ADR1001 - Ovenized Voltage Reference System

[mawyatt]

Below you may see the schematics of the output buffer as is used today - V1.
The critical parts I see are the R1=10k, and the missing opamp's feedback blocking (like 1k+1n).

I plan to make changes in it - see the new V2.
Not sure whether it could help with the noise, however.
Anyhow - I will change the R1 to 1k, add the feedback RC (R5, C7), add the R6 and remove the tantalum 1.5u.
Also what opamp to use?
Any hints would be welcomed..

PS: the TVS is P4KE12A (next time perhaps P4KE15A, to be not so close to the 10V)..

PPS: also thinking to lower the oven temperature, to something like 45degC..

Another thought, you could use a couple 1N4004 on the output, one to ground and the other to the supply voltage. This protects the output transistors and the op-amp. To further protect the op amp a couple 1N4148 diodes (or low leakage if you prefer, but probably won't matter) to ground and supply voltage on the negative input to the op-amp, the 1K or whatever you use here, acts as a current limit, so this should do a good job of protecting the op-amp.

Best,

[Andreas]

The question is whether the TVS diode is not starting to fire (thus generating noise) when so close with the stand-off voltage to the 10V.
A good topic for a measurement when you have the LNA etc handy..

Hello,

obviously I had not measured noise again after inserting the TVS for my ADR587LW 10V references.
So I repeated noise measurement and compared it against a measurement before inserting the transient zeners.

https://www.eevblog.com/forum/metrology/ad587lw-10v-precision-travel-standard/msg4976209/#msg4976209

So at least for the ADR587LW with ~4uVpp noise there is no measureable change.

with best regards

Andreas

[iMo]

What could be the experiment with the TVS:
1. wire the TVS as a zener with cathode resistor, with say 1k cathode resistor (could be a different value)
2. change the voltage from say 9V to 14V (with your 11V TVS), step 0.5V for example
3. measure the noise at the TVS' cathode at each voltage.

[Andreas]

What could be the experiment with the TVS:
1. wire the TVS as a zener with cathode resistor, with say 1k cathode resistor (could be a different value)
2. change the voltage from say 9V to 14V (with your 11V TVS), step 0.5V for example
3. measure the noise at the TVS' cathode at each voltage.

Hmm,

1K against 1K input resistance of the LNA gives 6 dB dampening in case the zener resistance is high (not conducting).
My preferred (low noise) NiMH Battery has 1.2-1.35 V voltage steps dependant on charge state.

with best regards

Andreas

[iMo]

Today the second measurement in the LAB1.
The numbers are close to the LAB2 second measurement I made a month ago (-1.8ppm).
This time the Fluke 8588A, calibrated in June (the same meter as in the first measurement in March).
Below the results - triggered with 10 samples, 10seconds each sample (8588A's default), total 100secs per measurement, made in sequence from A to J. The Mean and the Standard Dev (in ppm) comes from the meter's statistics.

There is some noise coming from the air drifts and thermal radiation off our bodies - as the first 6 measurements were made as we were sitting aprox 1m off the wiring. The "minus" measurements were made with us aprox 2m off the wiring.
As an example of that influence is the measurement "X" were we were waiving with hands while discussing aprox 1m off the wiring (not included into the results).. 

Below also a measurement with my 34401A over last night.
As the next step I am going to add the 1k resistors into the inputs of the OPA189.

PS: in the picture below you may see the Fluke shows STDDEV in Volts, the resolution was only 0.1uV though, therefore we switched into "ppm" mode, where, luckily, we got by one digit more

[iMo]

After I had spent a day with experimenting I wired the output buffer as you may see below.
I replaced the OPA189 with an ADA4523. On the first glance I do not see any difference between them, both behave the same way, output voltage the same, the noise needs longer observation (but I doubt I would see some difference in my noisy setup).
I also put a 100n foil in the feedback with 1k resistor in series.
The output voltage is the same as before, we will se how the new ADA4523 will do.

[iMo]

The output buffer - after a series of experiments I settled with the V5 - see below. I have not seen a difference using the ADA4523 with compensation there thus I continue with the OPA189.

Below also the latest aging report.

The measurements are already challenging the capabilities of my 34401A pretty much. Without the TC compensation and having the DMM 24x7 on for at least 2-3 days after a week powered off the results will be much hairy, imho.

What I've learned in the last 10 months:

1. the TC of my meter is aprox +0.5ppm/C in 38-46C internal DMM temperature (I have the LM35 sensor glued over the ADC)

2. the humidity (?) after the meter is powered off affects the values +2..+3ppm and slowly settles down after 2-10 days after the meter is powered on

3. there is a random walk which is typically <1.5ppm peak-peak within an overnight measurement, the running STD (over last 100 samples, 100NPLC) shows typically 0.7-1.2uV (that includes temperature changes), in quiet periods 0.6-0.8uV

4. the ADR1001#1's voltage dropped down by some 2-2.5ppm in last 10 months, and its TC looks like -0.14ppm/C

5. the difference between a battery vs an ac_mains powering the Vref is not visible in my case (here the reference pcb sandwich is isolated by ~1.5cm thick foam inside the box, and its internal temperature measured by an LM35 typically rises up by 1.2C when powered off the transformer placed in the same plastic box with air vent slots)

6. I messed with the R1 values from 1k to 47k. I saw none difference in the measurement noisiness except a small drop of the output voltage, perhaps 1-1.5ppm with the 47k.

Also, it would be great if our better equipped contributors here would be kind enough to analyze in detail the performance of the AZ opamp buffer as depicted below, for the stability and noise (and will suggest improvements).. In the next Vref I would put an 1k into Q2 base, and not sure whether to filter out the AZ switching at the opamp's output..).

[Andreas]

Hello,

I would add a small frequency compensation capacitor across output and negative Input of OP-Amp.

with best regards

Andreas

[iMo]

Below is the LTspice Stability Analysis of the V5 output buffer without a comp cap and with a 1nF and 10nF comp cap.

It looks to me the best stability is without a comp cap..

Made with the ADA4523 as the OPA189 model does not work here properly..

Also attached is the .asc file for the simulation..

PS: set the Gmin=1e-16 in the sim settings..

[iMo]

A hundred times faster sim version below..
The above slow math came from switching power supply stability analysis by ADI, afaik..

[dietert1]

What do you derive from your V(a)/V(b) simulation about the Bode plot of your output stage as a whole?
I think with that 10K || 1.5 uF at the output there will be a bandwidth of about 10 Hz and at any frequency above that there will be a phase shift approaching 90°. Above 10 KHz or so the phase shift will approach 180° due to opamp internal compensation. What you want to see is whether the gain above 10 KHz is enough for oscillation and how the direct feedback capacitor cures the problem.

Regards, Dieter

[iMo]

With the current setup in V5 (none fb cap) the large output capacitance (1.7uF) keeps the phase margin around 32deg which is rather low.
In order to improve the stuff one needs, for example, to lower the output capacitance (ie 100nF) and increase the R5*C_fb.
See below with R5=4k7 and C_fb 10p..100nF. With 1nF I can get 53deg already and with 10nF 85deg.

[dietert1]

Maybe the average reader has similar difficulties like me. For stability we are talking about noise gain, that is small signals and the output impedance of your buffer is about R3 = 25  Ohm. 25 Ohm * 100 nF gives 2.5 usec (low pass) and 4K7 * 1 nF gives 4.7 usec (high pass), so there is some overlap.

Regards, Dieter

[iMo]

The sim above is small signal as well. The AC=1 is just a placeholder (that AC analysis does not work the same way as the transient analysis).

The entire stuff is stable when at 0dB gain (left axis) the phase margin (at the right) is larger than say 45deg (the diff against 0deg or 180deg), people say.

For example the red_gain (with C_fb=1nF) crosses 0dB at 23kHz and the red_phase margin there is 53deg.

[mawyatt]

A quick look at open-loop responses of closed loop systems that are approximately 2nd order in behavior, is the open loop gain zero crossing should be ~-20dB/dec (-6dB/oct) for good stability. If the gain is dropping faster than -20dB/dec then the stability will begin to suffer.

If you look at the plots, the red trace is dropping a little faster than -20dB/dec and thus the PM is beginning to decrease as indicated at ~53 degrees which is just above critical dampening (~0.7 @ 45 degrees), where a PM of 60 degrees in overdamped. The blue and green traces are underdamped and likely to oscillate.

Best,

[iMo]

Ok, I've opened the box and I've made perhaps the final improvement..   
I've put a 22nF foil as the opamp's Cfb, and removed the 1u5 tantalum I had there at the output.
See below the sim - there is some safety phase margin for an additional capacitance at the output, if any.

I've put there in the sim the common mode choke I have there in hw from the very beginning as well.

Btw. - I've had a naive look at the output with my DS1062 (1mV/div) before and after - none oscillation visible.

[iMo]

After running for two days since the change (removed the 1u5 tantalum at the output and inserted the 22n opamp feedback capacitor) the output voltage is aprox 2ppm higher than it was before.

That could be because:

1. the tantalum cap had large leakage (less probable considering the output impedance is pretty low close to DC), or,
2. the new 22nF foil (not a fresh wima but an older used one off the junk box) has some leakage (it should be around 20nA based on simulation such it increases output +2ppm), or,
3. the changed stability settings have made it somehow, or,
4. the AZ switching applies somehow.

Still there is the question around the AZ switching and the potential artifacts, now the inverting input is wired to the 22nF capacitor directly.

Not sure what from that might apply here...

PS: added Z_out sim, vertical axis is impedance in Ohms, the 1k in the base of Q2 is not in hw yet, and it has none influence on the stability or the Zout sims.

[Kleinstein]

The impedance, especially impedance mismatch between the input can effect the input bias current. It should not be enough to cause 2 µV of difference, but worst case it may.
There may also be an effect on the offset directly.
Foil capacitors have usually little leakage, much less likely than leakage with a tantalum cap.

P.s. the Z_out simulation is just stopping where it may get problematic.
For the lower frequency part it may want some extra 1-10 ohm + some 100-300 nF at the output to dampen the maximum in the impedance. This could reduce ringing. Maybe just a resistor in series with the 220 nF would be OK.

[iMo]

Below the Z_out till 100MHz and with 0.01ohm (green) and 10ohm (blue) in series with the 220nF at the output.

[Kleinstein]

The dip at some 1 MHz is not that bad. It is just the output inductors and 220 nF capacitor, with the OP-amp FB part already with little effect.
10 ohm has barely an effect at the impedance maximum, but it helps a little. At least there is not extra peaking besides the transition from inductive to capacitive at some 10 kHz.

The 22 nF FB capacitor seems to make the amplifier quite slow and thus a relatively high output impedance (e.g. some 1 ohm at 100 Hz, e.g. some 1.5 mH effective inductance). One could consider a smaller capacitor, like 4.7 nF shift the curve to high frequency, lower impedance.

[iMo]

Z_out - the 4n7 FB capacitor and 10ohm (blue) and 0.01ohm (green) in series with 220nF at the output.
Stability - 4n7 FB and 10ohm in series with the 220nF.

[Andreas]

Not sure what from that might apply here...

The question is: how much is the voltage difference between ADR1001 output and buffer output.
I would als try a 100nF directly at the VOut terminals. (after the 10uH).

with best regards

Andreas

[iMo]

I would not mess with the ppm diff at this stage..
The priority is to design the output buffer first. I have to understand better how to approach the stability/compensation of the opamp vs. the output impedance/load, etc. 

[iMo]

Below the Bode for ADR1001 output amplifier, Cout 100nF/1u/10u, and also with a 1nF feedback capacitor with the same Couts. We do not have official information on the chip yet, the model comes from 2022.

PS: added the 1.7ohm output resistor I assume it is there in the 2022 sample.

[iMo]

A week back I did another change - see the V7 schematics below.
No change in the noisiness, nor in the output voltage after a week of running.
It seems that adding the C5 capacitor (4n7 or 22n foil) into the OPA189's feedback there simply moved the DC output by some 2ppm up.

Also I doubt we need the R1/C2 input low pass filter there. We have got the low pass filter after the zener's opamp inside the ADR1001 (see the 3k/9k3 divider with the 3u3 foil).