passive component tolerances for op amps used in front end of 18-bit ADC

Another option (which might be cheaper, although not simple) would be to use an explicit temperature sensor to tell your system how much error it should compensate for. Add some calibration, with maybe polynomial interpolation. You wouldn't need much accuracy in sensing the temperature. I don't know, I've never tried it, but it could work.

That's a good idea, Bill! It would be possible to temperature-cycle the PCB in an environmental testing chamber and add calibration for known errors.

Obviously, I need to have some sort of "reference signal" against which the system is calibrated. This could be provided by a function generator and defined at a certain temperature. The trick is finding an appropriate reference signal. Obviously, calibration would be device-specific.

The Vishay custom-made resistor networks look very interesting:

I would wonder how expensive it would be to create such a resistor network with a TCR of

Susumu makes 15 PPM resistors, available from Digikey for around 35 cents each. We tested 200 ohm and 390K parts, saw -8 and +5 PPM respectively. You can buy better resistors, well below 1 PPM, but they're expensive.

You can get affordable 1 PPM/K voltage refs, too.

If you are willing to temperature cycle, you can measure the temp and do software calibrations, but that's expensive too.

You can also occasionally mux in a known zero and a known 1 PPM-class reference, and software calibrate continuously, fixing all the errors except linearity. Then you can use junk 25 PPM resistors.

Some delta-sigma ADCs have a wide-range PGA on the front end that's better than anything you can make. They have a near-zero drift relative to the voltage reference. But slow.

Too bad people don't make SAR ADCs with that quality of PGA built in. I could sure use some about now.

John

Thank you so much for your response, John!

It's neat to realize that the part performs better than the specifications!

Some esoteric resistors seem very much suited for military or other hi-rel applications. Definitely not for mere mortals, and extremely difficult to source.

That's a great idea, John! So suppose that the 1 PPM-class voltage reference is 2.5V. Since this is a known value, measuring the difference between 2.5V and the voltage at a given temperature would give the change in voltage caused by the change in temperature. But how would I relate this to changes in resistance? I would suppose that the change in voltage would be related to a change in the voltage gain of the op amps comprising the analog front-end.

Would I have to switch the mux at every sample collected by the ADC, or could I simply calibrate at the beginning of each measurement? I would say that the calibration could be done at the beginning of every measurement.

Have you seen the AD7262 from Analog Devices? It only has a resolution of 12-bits, but perhaps it might be interesting?

Where can you get resistors that are better than 1ppm? While I've worked with resistors that are specified as better than 1 ppm/°C, about the best stability I've seen are resistors like the Tinsley reference resistors that are specified as 2 ppm/year

Bret Cannon

I believe that Digikey has these resistors:

Example part number: Y1624-1KCT-ND

Ah, I think that I've figured it out, John. It's simply another calibration. The voltage difference between actual and calculated reference is then subtracted from each sample taken by the ADC.

Apparently the temperature coefficient is listed as ±0.2ppm/°C for this particular part number.

You could mux in zero volts, digitize, then select the good vref and digitize. From that, calculate a zero offset and a gain cal factor, and apply them to actual input measurements.

You'd need to run a cal sequence often enough to catch any temperature drift errors. Every minute or so would be OK in a benign environment, where radical temperature transients aren't expected. More often if possible.

That's very nice. I want it in 16 bits!

John

You get less roundoff error by multiplying by zero and adding what it ought to look like.

Cheers

Phil Hobbs

Yes, that makes sense. Thanks, Phil!

Sounds good, John.

In my experimental application (environmental ground-based remote sensing), measurements are often taken on the command of a datalogger, which sends an instruction to the measurement device across an SDI-12, serial or CAN-bus. So as long as the calibration occurs immediately before taking a measurement or within less than a minute, it should be okay.

It would be great to find such a version of this part in 16-bits!

Thanks, John! This ought to get me going with respect to the analog front-end design.

Phil was teasing. He does that sometimes.

John

Isn't that how the IPCC climate models work? ;-)

Cheers! Rich

Yes, they all give different numbers. It just shows how complex environmental physics really is.

I think that I read too much into this. Hmm... "adding what it ought to look like"

;-)

this=20

Crickey, what is it made of?

--=20 Transmitted with recycled bits. Damnly my frank, I don't give a dear

----------