Characterising low frequency noise?

[splin]

OK thanks splin - it's taken me a while to get round to chewing through that Can I ask some dumb-ish questions? - I can't quite replicate your calcs atm - I'm probably missing something...

Firstly, where does the sqrt(20) factor come from?

Well you've probably worked it out by now but he paralleled 20 ADA4898s to get the voltage noise down to 0.201 nV/sqrt(Hz), from page 3 of http://www.hoffmann-hochfrequenz.de/downloads/lono.pdf

He clearly knows the disadvantages but must have miscalculated the i/p capacitor impedance:

That does not come for free: at the same time, current noise gets worse by the same factor. This
limits the usefulness of this amplifier to low impedance signal sources such as power supplies,
voltage references or control voltages from op amp outputs etc.

Also is there some easily available arithmetic which approximates total rms/p-p noise at the 1/f end given a couple of points on the 1/f graph, a bandwidth (or the upper end of the bandwidth anyway), and an assumption that the line is straight-ish? Or alternatively could you get the noise results like this out of spice?

TIA, Alan

Take a look at AoE 8.13.2 to 8.13.4. The latter describes how to calculate the 1/f noise corner from a couple of datasheet noise density values, one above and another below the corner frequency.

Or take a look at AD's Tutorial MT-048 'Op Amp Noise Relationships: 1/f Noise, RMS Noise, and Equivalent Noise Bandwidth'

[alanambrose]

>>> he paralleled 20 ADA4898s

Ah yes, duh.

Thanks splin for the links, very useful. More studying

Kleinstein: Yeah agree you are dependent on the model. I believe the ones I've been using do - as you can change the input filter capacitor and see a substantial effect on total input-referred voltage noise due to current noise. Having said that I must have done something wrong in that run (picked up the noise number from the wrong point I think). I've updated the previous graphic.

Alan

[splin]

The input current noise is exactly why Hoffmanns circuit is not good and only of limited use with very low impedance sources.
At low frequencies the input filter will not be useful - but at higher frequencies (e.g. > 100 Hz)  it can work. The problem is not the 10 K resistor but to small a capacitor.

I can't find the circuit in the AD4898 datasheet (maybe newer version), but usually the amplifier under test is used to do a first amplification (e.g. 100 or 1000 fold). So the following stage is far less critical.  The AD743 is well suited for such a filter with 1 µF and 1.5 M, as it has the low current noise of a FET amplifier: 7 fA at 1.2 M (which is about the impedance at the crossover frequency) give some 8.5 nV /Sqrt(Hz) of noise at that low frequency. At higher frequency it gets less.  Already at 3 times the lower limit the current noise is less critical than the voltage noise. One could have used a larger cap though.

With an amplifier like the AD743 it does make sense to use a few in parallel. It still does not need low impedance sources.

AD have indeed changed the datasheet by removing the offending section from rev E in response to someone pointing out the error on their community discussion board (not me - it was before I spotted it). I've attached the relevant section. As you can see the UUT, the ADA4898-1 is run at unity gain hence the problem.

It's a bit disappointing that they simply removed the section rather than measuring the .1 to 10Hz noise properly for a device touted for its low noise characteristics but I guess they are short of resources and the datasheet does include the voltage noise graph so you can at least work out the integrated noise yourself.

Splin.

[EDIT] Added unity gain statement

[alanambrose]

Are these simulations believable? C/R/Op Amp variations in LT Spice with 10mHz-10Hz and 100mHz-10Hz bandwidths.

Alan

[Kleinstein]

The values look plausible. There might be a small amount of 1/f current noise with AZ OPs like the LTC2057. So for the very low frequencies and small caps the noise might be slightly higher.

[splin]

It depends on whether you believe the datasheets - if they are wrong the SPICE models are unlikely to be correct. Take a look at AoE 8.13.6, page 569:

A caution: datasheet values for input current noise are sometimes seriously in error, evidently because the manufacturer didn't measure it, believing that it was accurately predicted by a shot noise calculation based on dc input current.

They accuse the datasheets for some bias-cancelled op-amps of specifying the current noise calculated from the bias-cancelled rather than the uncancelled bias current values with results 10X too low - see 5.10.8, page 327. They cite the LTC1012 as an example with a spec of 6fA/sqrt(Hz) at 1kHz; they measured 55fA/sqrt(Hz).

They found the discrepancies between published and measured current noise for many auto-zero op-amps were much worse - for example, the LTC1049 measured noise being 50X spec and the OPA2188 100X!

Worst was the MCP6V06 which specifies 0.6fA/sqrt(Hz); it doesn't state at what frequency but the datasheet claims it has no 1/f noise. AoE figure 5.52, page 336 shows the measured current noise density to be 230fA/sqrt(Hz) from 100Hz to 4k with large peaks (up to 10000fA/sqrt(Hz) above 4kHz, then reducing to 38fA/sqrt(Hz) at 50kHz!

The attached shows a few of their measurements - hopefully not enough to infringe copyright. There may an issue with exactly how they made their measurements but I have no reason to doubt them.

Unfortunately they haven't measured the LTC2057 - anyone set up to measure one?

[alanambrose]

Ah yes, very good point - many of the datasheet current noise numbers are nonsense. OK that makes the problem a little harder - I need to build an op amp noise measurement set-up first - then screen the likely op amps...

Alan

[alanambrose]

OK some more simulation results and a schematic for your thoughts. I'm targeting this as a first base solution here without pushing the envelope. This is based on Herr Hoffman's design with (apparently) lower noise op amps and an RC filter which should give a lower noise contribution from current noise. I don't want to start on a FET front end or any other more creative solution at this point - but would like to just have a baseline to check against. I'm planning to use MLCCs and live with any piezo drawbacks also. This design is targeted at low impedance sources and low frequency noise e.g. voltage refs. Oh, and I'm planing to take a flyer on LiPo batteries as a power source.

Apologies if this is old news but I just listened to the 9-part TI video series on opamp noise - I thought it was a great complement to the AoE noise chapter.

Alan

[Kleinstein]

The power supply is not that critical. At low frequency the PSRR of OPs is really good, so not problem there.

However using BJT based OPs in parallel and AC coupling at the input might be tricky. The main limit is current noise and even a single OP27 might need a really big capacitor to give good results at 0.1 Hz.  So there is rather little use in paralleling more BJT OPs unless you go for several mF of electrolytic caps. I doubt it gets viable to use even a single LT1028 at less than 0.1 Hz with AC coupling.

Paralleling OPs might be an Option for high quality JFET based OPs, as they have low noise current and the noise current is not going up as 1/f. So 10 JFET based OPs might be more viable - here things get better with more in parallel. So it might be worth using sockets for the input OPs, so one can relatively easy change and test the OPs in circuit.

MLCC caps should not be the lowest grade / highest capacitance - they could have a equivalent to Barkhausen noise (domain jumps) as they can have ferroelectric parts. So it could be difficult to get the really high capacitance.

The input AC coupling should not set the lower frequency limit - so the resistor should be larger and an extra filter after amplification used to set the lower frequency limit. This is lower noise and more flexible. The filter could be digital in software as well. One likely needs several minutes for settling anyway for thermal reasons.

The amplifier circuit still misses input protection ! this could be something like 2 anti parallel diodes at the OPs input to GND or the inv. input. Otherwise one could easily break the OPs when connecting to a low impedance source. Similar there should be a way to charge the input capacitor in a slow controlled way without overloading the source.

I would consider less than +-15 V for the supply (e.g. +-10.8 V (9-12,6 V) = 3 Li Cells each). This means less heat and no need for regulation.  At least the input part does not need that much amplitude and not many signal destinations (Scope, ADC cards, Soundcard, DMM) need a high voltage. I am also not sure about so much amplification.

[Andreas]

Hello,

some annotations:

The LT1007 is marginal at 1 MHz and amplification = 10 (5-8 MHz bandwidth amplification product).
Perhaps the LT1037 would be a better choice.
But in my personal opinion it is a better idea to have 2 different amplifiers for <10 Hz and >10 Hz.
So you can select the optimum OP-Amp for the task.

I am also unshure wether parallelling four LT1007/LT1037 is a good idea together with the 1K resistor.
Above 1.5K there are better choices for OP-Amps:
http://cds.linear.com/docs/en/design-note/dn140f.pdf
And 4 LT1037 in parallel would give a maximum resistor of 375 Ohms if chosing optimal input impedance.

The noise simulation seems also to give a too high value for the 0.1 - 10 Hz noise.
I get below 200uVpp (around 24 uVeff) noise floor with 1K + single LT1037 input stage when measuring a NiMH battery over 100s measurement time.
(with the typical 10s measurement time its usually in the 100-150uVpp range).

By the way: you should connect all resistors properly in your cirquit diagram (R27-R30).

with best regards

Andreas

[alanambrose]

Thanks guys,

Lot's of sound advice as ever

Let me mull over those points and do some more thinking and maybe some more simulations.

Duh, point taken about paralleling BJTs.

>>> The input AC coupling should not set the lower frequency limit

This occurred to me this later in the afternoon - I did that because lots of other designs seem to

>>> By the way: you should connect all resistors properly in your cirquit diagram (R27-R30).

Good point Ah that's curious - this is Diptrace and it passes the erc without those connected !

Alan

[alanambrose]

Errr,

Does this THAT300 front-end sound plausible? The model gives 0.036uV RMS for 0.01-10Hz and 0.032uV RMS for 0.1-10Hz. This is using the 'THAT 300-Series Noise Performance Optimized' model from here:

http://www.thatcorp.com/Device_Models.shtml

My spice file is here if anyone is interested:

http://anagram.net/nuts/LNA/THAT300%20front%20end.zip

Regards, Alan

p.s. I calculate R1 should be 16K for correct ENBW rather than the 10K6 shown in this model. Doesn't make a huge difference to the noise though - 37nV RMS 0.01-10Hz.

[zlymex]

Errr,

Does this THAT300 front-end sound plausible? The model gives 0.036uV RMS for 0.01-10Hz and 0.032uV RMS for 0.1-10Hz.
.................

It can well be true but not very low. The noise of my meter is 0.015uV RMS for 0.1-10Hz by using ADA4528-2(two opamps in parallel).

[Kleinstein]

It looks like the noise current might be quite high for the 1500 µF input capacitance.  It also depends on the 1/f noise of the transistors - could not find a note in the datasheet.

The discrete transistor circuit has the same limitation on input current noise. It might be slightly better than some OPs (like LT1007) as there is no bias canceling circuit, but there is still the inherent limitation for BJTs. The THAT transistors are not special in that way. They don't even have a high h_fe which would allow a slighly lower current noise.

So the amplifier circuit might be as good (low voltage noise) and bad (high current noise) as an LT1028 OP. So no really suitable for an AC coupled input at low frequencies.

A good AZ OP might be the better option - just have to find one with relatively low current noise. The lower one goes in frequency, the more it's in favor of an AZ amplifier, as there is no 1/f part.

[acbern]

Using BJT discrete input stages by far delivers the best noise level (MUCH better than any opamp and also better than JFET discrete amplifiers if properly selected). getting to 0.5nV/sqrt hz is certainly possible with a single diff. transistor stage, with paralleling even less. You pay this by relatively high current consumption and also you need to take care of drifts (precision low drift resistors, tight thermal coupling...). And only works with low souce impedance, and their input impedance is low.

[Kleinstein]

The problem is the required low input impedance. The amplifier circuit looks like it is made for something like a 50 Ohms input, but the 1500 µF input capacitor gives about an 1 KOhms impedance at 0.1 Hz and more at lower frequencies. So below about 1 Hz the main noise source would be input noise current time the impedance of the capacitor. Better values would be possible with using only one pair if transistors or reducing the current to a lower value. Just like with BJT based OPs, paralleling BJT inputs only makes sense for very low input resistance.

At low frequencies current noise is even more important, as the 1/f cross over for the current noise is usually higher than for the voltage noise. So at frequencies the suitable input impedance is lower by something like a factor of about 10. So even 50 Ohms might be to high for the amplifier above.

Not all BJT spice models include all noise effects. I have a little doubt that the noise is that low. This would require very low 1/f noise levels. Especially the only marginally higher noise for the 10 mHz lower limit is suspicious for a non AZ amplifier. Normally I would expect about  So it would be interesting to see the noise spectrum, not just the integrated value.

[alanambrose]

Hi,

FFS I just wrote a long post and the web browser lost it. Pah, start again. This is the abbreviated version:

OK this just went to FAB:

http://anagram.net/nuts/LNA/NoiseAmp3.pdf

Now includes input protection, input current limiting, capacitor charging, peak measurement etc. It draws heavily on the designs from Hoffmann, AN159, zlymex.

Alan

[TiN]

Hm, I'm interested far enough to design and make a PCB for it, if ETD time within before August is OK
How about that?

[alanambrose]

Ah thanks TiN - my longer message (before the browser lost it) said I have actually received the PCBs back already, and ordered the parts, but I'm away for a week so won't have time to make it up for a bit. I'll have a few spare so happy to send out. You might want to wait until I have verified the design though

Alan

[TiN]

Well, we can go in parallel and I'd route fixed version of your design with all caveats emptied already.

I wanted to do own preamp from long ago, but have too many unclear questions on which way to go, while your design seem to be a good in between solution and I have most of LT parts handy.

[Kleinstein]

I have no checked the exact data, but to me it looks like the current for the BJT amplifier stage is too high - thus very low voltage noise, but a rather high current noise. This is especially true for the 10 mHz lower limit -here the current should be considerably lower (e.g. a factor of 3-10) than for the 100 mHz lower limit as the effective input impedance (e.g. the coupling cap) goes up.
The exact value of the best current also depends on the quality of the transistors, especially the 1/f part of the current noise. I am not sure the transistors are really good ones for this purpose. The models used for simulations might not be valid for the 1/f region - so the simulation might be way to optimistic. The only marginal higher noise with the 10 mHz lower limit suggests this very much. Normally I would expect something like 1.2 to 3 times the noise with the lower limit, not just 12% more.

The 2 series resistors for input protection also add quite some extra voltage noise - this might be a problem at the upper frequency limit (10 Hz). So possibly more noise from these resistors than from the amplifier itself. R10 does not really help very much at that position, so I would leave it out.

I don't think you will need D1 and D2. The two diodes to GND should give enough protection.

R30 and R31 (and thus R32 too) should be higher values, e.g. 100 - 220 Ohms. The noise level here is not that critical anymore and 30 Ohms might be too much load to the OPs.

The second 10 mHz cut off should be switchable to 100 mHz as well - at least if the final filtering is not done in the digital domain.

In most aspects the circuit zlymex showed should give better performance, especially if the used AZ op is really that good or at least meets it's specs on current noise. This is especially true for the 10 mHz range.

[alanambrose]

Interesting points.

(edit: corrected FET to BJT)
 
I didn't check that the operating point for the BJT front-end is optimal (I meant to but forgot). But the board layout is the same, of course, so the values can be adjusted later. The model I used was the THAT300 spice model for noise so presumably it is close - how close I guess we'll find out. The close 10mHz and 100mHz noise values suggest to me that something more could be optimised, but I have not checked atm.

Re the R8/R10 series resistors you may be right, I didn't include in the SPICE model. I posted my model earlier if anyone feels inclined to adapt it.

Agree D1 and D2 are overkill but over-engineering is fun no?

After modelling a dozen or so op amp-only front ends in SPICE (although not I think zlymex 's choice) i came to the conclusion that I should try a BJT front end which seems to give an order of magnitude improvement without much work.

At this point, I'm aiming for a good shot at a workable noise amp here rather than the ultimate solution. I expect to do a bit more work when it comes time to test and characterise. I'm thinking I have a workable board for this even if I have to make a few modifications.

Regards, Alan

[Kleinstein]

The THAT300 is a BJT array - not JFETs. So expect similar problems with current noise as with other BJT based amplifiers. For die simulation it would be interesting to show the noise density and not only the integrated RMS value. This will show if 1/f noise is at least included (could still be to low). There is the option to use essentially the same circuit with low noise JFET pairs (e.g. LSK389) as well. Depending on the FETs, offset adjustment might be a good idea.

D1 and D2 also add to the current noise, though not that much with really low leakage diodes. So at least with a JFET input stage leaving them out helps.

[alanambrose]

>>> The THAT300 is a BJT array

Yep my mistake, have changed my previous message so it is less confusing And you are right that the 2x500R protection resistors are not helpful for noise.

A.

p.s. THAT300 noise density below (2 parallel BJT pairs but without protection stuff)

[Kleinstein]

The curve shows a very low 1/f noise crossover - something like 0.1 Hz. So either an exceptionally good transistor or poor model.  If this is for the amplifier with input capacitor, the 1/f part could be just from the capacitor - so maybe the model is missing the 1/f part all together.