DIY nanovoltmeters and ultra low frequency noise preamplifiers, who did it?

[branadic]

Quite an interesting topic that is currently being discussed on xdevs chat. Nanovoltmeters are great to intercompare multiple references with the 3-cornered had method. But they are expensive, a used 2182A is almost the price for a new one.
There are some very nice inspirations out there, even for low noise preamplifiers.

Senis NVM-01 NanoVolt Amplifier

Ultra-low Noise Chopper Amplifier with Low Input Charge Injection

Ultralow-noise chopper amplifier with low input charge injection

Ultra-low noise chopper amplifier with low input charge injection

Looking into seismic exploration even more interesting solutions can be found, one of the most detailed one is:

Ultralow Noise Low Offset Chopper Amplifier for Induction Coil Sensor to Detect Geomagnetic Field of 1 mHz to 1 kHz

Who has interesting links to papers and articles? Someone DIY'ed something similar already, probably with a fixed input range?

-branadic-

[David Hess]

That is pretty standard stuff for low impedance measurements.  Chopping is needed to control 1/f noise at low frequencies where the measurements need to be taken.  Jim Williams posted some examples of how to use a chopper stabilized amplifier to remove the 1/f noise of a precision amplifier to get the best performance of both.

Didn't they make super low charge injection chopper amplifiers using LDRs?  I haven't tried it yet but photo-FETs could be used to do the same thing.

[branadic]

That is pretty standard stuff for low impedance measurements.

Down to 3 mHz?

-branadic-

[Kleinstein]

There are Choppers based on LDRs, but due to the resistance of the LDRs the noise tends to be relatively high. So this is more like the 5-10 nV/sqrt(Hz) level and higher. There can still be some effective charge injection from cross conduction, as the switching is slow.

There are nV amplifier based on JFET switches and a transformer at the input - these can get really low in noise (well below 1 nV/sqrt(Hz), but charge injection needs a carefull trim to get it low. This could be a way for a DIY solution - it is old style, but relatively simple. AFAIK Keithley has such an amplifier for the K2001/2002. TiN can probably tell more on these.

There is some App. note from Maxim on a chopper amplifier build with CMOS switches at the input, they do bet pretty low noise, but not so shure about the charge injection.
An improved version could be found somewhere on this web page: http://www.janascard.cz/aj_Zakazkova_vyroba.html#USB

I have tried a similar (but using full wave chopper with 4 switches and thus lower noise - a bit like the input stage of the DA1281 DMM) circuit on the bread board, to get an idea if it works OK. The results were not that bad, though the bread board is obviously not the way to do nV stuff. So the main point was that is was stable and not oscillating and the charge injection spikes not excessive.
Compared to the ready made AZ OPs the switches are usually not as well matched and not that well fit for the purpose. On the other side a DIY chopper can use a relatively low chopping frequency and larger coupling / filtering capacitors that are not available on a chip. A lower chopping frequency gives less current noise and a JFET based amplifier can have an 1/f noise limit lower than the CMOS process used with the AZ OPs. Especially DIY an extra adjustment, e.g. to minimize charge injection, is also possible.

For my purpose a simple AZ OP was good enough - they are not that bad and could make a pretty simple system if the source is low impedance.
So the low noise chopper is at hold for the moment. I want to get the meter part working fírst.

Chopping works pretty well also to very low frequencies. The slight residual 1/f noise one sometimes sees with AZ OPs seems to come from a limited gain of the chopper part and than relatively high 1/f cross over of the main amplifier to start with. A seprate build amplifier can be better in this respect. The main limit are thermal effects.

[David Hess]

That is pretty standard stuff for low impedance measurements.

Down to 3 mHz?

Absolutely, which is why good chopper stabilized amplifiers have a 0.01 Hz (or DC) to 1 Hz or 10 Hz noise specification.  Low noise precision amplifiers do not usually bother going down to 0.01 Hz or DC because flicker noise dominates and it would be a misapplication of the part.

Ultimately the lower frequency limit is determined by the integration time of the measurement.  If you want to measure 0.1 Hz noise, then the measurement must last 10 seconds.  For 0.01 Hz, that becomes 100 seconds.  And that is why noise at these low frequency is important for low noise measurements, and why flicker noise which increases as the frequency decreases is such a problem.

Back when I was measuring noise of my low noise differential preamplifier which used a pair of LT1028s and an LTC1151 dual chopper stabilized amplifier, I used a stopwatch to time 10 seconds of sampled measurements.  The resulting 0.1 Hz to 10 Hz spot noise measurement matched the integrated noise curves of the LT1028 and LT1151 exactly.  Noise was low enough that I could measure low values of resistance from the Johnson noise of the resistor.  I found this rather unsettling but God never struck me down for it.

[Kleinstein]

Noise was low enough that I could measure low values of resistance from the Johnson noise of the resistor.  I found this rather unsettling but God never struck me down for it.

Seeing the normal Johnson noise is one part, looking for the resistor excess noise (which is additional 1/f noise of the reistors when used with voltage) is another part where a good low noise chopper can help and it also needs to be low current noise. It is a bit unsettling when a supposedly precision resistor fluctuates.
There is no absolute need to do this with DC, but the low frequency way is closer to the definition.

[David Hess]

There are Choppers based on LDRs, but due to the resistance of the LDRs the noise tends to be relatively high. So this is more like the 5-10 nV/sqrt(Hz) level and higher. There can still be some effective charge injection from cross conduction, as the switching is slow.

There are nV amplifier based on JFET switches and a transformer at the input - these can get really low in noise (well below 1 nV/sqrt(Hz), but charge injection needs a carefull trim to get it low. This could be a way for a DIY solution - it is old style, but relatively simple. AFAIK Keithley has such an amplifier for the K2001/2002. TiN can probably tell more on these.

I wonder if a varactor amplifier like the Philbrick P2 or Analog Devices Model 301 could do it.  Varactor operation removes flicker noise but there would need to be some way to cancel drift.

[David Hess]

Noise was low enough that I could measure low values of resistance from the Johnson noise of the resistor.  I found this rather unsettling but God never struck me down for it.

Seeing the normal Johnson noise is one part, looking for the resistor excess noise (which is additional 1/f noise of the reistors when used with voltage) is another part where a good low noise chopper can help and it also needs to be low current noise. It is a bit unsettling when a supposedly precision resistor fluctuates.
There is no absolute need to do this with DC, but the low frequency way is closer to the definition.

I knew I was limited to low values of resistance consistent with the input current noise, and since I knew the amplifier's input noise and the test resistance, I verified that the added noise was consistent with the Johnson noise.  I was using wirewound and metal film resistors so I did not expect any excess noise and indeed, I never saw any.  Unfortunately I did not think to test any low value carbon composition resistors.

My design was not quite up to the nanovolt resolution being discussed but it was within shooting distance and could have been improved, especially today with better chopper stabilized amplifiers and a discrete input stage.

[branadic]

Absolutely, which is why good chopper stabilized amplifiers have a 0.01 Hz (or DC) to 1 Hz or 10 Hz noise specification.  Low noise precision amplifiers do not usually bother going down to 0.01 Hz or DC because flicker noise dominates and it would be a misapplication of the part.

0.01 Hz is an order of magnitude higher to the given corner frequency of 3 mHz in the paper, while "m" means milli, that is 0.003 Hz.

My R9211 goes all the way down to 10 mHz, but is pretty much limited by its own 1/f-noise, which makes it impossible to characterize ultra low frequency noise of voltage references.

-branadic-

[David Hess]

Absolutely, which is why good chopper stabilized amplifiers have a 0.01 Hz (or DC) to 1 Hz or 10 Hz noise specification.  Low noise precision amplifiers do not usually bother going down to 0.01 Hz or DC because flicker noise dominates and it would be a misapplication of the part.

0.01 Hz is an order of magnitude higher to the given corner frequency of 3 mHz in the paper, while "m" means milli, that is 0.003 Hz.

My R9211 goes all the way down to 10 mHz, but is pretty much limited by its own 1/f-noise, which makes it impossible to characterize ultra low frequency noise of voltage references.

That is right, but manufacturers are not going to test even lower because of the time it takes.  The DC or 0.01 Hz specification is sufficient.  Chopper stabilized amplifiers have flat flicker noise, or at least it is suppose to be flat.  If someone cares about this then they need to qualify the parts themselves which should not be too taxing.

[dietert1]

Zeroing for voltage reference comparison should be at multiple time scales. The usual time scale for low frequency noise specs of 0.1 to 10 Hz can be covered by the DMM Autozero. For longer time scales one can use one zero channel of a low thermal EMF relay multiplexer.
The 2182A is meant for high speed low noise measurements. If i remember right, a 2182A drifts about 20 or 30 nV, so a relay MUX can help there. We also have one of the old relay chopper nanovoltmeters (Keithley). It is stable to 1 or 2 nV, but of course it won't be running for months or years.
A cheap and low power consumption setup i am using has 2 HP 3457A voltmeters in parallel and works quite well when combined with MUX zeroing.

Regards, Dieter

[Kleinstein]

The varactors usually have a thermal drift about comparable to a normal diode / PN junction. So getting rid of thermal dirft is prettty hard. The varactor amplifiers are more a thing for low bias and the option to have them isolated from the rest. I don't think they are very low noise either.

It does not even need carbon resistors to see resisor excess noise. To see the excess noise one needs to apply some voltage, so it is not allways there.
Even some thin film resistors show it in a measureable amount, if one looks at low enough a frequency. I did a test to confirm low noise of a good type, but forgot to test the ones I highly suspect to show more noise, because they showed extra 1/f noise in another circuit.

I just used an AZ OP, even a chep one (MCP6V51) and it does perform reasonable well. It just needs time to measure at 10 mHz or so.
Some of the modern AZ OPs are pretty good. They may still show some very low frequency 1/f noise, but even that is often no to bad and it is hard to tell appart from thermal effects of the setup. Thermal EMF and temperature fluctuations can also cause some 1/f noise.
There was another thread on AZ OP testing.

[guenthert]

[..]
Nanovoltmeters are great to intercompare multiple references with the 3-cornered had method. But they are expensive, a used 2182A is almost the price for a new one.
[..]

     The current generation is fairly expensive, but I've seen a bunch of Keithley 182 (as recently as 2014 still good enough for the Laboratoire national de métrologie et d’essais [1]) for a quite reasonabe price on e-bay.

     The old not-quite-nanometer Keithley 181 can be had for little money (the original input cable would double its value 

     When going the DYI route one mustn't underestimate the effort in getting the mechanical construction (RFI shielding / minimizing the effect of thermal EMF) right.

[1] https://www.bipm.org/documents/20126/47750471/EURAMET.EM-S41.pdf/39f8462f-4bb8-09ef-67d1-661535764a36

[David Hess]

The varactors usually have a thermal drift about comparable to a normal diode / PN junction. So getting rid of thermal dirft is prettty hard. The varactor amplifiers are more a thing for low bias and the option to have them isolated from the rest. I don't think they are very low noise either.

Bob Pease reported the same thermal drift that you would expect of a good discrete design, which of course the P2 was, but low frequency noise better than commonly available.  Something would have to be done to fix the thermal drift; I was thinking of chopping the varactors but that defeats the purpose of using them except for their virtues which you identify.

Some of the modern AZ OPs are pretty good. They may still show some very low frequency 1/f noise, but even that is often no to bad and it is hard to tell appart from thermal effects of the setup. Thermal EMF and temperature fluctuations can also cause some 1/f noise.

The above point cannot be overstated.  Noise from thermal EMF will overwhelm low frequency noise of even non-chopper precision amplifiers unless careful design and construction is used.

[e61_phil]

I've seen a bunch of Keithley 182 (as recently as 2014 still good enough for the Laboratoire national de métrologie et d’essais [1]) for a quite reasonabe price on e-bay.
[1] https://www.bipm.org/documents/20126/47750471/EURAMET.EM-S41.pdf/39f8462f-4bb8-09ef-67d1-661535764a36

The question is what do you want to measure. For a 10V comparision a Nanovoltmeter isn't really necessary. In many old publication they used a Fluke 8842A 5.5 digit meter with 100nV resolution. That is 0.01ppm at 10V and should be fine. The resolution may also be increased with averaging, especially if you average over 100s.

[iMo]

I wonder, myself, what would be the practical limit when comparing two references with that "simple single opamp" null-detector/meter and a 34401 as the meter. I messed with it years back, a chopper opamp (LTC1050 I think with 1000x amplification). With 1mV resolution dmm (at that time) I saw ~1uV diffs, imho.

[Kleinstein]

For a Nullmeter you don't even need a 34401 or similar full fledged meter for dispaly. Just a simple ADC (e.g. ICL7106 for a dispaly) or even an µC internal one is sufficient. The scale does not have to be very accurate if it is only about a small residual different or just to adjust to zero. So the ADC and refrence part is not that importand and I would consider something like a MCP3421  (18 bit SD with internal reference in SOT23-6) sufficient for a null meter and it is easy to power from battery.  There are other similare chips too.

The point with chopper OPs is, that one may need some additional filtering at the input and some protection. With just the OP input, the AZ inputs can reacto to the source impedance in the MHz range. So things like cable inductance and parasitic capacitance can effect the offset and that is not what you want from a meter. So it  needs some extra filtering to suppress this effect.
Here it is not about seeing 1 µV differences, but more like 100 nV relatively fast and 10 nV with a little time and ideally even 1 nV after averaging over longer time (e.g. 100-1000 seconds). I would consider a meter worthy nV meter if you get the noise down to the 20 nV/sqrt(Hz) range and the stabilit over a an hour to better than 100 nV, preferrably the 10 nV range. The LTC1050 OP is still a bit noisy. AFAIK the Keithley 2000 uses this as it's input and it should be the main noise source in the 100 mV range. So to really call it a nV meter it would need a slightly lower noise OP.

The other point is than to not get much other errors from thermal EMF and from the protection.  The simple protection with a resistor in the 100 K range (e.g. like used with the 34401 and other meters) would add allready significant noise, and in combination with a bias in then 100 pA range also quite some offset. So the protection usually needs to be more than just a resistor and clamps at the input.

[Mickle T.]

It was my old quick-and-dirty project of micro/nanovoltmeter for academic research lab. Based on reversed and totally reworked Yokogawa chopper DC preamp.

https://youtu.be/xlaAz28XRSQ

[MegaVolt]

Design of ultra low noise amplifiers.pdf  (figure 5)

[RandallMcRee]

The circuits posted so far all feature nanovolt sensitivity but differ in some important respects that create traps for the unwary. I know that most (all) of the posters know these things but for the benefit of my past/future self here are the important distinctions:

1) Is it AC or DC coupled?
  To put it simply AC coupling makes for an easier circuit but, as should be obvious, is quite restrictive. E.g. you cannot measure the difference between two 731Bs as you can with a nanovoltmeter.

2) Is it chopped or not?
  Chopping generally deals with 1/f noise admirably but has other tradeoffs with poor voltage noise, current noise (or both).

3) If it is not chopped then, typically, you can still use other techniques to overcome the 1/f noise. So, for example, here is an instrumentation amplifier with good nanovolt sensitivity:
    https://www.analog.com/en/analog-dialogue/articles/low-noise-inamp-nanovolt-sensitivity.html
  You can then use so-called bridge excitation techniques (LT AN-96) to remove the IA amplifier offset. That can be as easy as multiplexing the two inputs and averaging.

4) Along with such details as Thermal EMF there is always gain drift and other resistor TCR problems that are not often mentioned (so another trap for the unwary). The AD8428 mentioned above, for example, is exemplary not just because you can build a 0.7 nV/√Hz IA but also because the precision resistors are embedded in the IC. Many of the circuits shown so far, cannot realize their full potential without buying 0.2ppm (say) matched resistors or resistor networks.

Let me know if any of the above is misleading/wrong.
Thanks!

[MegaVolt]

AD7195  RMS noise: 8.5 nV at 4.7 Hz

[MegaVolt]

ADS126x  Noise: 7 nVRMS 2.5 SPS

[guenthert]

I've seen a bunch of Keithley 182 (as recently as 2014 still good enough for the Laboratoire national de métrologie et d’essais [1]) for a quite reasonabe price on e-bay.
[1] https://www.bipm.org/documents/20126/47750471/EURAMET.EM-S41.pdf/39f8462f-4bb8-09ef-67d1-661535764a36

The question is what do you want to measure. For a 10V comparision a Nanovoltmeter isn't really necessary. In many old publication they used a Fluke 8842A 5.5 digit meter with 100nV resolution. That is 0.01ppm at 10V and should be fine. The resolution may also be increased with averaging, especially if you average over 100s.

     Well, in old publications they used a galvanometer. 

The measurement will be noise limited.  I couldn't find noise figures for the 8842A, but it's 24h accuracy is specified as 50ppm of reading + 20 counts in the 20mV range, which one can take as a hint.  Nanovoltmeters fare better (the old Keithley 181 can be expected to be an order of magnitude better there).  But given the noise of the sources, there might be little to gain.

[guenthert]

For a Nullmeter you don't even need a 34401 or similar full fledged meter for dispaly. Just a simple ADC (e.g. ICL7106 for a dispaly) or even an µC internal one is sufficient. The scale does not have to be very accurate if it is only about a small residual different or just to adjust to zero. [..]

      Quite so.  For a true null meter (error detector) the scaling error is of little importance (and the offset one can calibrate oneself easily).  Comparing then two voltage sources gets awkward however, if they are stable, but not quite close.  One can use the potentiometer method (using a third source) to determine the difference, but that's a bit inconvenient (Bureau of Standards used long time ago a Brooks comparator for that).

[Kleinstein]

For a nV meter the ADC noise is the less important problem. It still help if the ADC is continuously sampling the input and not just reading the input half the time and do an internal zero the other half. This would double the time needed because it increases the effective BW for the input noise and amplifier noise.

So there can be an advantage in using an SD ADC (they usually do near 100% sampling) and not an ADC like the ICL7106 that does only some 25% sampling. The HP34420 (otherwise a good nV meter) is in this respect not ideal.  The HP typical concept with 50% input sampling and 50% zero is not the best choice for a nV meter.

The noise of the source is another point. Low BW and good filtering to keep the noise BW for the source noise may be more important than a super low noise amplifier, when it comes to just comparing 2 higher voltage references.  There are also cases where the source is low noise (e.g. checking thermal EMF) though.

Modern choppers don't have necessary high voltage noise. This was a point in the past, especially when doing only one sided switching (e.g. signal and zero) instead of polarity switching.

The right balance of voltage noise and current noise depends on the signal impedance and this can vary quite a bit.  When the nV range matter the sources are usually relatively low impedance, as otherwise the resistor noise of the source would become significant. So the current noise may not be that important for a nV meter than it is with a more normal voltmeter good for µV resolution. Anyway there is not 1 nullmeter to fit all.

[e61_phil]

The measurement will be noise limited.  I couldn't find noise figures for the 8842A, but it's 24h accuracy is specified as 50ppm of reading + 20 counts in the 20mV range, which one can take as a hint.  Nanovoltmeters fare better (the old Keithley 181 can be expected to be an order of magnitude better there).  But given the noise of the sources, there might be little to gain.

50ppm of µVs is in the pV range. So no one cares about the accuracy here. More important might by the floor spec of 2µV. But that should also include linearity and so on.

I don't have a 8842A, but a HP3478A and will give it a try.

[Echo88]

Very interesting project Mickle T! 
Can you give some more details about it? Like:
What was the used Yokogawa Preamp?
Is the used transformer a standard signal transformer?
When i looked at the signal shaper i thought that ive seen that somewhere ( https://www.eevblog.com/forum/metrology/em-electronics-model-n11-dc-nanovoltmeter/msg975473/#msg975473 )and indeed, there are the documents on the first picture. 

Back to topic:

https://arxiv.org/pdf/1708.06311.pdf Temperature controlled IF3602-based amplifier. It shows the limitations of a very low noise, non-chopped, JFET-amp.

[Gerhard_dk4xp]

Who has interesting links to papers and articles? Someone DIY'ed something similar already, probably with a fixed input range?   -branadic-

I have made a quick&dirty chopper stage based on EPC2038 GaN FETs.
They feature really small capacitance and small channel resistance.
That was more to see if I could solder them at all b4 I make a larger
time investment. They are nekkid chips  0.9mm * 0.9mm with 4 solder bumps
below.
Soldering was easier than expected with a little bit of hot air. A little little bit,
or one might not find them again.
I have not yet done any characterisation, I'm currently more interested into
microwaves.

I can bring the board to Stuttgart tomorow.

[dietert1]

This thread starts with a statement like "Keithley 2182A is desirable yet expensive". My experience is a little different: The 2182A resolution is 1 nV in its lowest measurement range (Ch1 +/- 10 mV). Compare this to a Keithley 148 with it's lowest full range of +/- 10 nV.
This is a severe design problem in the 2182A. The 1 nV is seven decades below full range and ADC stability is only good for six decades. With a lack of gain ADC noise or thermal variations do matter.
Of course the 2182A has other advantages, but null stability isn't the main strength. An ambient temperature transient of 1 °C can give you a 50 nV deviation. Even when running a 2182A inside a thermal chamber at +/- 0.1 °C for 10 days, it was walking about +/- 10 nV.

Regards, Dieter

[3roomlab]

the K 148 svc manual have a nice section "circuit description"
i reprinted that section. the original pdf is >4mb
it can be found inside ko4bb archives

i was quite curious how LISA/LIGO did their voltage reference, they have some measurement reports done which span down to 100uHz, but i dont know what they use to measure.

[Gerhard_dk4xp]

> a human being is like a 80w lamp heat source walking around a precision lab,
> sheds about 20g of water an hour on all the precision electronics around him/her.
> If a precision instrument has a voice, it would say get away from me you #@$%#$^ .

A human is about 100K, 1/4W 
(40y ago in the newspaper Elektor, under the title Electorture)

Cheers, Gerhard

[2N3055]

> a human being is like a 80w lamp heat source walking around a precision lab,
> sheds about 20g of water an hour on all the precision electronics around him/her.
> If a precision instrument has a voice, it would say get away from me you #@$%#$^ .

A human is about 100K, 1/4W 
(40y ago in the newspaper Elektor, under the title Electorture)

Cheers, Gerhard

For the purposes of heating/cooling calculations for the buildings a rough thermal output figure of 80-100W per person is estimated AFAIK.

[branadic]

This thread starts with a statement like "Keithley 2182A is desirable yet expensive". My experience is a little different: The 2182A resolution is 1 nV in its lowest measurement range (Ch1 +/- 10 mV). Compare this to a Keithley 148 with it's lowest full range of +/- 10 nV.

Wasn't it you calling a Prema 2080 scanner a piece for a museum recently? Now you draw comparisons between K2182 and K148? Really?

-branadic-

[dietert1]

branadic, you don't understand once more.
I am not reporting a bad design of a K148. Nor did i propose using a K148. We have one and although it works well, i would never try to run it more than some hours as its mechanical chopper is made from unobtainium. And it's a bit slow for noise studies.
I was reporting a design problem of the K2182A. The lowest range of the K2182A should not have been +/- 10 mV but +/- 1 mV. Should be fairly easy to mod this and repeat the tests. Also i will try and replace the black plastic covers by metal ones.

Regards, Dieter

[Kleinstein]

The Keithley 2182 service manual shows 2 amplifier stages, both capable of doing a x 100 gain. So in principle the HW should be able to have a 1 mV range.
The ADC in the 2182 is supposedly nearly identical to the K2010 7 digit meter. It is still similar to the K2000 but quite a bit better performance. The ADC gain stability should be a smaller problem if only a small fraction of the range is used. So I am not convinced that the ADC part is actually limiting. Chances are that already in the 10 mV range the amplifier part is limiting and not the ADC. Otherwise they would have likely enabled the 1 mV range already there in the HW.

Besides the ADC and amplifier there could also be thermal EMF caused noise, e.g. from the protection part.
The input protection with MOSFETs is somewhat sensitive to thermal effects, likely more than the ADC. The effect of the thermal transient is likely not from the ADC part, but from the amplifier (e.g. thermal EMF at the resistors or the switches) or the protection. The parts are in the extra box, but may still see same gradients. For my voltmeter I use a similar protection and it looks like 1 input shows quite some thermal effect (~ 500 nV for the effected channel, < 50 nV for the other).

The Keithley 148 is more like a dedicated nV meter for lower source impedance. The 2182 is still relatively high impedance and less speciallized.

A point to check with the K2182 may be if it also suffers from the slightly odd "bump" in the noise vs PLC curve like many other Keithley meters (e.g. DMM7510, DMM6500, K2002 ). This or may not be the case.

[dietert1]

The input stage should get metal shields. The existing plastic covers serve to prevent air drafts, but don't keep away electrical noise, e.g. from the nearby mains transformer and the VFD display. That design doesn't appear right. If you ever pulled one of the boards from a HP spectrum analyzer, you know what i mean. I already replaced the original IRF610 mosfets of the protection circuit by smaller 800V parts i recently bought. The smaller parts pick up less noise, it showed up in the numbers.

For a first test i added a 5R1 resistor parallel to R641 and R717 to increase the gain of the Ch1 input stage by a factor 10 - from 100x to 1000x. That modded 2182A started without complaints and appears to work well. Polarity reversal (FAZ=1) works but signals don't look nice, that needs some work. Most likely ACAL won't work now. For logging the display will be off anyways, with no wrong numbers visible..

The input stage JFET quad exhibits offset between 0 and 50 uV (FAZ=0, without covers). With all covers on, after some minutes it reads a pretty stable 20 uV. So these selected JFET quads are pretty special, There was a thread about 2182A JFET selection here: https://www.eevblog.com/forum/metrology/fet-selection-for-precision-circuits/msg1495213/#msg1495213.

Regards, Dieter

[Kleinstein]

An offset of some 50 µV looks rather low for matched JFETs.  Is it possible that this includes already a digitally subtracted constant part ?
AFAIK there is only 1 pair of amplifier jfets (SK170 AFAIK) for the input stage. The other 2 pairs with heat shrink are for switching / polarity reversal.
Looking at the pictures to get an idea about the circuit, it looks like there is some switching at R700 + U641 not shown in the block diagram, that may actually be some kind of coarse digital offset adjustment.

With a much higher gain one may have to do a better offset adjustment for the input JFETs. For the FET offset I would expect some 1-10 mV, which may be still acceptable with a gain of 100, but too much for a gain of 1000.
The offset translates to an AC background at the amplifier output (TP651 ?) and less amplitude would make the settling time less critical.

Instead of tweaking the existing amplifier without a full schematics, it may be more useful to build a separate one with a similar structure. The parts used are not that exotic. It looks like the LM394 transistor pairs are used for a 2nd amplification stage after the JFETs. Just the FETs as shown in the simplified block-diagram would be too little gain for the chopper part. With the initial gain of the FETs in front it should be good enough to have more normal transistors for the 2nd stage. So chances are the LM394 is overkill here. The point would be more some offset adjustment (trimmer or DAC) to ideally keep the ripple small.
One point I don't like very much with the K2182 amplifier is the odd mixed type chopper:  CMOS (2N7000 ?) for the input and JFETs (Q622+Q623) for the feedback. To me this makes not much sense. The compensation in the gate charge would likely be better with 4 switches of 1 type (e.g. 4 x MOSFETs, 4 x JFETs or a CMOS switch chip).

[dietert1]

I do have full schematics reengineered by Tin. It was incomplete but showed a lot more details than the service manual. It appeared in his youtube video "2182A trouble shooting". We can see that the LM394 is used as a cascode for the input JFET quad, so the FETs operate at constant Uds and at constant current. Ugs is about 0.3 V. The 5 mA current source is based on a TL431, an opamp and another JFET with a 499R source resistor for current measurement.
U614=MC34081 follows after the JFET quad, so there is enough gain to run the input stage at 1000x.
I wrote the JFET quads are very special and they are, once more unobtainium. Their combined offsets are really much smaller than 1 mV - a requirement for using FAZ polarity reversal at +/- 1 mV full range without losing dynamic range and it is fulfilled. I saw the 20 uV times 1000 on the scope, monitoring TP651. At G=1000 polarity reversal gives 40 mV there.
One can recognize that the JFET quads were combined from pairs already available as parts of the 2182 non A predecessor. Basically they should have used two constant current sources to drive each pair separately, but this was omitted. One can use small resistors like 5 to 12 Ohm below the common source of one pair to balance currents (2.5 mA each pair).
They made a circuit with nine interconnected discrete transistors in linear mode, of course prone for RF instability. There should have been some 33R resistors, e.g. to separate the JFETs from their cascodes and to spearate the JFET pairs from each other.

Regards, Dieter

[Vtile]

Please, send all those poor and unsatisfying HPs & Keithleys to me for proper disposal. 

Ps. As spoken, quad JFET I would suppose it does refer to quad JFET transistor at one die / TO-package.

[Kleinstein]

In the 2181/2182A they use 2 or 4 TO92 package JFETs as a matched set. So not even dual JFETs in one case, but measured and sorted parts. These are than coupled with some heat shrink.

[dietert1]

Yes, and the coupling is once more pairwise and less than perfect. Maybe they put the heat shrink after soldering the parts or the heatshrink softens during soldering, so there are air gaps. I remember remarks in a DIY low noise amplifier description that 2SK170 should be mounted on a heatsink to improve low noise operation at 10 or 100 Hz. SMD parts like the "new" TI JFE2140 should behave better in that respect.

I think the 2182A can be used down to 1 nV when operated inside a thermal chamber, with stable and noise free mains supply and if the zero gets recalibrated regularly or continuously. 2182A ACAL does not recalibrate the zero.
When used with a low thermal multiplexer, the MUX can provide either a relay for polarity reversal or a low thermal short on one channel. With continuous recalibration one can probably maintain the zero to 1 nV rms. Noise is about 1.3 nVrms for 5 sec readings (averaging 40x 2 PLC with FAZ=1).

Regards, Dieter

[Vtile]

Thank you for clarification, I was scratching my head as I was wondering if I have seen a such quad package in my recent newbie journeys to some 1970's instrumentation amplifier papers or not. Propably not.

My dilletant newbie mind is now asking if there is any sense to use as high as possible operating voltage Vce/Vds for transistor or differential pair as possible  (eg. 30V instead of 5V) as I faintly do remember that it is better for overall performance of amplifier stage. Most probably the answer is that it depends... I think I really need to read my book(s) again.

Anyhow, thank you again for interesting discussion.

[David Hess]

Tektronix was fond of using two metal can JFETs mounted in a block of aluminum where the highest precision was required.  A metal clip in the shape of an S used to be available for doing the same with plastic or metal parts.  Tektronix matched JFETs down to 5 millivolts and 10 microvolts per degree C in this way.

My dilletant newbie mind is now asking if there is any sense to use as high as possible operating voltage Vce/Vds for transistor or differential pair as possible  (eg. 30V instead of 5V) as I faintly do remember that it is better for overall performance of amplifier stage. Most probably the answer is that it depends... I think I really need to read my book(s) again.

There are a couple of reasons but you might be confusing the reason with controlling the emitter/source current.

Their is an advantage to using as high a voltage biasing the emitter/source current as possible so that the emitter/source biasing resistance is as high as possible, to improve alpha of an emitter/source follower or common mode rejection of a differential pair, but replacing the emitter biasing resistance with a current source gives the same advantage at lower voltage but higher cost.

Higher Vds and Vce keeps junction capacitance low, so where this matters, a higher voltage will be used, however in JFETs, higher Vds increases gate current through impact ionization (?).  Robert Pease mentioned rediscovering this at one point, and JFET circuits optimized for low gate current deliberately use low and controlled Vds.

[dietert1]

As this thread is about DIY: Meanwhile i "coated" the two covers of the K2182A input stage with 0.5 mm copper sheet. Need to do another two.

Regards, DIeter

[mawyatt]

Tektronix was fond of using two metal can JFETs mounted in a block of aluminum where the highest precision was required.  A metal clip in the shape of an S used to be available for doing the same with plastic or metal parts.  Tektronix matched JFETs down to 5 millivolts and 10 microvolts per degree C in this way.

Robert Pease mentioned rediscovering this at one point, and JFET circuits optimized for low gate current deliberately use low and controlled Vds.

Way back had seen some of the Tek Aluminum blocks (think they used them in the low noise diff amp plug-in, 7A22) and many used the "S" type clips which held a TO92 or TO18 can.

Bob Pease was a master at finding unique circuit/device features & niches. Remember Bob Widlar trying to beat JFET bias currents with his brilliant bipolar designs, fun times back then

Best,

[Kleinstein]

Today JFETs in a metal can are relatively expensive and even TO92 is getting rare. The main case you get is SOT23 and similar SMD ones.
The small size case has also some advantage getting the fet pair close together.  However measuring the FETs before soldering is a bit fiddely. For a well matched pair one would need to measure quite few (like 20).  I don't think the unmatched duals are a great help, more like using teh 2 in one case in parallel to get a larger area jfet.

[dietert1]

I thought about measuring JFE2140 pairs. They come in SO8, so i ordered a test socket ("programming adapter"). Their offsets are pretty low already from factory. Maybe one can combine two pairs with opposite offsets and glue one on top of the other to use them as one balanced pair.

By the way the 34420A has a FET pair labeled SNJ3600X05. Probably an Interfet N3600l geometry that has a "typical" spec of about 0.5 nV/sqrt(Hz) down to 0.01 Hz. The chip is 1.838 x 1.838 mm. Another part number was
Interfet IF3602 - without pair matching, in TO78 can.

Regards, Dieter

Edit: The noise spec is at 0.01 KHz, not 0.01 Hz. My Error. And the datasheet has a pair spec of 100 mV, that is near useless, "differential gate source voltage".

[Gerhard_dk4xp]

8 pcs. of IF3602
Wonder why it pops up everywhere this week.
There must be "some" matching. It could be even worse.

>  "typical" spec of about 0.5 nV/sqrt(Hz) down to 0.01 Hz.

Sorry to burst that bubble.   0.01 KHz.
And the noise does not follow 1/f , there is an unexplained plateau.

Gerhard

[Vtile]

The small size case has also some advantage getting the fet pair close together.  However measuring the FETs before soldering is a bit fiddely. For a well matched pair one would need to measure quite few (like 20). 

Well as a some small middle project I created a small tool for testing sot-23 transistors with ease. As shown in pictures it continues the 'metrology' tinfoil traditions. Contact surface is build with manhattan technique on cloths pin jaw, it seems to work flawlessly. It can be also used for other surface mount devices like resistor, which was my first DUT, with two different un-calibrated technologies so the non-correlation to standards can be guaranteed.

I should attend to 'Nanoamps and below like ninja' -thread with that galvanometer in Leeds & Northrup test set .. it's just crazy. Something verified with Keithley 197A Microvolt meter




Next test component was with MMBF4117 and 5$ component tester, unfortunately that JFET is just too much to tester, I need to ask about it at $5 tester thread...


I swapped to BC847C and that is in operation limits of this atmega wonder.


Last test was conducted with Meratester FET multitester, the resistance between measuring pads is over 10Gohms.


Construction is simple and by ordering gold plated PCB one can get rid of oxidation problems. That is the reason I tinned the pads, so I can remelt them easily with flux. The caps between pads are filled with cyanoacrylate (cheapest variety, wonderfull material), then swept with small wooden spatula to flush (another half of the clip). When quick glue were driedthe surface was cleaned with exacto-knife. If one needs to make temperature variation one could ie. glue a big thin SMD resistor at the another jaw and make some form of heat controller to it (I think I will do that at some point .. maybe). I can see also a future of making myself a clamp on heater for resistor selection with same type.



Hopefully this post gives some diy ideas.

 

edit. pictures attached with PC.

[David Hess]

Today JFETs in a metal can are relatively expensive and even TO92 is getting rare. The main case you get is SOT23 and similar SMD ones.

The small size case has also some advantage getting the fet pair close together.  However measuring the FETs before soldering is a bit fiddely. For a well matched pair one would need to measure quite few (like 20).  I don't think the unmatched duals are a great help, more like using teh 2 in one case in parallel to get a larger area jfet.

I was thinking close proximity of the SOT23 packages and then attaching an aluminum block to their tops with thermal epoxy would work.  Or maybe use 4 transistors in a discrete thermally coupled cross quad?

Precision monolithic JFET pairs are in theory still available from Linear Systems, and some others, like the LS843 with 1 millivolt offset and 5 microvolt per degree drift.

[Vtile]

I do know this is a wrong thread and should be on projects or... , but here is another transistor holder design (wip). Again with proper PCB this would be much better.

[MegaVolt]

is another transistor holder design (wip).

There are adapters for sot23.
https://aliexpress.ru/item/32808842375.html

[Vtile]

is another transistor holder design (wip).

There are adapters for sot23.
https://aliexpress.ru/item/32808842375.html

Thank you for a hint, it seems my google fu was weak. However, I do have an urge to add link here to Keithley nanovolt products. 

[CurtisSeizert]

By the way the 34420A has a FET pair labeled SNJ3600X05. Probably an Interfet N3600l geometry that has a "typical" spec of about 0.5 nV/sqrt(Hz) down to 0.01 Hz. The chip is 1.838 x 1.838 mm. Another part number was
Interfet IF3602 - without pair matching, in TO78 can.

Regards, Dieter

Edit: The noise spec is at 0.01 KHz, not 0.01 Hz. My Error. And the datasheet has a pair spec of 100 mV, that is near useless, "differential gate source voltage".

I recently built up a lf noise amp using a cascoded IF3602 long-tailed pair input stage feeding an ADA4625-2. Here is my noise floor measurement from that.  It is about 30 nV p-p 0.01-10Hz (The input HP filter uses C=34uF, R=10M, so -3dB is actually about 0.02 Hz).  There are some other noise sources in here (mostly the HP filter resistor at low frequency), but it gives a good indication of the 1/f corner being around 1 Hz.  Interfet's LTSpice models for the IF3602 are quite accurate from my (limited) experience with the part.  The source-drain voltage for the part used in this amplifier was at the low end of the spec, around 520 mV from memory (spec lower limit is 0.5 V), vs. 1.5V in the model.

I was nervous about the barn door matching spec for these, but I don't think they use much of it.  The two samples I got were 2 mV and 6 mV differential with Id of 2.3 mA.  I may need to order more of these, but unless I win the lottery I am not going to be doing a statistical analysis on Interfet's process control  I am working on a revision of the amp with some improvements, and I will upload the design files when I can confirm that it works if anyone is interested.

Curtis

[TiN]

Planning to join up for nV project challenge?. 

[CurtisSeizert]

Planning to join up for nV project challenge?. 

I thought about it, but I do not really have the time or expertise to do that scope of a design and deal with the offset issues that come with making such a system suitable for low level DC measurements.  I was considering the route of an autozero implementation of the IF3602-ADA4625 hybrid amp

Curtis