10 V reference based on LM399, LT1001, and LT3042 in bootstrap configuration

[ejd.pol]

Hello,

I wanted to share my latest (small) adventure in making a voltage reference.
I had wanted to design and make one myself for some time,
and made a lot of attempts on paper and even in KiCad, but nothing was constructed.
Target accuracy: about 5 digits, more is a bonus.

A couple of weeks ago, I got a very interesting idea, one that I had not seen before.
Once I started to really think things through, the idea became only more attractive.
The idea is to use an LT3042 as output buffer in combination with an LM399, in bootstrap mode.
And there is also an opamp with some precision resistors to make the output 10 Volt.

The combination of an LM399 and an LT3402 has been proposed before, see for example
https://www.eevblog.com/forum/metrology/lm399-based-10-v-reference/msg4777118/#msg4777118
but as far as I know not yet in a bootstrap configuration.

Using the LT3042 as output buffer gives excellent regulation and current limit on the terminals.
And there is plenty of current available to supply not only the LM399, but also the opamp.
Once the idea is clear, the schematic almost drew itself!

[see schematic]

The output voltage of the LT3042 is set by the output of the opamp.
Both inputs of the opamp are approximately 7 V, so no problem with a supply of 10V there.
The output of the opamp is connected with a 47k resistor to the set pin of the LT3042.
The set pin is connected internally to a precision current source of 100uA.

The voltage on the set pin is taken by the LT3042 as reference for the output.
So the voltage at the output of the opamp is 100uA * 47k = 4.7 V lower than the regulator output.
Assuming the regulator output is indeed the intended 10 V, the output of the opamp is 5.3 V.
That means also from the output perspective, the opamp can be supplied by the 10V regulated output!

In other similarly bootstrapped power supplies, the required voltage drop on the output of the opamp
is provided by a zener diode, but courtesy of the LT3042 current source, here a resistor suffices nicely.

The input section of the opamp is constructed such that Kelvin connections to the banana sockets can be made.
The use of three wires per banana socket is not strictly necessary,
but I wanted to have maximal flexibility, as that part of the schematic was the least certain for me.
The trim pot is used to fine-adjust the output voltage.
It is connected such that the wiper resistance does not play a role in the output voltage.

Will the bootstrap fly? Well, the 100uA current source is supplied by the input of the regulator,
which in this case is the output of the LM7812. So the current source will always work, no matter what.
Now even if the opamp would provide a short to ground at start up (which is the worst case),
the output of the LT3042 will rise to 4.7 V, and that is more than enough to get out of the 0 V corner.
So the circuit will reliably settle at the intended 10 V output level.

I had a few LT3042 modules lying around (I knew they would come in handy sometime!),
and one of these modules was prepared by removing four resistors.
Firstly, the 33k set resistor (blue-green) needs to be removed, otherwise it just won't work.
Also, I removed the resistor that drives the power-on LED, which simply saves a bunch of current.
The module will be built into a box, and so the LED will be invisible anyway.
For good measure, I also removed the two resistors connected to the PGFB pin.
That reduces wasted current to a minimum. In the schematic, that PGFB pin is connected to Vin.

The location of all four resistors is indicated in the picture below:
The must-remove in red, the recommended-remove in orange, and the if-you-feel-like-it-remove in yellow.
The original picture, which was quite helpful, was found in another EEV blog thread,
but I must apologize not to provide proper reference, I can not find it anymore.
In reality, the last three resistors were removed only after the entire reference was up and working.

[see module]

Next I made a board layout in KiCad for a prototype reference on a small experimenter board.
Making the symbol and footprint for the LT3042 module was simple, and the enite schematic fit
quite naturally on a standard 30x70 mm board with 10x24 holes.
The traces on the back side use only perpendicular directions, and are constrained to the 0.1" grid,
so they are easy to make with bare wires. The traces on the front side are free-format,
and will be made with insulated wire.

[see layout]

Actual construction was quite easy, and the following pictures should give
a fair impression of how the prototype board was assembled.
For the moment, the opamp, the precision resistors, and the LM199 are socketed.
Soldered connections will be used at a later stage.
The small green board is just to emulate the intended Kelvin connections to the banana sockets.

[see bottom, top, complete]

The first LM199 I tried produces 7.07 V, so in this case, I added a third resistor to the 5+2 k.
For now, I have used "normal" 0.1% accurate resistors. Not because of the precision per se
(the trim pot takes care of that), but because I expect/hope that the TC and drift are better.

I can check and trim the output value by measuring it with my Fluke 8840A.
Stability is good (even with the new LM199), no drift has been seen yet.
Total component cost: about $20!

So now I am at the point where I am considering how to build the reference into a box.
The plan is to use two 9 V rechargeable batteries as supply.
I have a plastic Hammond case that seems to be made for this purpose (the batteries fit precisely
when positioned on a side), but I also have a somewhat clunky metal box.

I would appreciate any and all feedback at this point.
Suggestions for further improvement would be greatly appreciated.
Should I design a real PCB, perhaps even with stress relief slots?
Is plastic sufficient, or should I go for the metal case?
Which type of banana sockets should I use? Etc., etc.

[Kleinstein]

The RC filter at the OP-amp output may be an issue, as it add phase shift to the LT1001 FB loop. This may lead to oscillation. One should plan for some extra fast local FB at the LT1001. Inside the loop the filter is anyway only partially effective.

Another issue is the trim port. Though relatively small the trimmer can still add significant drift, as trimmers often have a rather poor TC and long time drift.

[CurtisSeizert]

If you are excited about the design, which it sounds like you are, you should absolutely get some boards made. I would offer a few suggestions for the design:

(1) I second that C5 and C6 to GND is probably a recipe for instability. It would be better to configure U2 as an integrator. That said, I don't know that there is a really compelling reason to bootstrap the op amp output if you've already got 12V on the board. The mark against it is that it will probably be hard to both compensate the loop and bypass the current source noise without courting instability. If you configure the op amp as an integrator, some of the current source noise * R1 will be at frequencies where you've got no loop gain, so you would need to bypass it. It is possible to compensate such things (see the screen capture - U22A does not oscillate here), but with a pole that low, you'd be looking at some big capacitors, and with the additional phase shift of the regulator and reference in the loop, it may be rough. That is why it's probably easier to just drive the set pin with the output of an op amp directly. Perhaps you can still get away with the LT1001 because it will be sinking current, but if you think that's playing too close to the edge, I'd recommend the OPA205.

(2) It usually makes things more convenient for testing to close all the loops without relying on external connectors. If you use resistors to do this, you can add minimal error to the remote sensing.

(3) If you want to use a significant fraction of the current output capability of the LT3042, you may want to consider four-wire connections.

(4) I'd add optional footprints for an RC damper between the Zener anode and cathode in case you want to try one with the ADR1399.

(5) Pay attention to the input and output capacitance requirements for the LT3042. I always lay these out like the eval board with four terminals for the output cap.

(6) You can probably get away with less quiescent current through the R3/R4 divider. This will let you get away with a smaller integrator cap on U2.

[David Hess]

Everything old is new.  It is a great circuit idea and I have used it many times in various forms.

[iMo]

I would highly recommend you to put a RC low pass at the 399's output (like 10k/3u3 foil or a bigger cap if you are brave enough), and the opamp should be with low input bias current (and low TC).

Also people say the 12V for the heater is rather low, the higher the voltage the better, afaik.

PS: the pot trimmer - as indicated above - all trimmers I tried show jumps in their resistance - I think it is caused by the thermal expansion of the wiper's material thus the wiper moves a little bit with each temperature change (not talking aging of that material).

PS: and you may simulate the stability of the entire circuit in the LTSpice..

[David Hess]

I would highly recommend you to put a RC low pass at the 399's output (like 10k/3u3 foil or a bigger cap if you are brave enough),

Use two much larger capacitors in series and use one to bootstrap the other to remove its leakage.

and the opamp should be with low input bias current (and low TC).

The LT1001 is about right for a 10k source resistance, and has much better precision than the reference.  The lower input bias current LT1012 would be a better choice with a higher input resistance, however noise and drift will be dominated by the reference.  Note that trimming the offset voltage of these parts *also* trims their input offset voltage drift.

[iMo]

I would highly recommend you to put a RC low pass at the 399's output (like 10k/3u3 foil or a bigger cap if you are brave enough),

Use two much larger capacitors in series and use one to bootstrap the other to remove its leakage.

That has been elaborated a couple of times here with the 399 filtering, afaik.. For the same tau=RC you would need two 2x larger capacitors, when assuming the leakage of a 2x larger capacitor could be also 2x higher, then the total leakage current will be the same or higher. But perhaps my assumption is not valid..

[Kleinstein]

The point of the 2 capacitors in series is that the center voltage is essentially following the reference and thus essentially zero voltage for the cap to the output. This can reduce the effect of leakage currents. However for a not so long time constant or not very large resistors (e.g. > 100 K) the leakage current of modern film capacitors is usually small enough not wo worry.

The filtering part is mainly usefull to help with meters that are not sampling all the time, but working in an signal - zero cycle , e.g. measuring about half the time, or even with with a scanner.
Even this gets alteady tricky with a 10 PLC and thus 2.5 Hz / 3 Hz cycle and thus a target RC of some 200 ms. Such long time constants would need something like 10 K and 20 µF.

There is usually no need to bootstrap the OP-amp supply as the PSRR is usually good enough. If wanted there are upgrades to the LT1001 / OP7 of old days, e.g. OPA205 or ADA4077.
Unless heavily loaded and thus extra thmermal effects the regulation of modern 3 pin regulators is already quite good. In most cases this should be good enough. If an OP-amp is used to regulate from the main reference, than there is no need for the fancy LT3048 and a simple LM317 would be good enough as a power stage.

The divider resistors for a 7 to 10 V step are the hard part.

[CurtisSeizert]

I would highly recommend you to put a RC low pass at the 399's output (like 10k/3u3 foil or a bigger cap if you are brave enough),

Use two much larger capacitors in series and use one to bootstrap the other to remove its leakage.

I second this. I used this scheme with two 47uF/16V Ta caps for filtering an ADR1000 output after stepping down to 5V and the effect on tempco was <10 ppb/K compared to that of the ratio tempco of the VHD200 I used as the divider. It was also cheaper and about the same amount of board space as a 2220 Rubycon multilayer film cap, which was an either-or dictated by height limitations for me. If you are going to do this, I would recommend using diodes as in the screenshot in my last post so that you don't dump a bunch of current into the non-inverting terminal of the op amp if Vcc suddenly gets pulled below that node.

I used a 2k49/7k5 divider for the main output and drove the center of the two capacitors with a 255k/750k divider to ensure the upper Ta cap would not be reverse biased at the tolerance limits of the second divider. In practice, the bigger caps were pretty close to nominal, so there was about 30 mV across the top capacitor. The leakage current was on the order of the Ib of the OPA2205A I used as a buffer.

Definitely use LTSpice for testing stability. If you have built one of these up, do some load steps and see how the output settles.

I also agree with Kleinstein on the resistors. You may go over budget getting ones that keep your tempco close to the specified value for the LM399, but swapping out the LT3042 for an LM317 (if you don't care about broadband noise) would save you a couple bucks.

[Andreas]

For now, I have used "normal" 0.1% accurate resistors.

Hello,

for a 5,5 digit source I would at least use 0.1% resistors with specified T.C. (perhaps 5 ppm/K or better).
Normal 0.1% resistors have up to 25ppm/K which will spoil the accuracy of the 1ppm/K LM399.
A 10 deg C temperature change might give up to 500 ppm output voltage change with opposite T.C.

Alternatively a temperature controlled housing.

Using a trimpot in series to a resistor of a precision divider is usually avoided.
I would use a 50k trimmer across the 10V output and an additional (semi precision) resistor from the wiper to the 7V tap.

And please clean the boards with IPA or at least denaturated alcohol to reduce leakage currents.

with best regards

Andreas

[David Hess]

for a 5,5 digit source I would at least use 0.1% resistors with specified T.C. (perhaps 5 ppm/K or better).
Normal 0.1% resistors have up to 25ppm/K which will spoil the accuracy of the 1ppm/K LM399.
A 10 deg C temperature change might give up to 500 ppm output voltage change with opposite T.C.

Are there no suitable matched TC resistor networks for the non-inverting 1.43 gain?  It has to be trimmed anyway because of the poor absolute accuracy of the reference.

[Kleinstein]

I don't know special resistor array the 7 to 10 V step. One can however use resistor arrays with equal resistors (e.g. 8 x from NOMCT or TDP, e.g. 2 K or 5 K)  3 in series plus 2 in parallel and 1 plus 2 in parallel gives a nominal 7 to 10 V gain. One would need some fine trim based on the individual ref. voltage anyway. Here a selected fixed resistor is preferred over a trimmer, at least for the coarse part.

For the gain stage the TC is only part of the problem. Depending on the intended use, the long term drift may be the larger issue.

[iMo]

I tried with the LT5400 four resistor pack in past - the 4x10k chip gives you around 10.5V and the 2x100k+2x10k chip even a bit closer to the 10V. The ratio TC is less than 1ppm/C, ADI says. That is the cheapest none-statistical-arrays-version I am aware of..

[ejd.pol]

Hello, time to give an update.

Many many thanks for all the feedback, it really helped me to understand and to improve the design.

Firstly: Instabilities and zener noise

Kleinstein and CurtisSeizert : thank you for your helpful suggestions!

When looking at the outer loop (LT1001+LT3042), it is clear that the output of the opamp will swing
with a large amplitude, simply because it amplifies the higher noise frequencies (mainly from the zener)
without sufficient feedback, due to the low-pass filter between the opamp and the LT3042.

I have hooked up my scope to the output of the opamp, and could see these amplified noisy frequencies.
I did not see oscillations, the output of the opamp did not saturate, so the outer loop seems to be stable.

I did want to keep the outer loop structure as-is, so I did add local feedback with a capcitor,
creating the integrator, as suggested. That works indeed like a charm!

I did consider to add a compensation capacitor, but that would only decrease activity in
the output stage of the opamp, but the input stage would still amplify the zener noise, right?
So for now, I did not add a compensation capacitor.

For good measure, I also filtered the zener voltage as suggested by iMo.
And when I included that RC filter into the schematic, I saw that both RC filters
can be constructed entirely symmetrical on the non-inverting (zener low-pass filter)
and the inverting (integrator) inputs of the opamp.
That seemed to be the ticket, keeping the opamp inputs nicely symmetric.

In both cases, I used C=1uF and R=100k. To have less influence from the opamp input bias currents,
I switched to the LT1012, as also suggested in the feedback. 
I did not implement the dual-capacitor trick, as the capacitor data sheet says
they have about 10^10 ohms resistance. That seems good enough for the moment.

Secondly: resistor divider

Andreas: thank you for the suggestions regarding the trimmer, I hope to have drawn the intended topology.

For now, I have used a fixed resistor network, and not implemented the trimmer yet.
That is mainly due to lack of space on the prototype board, as I needed to make room for two capacitors!
The 3K and 7K resistors in the schematic are at the moment each composed of three fixed resistors.
That brings the output voltage to 10.0003 V, which is good enough for now.

The trimmer will be added in a later stage, when also higher grade resistors will be used.
Not sure yet which way to go with the resistors.

And yes Andreas, thank you for the kind hint, I did clean the flux with some alcohol.

Thirdly: Pre-regulator voltage

As the power source will be 2 batteries of 9 V, I do have some headroom to raise the output
of the pre-regulator. As the batteries are rechargeable, they do not actually reach 9 V,
so I do want to have sufficient voltage drop to accommodate battery-droop and regulator drop-out.
So I plan to raise the output of the pre-regulator to 14 V.

As it happens, I do have some pre-made modules with an LT1129, which would do nicely.
It has low-drop-out and plenty of current to give. I do not have implemented that yet,
but it is a serious contender.

Attached below are updated schematics, layout, and construction.
Work is in progress! (Albeit somewhat slowly...)

Some other responses:

Post #2, CurtisSeizert:

I do plan to use 4-wire Kelvin connections (as mentioned in the original post).
That is why the voltage divider, the zener, and the output are not connected in the schematic.
The small green board is to make these connections temporarily for measurement purposes.

And yes, a real PCB is brewing.

post #3, David Hess: Indeed, the bootstrap configuration itself is certainly not new,
as seen many times earlier. The are plenty of "other similarly bootstrapped power supplies"
as referred to in my original post, for example, Walt Jung's superregulator.

Also noteworthy is that the possibility of driving the set pin is mentioned in the LT3042 data sheet.
Quoted from the applications information section:

For applications requiring higher accuracy or an adjust-
able output voltage, the SET pin may be actively driven
by an external voltage source capable of sinking 100µA.
Connecting a precision voltage reference to the SET pin
eliminates any errors present in the output voltage due
to the reference current and SET pin resistor tolerances.

Kind regards, Evert-Jan

[Andreas]

Hello,

topology is ok.

But the trimmer should be lower ohmic than R5. (would suggest 50k)

with best regards

Andreas

[ejd.pol]

Small update: By changing R1 from 47k to 30K, the leakage current through the integrator cap is reduced considerably! 

[ejd.pol]

Hi Andreas, yes indeed: that way, variations in the trimmer resistance have less effect on the resistor-divider! Suggestion gladly accepted. 

[iMo]

Would be interesting to research (perhaps better equipped members may try) how is the largest value of the resistor in the input RC (based on a specific opamp of course), such it will not compromise the typical 399 performance much..
Let us assume the capacitor there is a film capacitor with 1u..6u8 and with reasonable low leakage.

Like: would an 1MegOhm (say w/ TC=25ppm/C, metal film) resistor there with a 6u8 wima foil and a modern 10-30pA opamp be still ok for a 4uVpp (in 0.1-10Hz, TC=0.5ppm/C) noisy 399??

[Kleinstein]

For the filter the TC of the resistor does not matter - drift here would only shift the cross over frequency and one would not care about a few percent here. There are 2 problems with a large resistor. One is the resistor noise. 1 M would have some 130 nV/sqrt(Hz) and thus noise comparable to the LM399 at some 20 Hz. The likely more serious problem is the effect of amplifier bias drift that converts to votlage dirft. With 1 Mohm this would be 1 µV from 1 pA of drift in the bias / leakage. One can get get slightly larger capacitors then 6.8 µF, though they do get a bit larger, but still not too bad and afforable compared to the reference.

A possible target for the cross over would be some 1 Hz cross over and thus R ~ 160 ms. This would attenuation 2.5 Hz that can be relevant for a DMM running with a 10 PLC AZ cycle. So some 160 K and 10 µF could be good enough for this.  A much lower cross over would have little advantage. The DMMs to test / use such a reference are usually good in averaging over the lower frequency noise and quite often the actual noise cut off is set by the meter, not by analog filtering.

[iMo]

So 1Mohm's thermal noise is comparable with LM399's noise, that could be acceptable as that results in 1.4 times more noise in the white area, the TC of the opamp's input bias should be less than say 1pA/C as that would result in 1uV/C drift which is 0.14ppm/C at the 10V output then (compared with 0.5ppm/C at 7V with the 399).
That sounds reasonable..

PS: out of curiosity I had a look into the vintage CA3140 opamp's DS - even this one could be used with the 399/329 and 1Meg/6u8 after trimming off its 2mV offset (40nV/rtHz in 1kHz, 2-5pA inp, 6uV/C inp, TC of inp current??).

[Kleinstein]

With 1 M if resistance there is also the current noise of the OP-amp that can become relevant. The BJT based ones may have current noise  in the 0.1 pA/sqrt(Hz) and more range that could become about as relevant as the noise of the 1 M resistor itself.
I would aim for a resistor in the 100-200 K range and a 10 or 22 µF capacitor instead of 2-5 µF and 1 M with a more tricky amplifier choice.
With a meter that supports high resolution based on 1 PLC cycles there is not even a  reason for that much filtering.

If noise is really an issue to next step up from the LM399 is the ADR1399, not so much excessive filtering.

[iMo]

It seems the LTSpice outputs some results with noise..
Tried with 10Meg and 3u3 in 0.1 to 10Hz for a couple of models..
You may try with reasonable RC values.

[Kleinstein]

With a 10 M resistor the cross over frequency is very low, but there is also quite some extra noise below the cross over. The 0.1 - 10 Hz range is a defecto standard to look at low frequency noise, but real world use may care about lower frequencies. So one should be careful optimizing the circuit for this noise specs as it may give poor performance for even lower frequencies.

When using a DMM with 100 PLC AZ mode from the average of 10 PLC conversions (4 seconds per point) one would be sensitive to noise in the range below some 0.12 Hz and some noise from the 2.5 Hz range. With averaging or addition digital filtering one can shilft the low frequency limit even  lower. The 0.1 - 10 Hz noise is not that relevant for this use case. Most of that band is suppressed by moden DMMs already.

It gets nearly impossible to filter the 1-120 mHz part.  Only the 2.5 Hz range is somewhat accessible for fitlering and here it makes little difference if the cross over is at some 1 Hz or 0.1 Hz. Due to the 1/f noise type the band (some 0.1 Hz width) around 2.5 Hz only makes up a small part of the total noise anyway. So even this filtering part is limited and may not be worth it, if it comes with more noise below 0.1 Hz (e.g. from thermal fluctuations).

Instead of large effort in filtering there is the alternative to use a better reference like ADR1399 that also helps with the really low frequency noise.

[iMo]

..Instead of large effort in filtering there is the alternative to use a better reference like ADR1399 that also helps with the really low frequency noise.

Yep, sure, we know, but imagine yourself we have the 399 handy only, no 1399 available today or tomorrow..
We want to know what is actually doable with RC filtering of the 399, provided we have an opamp which allows large R..

[Kleinstein]

If you looks at the reference of maybe different between 2 references with a DMM or similart. That DMM can already can do quite some filtering averaging.
Filtering in the analog domain mainly makes sense for those frequencies that the DMM can not handle well. For a DMM running a 1 or 10 PLC AZ cycle this would be mainly the 2.5 or 25 Hz range, that comes in from spending half the time on a zero reading and not actually looking at the signal. I see little use in fitlering out the even lower frequencies in the analog domain, if the DMM can do it as well and as essentially no extra costs and better accuracy.

PS: out of curiosity I had a look into the vintage CA3140 opamp's DS - even this one could be used with the 399/329 and 1Meg/6u8 after trimming off its 2mV offset (40nV/rtHz in 1kHz, 2-5pA inp, 6uV/C inp, TC of inp current??).

The CA3140 would make no real sense, as it has way to much 1/f noise. It can be OK at 1 kHz but is probably way more noisy then the LM399 at 10 Hz and below.