual voltage of the EMF is unimportant. I'm simply looking for time-dependent *changes in differences* between voltmeters.

So to this end, I imagined that each set of voltage measurements created a vector in three-dimensional space (e.g. [DMM #1, DMM #2, DMM #3]). If each DMM was properly calibrated, the angles between each unit vector and the vector created by the collection of voltage measurements should be the same. But as these DMM's are not at all calibrated, the vector should be skewed off to one side a bit.

Now if the Vrefs of these units drift at the same rate, these angles should stay about the same. But one if drifts more or less than the others, the composite DMM vector will start to point towards or away from that DMM's axis.

Attached is a PDF of the spread sheet.

[CaptnYellowShirt]

I'm sitting here kicking myself. I really should have reversed the EMF and taken the averages between both measurements as inputs to the above spreadsheet.

[Andreas]

Hello,

why not simply doing a graphically representation of the drifts over time (normalized to ppm).

With best regards

Andreas

[CaptnYellowShirt]

That's probably a simpler way of doing it. But my thought there was the source drifts too. The angles idea was a way to remove the absolute value of any components.

[iTist]

Hi,

How is the long time drift measured correct?
I saw plots where the References where measured once a day.
I measure my LT1027 5V Reference aprox.every 10s and get an variation between 10µV on a good day and 20µV on a bad day(over 10000 samples)
Using my Agilent 34461A meter i can get Data with 8 1/2 digit resolution via the webinterface. So my record was 7µV !!

I use batteries and the reference is in an metal box which is also is in a styrofoam (don't know the correct english name) box.
The ambient temperature is between 13°C and 15°C.

Where is the difference between very low Frequency Noise and Drift?

Thanks

Greetz

Oliver

[Conrad Hoffman]

Long ago, when I did my 1027 based voltage reference, I measured it against a bank of ovenized standard cells every day for over a year. I can't remember the exact results, but two things stood out. 1) The plastic version of the part was lousy compared to the metal can version. It never had the same stability. 2) With either one, even a small PCB, mounted at 3 points, contributed voltage changing stresses. I think it was relaxation with time, and temperature induced changes. It's desirable to solder only the pins that are used (float any others) and do compliant 3 point mounting of the PCB. Don't use offset pins- do that elsewhere.

Building a good reference is as much a mechanical exercise as electronic!

[Andreas]

Hello Conrad,

unfortunately I could not get a metal CAN LT1027.
So the question: Is the metal can version less sensitive to PCB twisting + bending than the plastic version?
Further question: Did you make any tests to hysteresis due to temperature cycling.
The LT1027 (in plastic case) has the lowest hysteresis that I found on any tested buried zener reference.
(below 5uV over a 10-40 deg Celsius range)
So it would be interesting if the metal can case has similar behaviour.

I usually solder only 1 pin (the GND pin) of a reference directly to the PCB.
The others two only with very thin wire (VEROWIRE).
If I twist the PCB the output voltage change should be less than around 5uV as a target value.


@Oliver: once a day is ok if the stability (standard deviation) of your measurement setup is well known.
More measurement values (together with environment data like temperature + humidity)
will enable better correlation calculations.
But be carefully with humidity: the time constant of the plastic housing for humidity is around 4-7 days.
(and not symmetrically).

I would say:
The noise is the statistical (not repeatable) part of your measurements.
(Reference noise, air currents, cabling resistance,  thermocouple voltages, EMI due to switcher noise)

Drift has a correlation either to temperature, humidity or time.

And hysteresis is the behaviour that depends on the history (higher or lower temperature, humidity in the past).

With best regards

Andreas

[Andreas]

I think eliminating all of the hermetic packages on high-precision parts was a "fools errand"-- and most of the semiconductor manufacturers copied each others behavior.

Hello Ken,

I think that the real reason was not copying but a decision due to ROHS.

And I still think that the "old" hermetically packages have better mechanical isolation from pcb stress than the new LS8-package.
They are probably only cheaper.

With best regards

Andreas

[Conrad Hoffman]

Yes, the metal can is less sensitive to stress, but also to some random changes I never pinned down. Might have been hysterisis, but I never specifically looked for that. Every now and then I get to check one of the old references I gave to a friend, and it hasn't shown significant drift in over a decade, at least to my ability to check it. It was built with the metal can! That's one of the advantages of building standards with no provision to trim them- you never lose the history.

[mycroft]

I could not find the "one month after data". Was it posted elsewhere?