Friday, August 20, 2010

Paper watch: Image sensor papers at the IEEE TED

The latest issue of IEEE Transactions on Electron Devices carries three (at first sight) interesting papers.

The first one is "eLeNA: A Parametric CMOS Active-Pixel Sensor for the Evaluation of Reset Noise Reduction Architectures". The abstract reads:
We present a novel complementary metal–oxide–semiconductor (CMOS) active-pixel sensor imager that incorporates different reset schemes to achieve lower reset noise levels. The sensor, eLeNA, features a 448 $times$ 512 array with a pixel pitch of 15 $muhbox{m}$, fabricated using a 0.18- $muhbox{m}$ CMOS process. Fourteen sections and five different reset methods were employed. Without using pinned diodes, we implanted structures for correlated double sampling. A noise of 6 $hbox{e}-$ is measured with a conversion gain of 49 $muhbox{V/e}-$. We will discuss various applications for the reset method that achieved the best overall performance, considering leakage current and read noise.
The second one is: "Simulation and Measurements of Stray Minority Carrier Protection Structures in CMOS Image Sensors". The abstract reads:
Recently, the rapid growth of CMOS technology has made it possible to integrate more periphery circuits into a CMOS image sensor. Although these periphery circuits improve image quality, they also lead to the generation of more stray minority carriers. Because the number of stray minority carriers is proportional to the frequency, the affected region increases with increasing operating frequency. Placing an appropriate absorber between the periphery circuits and the pixel has traditionally been accepted as the best solution for this issue. Four protection tactics were simulated in software and verified in a fabricated CMOS image sensor. The imager was fabricated using TSMC 1-poly 6-metal 0.18-$muhbox{m}$ process technology. On this chip, ten noise sources outside the pixel array were used to verify the effectiveness of the protection tactics in off-array tests, whereas in-pixel noise sources were used in in-pixel tests. To quantify the influence of stray minority carriers in the off-array test, the maximum depth of an affected region (DAR) was measured in a processed binary image. The off-array experimental results revealed that the DAR increased with either an increased operating frequency or a decreased separation between the noise source and the pixel array. The DAR of the affected pixels can be eliminated up to 48.1% and 23.8% by using the N-well and N-diffusion guard rings, respectively. The in-pixel experimental results have shown that the N-diffusion digital pixel implementation reduced the noise by 63.2% while only increasing the area by 10.68%. Detailed information about the effectiveness of different protection tactics in an imager design was collected in this paper. This paper can potentially provide a reference to help imager designers choose an appropriate protection tactic.
And the third one: "Per-Pixel Dark Current Spectroscopy Measurement and Analysis in CMOS Image Sensors". The abstract reads:
A per-pixel dark current spectroscopy measurement and analysis technique for identifying deep-level traps in CMOS imagers is presented. The short integration time transfer gate subtraction experimental technique used to obtain accurate results is described and discussed. The activation energies obtained for molybdenum (≈0.3 eV), tungsten (≈0.37 eV), and the phosphorus-vacancy (E-center) (≈0.44 eV) trap levels in silicon match published results measured with other techniques. The Meyer–Neldel Relationship (MNR) was observed between the Arrhenius preexponential frequency factor and activation energy. The trap capture cross-sectional calculation methodology using the MNR is presented. The cross sections of molybdenum, tungsten, and the E-center were calculated as ≈1 × 10−16 cm2,
≈1.5 × 10−16 cm2, and≈2.5 × 10−16 cm2, respectively, at 318 K. The data obtained suggest electric field enhanced emission, and Poole-Frenkel barrier force lowering of E-center defects occurs in the pinning implant regions. It is proposed that a changing Fermi level results in the correct activation energies being obtained below half the band gap and that the dark current measurement process is affected by the measurement time result of statistical mechanics. It is also tentatively suggested that, in this case, the observed MNR is a geometric relationship and not due to a physical process.
In case the abstract is a bit confusing, the authors conclude that: "The measurement and analysis technique presented could be useful to the image sensor industry for diagnosing fabrication plant contamination. It also has potential applications for the study of radiation induced traps in CMOS and CCD imagers."


  1. I am curious, does anyone have an opinion about the discussion part of the dark current spectroscopy paper?
    Does this theory seem correct to you?

  2. I am not an expert on that particular topic, so I myself cannot comment without going into more detail through their explanation.

  3. I see at least two points that need to be clarified:
    - unusual dark current activation energies are reported (below midgap). To justify these values, the author use the fact that the Fermi level can exhibit activation energy lower than 0.56eV. The problem is that the temperature dependence of dark current (i.e. the dark current activation energy) is not the same physical parameter than the Fermi level evolution with temperature (i.e. the Fermi level activation energy). I think that there is a mix-up between free carrier generation rate in the depleted zone and the Fermi level (which are related but not the same).
    - the second point easier to verify is the fact that to justify their theory the authors state that the generation rate of electron in the photodiode (eq 10) is equal to the electron density at the thermal equilibrium (eq 11), which is not true, and can be simply verified by the fact that the eq 10 dimension is e-/s and the eq 11 dimension is e-/cm^3 (if we neglect the obvious error on the redondant T^3/2 factor). This clearly shows that the authors think that the generation rate in e-/s in the depleted region is exactly the same thing than the number of filled electronic states at thermal equilibrium (in the neutral region for example).

  4. Hi,

    I'm the author of the 3rd paper and I welcome all comments and criticisms. Thanks for your interest in the paper. Sorry for finding this comment thread a little late - I don't expect any of you to re-read this.

    I proposed the theory in the discussion part of the paper to explain the measured results that I achieved. Activation energy results less than half the band gap were obtained on two different image sensors which matched the known activation energies of Mo and W. Mo and W were also known to be present in both image sensors. Given this fact I tried to explain the results with the proposed theory.

    First, it is important to realise how 4T pinned photodiodes work. Pinned photodiodes are depleted of electrons by the application of the reset potential and then isolated from all contacts during integration. Therefore, this is not the same as a reguar p-n diode with contacts on both sides at constant applied bias. The generation of photoelectrons or dark electrons, changes the bias in the photodiode which means that in the dark, the generation of electrons changes the generation environment in the PD. This means the first generated electron sees a different environment from the second, and so on.

    Under these transient non-equilibrium conditions activation energy analysis becomes difficult. The analysis is heavily based on pages 465-475 of Shockley's book which shows that the measured activation energy depends on the fermi level under similar strange conditions when the fermi level of a semiconductor is changing, which also applies to the conditions in a pinned photodiode (see above).

    On your second point, the equations are scaled according to the relative size of the defect in the PD, which according to my maths, means that cm^3 is cancelled. Moreover, T^3/2 factor cancels as (10) and (11) are equated.

    The key point of the paper is really that I found that short integration times ~10ms give activation energies which correlate to traps and that they are less than Eg/2 because of the influence of Fermi statistics. It is proposed that this is because longer integration times produce too much change in the conditions in the photodiode such that the equilibrium assumption of statistical mechanics falls apart.

    I'd suggest to try measuring the activation energies with ~10ms integration times and see what you get for hot pixels.

    For any further comments/discussion, please feel free to email me.




Note: Only a member of this blog may post a comment.