Saturday, June 08, 2013

Concert Lasers Damage Image Sensors

There is a growing number of videos showing how image sensors can be damaged by concert lasers:







One would expect the excessive energy to burn color filter first, while the underneath layers be able to withstand quite a significant heating. But it looks like the a whole column and/or row stops working, pointing to the electrical nature of the damage. The laser is green and should not cause the oxide charge accumulation in MOSFETs and STI. So, what can be the mechanism of such a damage?

My first guess was that a large photocurrent on the transistor diffusions in the array exceeds elecromigration limits blows some metal or via. Most CIS processes offer about 0.5mA per um of metal width for the array metals. Then, assuming the metal width in RED and Canon large sensors is 0.2um, the maximum allowed current should of order of 100uA. To get to this current, the photon flux should be about 2.0e15 ph/s per pixel. Seems way too much for these concert lasers.

And even if we managed to reach the electromigration limit of 100uA, the metal is supposed to last a long time at that current, such as 10,000 hours or more, assuming the sensor is colder than 70-80C. So, my next guess was that these large sensors get much hotter in video mode. If my memory serves me, the electromigration increases 3 times per 15C of temperature rise. Still, this does not sound enough to fit to the electromigration theory.

So, the next guess was that the laser heats the pixel locally, to 200-250C or so and then the electromigration limit gets exceeded. Currently, this seems to be my best guess. If true, the workaround should be to put the current limiters at each column and row, and also to the pixel VDD lines.

Anyone has a better explanation for the damage?

Update: Yet another Canon video of laser show published at Youtube on June 21, 2013, showing a similar column and row damage.

39 comments:

  1. Reset lines may be opened? Row reset driver fets column amplifier reset fets overloaded, fail

    ReplyDelete
    Replies
    1. But how can it fail permanently? To me, the driver should recover as soon as the load is back to normal.

      Delete
  2. The Red EPIC camera is getting hit with the full poutput of an Argon ion(?) laser that's probably Class3B (< 0.5W) or Class 4 (> 0.5W) with a diverging beam. Both classes are used in big concert venues (up to 40W!). You can see the beam divergance before the "hit" to make the scatter so these are probably Class 4 lasers.

    Assuming this is a Class 4 1W laser (it could be more than an order of magnitude bigger than that). A quick calculation (if I got it right) gives 2.7 x 10^18 photons per second for a 1W 532nm laser.

    http://www.wolframalpha.com/input/?i=532nm+%2F+%28h+*+c%29&a=UnitClash_*c.*SpeedOfLight.dflt--&a=UnitClash_*h.*PlanckConstant--

    The whole ~10 to 20mm diameter beam is taken by the aperture of the camera (this is pretty much a worst case situation and something that should have been planned for) with the spot image on the sensor covering 1% of the pixels vertically and horizontally (order of magnitude handwave but look at the little black spot on the video around the time of the damage). So that spot covers perhaps 1000 pixels (i.e. 5120 x 2700 x 0.01 x 0.01).

    Which gives a flux per pixel of perhaps 3 x 10^15 photons per second :-)

    A smaller spot or more powerful laser would easily beat that by and order of magnitude or more.

    I rather suspect that thermal effects might be larger that you guess too. Dumping a watt or more into a spot tens of pixels in diameter is going to locally raise the temperature very rapidly especially as most of it will be absorbed in the top 10 microns or so.

    The final thought: the structures at the surface of the sensor are of a similar order of magnitude to the wavelength of the light. With coherent light at the right orientation you might get "standing wave" effects with very high electrical field strengths in the metallization that might heat the metallization preferentially. Just an idea.

    ReplyDelete
    Replies
    1. Regarding the thermal effects alone, I'd expect them damaging color filter first. So, we would see lines and spots of burned color filter everywhere. But it appears to be not the case...

      But from your calculations it certainly can be a combination of heating to 200C and the electromigration.

      Delete
    2. Upon more thought, if indeed the flux is that high, there is no need to assume heating to explain the things. Electromigration alone could do this. The current from all pixels along the column, the row or vdd line is added up. It looks like the current limiters on row, column and pixel vdd lines could save the chip.

      Delete
  3. 1. The bad music puts the pixel on edge and the light pushes it over.
    2. The laser light is too bright. (Remind me to bring my sunglasses to the next concert)
    3. There is bipolar latchup occurring somewhere in the pixel.
    4. Some gate voltage is operating near a reliability limit, and light induced excess charge results in avalanche breakdown of the gate oxide.

    I favor 3 over 4. Designers have forgotten about CMOS latch up generally speaking. Still, a real analysis requires a pixel structure and TCAD sim.

    ReplyDelete
    Replies
    1. Me too thought about latchup. But one needs a pmos for thyristor-like structure. There are pmos devices on the edge of the array, in the readout or row driver. However, the videos show damage in the middle of the imager. I can't find a mechanism for the latchup in the middle.

      Delete
    2. Latchup can occure when there is a 4-layer structure (NPNP) and absence of current limitation. In 4T pixel design, there is no 4-layer structure to my knowledge even with substrate connection. If we look at the damage, it's alomost always loss of line or column, for me it should be the metal to metal breakdown. By taking the space between metal line in sub-micron process, the electrical field is not low.

      Other hypothesis is that field oxide breakdown. This was one of main limitation of fiber optic transmission range, because too strong laser power can create micro explosion inside the fiber and make it damaged.

      Vladimir, you can find some videos on youtube showing point-like blemish which looks like ML/CFA damage.

      Yang Ni

      Delete
    3. Yes, for the latchup we need gnd, vdd and the pnpn structure. One can find this combination at the edges of the array, close to row or column circuits, various guard ring, etc.

      I did not hear about light-induced metal to metal breaknown. What is the mechanism of that?

      Delete
    4. You need to look in a 3 dimension path for a possible culprit, not just vertically. For example, plenty of lateral npnp or v.v. structures. But you need something that goes to a ground plug or heavily doped substrate ground as well. Without knowing the structure exactly, it is difficult to speculate. Also, as I said elsewhere, if there is latent defect, it gets much much harder to sort out the culprit.

      Delete
    5. One advantage of nmos logic over CMOS one is that it does not have a latchup. It sits in my head since college years, I did not give it much thought since then.

      Delete
    6. I don't know for the RED sensor but many of the Canon SLR sensors are made with p- epi layers on n++ substrates. There you have vertical pnpn structure which can form the latch-up thyristor:
      - First bipolar by p+ pinning layer / photodiode n-implant / p- epi layer.
      - Second bipolar by photodiode (or other n+) / p- epi layer / n++ substrate.
      The thyristor switches on by the high light intensity, and then sources a very high current until something burns.

      Delete
    7. Been there, considered this. The first bipolar has its emitter and collector shorted, that is, p+ pinning layer is shorted to the p-epi. In this configuration I can not see how the thyristor can be fired up.

      Delete
    8. If I assume a 4T pixel, the breakdown of the selection transistor may have occured shorting column to gate and line driver explaining column and line dysfunction. I could also suggest that the driver of the select transistor is on the right ... The two columns may be eventually an artifact due to de-mosaicing.

      Delete
    9. But why the strong light initiates the select transistor breakdown?

      I can imaging that under the strong light all the diffusions go negative to -0.6V, may be. Then, the select transistor gets 0.6V overvoltage. Then the same should happen to transfer gate too. Then we should assume that the the photocurrent is strong enough to provide energy for oxide rupture. Can it be enough for the breakdown? In my experience, 0.6V overvoltage is too small for that. But given the large statistics in the array, who knows...

      Delete
    10. I agree that it is not enough for breakdown. The select transistor is the only one having its channel regularly pinned to ground . In the middle of the array the transistor is in high impedance and the photosensitive poly-silicon composing the gate may accumulate enough charge to breakdown the MOSFET oxide ...

      Delete
    11. The poly gate does not accumulate photocharge, not in visible band, anyway. The electron-hole pairs in the poly just recombine, as minority carriers do not have an escape path. If one shines a far UV light, the high energy carriers might jump over the gate oxide barrier. It's about 3eV or 3.6eV high, if I remember correctly. Needs a lot of energy to go over it.

      Delete
  4. The funny thing is, this happens only once! At least according to comments. So what could be a one time phenomenon? A weakest link?

    ReplyDelete
    Replies
    1. Actually, I think this is an important observation...not all sensors fail, and when they do, it seems a single pixel fails despite a sweep of the spot across many pixels.

      That does point to a sort of latent defect and not a general problem. Tracking this down could be very difficult since under the latent defect model, something in the pixel structure is not what was intended by the design. Physical failure analysis would be warranted.

      I mentioned this to a friend who produces image sensors and his comment was essentially "I think this voids the warranty."

      Still, with cameras it is just a cost factor and nuisance, with automated navigation systems, and other life-safety systems it could be disastrous.

      Delete
  5. Should we not first analyse what defect is occuring before trying to explain it? In the first video, here's what I see:
    1) a column is broken
    2) it looks like there's a point defect around 20% from the bottom
    3) at that location and to the left, there's a broken row.
    4) at the right, there's no broken row so it seems, but at some moments I do see something that looks like a broken row, could be an effect that is related to saturation
    5) at the right, all of the array exhibits reduced sensitivity

    The EPIC probably has row drivers on both sides which could explain some level of asymmetry from left to right.

    The asymmetry on the column isn't clear to me and would probably need a column read line the pixel supply to be defective somehow.

    The explanations in this blog all evolve around destroying a metal wire through electro-migration, but then we should also have a theory as to how an "open" on a wire would cause this kind of artefact.

    We can help RED in this blog, they just need to show their pixel layout (wires only) and explain how row control signals, pixel supplies and column lines are connected to the periphery. Chances are that the fact that a localized defect leads to such a wide spread impact on the array, is due to small mistakes in their array design methodology.

    ReplyDelete
    Replies
    1. Or perhaps localized defects introduced during processing?

      Delete
    2. I do not believe RED shows us its pixel, but Chipworks almost certainly did a reverse engineering of Canon 5D-MkII sensor, shown in the 3rd video. May be Chipworks can share the pixel layout with us.

      Delete
  6. Here someone destroys two cameras and half of the living room ;-) :
    http://www.youtube.com/watch?v=EoLR2LzHO-M
    (minute 3 and 6, 3 pixel defects and one sensor blown completely)

    More camera toast, but that is not surprising with 2W Laser:
    http://www.youtube.com/watch?v=KcazcEDF0WY

    I never liked laser beams in my eyes at shows... no matter how weak and how short.
    They easily burn stuff even in the 10mW range:
    http://www.youtube.com/watch?v=woiTedSKPrk
    http://www.youtube.com/watch?v=d1pXhhtUaso


    ReplyDelete
  7. Another possibility is that the pixel is hit by the light during the short time when one of the gates is turned on -- say the transfer gate is high for charge transfer from photodiode to floating diffusion. This is a time when fields are high, but because of the short time period would in normal operation have little impact on the device. But if you suddenly swamp the pixel with lots of free carriers it could create a hot carrier disaster and possibly destroy the gate oxide. This need for the triple witching of (a) the appropriate short duty-cycle clock alignment, (b) the sudden illumination by the high power laser and (c) a possible marginality due to a local defect would account for why it is a rare and single pixel event.

    ReplyDelete
    Replies
    1. Usually, transfer gate has LDD on the FD side, so the field there should not be bigger than in regular transistors. Also, the length of transfer gate is usually longer than in other transistors in the array, thus it should suffer less from hot carriers than the other transistors.

      Delete
  8. Clearly, these grooves were just too groovy for the relatively square image sensor to handle. It's tiny synapses were blown, as one would expect of something so enamored of rigid periodic structure. Sometimes you just have to let your diffusion float with the music, bro! Feel those beats overflowing vertically into your substrate, man!

    :)

    ReplyDelete
    Replies
    1. Heavy metal man; heavy metal.. thats the answer. Sometimes being negative helps too..to cancel all those positive species..

      Delete
  9. Assuming the laser light is pulsing with very fast edges then that edge (when converted to electrons) could jump through parasitic capacitance and put electrons in all sorts of places they shouldn't be.
    John

    ReplyDelete
  10. This may be something like heavy ion in space application, large photoncurrent discharge some memeory node, like SPI, so change the background (offset) of the image. In some cases, it triggers latchup, similar as SEL and damage some driving circuitary (thus the row loss), or a particular readout chain (thus the column loss).

    Would be intersting to see how this affect some rad-hard CMOS sensors.

    ReplyDelete
  11. the green color at 532 nm is obtained with the doubling an 1064 nm laser with an efficiency that remains below 50%. the filtering the infrared wavelength is not always that good and you have to add this energy to compute the whole figure. the bayer filter is perhaps even nearly transparent at that wavelength. the peak power is even quite large because the laser is not continuous but flashes perhaps at kilohertz rate.

    ReplyDelete
  12. At my last company we had an RMA of a part that was used in a 3-sensor pro-camcorder/ENG camera that had what looked like melted pits a few pixels across in the array area (in a few locations).

    We could never work out the cause of the failure, but speculated that it had been pointed at a laser.

    By coincidence, I was backstage at a dance music festival one time and saw that camera model being used.

    ReplyDelete
  13. Interesting story. Were all 3 sensors melted at about the same spots? Or just one of the sensors?

    ReplyDelete
  14. An 3 Imager & Broadcast application will implicate the Sony B4-Mount and with this an 2/3" Imager Format. This results in an Pixelsize of approx. 5µm² vs. Fillfactor.
    Common to all will be the colour separating prism - manufactured only by Canon and Fuji. Main orientation for the Imagers will be: Red on Top, Blue on Bottom and Green in direct optical Axis. (see: http://www.alt-vision.com/op400.jpg)
    This is the reason, why in most cases only the green Imager is affected.

    I dont think that the 1064nm pumping wavelenght is involved in damage. Most Broadcast cameras have an sharp IR-Cutoff Filter at approx. 700nm.

    Please keep in mind, that some of these professional lasers are not operated only in CW! They use techniques as Q-Switching and/or Mode-Locking to increase visibility. The resulting average-power maybe in the 10...100W range, but the Peak-Pulse Power will be in kW region.

    This plus a typical wideangle-shot ( wich increases the depth-of-field dramatically ) will focus the beam perfectly on only one Pixel. I guess, an microscopy investigation will show ablation or sputtered material.
    I'll try to fetch one of these destroyed imagers ;-)

    There only a few manufacturers, wich can rework an damaged Optical Block. Grass Valley (Breda/NL) will do this for their customers and replaces and realigns only the damaged part. Cost approx. 7.500€ / Imager and add the same, if separating prism is affected too.

    Actually we have an pixel defect, where only 5 Pixels exactly vertical aligend are affected. Production facility believes, this maybe an laser affected fault - but this is not confirmed till today.

    ReplyDelete
    Replies
    1. Minor correction. A lot of separation prisms are made by Nitto Optical in Japan. A very large manufacturer os prism assemblies (prisms with aligned CCDs is Toshiba) Optec in Italy makes a few special prisms and many optical companies in China will make them on special order.

      Some of these are built red first, then blue, then green straight through.

      Delete
  15. forgotten remark:
    Sensitivity of High-End DLP projectors against laser-rays is in the same or higher range compared to CCD or CMOS sensors!
    Always keep their lenscaps attached when lasers scanning!

    ReplyDelete
  16. Hmmm... 9-11 Mv/cm is GOX ten year breakdown. Say 20A GOX on imager 5-6v for 400usec or so. Not sure about local heating of imager, this may shorten GOX breakdown times but 5-6V may be hard to get since recombination kinda limits us. But, if we inject enough e- to sub to locally invert channels, especially in row dark column area, maybe there a row driver NCH gets turned on when its gate is low (and the compliment PCH on) to blow the row driver.

    ReplyDelete
  17. TrueSense patent on CCD latch-up work-around for bright light damage during electronic global shutter operation

    http://image-sensors-world.blogspot.ca/2012/05/failure-mechanism-in-ccd.html?m=1

    ReplyDelete

All comments are moderated to avoid spam and personal attacks.