Full Well Capacity

[marcmccalmont]

Since photosites are shrinking and per pixel DR decreasing, why can't the full well capacity be increased? If the well is a capacitor and its capacitance is a function of surface area why not add a layer of capacitors and run them in parallel?
Marc

PS I know if it was that easy they would be doing it but just trying to grasp the limitations in sensor design

[PierreVandevenne]

Wouldn't help because, when you think about it, what matters is the capacitance that arises from the formation of the depletion region: putting larger capacitors behind it wouldn't help. And given that's layer against layer, I can't see of they could be run in parallel. It's not as if there was a collecting zone and a wire cleanly transferring electrons to charge a capacitor. What could improve well capacity per unit of area of a silicon based sensor would be a "better" doping method. But that's pretty esoteric research (I remember reading about an advanced sensor which had improved doping but unfortunately low stability). You'd think that making the depletion region larger would increase it's capacitance, but it wouldn't help with the transfer of charges and would also generate more dark current. The limit is QE (wich can be very high), physical thickness and then everything that relates to the transfer/conversion of charges: it's much better to halve the noise than to double the well capacity for a given area anyway (assuming it could be done).

[ErikKaffehr]

Hi,

Marc, thank you for asking!

Pierre, thank you for explaining. Just another question, halving read noise would not affect shot noise, is that correct?!

Best regards
Erik

Wouldn't help because, when you think about it, what matters is the capacitance that arises from the formation of the depletion region: putting larger capacitors behind it wouldn't help. And given that's layer against layer, I can't see of they could be run in parallel. It's not as if there was a collecting zone and a wire cleanly transferring electrons to charge a capacitor. What could improve well capacity per unit of area of a silicon based sensor would be a "better" doping method. But that's pretty esoteric research (I remember reading about an advanced sensor which had improved doping but unfortunately low stability). You'd think that making the depletion region larger would increase it's capacitance, but it wouldn't help with the transfer of charges and would also generate more dark current. The limit is QE (wich can be very high), physical thickness and then everything that relates to the transfer/conversion of charges: it's much better to halve the noise than to double the well capacity for a given area anyway (assuming it could be done).

[PierreVandevenne]

Shot noise as the result of the Poisson process that leads to the variability in the flow of incoming photons is indeed independent of the rest (which is not to say there aren't other elements in the chain that could also be see as Poisson processes and have their own "shot noise"). If one wants to maximize SNR, one needs more samples. If counting photons is the goal, higher QE (higher chance of detecing the photon), real aperture (getting more photons at the input size), real focal length and sensor area (how the collector spreads the photons on the sensing area) and integration time are the main factors.

Shot noise as defined above is a property of the signal all the rest are the properties of the measuring instrument. But I am sure you know all that...

My comment on the read noise was more on the practical/industrial/business side of things. Given that sensors with a well capacity of 25000e and a read noise of 14e were quite frequent a few years ago, it makes more sense for manufacturers to pursue an attainable reduction of the read noise than to focus on esoteric doping. Likewise, maximizing the effective sensing area by minimizing dead space between actual sensels and trying not to lose photons that have already been captured by optimizing micro-lenses were, and possibly still are, areas where big concrete gains could be made. And that's without even considering the additional issues they have to deal with when using CMOS sensors.

Of course, that doesn't exclude a new wonderful doping method that would increase the depletion area's "storage" ability (I don't know how close we are from the theoretical limit, my guess is "not far" because we already have very high QE, but you never know) and conceptually Marc's question makes a lot of sense.

Note: my comments are based on a fairly decent understanding of how things work in the base theoretical sensor. While I know it is trendy to be "authoritative", I make no such claim. ;-)

[ErikKaffehr]

Hi,

Just to make thing more clear:

- Quantum efficiency corresponds to sensivity. Higher QE means higher sensivity
- FWC essentially defines shot noise (once FWC is fully utilized, that is ETTR exposure)
- Read noise only affects the darkest parts of the image

So, reducing redout noise would not have any significance on sky for instance on a well exposed image, but could have large impact on rendition of the darks?

Is this correct, or have I missed something? I'm aware of the fact that I ignored pixel nonuniformity.

Best regards
Erik

Shot noise as the result of the Poisson process that leads to the variability in the flow of incoming photons is indeed independent of the rest (which is not to say there aren't other elements in the chain that could also be see as Poisson processes and have their own "shot noise"). If one wants to maximize SNR, one needs more samples. If counting photons is the goal, higher QE (higher chance of detecing the photon), real aperture (getting more photons at the input size), real focal length and sensor area (how the collector spreads the photons on the sensing area) and integration time are the main factors.

Shot noise as defined above is a property of the signal all the rest are the properties of the measuring instrument. But I am sure you know all that...

My comment on the read noise was more on the practical/industrial/business side of things. Given that sensors with a well capacity of 25000e and a read noise of 14e were quite frequent a few years ago, it makes more sense for manufacturers to pursue an attainable reduction of the read noise than to focus on esoteric doping. Likewise, maximizing the effective sensing area by minimizing dead space between actual sensels and trying not to lose photons that have already been captured by optimizing micro-lenses were, and possibly still are, areas where big concrete gains could be made. And that's without even considering the additional issues they have to deal with when using CMOS sensors.

Of course, that doesn't exclude a new wonderful doping method that would increase the depletion area's "storage" ability (I don't know how close we are from the theoretical limit, my guess is "not far" because we already have very high QE, but you never know) and conceptually Marc's question makes a lot of sense.

Note: my comments are based on a fairly decent understanding of how things work in the base theoretical sensor. While I know it is trendy to be "authoritative", I make no such claim. ;-)

[PierreVandevenne]

- Quantum efficiency corresponds to sensivity. Higher QE means higher sensivity
- FWC essentially defines shot noise (once FWC is fully utilized, that is ETTR exposure)
- Read noise only affects the darkest parts of the image

So, reducing redout noise would not have any significance on sky for instance on a well exposed image, but could have large impact on rendition of the darks?
Is this correct, or have I missed something? I'm aware of the fact that I ignored pixel nonuniformity.

QE: yes: efficiency of photon absorption and their conversion to an electron-hole pair

shot noise: I wouldn't say it is defined by FWC as it is an observed phenomenon, one could say a fundamental property of nature. Of course, the bigger the actual sample, the lower the shot noise will be. FWC plays a role, but only if it is exploited.

read noise and FWC essentially define the dynamic range of the sensor. An hypothetical sensor with a FWC of 25.000e and a read noise of 10e will discriminate at most 2500 levels and have a DR slightly above 11 bits. Halving the read noise to 5e will will give a DR slightly above 12 bits. Reducing it 1e will give you roughly 14.4 bits of DR. Most of the gains in DR in the recent Sony sensors have been achieved by reducing read noise. (also ignoring their pixel non-uniformity work, which is of course very important from a practical point of view in a CMOS sensor)

shot noise: we'll have to live with it forever, the best we can do is to expose fully to best exploit the linear response range in the sensor (as in ETTR). Dark areas will suffer equally but it is of course better to have 13 bits/stops above the floor thn to have 9 bits/stops above it. For photographic applications, a high DR will allow the manufacturers to cut the signal and give us a visually nice uniform darks with room to spare. read noise: going from 2 to 1 yields the same improvement as going from 10 to 5.

[EricWHiss]

Interesting topic and information  -  of course I wish the sensor manufacturers would rather go the other way - bigger sensor footprint with bigger sensels and not just for SNR but also to avoid diffraction effects.  I'd like to see a 60mp full frame 6x6 or 6x7 sensor.