I think I have enough of the design there to get the idea across.
I am trying to make a trans-impedance amplifier that has basically zero phase shift at 100KHz to 200KHz. The JFET in the front end is needed to get the noise low enough. Can anyone improve on this topology?
R1 is forced to be largish by the need to keep the noise current well below 1pA/sqrt(Hz).
Composite TIAs never work as well as one expects. For high feedback resistor values, I'd far prefer something like a BF862 follower running into an LT1028, with the drain bootstrapped and a sub-Poissonian current source in the source lead. That'll get rid of the input capacitance, which will help noise and stability a lot. (Depending on your accuracy requirement, a BFT25A might be a better input stage.)
You might very well need a common-base stage in the front, in order to keep the Cdiode*Rf phase shift from being a problem--feedback reduces it, but it doesn't go away altogether.
Once your sensitivity to all the temperature-variable parameters (e.g. GBW and Cdiode) is reduced, you'll know the phase shift vs frequency quite accurately, so a little lead-lag network somewhere will trim the last couple of degrees of phase nicely, I expect, without needing trimpots or other such things.
True--I should have remembered. I've been designing with the ADA4898-1 recently, which is a lot like the LT1028 except with 45 dB lower DC gain, no guaranteed specs, and no funky noise peak AFAIK. It's also a bit slower and much cheaper.
The gm of a BF862 is 0.035 minimum, so the a few picoamps of noise current aren't going to be noticeable. Anyway, I thought you said you needed a large FB resistor--1 pA/sqrt(Hz) is 16k or thereabouts, which has a 1-Hz noise voltage of over 16 nV, so even 2 nV is a nit--about
0.07 dB. Of course it does cost you quite a bit of supply current, but that's life at 300 kelvin.
I really don't like using two op amps inside the same FB loop. It's usually done in an attempt to wring the last 0.5 dB out of some performance metric, and IME it always trades off something else that winds up dominating. The strange open-loop transfer function usually messes up the settling time something awful, for instance, and it's also common for input nonlinearities to be much worse than expected, especially in the presence of overshoot, which photodiode amps are prone to on account of the diode capacitance.
What sort of photocurrents are you looking at? Below 10 uA?
My main suggestion is to overdesign the first stage so you know exactly what the phase is doing between 100 and 200 kHz, and then put in a little (fixed) RC tweak to flatten it out. It should be only a few degrees.
Only if you use the wrong circuit topology. A common-base stage gets rid of that nasty noise peak very handily. What's your PD capacitance?
Understood. If you tell me a bit more about your photocurrent and photodiode choices, I can be of more help. Have you thought about the _optical_ error sources? Generally speaking, measuring an optical signal to an absolute accuracy of 1% is good going.
Hmm. With a 50k feedback resistance, and assuming you mean the overall closed-loop voltage gain, that comes out to a total summing junction capacitance (amp, pads, and photodiode combined) of something like 0.6 pF, which is a bit on the small side. Is that perhaps the transimpedance or the open-loop transfer function of just the amplifiers that you're quoting?
50k is overkill for that photocurrent--I suspect that you might be making life unnecessarily difficult for yourself. At 30 uA, 6k will get you within about 1 dB of the shot noise. That'll also push the noise peak 8 times further out in frequency, which is a big help.
If a fixed tweak isn't good enough, you can probably do just as well by controlling the photodiode reverse bias with a DAC driving a capacitance multiplier. That'll change the phase shift without causing a lot of noise. NB: the PD bias supply's noise looks just like TIA voltage noise, and gets multiplied the same way.
Have you thought about the phase shifts due to the photodiode RC?
I haven't followed the whole thread, so I don't know whether it will fit, but I've used with good success active feedback in a somewhat similar case (needed stable sub milli degree 30dB amplifier @ 100kHz).
It will cost you a sqrt(2) noise voltage degradation though.
Thanks Phil, I realized how 'stupid' this question was this morning. (In other words I figured it out for myself.)
Isn't keeping the phase shift really small the same as keeping the gain really flat? (Kramer's-Kronig relations and all that.) And you'll need to keep the gain flat way past 200kHz to keep the phase shift small at 200kHz. Is it a Bessel filter (Q) that has minimum phase shifts?
For a minimum-phase network, yes it is. Lumped element ladder networks are minimum phase, but anything with actual time delay isn't, and neither are all-pass networks, which have phase delay but no amplitude rolloff.
Thanks, I guess I was thinking about circuits that had some gain (or loss?) in them. I'm not very clever and so I would take a brute force approach. For flat gain and no phase shift keep all the poles far away from the band of interest.
OK I guess I can imagine that. (the all-pass filter example from above.) So is flat gain a necessary but not sufficient condition for small phase shifts. Or to ask it another way, can you build a circuit that has gain changes, but no phase shifts? (Not that I can imagine a use for such a thing.)