Access Official Technical PDF

Share Summary:LinkedInEmail

Detecting a single photon with a silicon image sensor is fundamentally a problem of measuring a very small electrical signal: in visible and near-infrared light, one photon can generate just one electron-hole pair. A single-photon CMOS image sensor based on repeated, non-destructive signal sampling offers a way to distinguish that charge without relying on internal avalanche multiplication. The approach could support low-light imaging where both sensitivity and readout speed matter, though its performance depends on balancing noise reduction against sampling time and charge-transfer effects.

Why single-photon detection requires very low noise

Every image sensor has readout noise, which can obscure the signal from a single electron. One route to photon sensitivity is to amplify charge inside the sensor before converting it to voltage. Single-photon avalanche diodes and electron-multiplying CCDs use this general strategy, but the source notes drawbacks including dark counts or dark current, after-pulsing, spurious charge, and excess noise.

An alternative is to reduce the readout noise itself. The design study uses 0.15 electrons RMS equivalent noise charge (ENC) as the threshold for reliably distinguishing single electrons. Instead of requiring each individual measurement to reach that level, the proposed pixel measures the same stored charge several times and averages the results.

How non-destructive sampling reduces read noise

The principle is related to the skipper CCD technique. In that method, charge is moved beneath a sensing gate and measured repeatedly without being consumed by each measurement. When the measurements are statistically independent, averaging reduces the random readout noise approximately in proportion to the square root of the number of samples. More samples can therefore improve sensitivity, but they also add readout time.

The proposed CMOS pixel combines two structures:

  • A pinned photodiode (PPD) collects photo-generated charge. Its pinning implant supports low dark current, while a separate, small sense node enables high conversion gain.
  • A buried-channel (BC) CCD structure stores and transfers the charge between gates so the signal can be sampled repeatedly.

After charge moves from the PPD into the buried channel, a sense gate measures it capacitively. A source follower buffers the gate voltage for readout. The charge is shifted between two storage regions and returned for another measurement; reset and signal levels are compared using correlated double sampling. Once enough samples have been collected, the charge can be read conventionally or cleared before the next exposure.

What the CMOS design changes

A major potential advantage over a skipper CCD is parallelism. CMOS image sensors commonly read a row of pixels at the same time using column-parallel circuitry. This can substantially shorten the total readout compared with architectures that rely on one or a small number of outputs. CMOS processing also allows a compact sense node and high conversion gain, reducing how many repeated measurements may be needed.

TCAD simulations and layouts for a 180 nm CMOS image sensor process were used to examine the concept. In one example layout, the estimated sense-node capacitance was 2.63 femtofarads, corresponding to a conversion gain of 61 microvolts per electron. Under the study’s assumed noise conditions and a 1 MHz sampling frequency, the modeled design required 34 samples to reach 0.15 electrons RMS ENC. The authors estimated a line rate of nearly 30,000 rows per second, equivalent to about 30 frames per second for a one-megapixel sensor.

These are projected results from simulations and design estimates, not measurements from a completed prototype. Actual results would depend on the CMOS process, transistor noise, achievable conversion gain, sampling rate, and implementation of the readout circuitry.

Key trade-offs: speed, charge transfer, and dark current

Sampling time and conversion gain

More samples lower the averaged electronic noise, but repeated measurements extend the time needed to read each row. The analysis shows why conversion gain is important: a higher gain improves the voltage signal associated with each electron and can reduce the number of samples needed. Faster sampling can also help, although transistor noise—including low-frequency 1/f noise—affects the benefit. Designers must assess these factors together rather than optimizing sample count in isolation.

Charge transfer and image lag

The charge must move many times through the buried channel. Traps in that channel can capture electrons, creating charge transfer inefficiency (CTI) if captured charge is not returned to its original signal packet. The study discusses this as a potential limitation, while noting that capture probability for small signals depends on trap behavior and may be lower than a worst-case estimate. Incomplete transfer from the PPD, known as image lag, is another consideration; established PPD design techniques can reduce it.

Dark current and capacity

Because the buried channel operates differently from the pinned photodiode, its dark current can be higher. Dark-generated electrons add noise that averaging cannot remove. Short exposures and cooling can help limit dark-current effects, with more significant cooling potentially needed for long exposures. The simulations also indicate that the buried-channel structure constrains full-well capacity: widening the channel can increase charge capacity, but may reduce conversion gain.

Readout architecture and practical implications

Repeated measurements create additional data that would be inefficient to send off-chip sample by sample. The proposed sensor architecture therefore places digital accumulation in the column circuitry, averaging samples before the resulting pixel values are multiplexed to the outputs. A dual-slope conversion approach could also support accumulation. This keeps the external data stream focused on averaged image values, while adding circuitry to the sensor’s readout path.

For imaging engineers and technical decision-makers, the concept is most relevant where very low-light sensitivity justifies more involved pixel and readout design. It illustrates how sensor architecture can trade time and circuit complexity for lower electronic noise. The approach does not eliminate other constraints: dark current, charge transfer, full-well capacity, and the target frame rate remain part of the system-level design.

Conclusion

Multiple non-destructive sampling offers a CMOS-compatible path toward single-electron readout by averaging repeated measurements of stored charge. Pairing a PPD with a buried-channel transfer structure may combine low-dark-current charge collection with the repeated sampling needed to reduce read noise. Simulations suggest that CMOS parallel readout could make this approach faster than comparable skipper CCD architectures, but prototype validation is essential. The practical value will depend on achieving the required noise and conversion gain while managing sampling time, charge-transfer efficiency, dark current, and charge capacity.

Access Official Technical PDF

About The Author