Fully differential amplifiers: Trimming and Feedback

[exmadscientist]

(Bear with me here for a bit, I'm a primarily digital guy who's been dragged into the world of high-performance analog design thanks to the dread curse of Cleaning Up Other People's Messes.)

I've concluded that a fully differential amplifier (FDA) such as the LMP8350 or AD8137 is probably the right approach to our current design challenge, but FDAs have a couple of problems:

• High power consumption, 10mA or so. This isn't ideal, but a FDA will replace enough other circuitry that I should be able to fit it into the power budget.
• High offset voltages, up to several mV. This is potentially a severe problem, but I don't see why it can't be trimmed out... except that nowhere have I seen a description of how to trim an FDA for input and output offsets. I'm particularly interested in trim circuits based on auto-zeroing opamps or similar, not trimpots... any trimpot that can be set can be mis-set and given the life these boards have ahead of them, I think trimpots are better avoided with careful designs.
• High bias currents, 1-5 uA or so. This can be worked with but is again not great.
• Feedback networks: manufacturers say they should be "symmetric", but what happens if the circuit I need requires a nonsymmetric feedback topology? How symmetric is "symmetric enough"? Which is more important: making sure feedback comes from both outputs or returns to both inputs? Both?

More information on what I really need to do, so you can tell me if I'm completely off base here:

I have a load that needs to be driven with a precise AC stimulus current of (let's say exactly) 2 mA p-p. The load resistance is variable but is in the 1k-2k range. The stimulus has frequency components up to about 2 MHz, and is produced by an existing DAC + buffer circuit as a differential control voltage that swings 2 V p-p between its terminals, symmetric about ground. +/-5V bipolar quiet analog supply rails are available, plus a few digital rails.

The load is a little bit nonstandard: we get the best performance when we drive it fully differentially, so if one end is at 3V, the other should be at -3V for a total of 6V p-p. If one end is, say, grounded and the other swings the full 6V, things don't work as well. (My boss would probably not be happy if I said much more about this oddball load, unfortunately; confidentiality, and all that). The circuit driving this load must therefore be fully bidirectional, sinking and sourcing current from both terminals. Importantly, when the stimulus signal is idling at 0 V differential, the error current through the load must be "very small". (I've asked another engineer for a more precise specification, but am yet to receive it. I am, for now, interpreting this as "<10 nA maximum at idle".)

To convert this 2V p-p control signal into a 2 mA p-p current stimulus, I've designed an improved Howland current pump circuit based on a FDA. A FDA seems like the right choice as they are very fast, low-distortion amplifiers and eminently capable of shoving 2 mA p-p into 2kOhms at 2MHz. I've put together Spice simulations and I get beautifully flat responses with nice, flat gain and phase response from DC up to MHz. (Simulations have used the LMH6651 model rather than the LMP8350, my current top pick, as apparently no model of the '8350 actually exists.)

OK, so what's the problem?

How do I trim out the offsets? When the input signal is zero, there's always a current flowing through the simulated loads -- sometimes of order 10 uA or more, which is not within spec. I suspect I'll need two trims, one to null output common mode and one to null input differential offsets, but the best way to do this is eluding me. I have not been able to find any references on nulling these fast amplifiers; evidently they're just not used in precision applications or something?

How do I reconcile the Howland's feedback topology with manufacturer recommendations for "symmetric feedback"? I've simulated two circuits: one is a bog-standard copy of the usual improved Howland, except that the load returns through the FDA's inverted output rather than to ground. This circuit has no feedback from the inverting output, except through the load directly. The second replicates the entire Howland feedback network at both outputs and connects the load between them. Certainly this is a symmetric network, but it doubles the number of precision matched resistors, already an issue for Howland circuits, and introduces a lot of added complexity. It also makes, as far as I can tell, no difference in the simulation results. Is the added complexity worth it? How could I quantify this?


Thanks for any insight or guidance anyone might be able to offer. Alternative suggestions for how to achieve my requirements are certainly welcome too.

[Jay_Diddy_B]

Hi,

Here is a model using the LT1995 that was published in EDN.

http://www.edn.com/design/analog/4439586/Gain-selectable-IC-yields-voltage-to-current-converter

There are some useful links at the bottom of the article.

The LT1995 contains the precision resistors that you need for this application.

I have built a LTspice model of the circuit:



The AC response:



Offset can be added like this:



The results are:



I have attached a zipfile with the models.

This circuit requires higher voltage rails, +/-15V, than you have. Otherwise it will probably meet your requirements.

Regards,

Jay_Diddy_B

[exmadscientist]

Hi Jay,

That's a very nice current source but it's still single-ended: one side of the load is grounded. The unusual load I have the distinct (dis)pleasure of dealing with strongly benefits from being driven truly differentially.

I've attached the best circuit I've yet come up with for the job. It's got fantastic AC performance over varying loads (the AC plot is for the configuration shown in the schematic, loads of 10 to 4kOhms; that covers the nominal operating range plus quite a bit of margin for general abuse by the end users). The DC plots show the response to a typical simplified stimulus into a load of 1k. The second one is a closeup of the end of the stimulus showing the offsets at various circuit nodes.

The fatal problem with this circuit is the -6.9uA offset current through the load (red trace, I(RL)). If that can be reduced to the nA range, I think this is good enough to go ahead and prototype. Probably related are the offsets: I(RL) peaks at +0.992mA and -1.006mA, not +/- 1.000mA. That's OK for this application, but the offset isn't.

I've attached the LTSpice files, including a few models for TI parts I've tried. LT1994 or LTC1992 can also be dropped in as qualitatively similar replacements, though they do not perform as well as the TI parts.

Suggestions? Alternatives?

[dom0]

Precision instrumentation amp over R5 -> integrator -> DC null output current?

[Jay_Diddy_B]

Hi,

I introduced another voltage source into the model, V3. This voltage source is to remove the input offset voltage of the LMH6551:



This improves the zero current performance.
I suggest looking for an amplifier with better input offset specifications.

Regards,

Jay_Diddy_B

[exmadscientist]

I've finally come up with a trimming scheme I'm happy with for the circuit I presented earlier. I've attached the schematic to this post.

Given that a fully differential amplifier is a two-output device, you might expect (well, I expected, anyway) that it would need two trims: one to fix up the output common mode and one to fix up the output differential mode. But what I found through experimenting in the simulator is that the thing actually performs much better if you have three trimming circuits: one, done in the straightforward way, to fix the common mode, but two differential trimming circuits, one each dedicated to the inverting and non-inverting outputs. Doing it this way allows lots of errors to cancel out to first order, since they're replicated on both sides. For example, I put a simple RC low-pass filter in each differential trim integrator of this circuit. They don't have perfect transient responses, but because they each have the same imperfect response, both summing junctions are affected approximately the same way and the associated error goes away. (Things could probably be improved by using a differential integrator or a non-inverting integrator for one half, but both of those options require matched capacitors for good performance, and that is one component sourcing nightmare I want to stay well away from.)

Precision instrumentation amp over R5 -> integrator -> DC null output current?

I simulated a few variations on this theme, and while it worked OK, the circuit below ended up working better, so that's the direction I've followed.

I suggest looking for an amplifier with better input offset specifications.

After considering a few different topologies for this current source, I've come to the conclusion that making a fast amplifier behave is probably a better option for this application than trying to make a precise amplifier play nicely in the MHz range. I simply haven't come across anything that's got the right combination of specifications to do the job as well as I'd like.

That said, two of the options I'm going to prototype are based on slower, more precise amplifiers, but they're a lot touchier in simulations, and don't look as good even when they're behaving well. But that's why you prototype rather than just trust the simulator....