Input stage mess

Discussion in 'Electronic Design' started by Fred Bartoli, Nov 28, 2005.

1. Fred BartoliGuest

Well, I've finally got doing that low noise preamp: the target is
200pV/rtHz, bandwidth from 0.1Hz (with provision for 1/f noise) to 1MHz.

Lots of interesting pbs to solve. One remaining thing is some stability
issue wrt to cable/generator impedance.
The input stage will be 4 to 6 paralleled Interfet's IF3602. They'll work at
a low servoed 1V VDS to minimze thermal noise problems.
Under those bias condition the JFet will show about 300pF Cgs and 180pF Cgd.
The Jfet is servo-cascoded so that amounts to and equivalent 480pF Cgs.
The closed loop preamplifier show an input impedance that has a painfull

Cgs w^2
Yin = -------------- with WT = 2 pi GBW and beta= feedback network
attenuation
WT beta + j w

This translate to an equivalent parallel network:

2 2 WT beta
Rin(w) = - ------------ - ------------
Cgs WT beta Cgs w^2

1 Cgs w^2
Cin(w) = --- . -------------------
2 w^2 + WT^2 beta^2

Now the figures:
designing for a 10 loop gain (WT beta) at 1MHz this give a low -12R for the
real part at high frequencies and will give a nice oscillation with the
input cable impedance (estimated between 300nH & 500nH).
As the generator impedance is low, the easy way to deal with this could be a
serie RC in parallel with the input (sort of 2-5 nF and under 12R resistor).
Unfortunatly a 10R resistor is 410pV/rtHz and 41pA/rtHz which translates to
about 130pV/rtHz across the 0.5uH cable inductance at 1MHz. A bit more than
I would like.
More, this 10R is dangerously close to the -12R value, which didn't
accounted for additionnal parasitic poles and will probably be lower.
Plus cable resonance with the input capacitance will rise the noise level.

I can't insert a damping resistor in series with the gate connexion for
noise reasons (200pV/rtHz is a super low 2.5R noise resistance).

I've thought of a lot of schemes for neutralizing Cgs, but found nothing
practical.

I also can't run the input stage open loop, which would solve this issue but
will raise some others.

Any idea?

2. John LarkinGuest

Can you post a schematic, real or simplified?

John

3. MarkGuest

add a C across the cable to resonant out the L.?

I don't undestand why your cable looks inductave anyway?

Is the source Z of the generator equal to the characteristic Z of the
cable?

The only way for the cable to look inductave is for it to be short
length (probably true) and for the generator source Z to be lower then
the cable characteristic Z.

Is the cable charateristic Z 50 Ohms? Whats the generator, a low Z
mic...is this for audio? Why go to 1 MHz?

Mark

4. Fred BartoliGuest

Here it is.

85mA 85mA
to to
125mA 125mA

V V
| |
+------------------------+--------.
| | | |
.--+ .--+-----+--------. | |
| .-.R3 | .-.R4 | /| | | |\ |
---| | ---| | >| /-|-' '-|-\ |<
---| | ---| | |-< | | >-|
| '-' | '-' /| \+|--+--|+/ |\
'--+ '--+ | \| | |/ |
| | | 2 V |
clamp --. | .-------------------' |
| | | | | | ___
| | | | | .---------------------|___|-.
10u | J1 |-+ | +-|J2 | | | R2 |
|| | | | | | | ___ | |\ |
in -+--||--+--->|-+ | +-|<-------+-|___|-. | .----|+\ |
| || | | | | |5mA R1 | 5mA| | | >-+-out
.-. .-. '---+---' | | GND | | GND .-|-/ |
| | | | | | V V | | |/ |
| | | | | .-----------------------. | |
1M'-' '-'1G | | Precision |--+--||--+
| | V | current mirror | .-.
GND | 160mA '-----------------------' | |
| to /| | |
| 240mA /+|-GND'-'
'------------------------------------------+-< | |
| \-|-+---'
| \| |
'-||---'

J1 and J2 are 4 to 6 paralleled IF3602, no source degeneration (noise),
running at 20mA/transistor.
R3 and R4 are one per transistor and 50R (1V across the Jfet, 1V across the
R)

The feedback path has _low_ impedance (1 fb path per jfet, 1R/27R).

As you can see, nothing terribly fancy. All lies in the details (like the
subject of this post).

For the problem of interest here, it comes from the differential input
voltage which rises 6dB/oct with a 90° shift wrt to the input voltage (due
to the open loop gain pole) and the JFETs capacitances across the inputs
(1.45nF between the preamp in+ and in-) which introduces another 90° shift
for the current injected at J1 gate, hence the negative resistance.
Exactly the same pb as an emitter follower loaded with a cap.

5. Phil HobbsGuest

Hmm. Looks like the origin of the problem is the overall feedback. Do
you really need the virtual ground? If not, how about making the input
afterwards?

Since your JFETs will be reasonably linear at that I_D, the feedforward
path will have low gain, and hence you can use a higher-noise amp to
generate it. This of course will require a bit of tweaking.

BTW are the 10-uf capacitor leakage and gate current really low enough
to use a gigaohm gate leak resistor-- 1pA ==> 1 mV? And do you really
want a 3 hour time constant on your bias circuit? Oh, and you have two
nearly perfect integrators in your dc feedback loop, which will cause
nasty behaviour.

Cheers,

Phil Hobbs

6. Frank MilesGuest

[snip]
Challenging!

I don't know if you can meet the noise specs by going down this pathway,
but if the series RC shunting the input adds too much noise, I think you're
going to have to bootstrap your input circuit to get rid of the negative RC
Zin. This will have to operate at high speed -- not at the slow speed of
the feedback loop. Of course there are limits to this approach as well, but
with the right topology you may be able to use a pure capacitance between J1's
gate and an appropriate virtual ground input node to provide a high frequency
boost to J1's source. This has been done, for example, in a few oscilloscope
vertical inputs to provide a wider-band input while reducing the dribble-up

Did you say anywhere what the spec had to be for input capacitance? Perhaps
a bit more HF energy can be stolen from the input? Some other games might
be played if that is allowed. From your resistor values it looks as though
Rin must be high.

Unfortuantely I don't see any way to use the capacitive feed-beside method with
your existing topology. It really needs a follower initial input stage.

This fixes your Zin problem but reaching your noise target will be fun.
I confess I haven't gone through the numbers to determine whether that is
possible with such a different structure. Good luck!

-frank
--

7. John LarkinGuest

Why is the input stage differential? Seems to me that doing that
increases the noise by 1.41 or something. And all those current
sources look complex and potentially noisy to me.

What are you using for the clamp? I have a similar problem... I need a
pA leakage clamp that can swallow some 10s of mA overloads.

John

8. Fred BartoliGuest

Thanks Phil.

Which virtual ground are you speaking about? The whole preamplifier is
non-inverting.

Hmmm, feedforward is one thing I forgot to think about. I can smell some
interesting idea here but right now I fail to see exactly what you have in
mind. Can you sketch something?

One thing I can't do is getting the Jfets out of the feedback loop. The
preamplifier will have 2 purposes:
1) investigate noise in a _very_ low noise power supply (200nV rms)
2) measure the supply transient recovery to an injected charge. The measured
recovery is about 200nV, so expected aberrations (electric and thermal)
should be lower than 50nV,... excluding noise
Unfortunatly the initial transient can be a few 100mV, up to about half a
volt, and will induce nasty thermal tails if the jfets were working open
loop.
Working with an hypothesis of 1mV/K offset gives 50uK for the stability, and
200pV/rtHz translates to 0.2uK/rtHz, so some serious power stabilisation is
in order there.

I've measured some of the Jfets at under 2pA gate leakage under these
conditions, much better than the given specs and this is for a few boxes, so
that's OK for me. I was pleasently surprised because these are pretty big
transistors and this will simplify my already enough complicated life.

The 10uF are specified at 25000s minimum time constant, i.e. 2.5 gigaohm.
The input DC component is about 10V and I've again measured some samples
much better. Again I can select so it's OK too.

And do you really
An unfortunate side effect of the low frequency requirements.
I've not shown some speed up circuitry which was not relevant to the pb.

The 10uF is mandatory because of the gate shot noise rising the input noise
at low frequency, and the bias resistor has to be 1Gohm because I obviously
don't want its current noise be greater than the gate noise current.
Oops, yes I obviously missed a zero somewhere

9. John PopelishGuest

Perhaps a resistor in series with the output opamp feedback capacitor
could cancel one of those poles.

10. Fred BartoliGuest

Yep. It indeed increases noise by a sqrt(2) factor and makes me double the
Fets number to regain it... and worsens the negative resistance number by
the same factor as a side effect

Anyway some other problems make this a preferred solution:
A single branch input stage will have to either have its source current
provided by the feed back network (if one use feedback around it), or
compensated with a current source.

The preamplifier will have 2 purposes (and 2 modes):

1) investigate noise in a _very_ low noise power supply (200nV rms). In this
mode the diff input stages provide good rejection of the common tail current
source variations. With a single sided input stage this noise will not be
reduced.

2) measure the supply transient recovery to an injected charge. The measured
recovery is about 200nV, so expected aberrations (electric and thermal)
should be lower than 50nV,... excluding noise
Unfortunatly the initial transient can be a few 100mV, up to about half a
volt. I obviously can't have the lower feedback resistor (0.2R!) stand half
a volt (and I want the power to be limited: low noise switching supply for
the isolated input stages, low thermal convection noise and blah...), so in
this mode, where I can average out the noise, the feedback network is
switched to a more current friendly 100R/2.7K per transistor, which will
represent a manageable total 20-30mA at 0.5V input.
This feedback network obviously can't provide the 100mA transistor current,
so a compensation current source would be required, with a too big impact on
LF noise.

Another pb for the single sided solution is that the source will have to be
servoed at 0V (current in the 0.2R fb resistor) which will make the gate
potential between -0.1/-0.5V, with potential leakage problems in the
protection device.
A few mV is all I can accept here.

Yet another pb is the looong input time constant and the diff input stage
will help slow thermal drift rejection.

And... and...

Probably something like this. As I can't have any series resistance, at
least at the begining of an input event, the diodes will have to cope with
high current pulses.

----+--+--------+------
| | |
V - |
- ^ |
| | /| |
| | /+|-'
+--+---< |
| | \-|-.
| | \| |
| +--------+-----> current limiting switch command
| |
V -
- ^
| |
GND GND

And also probably some switched current limiting resistor (I can have a
permanent one due the noise specs), because I have to protect the supply
under test from the 10uF cap charge (and discharge).
Unless I discover the miraculous depletion transistor with ultra high Gfs
and low Idss.
Maybe I can have my little daughter write this down on her letter to SK, but
I guess she'll fell upset if I have that nice transistor and something else
is missing.

Do you know such a beast?

Hmmm, the IF3601-02 I have measure 1.9ohm RDSon. Maybe I should look a bit
further, but not too much hope...

11. John LarkinGuest

http://www.thinksrs.com/products/SR554.htm

John

12. Fred BartoliGuest

"John Popelish" <> a écrit dans le message de
I don't know why but some messages I answer are not quoted so I have to
manually add the quote sign, like below, to keep the message understandable.
Anybody knows why?
I've first thought of this, but this is just pushing up the GBW product and
when the loop gain falls again, guess what...
On the same note I've thought of limiting the gain in the pass band to have
it uniform, like loopgain = constant (50) from LF to the corner freq. But
the outcome is the same at HF.

13. John PopelishGuest

So where are the natural poles in this circuit, not counting the one
at zero for the output integrator? I would put the zero at the lowest
one of these poles. Setting this feedback zero would also probably
include increasing the size of the integrator capacitor, to reduce the
low frequency gain so that the second pole occurs beyond unity gain.

14. Fred BartoliGuest

Yup. From the whole bunch of pb I could list I'm pretty safe to say that
this is probably the limit of what's doable at room temperature.

Hmmm, now that sounds interesting. That really is a clever trick, if I
understood correctly what you meant.
Do you have any specific scope reference or schematics to look at?

Plus, in my case, the following current mirror and summing, neutralize this
current injection (from the loop gain POV).
Really nice!
Not at all (well in AC). The source is low impedance (indeed, it is a low
noise one).
The (almost only) limiting factor is the cable series inductance, up to 1
meter length requested, but I hope I can cut this by half.
What other nice thing do you have in mind? I've made me curious

For the high value resistors, as I explained in my answer to Phil Hobbs, the
1G is there to achieve the specs at the low frequency end.
The 1M at input is here because I expect to use its noise to check the
connected-unconnected state and it is also a weak pulldown for the DC
blocking cap. I also don't want current induced contact noise, so I reduce
it to the minimum level.

Maybe I can see some solutions.

Time to check...

15. Phil HobbsGuest

Well, your amp is going to take a long time to stabilize when you turn
it on, just based on that 10,000 second time constant. With a 15V power
supply, the bias loop will slew at (at most) 15 mV/s.

If all you want is a noise measurement, you can use the correlation
trick: put two amplifiers on it, and cross-correlate their outputs. The
amplifier noise goes away and the true signal survives. Gives you a
nice mean-square noise measurement. You can set the measurement
bandwidth and integration time separately to get the measurement you
need--e.g. you can measure the noise in 1 MHz bandwidth but integrate
for 1 second, and get a 60 dB noise improvement.

You can get rid of the thermal tails pretty well by biasing the JFETs at
their peak power points.

Alternatively, you might consider turning the problem inside out: use a
normal op amp to do most of the job and use the JFETs to measure its
differential input voltage, so that the *JFETs* are the feedforward
path. If you can match the gains of two signal paths accurately, you
can get the low noise of the JFETs and the stability of a normal op amp.

The noise suppression you can get with that trick will be limited by the
gain matching and phase shift between the JFET feedforward path and the
normal op amp path.

(Anybody done that before? I might have invented it just now.)

Cheers,

Phil Hobbs

16. Fred BartoliGuest

I've seen it. Unfortunatly its nasty input impedance, 3R at 1Hz, 0.2R at
0.1Hz, won't fit the bill.

I wish it would. Well, in fact no: No pb, no contract...

17. Fred BartoliGuest

I don't remember if I've already said this, but I have a fast settling mode
to keep this time in the seconds area. The only unknown is dielectric
absorption, but a quick check will say all.

I remember you already mentioned this but we also want the spectral density
repartition in order to be able to diagnose the causes.
Already done. I also wish I could bias them at their temperature
compensation point but the high current value render this impractical :-(

Again I'm not sure I see all the details of what you have in mind. Can you
elaborate?
But wait... Hmmm, you're giving me another idea, or maybe the same, just
pushed a bit further, or presented differently...
If I reduce the overall GBW product I rise the resistance negative part and
also reduce the maximum possible frequency of oscillation, hence increase
cable inductance tolerance (is this enough, I'll have to check this point).
Combined with feedforwarding the JFETs signal we could hold a good starting
solution.
Combining all this with a bandwidth limited (since we have ffwd) version of
the Frank Miles' "source boost" trick to get rid of the last bit of negative
resistance and it could well be a perfect match to what I looked for.

I'll look into all that tomorrow. It's time to bed now.

18. MarkGuest

Why do you need BW to 1 MHz to check a power supply output?

Mark

19. Fred BartoliGuest

Yes, why...

Well, because...

Seriously, because those power supplies are part of an ultra sensitive high
definition, real time, video imaging system where some minor aspects of
noise have huge impact on the image quality. Spatial and temporal coherency
of noise, as sampled by the imager, is of primary concern and effects well
below an LSB of the 16b ADCs are visible and annoying.
And when the final customer pays \$5M for one system he simply wants the very
best quality

20. Ken SmithGuest

I've done things like that and have a couple of suggestions:
note the points A and B.

(1)
If you add a switch and to switch in a lower resistance between A and B,
you can get the circuit settled more quickly.

(2)
If you add a series RC from A to B you can add a zero to the bias point
servo. You can use this to get rid of peaking at the gain cross over
point.