LM399 based 10 V reference

[branadic]

The air pressure here at my place in London dropped from approx 1018 mbar on Wednesday midday to 995 mbar by midday today - that's a 2.25% change, not insignificant.

I agree with that. So even though I haven't monitored the ambient myself there is still the possibility to compare to online weather data. So here are the data available on the web for my location in the time I measured.

[enut11]

In Reply #533 I mentioned that I purchased four LM399 chips and that I wanted to run a simple test to see which was the best to run with in a ref circuit.

Also, IconicPCB mentioned that "I think this comes to the question of "pre ageing " components... it does not necessarily work.
Without soldering.. there is no way of knowing just how the component will behave post final assembly."

Nevertheless I did not want to just pick one at random so I ran a 12Hr bedding-in test on each one with the following results:

LM399a: Was quite stable for most of the time but showed a drop-off towards the end. Practical variation was ~30 uV.
LM399b: Was more stable for longer?. Practical variation was ~20 uV
LM399c: Showed more instability over the 12 hours. Practical variation was >40 uV
LM399d: A bit better than 'c' but still a large variation over the 12 hrs. Practical variation ~40uV

Which one would you choose to be part of a buffered LM399 10v voltage reference?
enut11

PS
1) The socketed circuit board that I used for testing these devices has been in use for some time and has logged up many hours testing other devices prior to this test
2) I did not do any temperature or humidity logging as I am not set up to do so. On average, temperature range was 25-28C and RH 60-70%
3) I used a blister pack around the test PCB to minimise the effects of air drafts
4) All measurements were done on my HP3456A using the Legacy HP XL Logger developed by @IanJ and modified by @Bud (available on the Forum)

[Edwin G. Pettis]

You cannot predict any future performance of these chips based on such a short test run.  According to Bob Pease, they usually ran these chips (for guaranteed drift rates) for 1,000 hours.  Bob gave me a set of four LM399 chips with guaranteed drift rates with charted drifts over the 1,000 hour period.  While some inference could be assumed after 250-500 hours with reasonable accuracy for the future drift rate, Bob said they never used anything less than 1,000 hours of operation to guarantee long term drift.  Generally, after 250-500 hours you could usually see any chips that were not likely to perform for a given drift rate in the future, these most likely would still meet maximum drift rate later though.  Another factor which also took time was noise, with time the noise could either go up, go down or stay the same, again many hours are needed to tell which way a chip was headed.  There is no reason not to use sockets for this initial drifting/noise test, the act of soldering the chip into a PCB could affect the zener's crystal lattice, care is required as is with the LTZ in this case.  Soldering/de-soldering a chip is not really recommended as it will likely affect stability.

[enut11]

Hi Edwin
At least a 1000 hours is only 42 days, not really that long. Looks like I need a few more mule boards for testing.

I am currently mapping the Vz to see if it shows any difference.

Thanks for the feedback. It will be interesting to compare my initial findings with those after 1000 hours.
enut11

[Kleinstein]

To see the stability of the reference, one should look at the reference voltage only, not the voltage scaled to 10 V. The resistors used for scaling from 7 to 10 V are a second thing, they can drift just as much as the LM399 - this is especially true for the very beginning, when there will be humidity released form the board. A short 12 h test could only identify rare very bad units with maybe excessive popcorn noise.
If at all the slightly higher noise level for the later 2 refs might be relevant, but for the low frequency noise there are also effects like thermal EMF effects and thus temperature fluctuations or air pressure variations that can contribute. So even for the relevant noise the tests might be to short and not well enough thermally shielded.

[Andreas]

Hello,

of course a 12 hrs test is too short.

On day 65 I would not have thought that LM399#CH7 is more stable than LM399#CH6
https://www.eevblog.com/forum/metrology/lm399-based-10-v-reference/msg402498/#msg402498

But when I compare today standard deviation of the measurements over 1 kHr then CH7 is about 0.25 ppm and CH6 is > 0.5 ppm
https://www.eevblog.com/forum/metrology/lm399-based-10-v-reference/msg796829/#msg796829

And the conditions are undefined.
- too many contacts (seebeck effect) due to "kleps" (I would prefer direct soldering on the PCB)
- pins of LM399 not shielded against air draft
- no decoupling capacitor (10-100nF) against EMI across the LM399 zener nor across the heater
- no measures to fix the tilting angles of the LM399 (yes it is sensitive against tilting).

By the way: do you know the T.C. of the HP3456A and the according environment temperature which was during measurement?

And perhaps you could increase resolution by measuring the difference of your references in 100mV range.
So you should be able to tell which reference drifts most against the average value of drift.

with best regards

Andreas

[enut11]

To see the stability of the reference, one should look at the reference voltage only, not the voltage scaled to 10 V. The resistors used for scaling from 7 to 10 V are a second thing, they can drift just as much as the LM399 - this is especially true for the very beginning, when there will be humidity released form the board. A short 12 h test could only identify rare very bad units with maybe excessive popcorn noise.
If at all the slightly higher noise level for the later 2 refs might be relevant, but for the low frequency noise there are also effects like thermal EMF effects and thus temperature fluctuations or air pressure variations that can contribute. So even for the relevant noise the tests might be to short and not well enough thermally shielded.

All noted, thank you very much. Need to improve my quality control

I have started to log Vz inside an insulated cardboard box. This should take care of the effects of air currents on the output.
enut11

Edit: The Vz seems quite noisy. Will have to look at noise reduction as per @andreas' suggestion

[Kleinstein]

It is already an free air space of more than something like 5 mm over the board that there can be convective and unstable air currents can appear, that can cause low frequency temperature fluctuations and thus via thermal EMF (e.g. at the OPs) lead to low frequency additional noise. So for really sensitive circuits one should have a cover relatively close on top.

It should not be so bad with the LM399, as the critical part is all inside the chip. Still at least the area right around the LM399 should be more closely covered as there is quite a temperature gradient and the pins are able to generate thermal EMF.

For the outside the cardboard box and the plastic container are only a partial fix. A much better solution would be a metal can, as they reduce temperature gradients and also reduce EMI.

The capacitors are the reference are more against EMI, they don't provide significant filtering as the output impedance of the LM399 is already quite low. looking at the output impedance curve of the LM399/LM329, I would even be careful with low ESR capacitance in parallel to the zener - something like 1-10 Ohms in series to the cap should be beneficial. Getting some filtering for the 10 V reference is a different thing - usually a capacitor at the OP and a resistor between the reference and the OP. This is filtering noise above something like 1 kHz.

[branadic]

Measurement is still running and it seems like there is really some pressure sensitivity as can be seen in the pictures below. However this is not the whole story.

I now have a sensor system with TM112, SHT25 and MS5611 that needs some code to run it synchronous with the dmm readings. This way I can directly plot the dependence from temperature, humiditiy and ambient pressure and compensate the readings for that to see the real drift.

[branadic]

I measured my LM399 with a Keysight3458A together with some additional ambient sensors (SHT25 for temperature and humidity, TMP112 for measuring temperature at selected points and MS5611 for temperature and pressure). Thanks to Andreas I'm able to get all measured values with an aquisistion software in one file.

In memoriam my reference is inside a styrofoam box, the reference itself is within an additional plastic case, that is filled with cotton. I used 5ppm/K SMD resistors for the 7V --> 10V translation, that have a crystal heater on top, to keep them all at same temperature (see also https://www.eevblog.com/forum/metrology/lm399-based-10-v-reference/msg219527/#msg219527).



First of all I ran all ambient sensors close together, to see if the single temperature values are tracking each other. This is true, but they show different absolute values. However, this is fine for relative measurements though.

After this measurement I put the TMP112, which at least has lower resolution, inside the styofoam box to catch their inner temperature.
Beside the popcorn noise of around 5µV I tend to see an influence of humidity in my measurements, which fits best with the change in reference voltage. So this comes either from my reference or from the 3458A.

Next step is to turn off the crystal heater and measure the change in behavior.

[Andreas]

Hello,

perhaps you should try to correlate humidity with a PT1-filtered humidity value.
typical time constants for DIP8-Epoxy-packages are 3-7 days.
(to find the best fit value I correlated several filtered values with different time constants).

Your divider resistors/DMM may vary.

With best regards

Andreas

[branadic]

perhaps you should try to correlate humidity with a PT1-filtered humidity value.

Why and how?
If I compare time domain data (humidity vs. reference voltage) you can see that there is no time lag but a visible similarity. The problem in the reference voltage over humidity plot is, that the noise and the popcorn noise are hiding the correlation somewhat.
I once replaced LT1001 in plastic DIP packages with CERDIP package, even though the photo still shows the plastic package. I still have no explanation why the reference should have a humidity influence, except the FR4 circuit board and the components mounted on it.

[Andreas]

How about the FR4 epoxy influencing the 7->10V voltage divider resistors?
What coating do the divider resistors have? Some (rare) are glass passivated others with epoxy.

With best regards

Andreas

[branadic]

I keep it like occam's/ockham's razor: "...Among competing hypotheses, the one with the fewest assumptions should be selected..."
So before I do any further hypotheses I will do additional measurements in the sense of an exclusion procedure.

That's why I turned of the crystal heater (thermal controller) that is heating my resistors and now measure how the reference behaves while still monitoring ambient parameters. There are several active controllers within the circuit: the power supply, the reference itself and the crystal heater for the resistors, each effecting the complete result of output voltage of the reference. The latter is now turned off, so in a few days I'm somewhat more wise about its influence.

[MisterDiodes]

Keep an eye on those SMD divider resistors - they sometimes can pick up FR4 board stresses that show up on the divider ratio output.  For precision, on FR4 we'll usually use TH resistors or if we have to use SMT they will go on something like a better stability Rogers pcb material, etc.

Just a suggestion:  You might want to monitor the '399 output directly, before the booster circuit - just to see if you have a resistor stability issue.

Andreas is correct - humidity effects are seen over many days or weeks / months.  The water absorption rate into plastics is definitely there, but fairly slow.  You normally don't see big humidity effects instantly on a circuit - unless you're using a humidity sensor, etc.

[branadic]

You might want to monitor the '399 output directly, before the booster circuit - just to see if you have a resistor stability issue.

You can be shure that this is one of the very next steps.

[Cerebus]

Keep an eye on those SMD divider resistors - they sometimes can pick up FR4 board stresses that show up on the divider ratio output.  For precision, on FR4 we'll usually use TH resistors or if we have to use SMT they will go on something like a better stability Rogers pcb material, etc.

The problem with discrete through hole parts is that it's difficult to get the thermal bonding that comes for 'free' with SMD parts. Also discrete parts aren't going to come from the manufacturer with a tempco tracking guarantee in the way that some SMD parts do, at least not without handing over many shekels to someone like Vishay or Edwin.

With that in mind, do you have any take on using DIP resistor networks such as Vishay's TDP series that do have tempco tracking guarantees. Is moving to a DIP package as good at avoiding board stress issues as moving to fully discrete parts, or is it a halfway house?

[branadic]

Today I present the results of the last 1111 hours of my LM399 with 10V booster with heater on that is on top of my 5ppm/K SMD resistors for the booster. Measured with our 3458A (100NPLC).

Mean: 10.0017711V
Std: 3.4µV

The total difference from mean is about +/-1.5ppm incl. all spikes, not bad at all.

Using the environmental sensors SHT25 and MS5611 I calculated temperature coefficient in the order of 0.8µV/K and humidity coefficient in the order of 0.5µV/%rH or expressed as dew point temperature coefficient 0.875µV/K using magnus formular. There is also some small amount of pressure coefficient in the order of -0.25µV/hPa.
I will try compensating the effects the next days.

-branadic-

[Andreas]

wow nearly 7 weeks (continously?) measurement. 

Hmm,

how much of the coefficients belong to the LM399 and which part is for the 3458A?
And how much do they add or compensate?

You need more (different) instruments/refererences to distinguish.   

with best regards

Andreas

[branadic]

wow nearly 7 weeks (continously?) measurement. 

Yes, almost continuously. There was only a small break of a few hours in between, where I couldn't measure the reference, but that's it.

how much of the coefficients belong to the LM399 and which part is for the 3458A?
And how much do they add or compensate?

I can't tell you. One sad thing is, that there is now multiplexer inside the 3458A and damn, I once sold my Prema 2080 when I got my Prema 5017 SC. That gear would have been perfect to compare multiple references on the 3458A.

Here are the real calculated coefficients instead of reading them out of a diagram using a linear fit:

Temperature sensor inside my styrofoam box with the reference:
TMP112 = 782.2192909802102e-09     10.00175078714701

Ambient conditions:
SHT25_Temp = 702.0809476528713e-09     10.00175529680888
SHT25_Hum = 252.2596185775486e-09     10.00175926263134
SHT25_Dew = 874.2006798035875e-09     10.00176195005415e+01
MS5611_Temp = 649.5007437887549e-09     10.00175629266645
MS5611_p = -252.3944060096974e-09     10.00201462316427

Next step, compensate for temperature, humidity and pressure.

-branadic-

[branadic]

I performed the compensation for temperature, afterwards for humidity and at least for ambient pressure. Attached are the single steps as a series of diagrams.

The compensation is resulting in:

Mean: 10.001790V
Std: 2.58µV

[branadic]

Last but not least the Allan deviation and the FFT of the compensated voltage.
Interpreting the charts by my self I would say, that only random walk and aging is visable.

Assuming a linear behaviour for aging/drift during the 1111h I can calculate a value of <2µV/1000h 

[cellularmitosis]

Very cool!

In order to calculate those coefficients, do you need an environment where changes can be isolated?  I.e. You can change temperature while holding humidity and pressure constant?   In any air conditioned scenario, humidity and temperature are going to be difficult to isolate.

[branadic]

In order to calculate those coefficients, do you need an environment where changes can be isolated?

No you don't. This is similar to compensating a sensor for ambient influences. You would measure its characteristic curve first. Afterwards you would keep the sensor steady and would change ambient conditions without changing the main variable the sensor is sensitive for. For example, measuring ambient influences (temperature, humidity, pressure) on a inclination sensor would be done keeping the sensor in its zero position.
This is the same with a voltage reference, keeping the zener in zero position without changing current or voltage across it equals the sensor in zero position. Now you can change ambient condition. All you need for compensation is a reference sensor that is only sensitive to the variable that you want to compensate.

I first compensated for the most significant variable, temperature. After this I compensated the temperature compensated values for humidity (the second significant variable) and at least I compensated the temperature and humidity compensated values for ambient pressure.

What I can't compensate for is drift, as I have no idea by what amount my reference and the 3458A drifted. Maybe the drift on one of them is compensating the drift of the other. The only way to get away with that are multiple references that you compare with each using one 3458A or having multiple 3458As measuring a single reference.

-branadic-

[Kleinstein]

To calculate a compensation for different influences one needs sufficient changes in the parameters. Usually one needs to have a balance of enough changes in temperature, humidity etc. to get a measurable effect and small enough changes to assume the effects are linear.

Compensating for one effect at a time (e.g. start with temperature as the possibly largest) is only a first approximation. It might be ok for a data-set that is very long. A more accurate calculation would use regression with all parameters in a single model. Some software could also give you an idea on how accurate the estimated parameters can be. However some of those error estimates also don't work - because they assume a wrong model.

Especially the effect of humidity can shown quite some delay and thus might not be well described by a simple linear (and instant) approximation. So the effect of humidity can be tricky.