Author Topic: Statistical arrays  (Read 9780 times)

0 Members and 1 Guest are viewing this topic.

Offline doktor pyta

  • Frequent Contributor
  • **
  • Posts: 276
  • Country: pl
    • RF Scientific
Re: Statistical arrays
« Reply #50 on: January 01, 2019, 09:43:59 pm »
It would be good idea to have individual resistors; T.C. measured.
It may contribute our understanding of the problem.
I have to think how to use my gear to make 8x 4wire measurements at two temperatures easiest way.

At this point my theory of best performance of layout v.2 is that resistors R4, R2 have equal temperatures.
Same with R1, R3.


Offline antintedo

  • Contributor
  • Posts: 15
  • Country: pl
Re: Statistical arrays
« Reply #51 on: January 14, 2019, 12:58:05 am »
I found some time to perform TCR measurements. So far only 1 sample of 10k NOMCA was tested.
1:2 divider built out of this sample measured 0.4ppm/K which seems to agree with the results below.

Code: [Select]
Sample #1
No  T1      R1          T2      R2          dT      TCRavg  TCR/TCRavg
1   26      10.00243    64      9.99943     38      -7.89   77%
2   26      10.00286    64      9.99750     38      -14.11  138%
3   26      10.00363    64      9.99877     38      -12.79  125%
4   26      10.00189    64      9.99913     38      -7.26   71%
5   26.5    10.00284    63      9.99927     36.5    -9.78   95%
6   26      10.00264    63      9.99840     37      -11.46  112%
7   26      10.00179    64      9.99839     38      -8.95   87%
8   26.5    10.00230    64      9.99862     37.5    -9.81   96%

Sample #2
No  T1      R1          T2      R2          dT      TCR     TCR/TCRavg
1   25      10.00220    67      10.00106    42      -2.71   73%
2   25      10.00235    67      10.00032    42      -4.83   129%
3   25      10.00193    67      10.00057    42      -3.24   87%
4   25      10.00160    67      9.99992     42      -4.00   107%
5   24.5    10.00193    67      10.00021    42.5    -4.05   108%
6   25      10.00160    67      9.99982     42      -4.24   113%
7   25      10.00205    67      10.00074    42      -3.12   83%
8   25      10.00214    67.5    10.00056    42.5    -3.72   99%
« Last Edit: July 04, 2019, 06:44:01 pm by antintedo »
 
The following users thanked this post: Andreas, e61_phil, EmmanuelFaure, doktor pyta, alm, razvan784

Online Andreas

  • Super Contributor
  • ***
  • Posts: 2410
  • Country: de
Re: Statistical arrays
« Reply #52 on: January 14, 2019, 08:25:20 pm »
Thanks for the measurement,
I never had expected that the stray between neighboured resistors is that large.
I had expected more likely a rising T.C. from one side to the other or a center symmetrical behaviour.

with best regards

Andreas
 

Offline branadic

  • Super Contributor
  • ***
  • Posts: 1486
  • Country: de
Re: Statistical arrays
« Reply #53 on: January 14, 2019, 08:41:18 pm »
I would have expected more equal values for t.c. between single resistors as the ceramic area is comparable small so are the resistors and they are manufactured in thinfilm technology, so multiple networks are processed in parallel and films are highly uniform over small areas.

-branadic-
Fluke 8050A | Prema 5000 | Prema 5017 SC | Advantest R6581D | GenRad 1434-G | Datron 4000A | Tek 2465A | VNWA2.x with TCXO upgrade and access to: Keysight 3458A, Keithley 2002, Prema 5017 SC, 34401A, 34410A, Keithley 2182A, HDO6054, Keysight 53230A and other goodies at work
 

Offline Magnificent Bastard

  • Regular Contributor
  • *
  • Posts: 124
  • Country: aq
Re: Statistical arrays
« Reply #54 on: January 14, 2019, 09:34:41 pm »
I found some time to perform TCR measurements. So far only 1 sample of 10k NOMCA was tested.
1:2 divider built out of this sample measured 0.4ppm/K which seems to agree with the results below.

Code: [Select]
No  T1     R1         T2   R2        dT     TCR
1   26     10.00243   64   9.99943   38     -7.89
2   26     10.00286   64   9.99750   38     -14.11
3   26     10.00363   64   9.99877   38     -12.79
4   26     10.00189   64   9.99913   38     -7.26
5   26.5   10.00284   63   9.99927   36.5   -9.78
6   26     10.00264   63   9.99840   37     -11.46
7   26     10.00179   64   9.99839   38     -8.95
8   26.5   10.00230   64   9.99862   37.5   -9.81

These results seem suspiciously variant (not the absolute values, but the relative TCRs).  Can you describe in detail how you made these measurements?
 

Offline antintedo

  • Contributor
  • Posts: 15
  • Country: pl
Re: Statistical arrays
« Reply #55 on: January 15, 2019, 12:13:41 am »
These results seem suspiciously variant (not the absolute values, but the relative TCRs).  Can you describe in detail how you made these measurements?
The resistor was soldered on a FR-4 SOIC to DIP converter and attached to a stabilized heater with a pretty large heatspreader in between.
The whole setup was put into a lightly insulated aluminum box.

For every measurement wires were soldered to corresponding DIP holes.
I used 34401A in 4-wire mode to perform 3 measurements: first at ~25C then at ~65C then again at ~25C.
Temperature readout was done with a thermocouple placed near the middle of the package. Plenty of time was given to reach thermal equilibrium after heating/cooling cycle and soldering.

After doing all 8 resistors, I repeated the measurements in the same order as first round of measurements.
I didn't write it down, but I think repeatability of resistor values between two runs was well below 10 ppm, otherwise I would have just discarded the results.
Room temperature measurements after the experiment agreed to 7ppm with measurements taken the day before. I also used VHP101 resistor box as a sanity check.

I think the main downside of this setup was mechanical stability and the need to solder to the board between every measurement, potentially leaving flux and skin contaminants on the surface. I will try to eliminate this and re-test, but it has to wait until the weekend.

I had expected more likely a rising T.C. from one side to the other or a center symmetrical behaviour.
Andreas
I was surprised as well. Out of curiosity I tested one more more sample today and got very different results. TCR was much lower, 3.7 ppm average +/- 1 ppm spread with no obvious patterns. It is certainly possible something was wrong about my initial methodology. This time I paid more attention to handling the board when soldering and fixed the board rigidly using bolts instead of springs.
 

Offline nfmax

  • Frequent Contributor
  • **
  • Posts: 854
  • Country: gb
Re: Statistical arrays
« Reply #56 on: January 22, 2019, 09:29:08 am »
Maybe the effect of thermally-induced strain changes from the mounting arrangement? All resistors are strain gauges!
 
The following users thanked this post: antintedo

Offline branadic

  • Super Contributor
  • ***
  • Posts: 1486
  • Country: de
Re: Statistical arrays
« Reply #57 on: April 10, 2019, 04:06:18 pm »
Since I was interested in the question if the NOMCA arrays are useful for a 7V --> 10V boost circuit and noone has yet answered the question by experiment results I've now ordered the boards with my design shown a few posts before. They will receive soon and I will then perform some measurements such as ratio measurements over temperature. If they perform well I can directly use them for my LTZ1000 references.

-branadic-
Fluke 8050A | Prema 5000 | Prema 5017 SC | Advantest R6581D | GenRad 1434-G | Datron 4000A | Tek 2465A | VNWA2.x with TCXO upgrade and access to: Keysight 3458A, Keithley 2002, Prema 5017 SC, 34401A, 34410A, Keithley 2182A, HDO6054, Keysight 53230A and other goodies at work
 
The following users thanked this post: Andreas, e61_phil, 2N3055, Grandchuck

Offline doktor pyta

  • Frequent Contributor
  • **
  • Posts: 276
  • Country: pl
    • RF Scientific
Re: Statistical arrays
« Reply #58 on: April 13, 2019, 12:03:36 am »
branadic,

I did not have time to test the idea but this is what I came up with (drawing).
Layout for this can be designed symmetrically and I think it could be further improved by paralleling similar dividers.

Offline dietert1

  • Frequent Contributor
  • **
  • Posts: 281
  • Country: de
    • CADT Homepage
Re: Statistical arrays
« Reply #59 on: June 25, 2019, 09:30:32 am »
I was looking at Nomca arrays for the 13:1 divider for a LTZ1000 reference and found similar results as above when using 3x 5K in series with 4x 5K in parallel. Division ratio TC was less than 1 ppm/°C for two units and 2.4 ppm/°C for one unit.

Now, when i look at the measurements presented above by antintedo for the TCs of single resistors in a Nomca array, i'd like to show where one may get by selection. This is only an example, but my conclusion would be that a Nomca array can be used for the 7V to 10V gain stage, if one finds a practical solution for free combinations. As soon as i have time i'll check this with my 5K Nomca arrays.

Regards, Dieter

PS: Sorry for the confusion, but those measurements were not from Magnificent Bastard.
« Last Edit: June 26, 2019, 05:01:13 pm by dietert1 »
 

Online Kleinstein

  • Super Contributor
  • ***
  • Posts: 6324
  • Country: de
Re: Statistical arrays
« Reply #60 on: June 25, 2019, 10:02:18 am »
Measuring and selecting TC is a slow process. There is an additional difficulty - soldering and board stress can be part of the TC. So it would need TC measurement after soldering and than adapting the layout to the appropriate connections.  :-//
So it is more like a very limited number of option one could chose from by adding bridges, like leaving out one resistor.

Getting 2.4 ppm/K for the divider is quite a bit. The 3 in series and 4 in parallel version is already averaging over 3 and 4 resistors so that chances are low to get such a poor combination with individual resistors at 5 ppm/K spread.

Just for a test one could try re-soldering the bad divider.

The other point is that often the TC is only the easier part. Long term drift may be of similar importance, with little chance to measure and select upfront.
 

Offline TiN

  • Super Contributor
  • ***
  • Posts: 4046
  • Country: tw
  • xDevs.com/live - 24/7 lab feed
    • xDevs.com
Re: Statistical arrays
« Reply #61 on: June 25, 2019, 10:15:43 am »
Perhaps a socket could be used for testing, to avoid heat stress to the package?
Like this one?
YouTube | Chat room | Live-cam | Have documentation to share? Upload here! No size limit, firmware dumps, photos.
 

Offline dietert1

  • Frequent Contributor
  • **
  • Posts: 281
  • Country: de
    • CADT Homepage
Re: Statistical arrays
« Reply #62 on: June 26, 2019, 05:22:34 pm »
Measurement of the TCs of the eight resistors in the array requires a 8x multiplexer and one temperature cycle. Yes, maybe the resistor array can't be soldered after the measurement without affecting it's behaviour. That needs to be decided by measurement.
Anyway, results better than 1 ppm/°C without selection may be good enough to try and control the temperature of the resistor array in a way similar to the heated zener reference.

Regards, Dieter
 

Online Kleinstein

  • Super Contributor
  • ***
  • Posts: 6324
  • Country: de
Re: Statistical arrays
« Reply #63 on: June 26, 2019, 07:53:42 pm »
One normally would not even need 1ppm/K performance in the LTZ1000. Without adjustment (R9) the TC of the LTZ1000 circuit is more like in the 0.2-0.5 ppm/K range which corresponds to some 20-50 ppm/K for the resistors. The individual R9 trimming would also include resistor TC.
However this is only the TC part - it does not help with long term drift.

I don't know how much the soldering effects the TC. Besides TC soldering can effect the short time drift and possibly hysteresis effects. These could be confused with a TC error in an early test. Still a significant drift / error in the TC test is not a good sign, even of not due to actual TC but hysteresis or aging.

I don't think temperature control is helping, as the TC is not the real critical parameter - it's more like aging that is critical and the TC is more used as an indication of stability, with the real thing very difficult to measure. Some burn in after soldering could help - though it would also hide things. Ideally one would measure before burn in and get suspicious if burn in changes too much.
 

Online imo

  • Super Contributor
  • ***
  • Posts: 1933
  • Country: 00
Re: Statistical arrays
« Reply #64 on: June 26, 2019, 08:11:40 pm »
While looking at the 34470A teardown, at 18:40-22:45 for example, they are using rather common resistor parts around the LTZ1000, imho.
« Last Edit: June 26, 2019, 08:28:14 pm by imo »
 

Online Kleinstein

  • Super Contributor
  • ***
  • Posts: 6324
  • Country: de
Re: Statistical arrays
« Reply #65 on: June 26, 2019, 08:27:06 pm »
The demands on the resistors around the LTZ1000 are not that high. It is only of one wants really low long time drift (e.g. < 2 ppm/a) that is really takes really stable resistors (still only some 50 ppm/a). AFAIK the 34470 is not the highest stability - more like having the LTZ1000 for low TC and low noise.

The more critical resistors are in a 7 to 10 V stage. Another point could be a gain stage or divider and possibly a resistor in an ADC, that can effect linearity due to self-heating.  Selft heating could increase the temperature of a resistor by something like 0.1-1 K. So an TC of 10 ppm/K could cause an INL error of some 1-10 ppm.
 

Online Andreas

  • Super Contributor
  • ***
  • Posts: 2410
  • Country: de
Re: Statistical arrays
« Reply #66 on: June 26, 2019, 08:29:43 pm »
While looking at the 34470A , at 37:45 for example, they are using rather common resistor parts around the LTZ1000, imho.

Ok as a volt-nut you have to count the vishay VSMP series (around $15 per pop) as rather common.

the TC of the LTZ1000 circuit is more like in the 0.2-0.5 ppm/K range

from where do you have such large values?
my results with long legs on the non-A version are below 0.06 ppm/K

https://www.eevblog.com/forum/metrology/ultra-precision-reference-ltz1000/msg2301414/#msg2301414

with best regards

Andreas
 
The following users thanked this post: imo

Offline branadic

  • Super Contributor
  • ***
  • Posts: 1486
  • Country: de
Re: Statistical arrays
« Reply #67 on: June 26, 2019, 08:41:38 pm »
Quote
from where do you have such large values?
my results with long legs on the non-A version are below 0.06 ppm/K

This are numbers I can confirm with 400k resistors unpopulated, short legs.

-branadic-
Fluke 8050A | Prema 5000 | Prema 5017 SC | Advantest R6581D | GenRad 1434-G | Datron 4000A | Tek 2465A | VNWA2.x with TCXO upgrade and access to: Keysight 3458A, Keithley 2002, Prema 5017 SC, 34401A, 34410A, Keithley 2182A, HDO6054, Keysight 53230A and other goodies at work
 

Online Andreas

  • Super Contributor
  • ***
  • Posts: 2410
  • Country: de
Re: Statistical arrays
« Reply #68 on: June 26, 2019, 09:13:18 pm »
Quote
from where do you have such large values?
my results with long legs on the non-A version are below 0.06 ppm/K

This are numbers I can confirm with 400k resistors unpopulated, short legs.

-branadic-

ok this confirms my opinion that it is not a good idea to shorten the legs on the non-A Version.
(with half shortened legs I had also a larger T.C. than before).

with best regards

Andreas
 

Offline dietert1

  • Frequent Contributor
  • **
  • Posts: 281
  • Country: de
    • CADT Homepage
Re: Statistical arrays
« Reply #69 on: July 01, 2019, 02:38:18 pm »
In the meantime i have done some multiplexer measurements on three more Nomca 1603 8x 5K arrays. They got mounted on SMD adapters with gold plated pins and cleaned in ultrasonic bath. Then i setup 4 wire R measurements using a HP3456A at 0,1 mA measurement current with averaging to get below ppm resolution. Temperature span was 18 °C. Error on a single resistor TC measurement is about 0,03 ppm/°C.

Results on TCs for each part (average( 8 ), stdev( 8 ), span( 8 )):

D     4,34 +/- 0,62 ppm/°C, max-min = 1,85 ppm/°C
E   11,23 +/- 0,63 ppm/°C, max-min = 1,74 ppm/°C
F   16,20 +/- 1,55 ppm/°C, max-min = 4,51 ppm/°C

According to my measurements those are pretty serious parts, well suited to produce temperature compensated voltage dividers. And it makes complete sense to measure TCs of individual resistors in order to determine a good combination - better than statistics! Combining  resistors properly each of my three parts should give me a 13:1 divider with less than 0,03 ppm/°C on the division ratio. At least that's the result of a simulation using the measured curves of the individual resistors.

Regards, Dieter

« Last Edit: July 01, 2019, 02:41:29 pm by dietert1 »
 
The following users thanked this post: branadic, antintedo

Offline branadic

  • Super Contributor
  • ***
  • Posts: 1486
  • Country: de
Re: Statistical arrays
« Reply #70 on: July 01, 2019, 02:50:52 pm »
Thanks Dieter,

since I'm still waiting for my boards to arrive, because of the inability of DHL or customs the boards were sent back to Elecrow, I ordered a zero force adapter

eBay auction: #282277360455

that received today to perform similar measurements. Need to perpare a cable for multiplexer of my Prema 5017SC.

-branadic-
Fluke 8050A | Prema 5000 | Prema 5017 SC | Advantest R6581D | GenRad 1434-G | Datron 4000A | Tek 2465A | VNWA2.x with TCXO upgrade and access to: Keysight 3458A, Keithley 2002, Prema 5017 SC, 34401A, 34410A, Keithley 2182A, HDO6054, Keysight 53230A and other goodies at work
 

Offline dietert1

  • Frequent Contributor
  • **
  • Posts: 281
  • Country: de
    • CADT Homepage
Re: Statistical arrays
« Reply #71 on: July 02, 2019, 10:19:48 pm »
Today i got more evidence for the concept of combining selected resistors within a Nomca 1603 array.

The part B that i mentioned above giving 2.4 ppm/°C showed less than 0.5 ppm change during a 18 °C temperature cycle now after rewiring. It's an upper limit because from the multiplexer measurements i expected a residual of -0,015 ppm/°C on the division ratio and i couldn't see anything with our HP 3456A. That part B is the only one in six that exceeds the 5 ppm/°C tracking spec in the datasheet with a span of 6.7 ppm/°C between measured min and max TCs.

Regards, Dieter
 

Online Kleinstein

  • Super Contributor
  • ***
  • Posts: 6324
  • Country: de
Re: Statistical arrays
« Reply #72 on: July 03, 2019, 06:36:53 am »
Whether selecting the resistors is only practical if soldering / mounting on a board does not change the TC significantly. For PCB use, there are usually also not that many option to choose from. There may be a few jumpers for options but usually only a few, especially if not many of the resistors are unused.

If soldering has only a limited effect it may at least be possible to test the resistors (if the contacts in an affordable socket are good enough) and than decide on the orientation of the chip. This may avoid a worst case (with a suitable use of the resistors) - which could already be quite a gain.
Selecting a suitable resistor combination is much easier with individual resistors - more like hard with an array.
 

Offline dietert1

  • Frequent Contributor
  • **
  • Posts: 281
  • Country: de
    • CADT Homepage
Re: Statistical arrays
« Reply #73 on: July 03, 2019, 08:23:49 am »
You are not asking, but i will tell you how to do it.

The construction looks like the thin film arrays of the HP 3456A. The Nomca arrays were put onto SMD to DIL Adapters with gold plated pins. On the carrier board there are gold plated socket strips. Wiring is underneath the carrier board with short soldered bridges, that i prefer over the plugged configuration bridges of the HP3456A. This construction isn't ideal in the sense that is has short copper tracks and wires and the pins/sockets in the path that are not tracking well during fast temperature changes, but the arrays are not soldered anymore after characterization.

I used a similar construction before with S102 resistors soldered on top of 4 pin modules so that one can use 4 wire measurements for characterization and not solder them afterwards. In that case the doubled pins give some redundance when the parts are on the carrier board afterwards. I also thought about making an array of S102 resistors and see how they compare. Individual TCs will be smaller, but temperature tracking will be worse, so finally the result may be similar.

Regards, Dieter

PS: One may even think about using TSSOP32 adapters for the Nomca arrays in order to have 4 wire connections.
« Last Edit: July 03, 2019, 08:31:58 am by dietert1 »
 

Offline Edwin G. Pettis

  • Frequent Contributor
  • **
  • Posts: 413
  • Country: us
  • The plural of anecdote is not data.
Re: Statistical arrays
« Reply #74 on: July 03, 2019, 06:40:29 pm »
The NOMCA arrays are not particularly good for certain applications:

!. Ta2N resistors are rather noisy and has 1/f noise as well, all above the theoretical resistor noise level.
2. While short term stability is not bad, long term stability is questionable especially at low PPM.
3. Shelf stability at 25°C is only fair, tolerance drift at up to .01%, ratio drift up to .002% for the first year, Vishay states no specification beyond 1 year.  If these exhibited good long term stability you can be sure Vishay would have put that spec in there.
4. Ratio matching of 5 PPM/°C is hardly impressive, other types can be below 1 PPM/°C including wire wound.
5. All factors affecting the specifications more or less apply to any resistor type being used.  This requires careful design to manage thermals and unequal heating even on a single substrate.
6. Sensitivity to soldering, per the data sheet, possible significant (and permanent) changes can occur, not only to tolerance but TCR and stability.

In the case of the current discussion on a 7V to 10V boost circuit, any deviation of resistor matching will be multiplied by the amplifier gain, this includes noise and drift of resistor specs over time.  While a NOMCA array may have an attractive price, its specifications for this particular application leaves a lot to be desired.  If you are expecting a long term stability and low noise 10V output over months, these are not the resistors to pick.  The Vishay data sheet is pretty clear on the specifications they do give, NOMCAs are not for this application.

S102 resistors also stink for this application.
 


Share me

Digg  Facebook  SlashDot  Delicious  Technorati  Twitter  Google  Yahoo
Smf