Precision gain matching of two instrumentation amplifiers

[ricko_uk]

Hi,
what is the best way to match the gain of two instrumentation amplifiers (clearly the same part type/number)?

The obvious one that comes to mind is to select two high precision resistors for the gain resistors. But are there other techniques?

To be clear the aim is to achieve the following ideal case (in practice as close to it as possible):
- feed the same signal coming from the same signal source to both INA and the outputs are exactly the same, to the point that if you invert one of them the output is a flat 0V

Thank you

[GerritMax]

Hope this might help https://easyeda.com/GerritMax/Transistor-Matching-Circuits-by-Ian-Fritz

And also watch this video as it shows you how to calibrate it.

[robzy]

(in practice as close to it as possible)

How close do you actually mean?

There are many causes of mis-match here, and its possible to spend $1000 on a solution and still not be close enough (for some people).

You need to define, in some way, what amount of error is allowed.

[ricko_uk]

Thank you both,

Robzy,
I don't have an exact figure figure because it is for various experiments I have in mind. Ideally at least within 0.1mV on the output after amplifying by at least 500 to 1,000.

I want to get it as close as I can get with general proven best practices that do make a difference if not implemented. I am looking for solutions, suggestions, tips that can improve it over just having high precision matched resistors and equal INAs.

Thank you

[magic]

Integrated in-amps use one external resistor for gain - make it adjustable by 1% or so and

[pwlps]

The obvious one that comes to mind is to select two high precision resistors for the gain resistors. But are there other techniques?

I'm not sure it's worth in your case but since you're asking for ideas here is one:
At low frequency it is possible to use a single in-amp,  chopping both input and output with switches (something similar to chopper-stabilized opamps).  The switching pattern might add a suitable delay on the output to let it settle and avoid transients at the chopping frequency.

[ricko_uk]

Thank you Magic and Pwlps,

Magic,
good and simple idea! But don't trimmers and pots introduces lots of noise?

Pwlps,
interesting one!

[David Hess]

I don't have an exact figure figure because it is for various experiments I have in mind. Ideally at least within 0.1mV on the output after amplifying by at least 500 to 1,000.

At the level of 0.1 millivolts, gain is not the only significant source of error; instrumentation amplifiers also have input offset or output offset or both.  That means that offset is first trimmed and then gain is trimmed, and in some designs input and output offset are trimmed separately.

Trimmer potentiometers can be used.  Offset can easily be trimmed electronically using DACs.

[bson]

good and simple idea! But don't trimmers and pots introduces lots of noise?

Not so much noise, the problem is the surfaces collect crap and may oxidize.  The result is they will need readjustment every so often.

It's better to determine what error is acceptable and design for it.  But that also depends on the gain.  Tempcos for example will cancel out in a divider, Rg||Rf = (tc*Rg)/(rc*Rg+tc*Rf) = (tc*Rg)/(tc*(Rg+Rf)) = tc/tc*Rg/(Rg+Rf) = Rg/(Rg+Rf).  In reality they reduce to the residual variance in TC, but as long as the resistors have similar material properties that's a very small quantity.

Actual values can be binned for.  Measure 100 of each resistor and then write a simple program to find a good set of pairs that minimize the gain disparity.  Have it calculate what that is, see if it's good enough.  You may also need to measure the amps themselves and take them into account.

[ricko_uk]

Thank you Bson and David

[David Hess]

When trimmers are used, then a series and parallel combination of resistors is included to minimize the required trimming range so that inaccuracy of the trimmer itself has a minimum effect on overall accuracy.

Some applications cannot tolerate trimmers and must use other methods which include electronic trimming like with DACs or custom installed series and parallel fixed resistors.

[coppercone2]

ina of single resistor gain set will have a gain accuracy specification that will hamper your efforts to match them initially with precision resistors , the part has its own transfer function multiplied by the resistance.

[bdunham7]

Are you working with DC and if not, what bandwidth?

[ricko_uk]

In some cases also with DC but 99.9% of the time just AC signals.
I am ok to design it just for AC if it relaxes the constraints and/or makes it easier to match the two INA outputs to a higher degree.

In terms of frequency, up to at least 100KHz, preferably 250KHz.

[exmadscientist]

You're going to have a tough time getting gain of 1000 at 100kHz from any instrumentation amplifier, much less a DC-accurate one. And it's all but impossible to match instrumentation amplifier errors well enough to even generate a good differential output, much less drive it down a cable.

If you're just building one, then trimming may be able to get you to where you want to go. But if this is for anything that will be produced in quantity, the trim quality required is probably going to be infeasible.

I'm also wondering if this is a bit of an XY problem. Providing a bit more information about the overall end goals might help you get better tailored answers.

[David Hess]

ina of single resistor gain set will have a gain accuracy specification that will hamper your efforts to match them initially with precision resistors , the part has its own transfer function multiplied by the resistance.

Many modern instrumentation amplifiers which use a single external resistor to set their gain have a very good temperature coefficient of gain.

[coppercone2]

I don't mean tempco, I mean gain error. I am not sure if parts that have super tempco stability and excellent gain linearity have good initial gain error. It means you need to program them differently initially. I.e. LT1167 has a gain error of 0.08% at G = 10 (one of the parts I used alot). And there is also gain linearity but I don't think its relevant here unless you are actively adjusting the gain. The linearity is only 10ppm, but the error is 800ppm. I would assume this error falls on a bell curve distribution. I think AD wanted to kill LT for this part (the power point is vehement about cost with their replacement). Can't believe they merged now.

If I recall I could not find parts that had low gain error when I was looking at the initial accuracy of the system (bounding adjustment loss accounting)

Given the excellent gain linearity, you could use trimpots I guess.

But I never needed to match differential channels before in an analog system, I would do all those trims later. This poster sounds like they want a initial analog match. Does not seem like a typical system. This is very tricky. I think you are most likely going to get pissed off trying to match it lol. Not to mention to match CMRR.

[bdunham7]

In some cases also with DC but 99.9% of the time just AC signals.
I am ok to design it just for AC if it relaxes the constraints and/or makes it easier to match the two INA outputs to a higher degree.

In terms of frequency, up to at least 100KHz, preferably 250KHz.

You haven't stated exact levels, but if we assume you want to take 1mV and amplify it 1000X to 1V and want an error of less than 0.1mV on the output, that means 0.01%.  A 0.01% error at 250kHz exceeds the specs of some pretty high-end DMMs even at the 1V level.  It also corresponds to 100 nanovolts of input noise or error. 

If you assume higher levels, 0.1mV implies an even lower error in percent, while lower levels and 1000X gain imply some pretty low input levels.  I don't know what state-of-the-art capabilities would be for this type of application, but your goals seem pretty ambitious. 

You mention what appears to be nulling the outputs, which makes things even harder than something like a DMM.  Every error in each amp--input offset, output offset, input noise, frequency response, phase shift any nonlinearity and more that I can't think of offhand--will show up in the difference signal.