How good or bad are TL431/432 type references?

[splin]

I need a cheap reference for an outdoor sensor which will be subject to temperature range of -30C to 50C. I'd like to achieve .1%, after initial calibration, over time (10 years) and temperature but that may not be possible with my budget, preferably < $0.50

TL431/432s are too cheap to ignore; the temperature coefficients aren't particularly good but they are cheap enough to allow a thermistor to be used to compensate (in software). The actual shape of the voltage against temperature curve is going to vary from device to device, so does anyone have experience as to how much improvement, if any, could be achieved without having to calibrate every device against temperature?

Power isn't a problem so perhaps it would be better to provide a crude oven by placing a pair of heater resistors either side of the reference. It may be possible to keep the reference within a 10C range at the cost of a transistor and a pair of resistors. The microcontroller will probably include a temperature sensor so a thermistor might not even be necessary.

More importantly TL431s don't have any long term drift specifications so does anyone have any data or experience to share?

Also there is a bewildering array of devices from different manufacturers including TI, Diodes Incorporated, NXP etc. at a variety of different prices. Does anyone have recommendations on which are better or should be avoided?

[danadak]

There is a method that can bring high precision to low precision engineers.

See attached.


Regards, Dana.

[splin]

There is a method that can bring high precision to low precision engineers.

See attached.


Regards, Dana.

Yes, I know you can do that but as I said:

how much improvement, if any, could be achieved without having to calibrate every device against temperature?

Calibrating every unit over temperature is a slow and relatively expensive process - there is no point saing a dollar on a voltage reference then having to spend more than that in calibration. Neither does it address the issue of long term drift given that re-calibration after installation is not an option.

[T3sl4co1l]

TL431's fine for 1%-grade work, but if you need it tighter, and the temp range, go with something like LM4041 or better (REFxxx?).

A better question: why do you need a reference at all?  What are you sensing?  Why can't it be done with a ratiometric measurement?

Tim

[danadak]

Calibrating every unit over temperature is a slow and relatively expensive process - there is no point saing a dollar on a voltage reference then having to spend more than that in calibration. Neither does it address the issue of long term drift given that re-calibration after installation is not an option.

Keep in mind that approach is run by the DUT, it cals itself at production
test. But it does take test time, especially time to sweep temp with a heater.
But if sensor has a fixed response curve shape one can do a step heater event and
use a two point measurement to gen the correction f().

.1% not a problem with this general approach. I am thinking a 10 - 30 sec test time,
if the device curve is a smooth repeatable curve. eg. it is log for example.

Then there is two current T measurement of a diode driven by a current source.

http://www.eetimes.com/document.asp?doc_id=1279718

http://cds.linear.com/docs/en/application-note/an137f.pdf


Regards, Dana.

[MagicSmoker]

I need a cheap reference for an outdoor sensor which will be subject to temperature range of -30C to 50C. I'd like to achieve .1%, after initial calibration, over time (10 years) and temperature but that may not be possible with my budget, preferably < $0.50

Analog Devices ADR504x (where x encodes the voltage, but not literally) and STMicro TS4061 (1.225V) are two examples of 0.1% accuracy voltage references which cost around $0.50US... at least if you buy them by the reel, anyway.

I would also suggest a ratiometric type of measurement technique (e.g. - Wheatstone bridge) to sidestep the need for an accurate reference in the first place, but if you are digitizing the signal from the sensor with an ADC (implied by the mention of "microcontroller") you are right back to needing a stable/accurate reference... for the ADC.

[Alex Nikitin]

I suppose in this case the initial tolerance is not that important as it can be calibrated out, but the temperature range and coefficient, as well as the long term stability and hysteresis are important. I would suggest the LM4030 series reference, which is characterized fully in these respects and the C grade is not that expensive  at $0.45 for 1K+ quantity (and with 30ppm/C max tempco would meet +/-0.1% requirement in +/-33C range from the calibration temperature).

Cheers

Alex

[T3sl4co1l]

I would also suggest a ratiometric type of measurement technique (e.g. - Wheatstone bridge) to sidestep the need for an accurate reference in the first place, but if you are digitizing the signal from the sensor with an ADC (implied by the mention of "microcontroller") you are right back to needing a stable/accurate reference... for the ADC.

An ADC is naught but an inverse potentiometer.  Set ADCREF == device_REF and you're done.  REF can be any arbitrary voltage that's sufficient for operation, usually VCC.

Tim

[Cerebus]

I would also suggest a ratiometric type of measurement technique (e.g. - Wheatstone bridge) to sidestep the need for an accurate reference in the first place, but if you are digitizing the signal from the sensor with an ADC (implied by the mention of "microcontroller") you are right back to needing a stable/accurate reference... for the ADC.

An ADC is naught but an inverse potentiometer.  Set ADCREF == device_REF and you're done.  REF can be any arbitrary voltage that's sufficient for operation, usually VCC.

Tim

Yup, ratiometric measurements are often a very cheap way of gaining a lot of precision for no money at all. If you look at the specs for high-end DVMs/DMMs that offer ratio measurements, the ratio measurement precision is often close to an order of magnitude better than the standard measurement precision.

One word of caution though, watch out for ADCs (sigma-delta and anything that says 'current balancing') injecting a series of current pulses into the signal being measured and the reference input. If altering your (nominal) VREF with noise like this will affect your measurement, buffer the VREF to the ADC (and account for any drift/gain/offset errors in the buffer) or pick an ADC without this problem.

[MagicSmoker]

An ADC is naught but an inverse potentiometer.  Set ADCREF == device_REF and you're done.  REF can be any arbitrary voltage that's sufficient for operation, usually VCC.

My understanding is that variations in Vref to an ADC can result in some strange nonlinearity errors, the severity (or lack thereof) depending on the specific ADC architecture (for example, limit cycle oscillations from the reference voltage straddling two bins). This isn't stuff one normally worries about... unless specs like "0.1% accuracy for 10 years" get thrown about.

[T3sl4co1l]

Yeah, excepting pathological cases of course.  Try to avoid those entirely.

Tim

[splin]

Thanks for all the responses so far. The application is for an ADC measuring current and thus an absolute reference is required.

The TL431 generally has a maximum deviation spec of 16mV (0-70C), 34mV (-40 to 85C) and costs around 5 cents (10K). 34mV is 1.36% which is pretty bad. The AT4341 improves that to 16mV, however the Diodes Inc. AP431i and AS431 variants (among others) are even better at 6mV (0-70C) and 10mv (-40 to 85C) which is 32ppm max. Better still it's even cheaper at only 3.7 cents!

10mV is 4000ppm or .4% so would still need to be compensated to reach the .1% target. If I can keep the reference above 0C with a couple of heater resistors then the 6mV spec is only .24% which isn't far off. As I said, 3.7 cents is very hard to ignore given that this is a cost sensitive application.

TL431's fine for 1%-grade work, but if you need it tighter, and the temp range, go with something like LM4041 or better (REFxxx?).

A better question: why do you need a reference at all?  What are you sensing?  Why can't it be done with a ratiometric measurement?

Tim

The LM4041C at 100ppm x 125C = 12,500ppm or 1.2% is very similar to the TL431 and costs around 15 cents and much worse than the AS431. I'm not sure why you suggested this one? It does specify typical long term stability though of 120ppm/1K hours which isn't great - 10 years drift could be 1230ppm assuming 120pppm * SQRT(10 * 24 * 365 / 1000).

The cheapest REFxxx which is better than 30ppm is the 20ppm max REF31xx but is relatively expensive at $.99. It might just meet .1%  from -30C to 50C but the datasheet only specifies 0-70C and -40 to +125C though it does have a long term drift spec of 70ppm/1k hours.

Analog Devices ADR504x (where x encodes the voltage, but not literally) and STMicro TS4061 (1.225V) are two examples of 0.1% accuracy voltage references which cost around $0.50US... at least if you buy them by the reel, anyway.

The ADR504xB is 75ppm so nowhere near as good as the AS431. The TS4061 is 35ppm so again not as good at more than 10x the cost.

I suppose in this case the initial tolerance is not that important as it can be calibrated out, but the temperature range and coefficient, as well as the long term stability and hysteresis are important. I would suggest the LM4030 series reference, which is characterized fully in these respects and the C grade is not that expensive  at $0.45 for 1K+ quantity

Like the other suggestions 30ppm is barely any better than the 32ppm AS431 LM4030C.

(and with 30ppm/C max tempco would meet +/-0.1% requirement in +/-33C range from the calibration temperature).

Cheers

Alex

But there is no guarantee of that. The dV/dT spec is based on the box method of max deviation/(125C - (-40)C). If you look at Figure 1, voltage V temperature, you can see that dV/dT between 0 and 90C may well be worse than 30ppm - possibly 55ppm if you use 30ppm x 165/90. Of course the actual shape of the voltage/temperature curves of devices could vary significantly more than the 5 typical units shown so it's impossible to be certain. The fact that they do show a better 10ppm 0 - 85C spec for the expensive A grade than 20ppm for the -40 to +125C range doesn't mean you can assume that the worst case C grade tempco is going to be better than 30ppm over 0 to 85C.

[RANT]
A typical example of a datasheet designed to be deceptive in my opinion - lots of data but one of the most important specs, temperature coefficient, is described as dV/dT but what they really mean is a maximum deviation of 12.375mV (for the 2.5V version) over 165C. That could be linear from -40 to +125C or it could all occur between 0C and 40C and still be in spec. If you are designing for worst cases then the 12.375mV is the only figure they are giving you. Why pay for $.45 for this when the AS431 is only $.037? And for applications where the temperature range is 0 - 70C the AS431 is way better spec'd with a guaranteed 6mV max deviation. (Other factors obviously apply in the selection process).

Linear get this right in their LT1009 spec which includes upper and lower voltage limits on the voltage/temperature graph as well as typical.

Even worse is the REF20xx which features a Vref/2 output as well as Vref. The temperature coefficient is spec'd at 8ppm max but over -40 to 125C - most likely because they would have to declare a worse figure for the industrial temperature range of -45C to 85C. You could argue that they don't want to bear the cost of extra testing, except that they provide specs for both temperature ranges for Vref to Vref/2 tracking so they almost certainly have the relevant data. I'm convinced the reason is they want to protect that 8ppm max tempco marketing headline for probably the most important selection criteria for references. Great for automobile equipment designers but not very helpful for just about everyone else. If your application is 0-70C then I guess you have to class this as a 19ppm part - unless you have sufficient clout with Ti to get more relevant guarantees.
[/RANT]

Can you tell that I'm getting more and more disillusioned by the quality of datasheets these days? Apparently full of information but practically all of it 'typical', often containing errors, large wodges of cut and paste from other datasheets which may be irrelevant/inapplicable and data which is contradictory or just unbelievable. In the latter category are the 20ppm/1khrs long term stability claims for plastic packaged parts including the LT004 and LT1009 - probably applicable to long obsolete hermetic metal packages but likely to be seriously misleading for current parts.

[MagicSmoker]

It's clear you've given your problem more thought than the average complainer thread starter around here. Unfortunately, if you want guaranteed stability in a voltage reference over 10+ years and a wide range of temperature then you best be prepared to pay more than 4 cents for it.

It's also hard for me to imagine a product that would have such demanding requirements for accuracy over time and temperature and not be worth a premium - to use a semi-lateral analogy, this is like putting a GPSDO in a bedside alarm clock...

[splin]

It's clear you've given your problem more thought than the average complainer thread starter around here.

Actually I hadn't done a great of research on suitable references beyond looking through the price sorted list of references on DIGIKEY. After going through many, many pages of TL431 variants I thought I'd ask for opinions. Most of my last response comes from checking out the suggestions in detail and why it took a while to respond.

Unfortunately, if you want guaranteed stability in a voltage reference over 10+ years and a wide range of temperature then you best be prepared to pay more than 4 cents for it.

Hehe.. you are probably right - but given that the 4 cent part doesn't see to be bettered, tempco wise, by most references below $1 I'm not sure. Especially since it appears to match (on paper) suggestions made here by some experienced posters.

It's also hard for me to imagine a product that would have such demanding requirements for accuracy over time and temperature and not be worth a premium - to use a semi-lateral analogy, this is like putting a GPSDO in a bedside alarm clock...

I'm not convinced that .1% over ten years is especially challenging - I'd bet there are lots of cheap multimeters that have bettered that using dirt cheap references including zener diodes. I could be wrong of course. Hardly of the same order as an atomic clock compared to a quartz movement.

The tempco clearly is an issue. Of course it would be easy to throw a $5 reference at the problem but that's not engineering. I originally expected to have to pay at least $1 assuming TL431s were rubbish but the AS431 specs surprised me and got me wondering if it could be used by helping it out a bit by spending a bit of the savings on compensation or 'ovenising'.

Another option is to use several references in parallel 4 or even 8 as they are so cheap, to statistically reduce the variances in drift and tempcos. By causing the combined temperature characteristic to approach the mean, temperature compensation would be more effective.

I'm not sure just well this will work without selecting each part from different batches, types or even manufacturers, as within any particular batch they are likely to all have very similar characteristics so the statistical improvement would be small. I'd expect that worst case long term drift behaviour would benefit more, as the ageing processes are more subject to random events. Anyone explored this route?

It would also reduce noise which is unspecified for the AS431 - and probably not very good given its low power.

[splin]

Calibrating every unit over temperature is a slow and relatively expensive process - there is no point saing a dollar on a voltage reference then having to spend more than that in calibration. Neither does it address the issue of long term drift given that re-calibration after installation is not an option.

Keep in mind that approach is run by the DUT, it cals itself at production
test. But it does take test time, especially time to sweep temp with a heater.
But if sensor has a fixed response curve shape one can do a step heater event and
use a two point measurement to gen the correction f().

.1% not a problem with this general approach. I am thinking a 10 - 30 sec test time,
if the device curve is a smooth repeatable curve. eg. it is log for example.

Then there is two current T measurement of a diode driven by a current source.

http://www.eetimes.com/document.asp?doc_id=1279718

http://cds.linear.com/docs/en/application-note/an137f.pdf


Regards, Dana.

Dana, I had considered this approach and it may be required anyway for other steps in the signal chain, though I'd prefer to avoid it if possible because of the extra work in setting up and characterising test rigs and procedures, documentation etc. It is also an opportunity for additional calibration errors to arise.

Heating/cooling the whole circuit board would have the major advantage of compensating for most sources of temperature dependent errors, but I had considered that blowing temperature controlled hot and cold air over the voltage reference alone (and possibly also the on-board temperature sensor used for compensation) could be a lot quicker and easier to arrange and the thermal response time of the reference might allow the calibration to be achieved in 2 or 3 seconds. I expect that three temperatures, including room temperature, would be sufficient.

I'm not sure of the relevance of the diode temperature measurment references you gave - a thermistor would be a lot simpler and cheaper (a diode would need an opamp given the low signal levels) given the modest accuracy requirements needed.

[tggzzz]

Another option is to use several references in parallel 4 or even 8 as they are so cheap, to statistically reduce the variances in drift and tempcos. By causing the combined temperature characteristic to approach the mean, temperature compensation would be more effective.

The presumption is that the drifts and tempcos are uncorrelated. That is questionable, doubly so for devices from the same batch.

Besides that, would using 8 devices really make any difference to your end system?

[Wolfgang]

... why dont you try an LM723. It has better tempco and also long-term drift specs. And it is dirt cheap.
Just as a not completely serious project of how stable it could be ...

https://electronicprojectsforfun.wordpress.com/silly-circuits/silly-circuits-a-heated-lm723-reference/

[dardosordi]

Good ol LM723 still kicking ass

[splin]

... why dont you try an LM723. It has better tempco and also long-term drift specs. And it is dirt cheap.
Just as a not completely serious project of how stable it could be ...

https://electronicprojectsforfun.wordpress.com/silly-circuits/silly-circuits-a-heated-lm723-reference/

Well apart from costing 7 times as much as the AP431S, having almost twice the typical 0 to 70C drift (2100ppm v 1200ppm), having four times the worst case 0 to 70C drift (10400ppm v 2400ppm) and the long term drift being horrendous at 500ppm/1k hours (not 50ppm as you state in your linked project, and not ppm/sqrt(hours) as usually specified for references)? 

To be fair, the Ap431S doesn't specify long term drift so yes, although it could be as bad or worse, I'd be amazed if it was. Most references are specified at 120ppm/sqrt(1k hours) or less.

It would be interesting to hear the experience of TL431 type reference's long term drift - surely someone here must have used them?

[EDIT] I just noticed that you quoted the correct 500ppm/1kh in the long term results section. It will be interesting to follow your results to see if the drift does in fact decrease over time. The spec may be referring to long term drift over the first few thousand hours or so rather than the implied linear drift over time - which would be suprising.

[Wolfgang]

Hi,

you found out that I filed it under "silly projects". Of course I knew that it was no premium reference, but for about 10 to 15 €Cents @100pcs its not a bad one either.

To the topic of long term drift:

When you look at all heated references you see a stong increase with all that have not been pre-aged. The same here, so no surprise.
I am curious how it behaves a) after a year and b) with lower chip temps, say 60 °C instead of ca. 85 °C.

I'll update my webpage when I have new data.