200 kHz 14 bits low-noise measurement

[TheLaserGuy]

Hi everyone !

I'm looking for a device which could measure some differences of voltages very accurately at a repetition rate of 200 kHz. Let me explain you the measurement I want to make : I will send a laser at repetition rate of 100 kHz on a photodiode, and I would like to record the shot-to-shot fluctuations of the intensity provided by the photodiode during at least 1 second (so 200 kS memory minimum to respect Nyquist and be able to look at the fluctuations in the frequency domain via FFT later on). Since the shot-to-shot fluctuations I'm expecting are in the range of 0.1%, I would ideally like to be able to resolve 0.01% (so 14 bits ADC minimum).

Since I will most likely publish this value in a scientific journal, I need to trust at 100% the measurement, without wondering if the fluctuations I measure come from the front-end of the oscilloscope or from the laser itself (which also mean that I don't want to use a RC filter to remove the DC component to maximize the dynamic range). Another plus would be to have the possibility to record this simultaneously on 2 (or even 4 would be awesome ) channels. Regarding the voltage, since it's a high power laser, it shouldn't be an issue : I will have the option of changing as much as I want the intensity on the photodiode

The repetition rate of my laser is really fixed, so 1 MS/s for example would be nice, but not needed. Though, the analog bandwidth of the oscilloscope should be at least 200 kHz, just to be sure that I won't have any significant attenuation of certain high frequencies. And last thing : regarding the linearity, there is not so much constraint on it since I want to look at variations in the range of +-1% maximum (which means that absolute accuracy doesn't matter that much neither).

Regarding the options available, I've found the following ones :
- PicoScope 4262, the nicest option in my opinion, just a bit expensive
- Digilent Analog Discovery 2, memory of only 16 kS (would still be ok, but not the best), and no specs on the noise which is the main problem
- Virtins VT DSO-2A10E, 13 bits @ 625 kHz sampling, no specs about the noise
- National Instruments PCI-6036E, no internal memory as far as I understand, and only 1 kS in the buffer (so even if they claim 10 kS/s, if the data are cut, I can't use it), but at least the noise is specified and looks ok-ish.

Any idea of something I could have missed (in the way I want to make the measurement, in the constraints or on the oscilloscope I can use to make it) ?

Thanks a lot !

[Kleinstein]

The required sample rate depends on the pulse lenght and BW of the photo-diode amplifier, not the repetition rate.  Depending on the detector I would more like consider 1 Ms/s the maximum with preferably more (e.g. 10 Ms/s). With faster sampling one can get away with slightly less amplitude resolution. So 12 Bit may still be OK, especially if more than the minimum number of pulses can be used.

I am not sure the FFT would be right way to look at the data. Chances are more like one would have to write it's own custom data handling routine.

[TheLaserGuy]

Thanks for your reply Kleinstein !

Regarding the detectors, I will use 2 different ones (for 2 different lasers, but same repetition rate and requirements) :
- PDA10DT-EC from Thorlabs, that I will use in non-biased mode, with transimpendance amplifier (https://www.thorlabs.com/thorproduct.cfm?partnumber=PDA10DT-EC)
- DET10A2 from Thorlabs, reversed biased, without amplifier (https://www.thorlabs.com/thorproduct.cfm?partnumber=DET10A2#ad-image-0)

I already tried with a 10 bits scope (1 GHz, 5 GS/s), but it turned out that I couldn't use the data because the front-end of the scope was too noisy, and even though I had more time sampling points (around 1000 per shot with the DET10A2 and connected with 50 termination, 1V max amplitude to remain in the linear part of the photodiode), I "lost" this resolution by having a noisy front-end (and adding more quantization noise). Hence why I thought that a "1 shot = 1 measurement" would reduce this, especially if I assume that the noise of the front-end remains on the order of a few LSBs.

The problem is that I don't have a 12 or 14 bits oscilloscope to test this, so I'm looking for a "cheap" replacement just to make this measurement. Would a 12 bits 10 MS/s low-noise front end ADC be better compared to a low-noise ADC 16 bits 1 MS/s for example ? Do you have an idea of what low-noise 12 bits ADC could do the job ?

For the FFT, I will of course want to save them because I would like to make a deeper process after the measurement (but I expect the main sources of noise to be in the 1-300 Hz and 3-6 kHz ranges).

[splin]

A much cheaper and possibly better solution; buy a NUCLEO-H743ZI2 board for $27.

The STM32H7 has 3 x 16bit 3.6MSPS ADCs with hardware oversampling:

ENOB  13.2
DNL    14bit equiv (differential linearity, equivalent to a 14bit ADC with 1 bit DNL)
INL     13bit equiv (integral linearity).
THD    -90dB

Two of the ADCs can be operated in dual interleaved mode so effectively at 7.2MSPs. With oversampling the resolution could be increased, in theory, to 15.7 bits @ 200kSPS. Oversampling by itself won't improve the linearity of the ADC but you could add a bit of dither to the signal using one of the H7's DACs which would.

The 480MHz H7 processor is pretty powerful with loads of RAM and would allow you to do a lot of data processing in real time.

Free software development tools are available from ST and the debugger interface is built into the NUCLEO board. Example software is available in the Cube software libraries to setup and run the ADCs so you shouldn't need to write any code to get started.

But, (there's always a 'but' isn't there!) there are some potentially significant performance limitations that aren't obvious in the datasheet. See my post here:

https://www.eevblog.com/forum/microcontrollers/stm32h7-adc-performance-ya-gotta-read-the-fine-print!/

And most importantly follow the link imo posted in reply #14. Specifically:

https://community.st.com/s/question/0D50X00009sV2JSSA0/fluctuations-in-adc-value-of-stm32h743-mcu

The linked thread discusses a serious noise problem with the NUCLEO H7 boards which turns out to be easily resolved, to give enough ADC performance for your needs.

There are some other NUCLEO H7 boards (eg. the NUCLEO-H755ZI-Q) which may be newer with later (better) silicon revisions and perhaps with the 743Zi board's 25MHz noise problem (from the ethernet circuitry) resolved. You'd have to do some research to find the latest NUCLEO.

[iMo]

I will send a laser at repetition rate of 100 kHz on a photodiode, and I would like to record the shot-to-shot fluctuations of the intensity provided by the photodiode during at least 1 second (so 200 kS memory minimum to respect Nyquist and be able to look at the fluctuations in the frequency domain via FFT later on)

What about to amplify the fluctuations only?
AC coupling with an amplifier, and then ADC (or with your oscilloscope)?
What is the lowest fluctuation frequency?

[Kleinstein]

Typical fast scopes can have quite a bit of low frequency noise. So a 1 GHz scope is the wrong instrument to look at a low noise 500 Hz signal. Below some 10 kHz a fast scope can be pretty bad.

The information about the laser noise would be in the difference between the base line and the usually short part for the laser shot. As most lasers are not 50% on 50%  this would be best looked at in time domain and not so good in the frequency domain.
If done right the analysis should suppress the low frequency components pretty well.

It depends on the length of the pulse if a 10 Ms/s 12 Bit ADC would be better than 16 Bit at 1 Ms/s. Ideally one would want a reasonable time resolution (e.g. more than some 10 samples) for the pulse, so one can really separate the time in the pulse and before / after.
I would avoid a slow detector or extra pulse stretching stage, as this would only spread the energy over more time and add more tail.

To resolve 0.01% it would be about 14 Bit resolution. With oversampling this would be some 16 perfect 12 bit samples. Actual ADCs / DAQ cards tend to have more noise than resolution limit by often 1-2 bits. However it depends on the details and front end.
At that accuracy level also the optics can become tricky: interference could change the intensity on the detector and the detector may not be that constant in sensitivity everywhere. Also dust could be a problem.

From the specs the PicoScope looks pretty good and not to expensive for real science. They also have a few other, slightly faster ones in a similar price range.

[Andreas]

AC coupling with an amplifier,

... or add a offset to the channel so that you see only the (clipped) peak values (amplified by the scope settings).

For me the measurement task is not clear: what do you really want to measure?
- The peak amplitude of every shot?
- or the (integrated) power of every shot.

The amplitude seems to be < 1V, obviously a naked photodiode loaded with a 50 Ohms resistor.
1000 measured samples per shot at 5 GS/s implies a pulse width of ~200ns so a minimum sample rate of 10-20 MHz right?
What is the rise-time of the photo-Diode. (in case you want to integrate the power?)

Did you use the 20 MHz Bandwidth limiter of your scope?
One of the resolution enhancement modes?
Is the Photodiode placed into a metal shielded housing? (cookies box).
If using unshielded connection lines and make a FFT I can see every frequency of the local FM-stations.

with best regards

Andreas

[TheLaserGuy]

Thanks to all of you for your help !

@splin : I looked into the NUCLEO and this noise of 300 LSB is just way too much for what I'm targeting, even though it's 14 bits and 3.6 MS/s. The thing is that I won't try to implement from scratch one of those ADC (no time, and not enough knowledges to do so). Do you know another board with a similar ADC without those bad decoupling issues (and which would not require too much debugging) ?

@imo : As I mentioned before, since I want to have the most accurate measurement possible, I can't just put a high pass filter, which would modify the noise I measure. Typically, the biggest contributions of noise I'm expecting (and that I already saw with the 10 bits oscilloscope) are in the 1-100 Hz (mechanical vibrations and air fluctuations), 100-300 Hz range (audio vibrations) and in the 3-6 kHz range (active stabilization of the beam by piezo actuators). So, because of all this, I can't just reduce filter out the signal by applying a high pass filter since it would affect the integrity of the measurement.

@kleinstein : Concerning the laser itself, it's a femtosecond laser (15 fs), so the duty cycle is almost 0. The problem for the analysis is that I don't want to remove low frequency components since it would affect the integrity of what I want to measure. I guess I could modify the termination of the photodiode in order to modify its RC constant. But in that case, using classical 50 Ohms BNC cables would make a impedance mismatch, so I don't know if I should go for this or not. Other possibility would be to use a faster photodiode to have a typical RC constant of 10 ADC points as you suggested. I guess that with a 100 MS/s scope, a 1 ns RC constant should be in the right range ...
For the oscilloscope, I thought to this only when I analyzed the data and since the laser is not working for a moment now, I can't take a new measurement right now with a lower bandwidth oscilloscope to see if it would change something. Next time I'll have the opportunity to make the measurement, I'll try with a slower oscilloscope. Another stupid thing I did when I took the measurement was to intentionaly limit the amount of power going on the photodiode to have only 10-20 mV of amplitude, and thus be far from any saturation for the photodiodes. I now realize that this was a bad idea, and next time I'll properly check the voltage I need before saturation (an thus maximizing the dynamic range).
And for the laser, in order to ensure that beam pointing wouldn't affect too much the measurement (and also to reduce the effects of the inhomogeneties on the photodiode surface), I used a strongly divergent lens just before the photodiode, to get an "uniform" intensity profile on the photodiode. And for the interferences, since it's either picosecond or femtosecond pulses (we have different pulse durations at different places in the laser), so even if we had interferences, the typical constant decay time of the Fabry Pérot would be much faster compared to the rise/decay time of any oscilloscope, so no worries on that side

@Andreas : I want to measure the energy in each shot, so integrated power would make more sense, but the big question is to know which threshold of voltage to take for the integration. When I did this first test measurement with the 10 bits 2.5 GS/s scope, depending on the threshold I had for the integration during the analysis, I ended up with completely different results since the amplitude fluctuations I was looking at started to be on the same order of magnitude as the front-end of the scope. Hence why I thought a high dynamic range "1 shot = 1 measurement" by taking the peak intensity only with a low noise front-end would make more sense and would be the most reliable for the signal integrity.
For the amplitude, as I said to Kleinstein, next time I will have the opportunity to make the measurement, I will first characterize properly the saturation of the photodiodes to know up to which voltage I should use the photodiode and maximize the dynamic range of the scope. But I guess it's going to be more in the 10 V range for both photodiodes (10 V max amplitude according to the datasheet for the PDA10DT-EC and 10 V reverse bias for the DET10A2). For the rise-time of the photodiodes, it's 1 ns for the reverse-biased one, and I don't know exactly for the other one (not in the datasheet), but the max bandwidth after amplification is 1 MHz according to the datasheet.
I didn't use the 20 MHz banwidth limiter to avoid to affect the signal integrity. The 2 photodiodes are mounted in a metal shielding (if you want to see : https://www.thorlabs.com/thorproduct.cfm?partnumber=DET10A2 and https://www.thorlabs.com/thorproduct.cfm?partnumber=PDA10DT-EC). Unfortunately, we don't have RG223 BNC cables in our lab, so I used simpple RG58 single shielded ones, but the FFT doesn't show any high frequency peak above 1 MHz (I made a measurement at 2.5 GS/s for 100 ms with a 1 GHz bandwidth, and nothing as strong as what you attached).

[Kleinstein]

For calculating the pulse power i don't see a problem with the threshold:  My intuitive idea would be to use something like half hight to trigger / get the time for the pulse start. The integration would than be from some fixed time before the pulse starts. For the end of integration one would use a time where the level is low, like 10%, or alternatively instead of a hard end a decreasing weight for the decay part.
For the integration one would also need a baseline for the dark state. This could be from something like some time (e.g. 10 times the pulse lenght) before the pulses. When using the baseline before the pulse, there is essentially no more sensitivity to low frequencies (e.g. < 10-50 kHz). So AC coupling and low frequency noise from the scope input is not a problem anymore. So no need to use a slower scope. A scope with a dedicated 50 ohms input could be an advantage, but I would expect that for a >= 1 GHz anyway.

Low frequency intensity variations would be still visible, but in the pulse hight, not necessary in the FFT over the whole signal.
In the FFT the interesting part is more like a little around the 100 kHz repetition rate.

Using just a photo-diode and terminated cable is probably OK. I don't think one would need a 1 ns pulse - something like 10-100 ns sounds more practical. If the sampling rate is high enough there is no need for the analog filter at the scope.
So the instrument of choice would likely a fast higher resolution scope - more expensive than the pico-scope, or possibly a dedicated DAQ card that can also to the simple data processing (may not be so simple to learn). By filtering stretching the pulse one adds a little noise, but could get away with a slower ADC / scope. There should be some information available from nuclear experiments for similar pulses from scintillators.

It makes sense that interference is less of a problem with fs pulses.

[splin]

Thanks to all of you for your help !

@splin : I looked into the NUCLEO and this noise of 300 LSB is just way too much for what I'm targeting, even though it's 14 bits and 3.6 MS/s. The thing is that I won't try to implement from scratch one of those ADC (no time, and not enough knowledges to do so). Do you know another board with a similar ADC without those bad decoupling issues (and which would not require too much debugging) ?

You must have missed the end of the thread - did you select "more answers"? The 300LSB noise was coming from the ethernet driver ic. Removing power to it by unsoldering a ferrite bead resolved the problem and noise went down drastically:

As you see - I've got it all well within spec, major half of samples are somewhat in 10LSB peak to peak range, while all of them within 30LSB.

That is a 16 bit ADC so 30LSBs p-p noise equates to 11.1 bits noise free. Oversampling by 20 improves that to 13.25 bits noise free. I say oversampling by 20 because I believe the 30LSBs pp noise was achieved when running the ADC at 2MSPS, with single ended input. Two ADCs can be run in interleaved mode to give 4MSPS - 20 x 200k.

At 3.6MSPS noise may be worse; the extra oversampling may more than compensate but it would need testing.

Using differential input mode should improve noise performance by about 0.5 bits. Ideally you would need a fairly fast ADC driver or opamps to handle the high speed capacitive switching load presented by the ADC input. There are loads of good application notes for ADC drivers. Eg. take a look at Fig. 80 in the ADS9224R (16 bit 3MSPS ADC) datasheet on page 51:

http://www.ti.com/lit/ds/symlink/ads9224r.pdf

[MegaVolt]

Any idea of something I could have missed (in the way I want to make the measurement, in the constraints or on the oscilloscope I can use to make it) ?

DMM7510

[exe]

What is the output signal level and impedance? What is the output waveform? I think, as it was suggested before, your measurement bandwidth should be way more than the "fundamental" 100 or 200kHz frequency.

Chances are, you can amplify signal to get above the noise floor of you measurement gear. However, building a fast and precise amplifier is a challenge on its own.

May be you need something like peak detect, and sample and hold? Again, this is a different challenge.

Last but not least, before going this endeavor and investing a lot of time make sure this is really what you need. I mean the measurements. May be, if you describe your research idea, we'll make better suggestions 

[Kleinstein]

The signal comes from a photodiode, likely directly tied to a 50 Ohms terminated cable and at least without much extra capacitance the amplitude can be quite large (e.g. > 1 V).
Without extra capacitance the pulse can be short - a few ns depending on the detector.

With such short pulsed, that could be stretched with come extra capacitance, one wants a relatively fast sampling. More like in the >100 Ms/s to a few Gs/s range. So the fast 10 bit scope already available may not be so bad, if looking at the data the right way and maybe with some pulse stretching. Chances are it gets expensive with more than 12 bits in the > 100 Ms/s range.

I don't think the OP wanted a way to build the ADC board. More like ideally get a suitable DAQ card or scope.

[Andreas]

Hello,

since the pulses are in the 15 fs range I fear that the largest problem is that there is no synchronisation between sender and oscilloscope clock. (otherwise a bandwidth in the 20000 - 50000 GHz range would be needed).
I fear that that what the TE is interpreting as "noise" is only the sampling jitter on the trailing edge of the filtered pulse. (due to too low band with of the measurement sensors).

For a evaluation of the signal something like a synchronous rectifier or lock in amplifier is probably a better solution than a scope.

with best regards

Andreas

[Kleinstein]

The light pulse is in the 10s of fs, but after the photodiode the signal would already be much slower, some ns. If need an additional capacitor in parallel to the photo-diode could be used to stretch the pulse even more.  For the low ns range the 5 Gs/s speed could in deed already be on the slow side and jitter a problem, if the scope does not have suitable BW limitation. 1 GHz with 5 Gs/s is still quite fast, so there can be some aliasing and sensitivity to jitter, if the pulses are really fast.

If looking for pulse to pulse scattering a lock-in-amplifier would not help very much. In addition the pulses are much below 50% duty cycle, which would be more suitable for boxcar averaging not a classical lock-in.
Ideally one would have a fast ADC with some FPGA / DSP for the processing to get the pulse high in real time. There may be such analyses for radiation detectors, though most of them with lower resolution.

With a limited budget my best bet would be with some pulse stretching and than using the existing 10 Bit scope, to transfer data packages of something like 1 µs before and some 100-200 ns for the pulse.