As BD pointed in comments, Caeleste publications page has been updated to include the latest CNES Workshop 2011 presentations. The most interesting one is "A 0.5 noise electrons CMOS pixel" by Bart Dierickx, Nayera Ahmed, and Benoit Dupont. The presentation explains the 1/f and RTS noise reduction principle by cycling the pMOSFET between accumulation and inversion:

150 inversion-accumulation cycles are averaged to reduce pixel noise down to 0.5e level:

The result was measured on the technology demonstrator based on 100um standalone test structure, ~7μm MOSFET area, pixel is used in CTIA mode with >1000μV/e- conversion gain:

Another new CNES Caeleste presentation is "High QE, Thinned Backside-Illuminated, 3e- RoN, Fast 700fps, 1760×1760 Pixels Wave-Front Sensor Imager with Highly Parallel Readout."

According to the paper, the noise is reduced by the sqrt(samples) or I/A cycles.

ReplyDeleteSingle sample noise is ~2e-. In theory, it should take only 16 cycles/samples to reduce the noise down to 0.5e-.

Why is nearly 10X cycles needed to get to this number? Shouldn't the noise after 150 of these samples be closer to 0.15e-?

the multi-sampling can only reduce the un-correlated noises. If the noise is concentrated in the low frequency band, the sqrt law cannot be applied.

ReplyDelete-yang ni

thanks Yang...

ReplyDeleteI thought the whole point of the paper was that by doing the I/A cycling, they are "uncorrelating" the 1/F noise.

Also, it would be nice to see the noise reduction vs. cycle frequency. Getting an accumulation layer is a notoriously slow process, which I assume is reflected in the much longer accumulation time vs. the inversion sampling time in the plots.

I don't know the details in this circuit. But it has been proposed in literature several approaches to cope with 1/f noise, including accumulation, gate pulsing, etc ... Visually this can be seen as you want to make fall the dusts on a wall, so you introduce some viberation to help/accelerate this process.

ReplyDelete-yang ni

If noise is correlated it is not really noise. In this case, you could simply remove the offset by double sampling techniques. Thus, making the noise uncorrelated resp. changing the frequency characteristics is not the real noise reducer in this approach in my opinion. Actually the noise power in the low frequency spectrum is rather lowered by the pulsing because the probablity of trapping events gets smaller. The oversampling simply averages the remaining noise sources.

ReplyDeleteOnly white noise is uncorrelated. Noise sources that are not white do not count as noise sources? Interesting philosophy. =D

ReplyDeleteInteresting blog posting & article: Moments in Time: The First Kodak Moment

ReplyDeletehttp://blog.appliedmaterials.com/moments-time-first-kodak-moment

nice interview, but what is the relationship with this discussion thread?

ReplyDeleteClarification. In the actual setup, we use a low-pass filter to do the "averaging". Thus, although the factor of over-sampling is 150, we do not nicely average these 150 samples (actually 75 for reset level and 75 for the signal level as this includes CDS), but we low-pass filter it. As we must keep the time constant of the LP filter short compared to the reset and signal times, we have time constants that are in the order of ~20...30 inversion/accumulation cycles. Using that, we should thus have a factor SQRT(20)...SQRT(30) reduction, which is "about" what we see.

ReplyDeleteAlso: we cannot exclude that in the 1/f noise there is a fraction that does not obey the McWorther theory. This fraction will not be reduced.

someone said "if noise is correlated it is not really noise". Well... you better stay anonymous. First of all in this context "correlation" means "correlation in time domain". 1/F, RTS, EMI, interferencese are strongly time correlated. And even pure thermal white noise, once passed through a lowpass or band filter, is time correlated.

ReplyDelete"un-correlation" means thus also: are consecutive samples on the same signal statistically completely independent? If so, the accuracy of the average of N measurements improves with SQRT(N).

Mr Dierick, thank you for nice presentation. What is signal on page 19? Is it integrator? Is it signal from page 17? You can give more details about LPF? In noise measurement you pulse transfer 1 time? Is noise on signal p.19 only 0.5mV? Factor reduction is SQRT(16), no? You also measure with real averaging 150 samples? Thank you for clarification.

ReplyDeleteBart,

ReplyDeleteExcellent work, this comes almost 20 years after your paper on RTS reduction by cycling from inversion to accumulation...

Congratulations on your work,

B. Dierickx, E. Simoen, “The decrease of random telegraph signal noise in metal-oxide-semiconductor field-effect transistors when cycled from inversion to accumulation”, J. Appl. Phys., vol.71-4, p.2028 (1992)

and this one ?

ReplyDeleteI. Bloom and Y. Nemirovsky, “1/f noise reduction of metal-oxidesemiconductor transistors by cycling from inversion to accumulation”,

Appl. Phys. Lett., vol. 58, no. 15, pp. 1664-1666, Apr. 15, 1991

Why this paper can not be found on Caelest site??

ReplyDeleteHi, the CNES conference is without proceedings, however, you can find the publication on our website still following this link:

ReplyDeletehttp://bit.ly/rK2ePJ

answer to "Mr Dierick, thank you for nice presentation. What is signal on page 19? Is it integrator? Is it signal from page 17? You can give more details about LPF? In noise measurement you pulse transfer 1 time? Is noise on signal p.19 only 0.5mV? Factor reduction is SQRT(16), no? You also measure with real averaging 150 samples? Thank you for clarification."

ReplyDelete-p.19 each of the "sawtooths" are 150 accumulation/inversion cycles. it is taken at the rightmost point of page 17. The pixel's reset happens just before the steep step. The steep step itself is due to the reset clamping. The sawtooth slope is not due to dark current but to the clamping AC coupling and the transients that happen in the pixel due to the reset and the well pulsing. The charge injection (transfer gate) happens in the middle of the ramp but as this is the dark situation one does not see that step.

-the LPF is a RC circuit with a buffer.

-p.19 the Y-scale is Volts. The spread of the individual "inversion samples" as represented by the green lines (which are powerpoint drawn) should be ~2 mVrms, being thus ~2 electrons RMS (as we have ~1000 µV/e-). The reduction to 0.5 happens afterwards after LPF.

-our setup did not do real averaging but lowpass filtering, see previous post in this forum.

-the observed factor of reduction is ~4. This approaches the expected factor of reduction see previous post in this forum.

-the PPT is here: http://www.caeleste.be/publicaties_bart/2011%20zps.ppt