Ultra Low Noise Reference 2DW232, 2DW233, 2DW23x

[Andreas]

What can you say about the 1N829A?

https://www.eevblog.com/forum/metrology/ultra-low-noise-reference-2dw232-2dw233-2dw23x/msg1065122/#msg1065122

The lowest values I had on some NOS Motorola parts.
Followed by new production APD devices.
Finally the NOS ST parts.

with best regards

Andreas

[zgxd]

2DW235 , DIY  10V   23ma   恒温50℃   9h  0.3ppm

[unseenninja]

What can you say about the 1N829A?

You will need to buy as many 1N829As as you can afford. Then connect them all to 7.5mA current sources and let them run for a few months. Once that time has passed, you can start logging the output voltage of all the diodes and begin the process of eliminating the diodes with popcorn noise from the group. After removing the obviously noisy or high drift diodes from the group, you can start to more closely evaluate the performance of the remaining diodes. Once you have a couple of good candidates, you can start varying the Zener current while cycling the diode in a temperature controlled chamber until you find the current which delivers the lowest temperature coefficient.

Expect this process to take months. There is a significant risk that you may find only one or two diodes which deliver acceptable results, or maybe even none at all.

Finding good 1N829A diodes is a rabbit hole.

[ch_scr]

As described here, I swept 99 2dw232 diodes, configured as "diode+zener in series".
As the full sweep-suite (3.5, 4, ..., 6.5mA; 15, 20, ..., 60°C; see pic 1) took ~11h, I changed diodes twice a day and initially just noted the observed zero-tc-temp vs. current manually in an excel sheet "along the way".
Now finally, I got around to pulling the whole sweep data as csv (all 240MB of it) and put it through a Python data processing.
The data was tagged if the fixture temp was in the "ramp" or "stable" phase, so I could easily use only the "temp stable" data points.
I just ended up averaging and putting a parabola curve fit onto them to extract the zero-tc points (that worked even, if it was somewhat outside the actually swept temp range, as long as there was enough "curvature").
As 7 currents for 99 diodes makes almost 700 pictures, exporting to individual png seemed not practical. Thankfully, the Python "Bokeh"-library allows to easily combine individual plots into one html. The second picture shows how that looked, blue being the plotted data points and red the parabola fitted onto them (where Python "curve_fit" could). That full html is 38mb still...
Since the turning-point of those fitted parabola are the wanted zero-tc points, the core data could now be extracted!
I used "y = a*(x-b)^2+c " as the parabola model to get the turning point directly as parameter "b". These "parabola model parameters" are the in csv attached. 240MB reduced to 58kB, somewhat magical if you do it for the first time.
What that looks like plotted shows picture 3, all 99 zoom-able plots are in the attached html file.
From this it seems reasonable to conclude a more-or-less linear relationship of zero-tc-point vs current for these diode.
I also made histograms at a few currents (amount of zero-tc-points in 5°C wide bins from 0-100°C) but no pattern emerged.

PS: please excuse most of the plots are marked "2dw236", there was an error in the script logging the data and all of the data got "mis-labled" that way 

[Korchahin]

As described here, I swept 99 2dw232 diodes, configured as "diode+zener in series".
As the full sweep-suite (3.5, 4, ..., 6.5mA; 15, 20, ..., 60°C; see pic 1) took ~11h, I changed diodes twice a day and initially just noted the observed zero-tc-temp vs. current manually in an excel sheet "along the way".
Now finally, I got around to pulling the whole sweep data as csv (all 240MB of it) and put it through a Python data processing.
The data was tagged if the fixture temp was in the "ramp" or "stable" phase, so I could easily use only the "temp stable" data points.
I just ended up averaging and putting a parabola curve fit onto them to extract the zero-tc points (that worked even, if it was somewhat outside the actually swept temp range, as long as there was enough "curvature").
As 7 currents for 99 diodes makes almost 700 pictures, exporting to individual png seemed not practical. Thankfully, the Python "Bokeh"-library allows to easily combine individual plots into one html. The second picture shows how that looked, blue being the plotted data points and red the parabola fitted onto them (where Python "curve_fit" could). That full html is 38mb still...
Since the turning-point of those fitted parabola are the wanted zero-tc points, the core data could now be extracted!
I used "y = a*(x-b)^2+c " as the parabola model to get the turning point directly as parameter "b". These "parabola model parameters" are the in csv attached. 240MB reduced to 58kB, somewhat magical if you do it for the first time.
What that looks like plotted shows picture 3, all 99 zoom-able plots are in the attached html file.
From this it seems reasonable to conclude a more-or-less linear relationship of zero-tc-point vs current for these diode.
I also made histograms at a few currents (amount of zero-tc-points in 5°C wide bins from 0-100°C) but no pattern emerged.

PS: please excuse most of the plots are marked "2dw236", there was an error in the script logging the data and all of the data got "mis-labled" that way 

Great job! I need to come up with something like this. I have a board with ADS1256 - 8chanel ADC. I haven’t checked its stability yet, but it might be possible to use it with an ESP32 and additional harness for a measuring stand. Thanks for the idea and the work done!

[guenthert]

[..]
I have a board with ADS1256 - 8chanel ADC. I haven’t checked its stability yet, but it might be possible to use it with an ESP32 and additional harness for a measuring stand. Thanks for the idea and the work done!

   I attempted something similar, monitoring the difference (series opposition) voltages of a set of 1N829 Zeners against a pre-selected one, but somehow managed to fry the (nice, including good reference and DC-DC converter to separate digital and analog supplies) ADS1256 based board from fleebay.  Don't recall what happened (either 6.2V went into the input expecting some mV or the supply voltage was exceeded).  That was a nice board with good resolution, but probably better part of a fixed system, then used by itself in a volatile lab. At least it turned out insufficiently robust for this fool.

[iMo]

..a 1k resistor in the inputs might help, there are protection clamping diodes at the inputs which would take 10-20mA worst case..
PS: it seems the ebay modules have usually got the 1k resistors there at the inputs..

[dietert1]

Here those ADS1256 modules are mainly used for precision temperature measurements, with a half bridge made of a 10K NTC and a 10K resistor.
On our (amazon, ebay) modules the input resistors came as 100 Ohm. Of course one can replace them by 1 KOhm as imo proposed. Another way to produce better input protection are two back-to-back depletion JFETs with a 200 Ohm resistor between their sources and each gate wired to the source of the other JFET. This will be safe up to about +/- 25 V.
A similar scheme with 800 V mosfets exists in the Keithley 2182A nanovoltmeter yet it needs more parts - not that interesting for a scanner.

Regards, Dieter

[iMo]

The DS says the max inp current is momentary 100mA and 10mA permanent (I think it is about the clamping diodes). So with 10mA and 5V analog supply it should withstand +15.6/-10.6V input with 1k (some adjusting would be needed because of the input impedance).

[dietert1]

Another hidden risk is increasing the supply voltage via one of the protection diodes. I mean with a small protection resistor one may feed enough current into one of the inputs to kill the device by over voltage. With the default 100R resistors 6.2 V may have been enough. Maybe one needs to clamp the supply with a zener.

Regards, Dieter

[Kleinstein]

As a step up from the back to back JFET protection there are depletion MOSFETs like BSP135 that could provide protection to some 500 V. It is less effort than the MOSFETs + PV gate driver as used in the Keithley K2182 / K2000 / K2001 / 2002.  Another advantage of depletion FETs is having no heat and thus not as much worries abouter thermal EMF. It would still need good thermal design to keep temperature fluctuations low.

For noise testing to a really low level the demand is quite high, when using direct an ADC. It still needs a very good reference with that ADC. It may be OK for the drift, but there is also drift an the ADC gain and it's reference.

For just the classic noise in the 0.1 to 10 Hz range maybe a little lower one could use AC coupling and an amplifier. The following ADC would not be that critical, though it should be reasonable fast and allow for good mains hum suppression.

For the much lower frequency part one would kind of need a good reference to compare to and maybe better only look at the difference between 2 reference (ideally close in the voltage). With a set of 3 reference and simultaneously looking at the 3 differences one could get noise data for the individual reference (at least the RMS part to some degree).
For this measurement the more suitable choice would be more an ADC chip for simulaneous sampling with 3 or 4 channels, with not that high demand on the individual ADCs. Part of the resolution would be needed to take a difference in the voltages into account and for this reason just a µC internal ADC could be a bit tricky (e.g. with fine trim to adjust for a common voltage). Simultaneous reading is not absolutely needed, but it would help to get a result faster and less effected from natural scattering. Popcorn type jumps could also be directly analysed in the time domain (not seen in the diference between the other 2 reference).

[iMo]

BSP135J protection..
PS: the D1 and D2 diodes are on-chip clamping diodes, R1 will not be used, there are several grades of BSP135 - J/K/L/M/N based on the Vgs_thr..

Model from Bordodynov's lib:
.model BSP135J VDMOS (Rg=5 Vto=-0.90 Kp=0.38 Lambda=0m Rs=3.5 Ksubthres=0.1 Mtriode=0.8 Rd=40 Bex=-1.9 Vtotc=-0.8m Tksubthres1=1m Trs1=3m Trd1=3m Cgdmax=200p Cgdmin=3p a=0.7 Cgs=60p Cjo=150p m=0.75 VJ=0.75 IS=10p N=1.2 Eg=1.12 Rb=1 Vds=600 Ron=60 mfg=InfIH2008)

[iMo]

..
For this measurement the more suitable choice would be more an ADC chip for simulaneous sampling with 3 or 4 channels, with not that high demand on the individual ADCs..

I remember times I was building ADCs with discrete components - a S/H chip with a quality capacitor, a DAC, a SAR TTL register, a comparator..

What about 3-4 Sample/Holds with a single ADC instead of the 3-4 simultaneously sampling ADCs??

[Kleinstein]

Absolute simultaneous sampling is not needed as the interesting noise part is the low frequency part. With proper anti aliasing filtering also switching between channels could also work. However SD ADCs are not that great with switching between channels as it takes addition time for settling. With SD ADCs there are chips available with true multiple ADCs inside (e.g. ADS131m04, MCP3913). Also just 3 seprate chips would work.
It is still quite some effort for just getting the noise also for lower frequencies.

With more time one could even just compare 1 pair of references at a time. It would however take longer (e.g. 3 times for only one part at a time and maybe another factor 2 from fluctuations in the noise that are reduced with reading at the same time).

[dietert1]

Using ADS1256_CMD_WAKEUP the ADS1256 can be configured to take 40 independent measurements of 20 msec = 1 PLC in one second with the timing determined by the MCU, e.g. at 25 msec intervals. So between every two samples there is a 5 msec pause to configure channel, PGA etc.
In the datasheet they claim that one can run the ADC 100 % of the time and still operate the MUX to get independent samples. I think it depends on interrupt latency and one needs to send commands for the next sample before reading previous result.

Regards, Dieter

[Kleinstein]

The ADS1256 can be reasonable fast switching channels. It may not even take 5 ms for settling between switching channels. It should be 5 or 6 values  at the 30 kSPS internal rate for the sinc5 filter part to wait an thus only some 0.2 ms alsolutely needed for the switching time. 3 steps and 5ms waiting 20 ms conversion would be 75 ms for 3 channels.

[iMo]

So wire all 1256' inputs to a single reference and look at the results while doing the fast scanning through the channels. Perhaps it would work fine. An easy check..

[Kleinstein]

One may not notice a glitch of the channels are the same voltage. For real test it would need different voltages.

In principle one could even get away with only 2 channels, the 3rd difference could be calculated in theory, though one may need to take into account the ADC and amplifier noise more carefully with less symmetry.

[Birb]

Not sure if this will be helpful, but after looking into a DeltaV against Current, it seems to follow
V(I) = a*ln(I) - I/b + c
where V(I) is the DeltaV and I is the current used, between two same temperature points.
unfortunately it does not predict the variation with temperature at all.

I wonder if around 5 points can be taken, then the plot regressed, and the zero TC neighbourhood be found quickly
though I note that this function is non-linear and is difficult to compute a fit.

[Korchahin]

I was hoping to select the 2DW232 with the highest TC and believe that it will be the least noisy. But apparently we will have to start with making an LNA (with DC-servo's?) to determine the noise. And also a good voltmeter. Someone can suggest a ready-made project for ADS1263 + ADR1399 (LM399) or similar OSHW projects. I guess I got infected with “volt-mania”.
Thanks guys for your help and quick response.

[iMo]

I guess I got infected with “volt-mania”..

I would highly recommend you to get a vaccination soon, even it might be late already.. 

[Kleinstein]

I was hoping to select the 2DW232 with the highest TC and believe that it will be the least noisy. But apparently we will have to start with making an LNA (with DC-servo's?) to determine the noise.

For just measuring the 0.1-10 Hz or similar noise there is not need for a good voltmeter. The method of choice for that frequency range is an AC coupled LNA. There are plenty of plans for such LNAs around and these are not that complicated. The actual measurement of the output could than be with a sope of any meter to get some 25 Hz of higher sampling rate if one does not care that much about the exact filter response at the upper end. The exact filter to use is anyway a tricky topic and not allways handled the same.

The tricky part is if one wants to look at noise for even lower freuencies and the longer time drift. Looking at the long term drift is what may want a good DMM. With the hoped for low noise level a direct voltage measurement with a DMM is not good enough. It could work with the difference of 2 references.

There is no reason to belief to get a lower noise with a higher TC. More of the opposite as temperature variations times TC add to the low frequency noise. For the 2DW232 the TC also depends on the current - ideally one can find a current with near zero TC and would regulation the tempeture to be at least somewhat stable to fight the 2nd oder part.

[iMo]

..Someone can suggest a ready-made project for ADS1263 + ADR1399 (LM399) or similar OSHW projects. I guess I got infected with “volt-mania”.
Thanks guys for your help and quick response.

I collected all that parts years back, but did not find one. But if you are good with KiCad - start a new thread, stay dedicated and focused, do not get distracted for some time, do not try go beyond 6.5digits - and people will help you as that has been discussed for years here already.

One 100x100mm large 4 layer ADC board with:
1. analog inguard front end (bootstrapped input opamps with a 1/10 divider, +/- 20V input)
2. ADC 1263 with 1399 reference (the 1399 is not easy to use as you will need 5V buffered)
3. temperature sensors
4. a data isolator (like ISO6xxx)

and the second 100x100mm 2 layer board with a complete 4 floating separate voltages (linear): +25V/100mA, +25V/50mA, +5V/100mA, +5V/300mA (for an mcu with a display for example), with onboard pcb transformers (even people will say their isolation is poor, perhaps good for those 6.5 digits).

The output of the first inguard ADC board  (like 5 digi lines) will go to an outguard MCU (any you have handy, Rpis, arduinos, esp32, 16b/32b MCUs, whatever which can do 64bit floating point in C).

[mawyatt]

I guess I got infected with “volt-mania”..

I would highly recommend you to get a vaccination soon, even it might be late already.. 

The "Doctor" prescribed a HP3458 along with a Fluke 732/734 as a temporary cure

Best

[zgxd]

2dw235  2024.3  DIY   23ma  50 ℃   10.000 199V   
2024. 5. 5     10.000 097V
2024. 5.19    10.000 128V
2024. 7.10    10.000 214V
2024. 8.18    10.000 251V
2024. 8.25    10.000 254V
2024.10.27   10.000 302V