A High-resolution Colloidal Quantum Dot Imager by Monolithic Integration


 Near-infrared (NIR, 0.7–1.4 µm) imagers have wide applications in night surveillance, material sorting, machine vision and potentially automatic driving. However, limited by the high-temperature processing and requirement of single-crystalline substrate, so far flip-chip is the dominant way to connect infrared photodiodes and silicon-based readout integrated circuit (ROIC) to produce infrared imagers, suffering from complicated process and ultra-high cost and hence limiting their widespread applications in the market. Here we report the monolithic integration of colloidal quantum dots (CQD) photodiodes with complementary metal-oxide-semiconductor (CMOS) ROIC, operating as a low-cost and high-performance imager. The CQD photodetector is well designed with a CMOS-compatible structure, demonstrating a response spectral range of 400–1300 nm, a detectivity of 2.1×1012 Jones at room temperature, a -3dB bandwidth of 140 kHz and a linear dynamic range over 100 dB. The CQD imager can identify materials, inspect apple scar and veins with a large size of 640×512 pixels and a spatial resolution of 40 lp/mm at a modulation transfer function of 50%. Monolithic integration significantly reduces the cost without sacrificing performance, thus providing huge potential for the ubiquitous deployment of infrared imagers.


Introduction
Imagers, by integration of photodetector array with silicon readout integrated circuit (ROIC) to convert incident light into electrical signals, are the foundation of light-sensing technology widely used in machine vision [1][2][3] , security monitoring 4-6 , bioimaging 7 and other emerging elds. For visible imagers, homogeneous silicon photodiodes are monolithically integrated with silicon ROIC, dominating visible imaging with ultralow cost and excellent performance. However, for infrared imagers, the fabrication of InGaAs, InSb and HgCdTe photodetectors need high quality single-crystalline substrates and high temperature vacuum processing, ruling out their direct integration with ROIC. Therefore, these photodetectors are exclusively heterogeneously integrated with silicon ROIC via wire-bonding or indium solder bumps for infrared imaging at exponentially increased cost and impaired performance [8][9][10] . The prohibitive price is the main reason that such infrared imagers have their limited application in military and industry and fail to penetrate into everyday life such for cell phones despite they could provide much valuable information unseen by the visible imager.
Colloidal quantum dot (CQD) is one kind of solution-processed nanocrystal possessing a size tunable bandgap covering from visible to long-wavelength infrared. Monolithic integration of PbS CQD with silicon ROIC promises high resolution and cost reduction, thus providing the most affordable way for large scale deployment of infrared imagers. Optimization of CQD photodiodes are well studied yet their integration with ROIC for imaging is much less reported. PbS CQD was rst reported to sensitize organic materials on TFT with 256×256 pixels (pixel size of 154 μm) for longer-wavelength imaging 11 . The organic/CQD TFT imager worked at high reverse bias (-5 V) with large dark current (>1 μA/cm 2 ), low external quantum e ciency (<20%) and limited -3dB bandwidth (~2 kHz). The large pixel size and poor pixel homogeneity limited the imaging resolution. Another report of CQD/graphene imager using CQD to sensitize graphene showed very high responsivity (~10 7 A W -1 ) and detectivity (>10 12 Jones) but suffered from low -3dB bandwidth (~0.16 kHz), large dark current and slow imaging frame rate 12 . Furthermore, its resolution was severely restricted by very poor pixel homogeneity due to the strong bias dependent resistance and photoresponse associated with the horizontal CQD/graphene phototransistors con guration.
Here we present a high-performance near-infrared imager by monolithic integration of top-illuminated CQD photodiodes and large-scale (640×512) COMS ROIC. The device structure and fabrication procedure of PbS CQD photodiodes were well optimized to ensure CMOS-compatibility, good homogeneity, low dark current and high external quantum e ciency. After integration, our CQD imager showed a high resolution of 640×512, a pixel size of 15 μm and a modulation transfer function of 40 lp/mm at a modulation transfer function of 50%, and it produced infrared images with quality comparable with the commercial InGaAs imager. We further demonstrated the applications of our CQD imager for matter authentication and vein imaging.
Design of CMOS-compatible CQD photodiodes For ROIC integration, photodiodes need to be top illuminated to enable incident light absorption and signal acquisition. Con guration of our CQD imager is schematically shown in Fig. 1a, where the photosensitive CQD devices are directly integrated onto ROIC with a transparent conductive oxide on top to enable top illumination. The depletion region of the photodiode close to illumination side prompts the e cient drift of photo-generated carriers by the built-in electric eld rather than diffusion as shown in Fig.   1b. However, most of high-performing CQD photodiodes reported in literature are bottom-illuminated [13][14][15] ; top illuminated CQD photodiode is rarely reported with high-quality junction and favorable stability possibly due to the di culty of depositing a high-quality transparent conductive oxide on top without damaging the bottom vulnerable CQD layer.
The structure of our PbS CQD photodiode is NiO x /PbS CQD/ZnO as shown in supplementary Fig. S1a.
We chose sputtered NiO x and ZnO as the hole and electron collection layer respectively for its proper band position, high stability, high reproducibility and easy manufacturability. Solution processed NiO x and ZnO layers have been extensively explored at the beginning of this project but eventually discarded due to its unsatisfactory reproducibility despite they often produced devices with high performance. PbS CQD was synthesized from the cation-exchange strategy 16,17 and are monodispersed with an absorption peak at 920 nm as shown in supplementary Fig. S1b-c. The sputtered NiO x and ZnO lms are super smooth with roughness of 1.3 nm and 2.1 nm and the ZnO lm is well oriented along [002] direction (supplementary Fig. S1d-f). The NiO x /PbS CQD/ZnO photodiode shows over 3 orders of recti cation but poor photoresponse (external quantum e ciency, EQE < 0.4%) in Fig. 1c.
We suspect that poor junction quality accounts for the extremely low EQE. We thus sought for drive-level capacitance pro ling (DLCP) and capacity-voltage (C-V) measurements to characterize our photodiode.
Since DLCP is only sensitive to the bulk defects and C-V reveals both interface and bulk traps 18 (supplementary Fig. S2a-b), the interfacial defect density can be estimated from N A,CV -N A,DLCP as 5.8×10 16 cm -3 in Fig. 1d. This value is much higher than that of the solution-processed ZnO nanoparticle/PbS CQD photodiode as 2×10 16 cm -3 due to the interface destruction during ZnO sputtering 19 . A thin layer of fullerene (C60) with strong C-C bond and consequently mechanical and electrical robustness was chosen to protect the interface. Thermal evaporation was selected for its gentle deposition and less damage caused to the CQD layer. After careful optimization of the C60 layer, the  Table 1. The EQE at short wavelength in the top illuminated mode is almost one order of magnitude higher than in the back illuminated mode, further proving that top-illuminated mode helps to charge collection because most of photo-generated carriers are swept by the built-in eld instead of circuitous diffusion. at bias from -1 to -0.5 V, revealing photogenerated carriers are completely collected by the built-in electric eld with the assistance of an external bias. Such a at photoresponse is important for real application because any abnormal voltage uctuation near the working bias from the ROIC would cause negligible interference on the signal intensity and thus enable authentic imaging. When the light power received by the device ranges from ~7 nW to 0.7 mW, the photocurrent is linearly proportional to the light intensity as shown in Fig. 2b. The measured linear dynamic range (LDR) is thus calculated greater than 100 dB at -0.5 V bias and 80 dB at self-powered state (zero bias) as shown in Supplementary Fig. S6a, restricted by unattainable weaker or stronger light source. The lowest detectable light intensity of the device at selfpowered state is about 1.9 pW at a signal-to-noise ratio of 1, identi ed by electronic noise measurement in supplementary Fig. S7. The estimated LDR of the device is over 150 dB, which is close to the measured LDR of 160 dB of silicon diode 26 .
The broadband spectral response at -0.5 V bias in Fig. 2c ensures that the device works over a wide spectral range from 390 to 1300 nm. The responsivity and EQE are as high as 0.46 A W -1 and 63% at 970 nm, respectively. And the device also has favorable performance at self-powered state with an EQE about 28% at 970 nm (supplementary Fig. S6c). The bumpy EQE spectrum is somewhat unexpected, and we ascribe the peak at ~970 nm to the excitonic absorption and peaks between 400-800 nm to optical interference. The frequency response is shown in Fig. 2d, suggesting the fast response of the PbS CQD photodiode with a -3dB bandwidth up to 140 kHz. Furthermore, the transient response of the device in  S6d). This photoresponse is fast enough to permit imaging application which is often working at a speed of 30 frames per second. Fig. 2f illuminates the frequency dependent current noise spectrum of PbS CQD photodiode measured by lock-in ampli er at -0.5 V bias within 10 -14 -10 -13 A Hz -1/2 . The total noise is contributed by shot noise, 1/f noise, and generation-recombination (G-R) noise as calculated in supplementary Method. The calculated total noise is plotted as red line in Fig. 2f, ideally tting with the measured noise spectrum. At frequency < 1 kHz, the noise is mainly originated from the 1/f noise due to the scattering between CQDs and at the interface between different functional layers 27,28 . At frequency > 1 kHz, the shot noise and G-R noise primarily determine the noise of our photodiode. Furthermore, the current noise at self-powered state is Our PbS CQD imager consists of CMOS ROIC and PbS CQD photodiode array by monolithic integration (Fig. 3a). The PbS CQD photodiode was deposited onto CMOS ROIC panel with 640×512 pixels array below 100 o C as described in supplementary Method. The wet etching method was used to prepare the imagers, including lithography and lift-off etching to expose the ports and separate imagers as shown in supplementary Fig. S9. The PbS CQD photodiode pixel array is de ned by CMOS ROIC bottom electrode array from fab in supplementary Fig. S10. The pixel size is 15 μm with a pixel pitch of 2 μm. The whole ow process has no negative effect on the PbS CQD photodiodes. After lithography and lift-off etching, the PbS CQD photodiode pixels basically keep the same performance as show in supplementary Fig. S11.
The schematic diagram of the layer sequence is in detail shown in Fig. 3b. The separated bottom metal electrodes de ne the pixel size and pitch. The top contact is a common transparent conductive oxide layer with good conductivity and transmittance. The other layers are continuous with almost no electrical and optical crosstalk, due to a 150-250 nm lateral carrier diffusion length of all layers and <1 μm whole device thickness 29,30 . The SEM image in Fig. 3c shows a cross-section view of our PbS CQD imager. The thickness of PbS CQD is optimal 600 nm with high EQE and low J d . The landscape layout of the image sensor in Fig. 3d indicates the function of each area. The pixel area is located in the center with analogto-digital converter (ADC), column multiplexer and signal output around.
Circuit diagram of one single pixel in Fig. 3e is a BDI (buffered direct injection) structure in our PbS CQD imager. Readout timing diagram of our PbS CQD imager in Fig. 3f controls the on-off of the transistors to reset the collection capacitor (CC), integrate photo-generated carriers and read out signals of pixels in sequence. The CC is reset via opening the gate of M 2 (RST in sequence) before light exposure. Next, the photo-generated carriers by exposure involved in the pixel are stored into CC via opening the gate of M 1 .
During the readout period, the signal voltages on the CC are rstly read out by opening the gates of M 4 (SEL in sequence) and M 6 (S/S in sequence). Then the sensing nodes are reset by M 2 (RST in sequence) and reset signal is read out by opening the gates of M 4 (SEL in sequence) and M 7 (S/R in sequence) for noise cancellation. Such a correlated double sampling are widely used in imaging to reduce the noise and improve image quality. By adjusting the integral time, the collection capacity is affordable to accumulate carriers excited by different-intensity light for high-quality imaging.   Fig. 4 by imaging the standard charts, such as classical 'Lena', 18% gray card, ST-52 chart and ISO-12233 test chart. All image-capturing setups are in detail described in Supplementary Method. The photograph of 'Lena' captured by our PbS CQD imager ( Fig. 4b) is of much higher resolution than that captured by the PbS CQD/graphene imager (Fig. 4a) 12 at the same condition. This is partially due to larger scale of our PbS CQD imager (640×512) than the PbS CQD/graphene imager (388×288), and partially due to the better pixel homogeneity in our vertical PbS CQD photodiodes than the horizontal PbS CQD/graphene phototransistors which suffers from bias dependent resistance and photoresponse. The homogeneity of our PbS CQD imager is slightly worse than the commercial InGaAs imager by imaging the 18% gray card under halogen lamps as shown in Fig. 4c. The grayscale root mean standard deviation (RMSD) of our PbS CQD imager as 3.38 is about 2 times higher than that of a commercial HAMAMATSU InGaAs imager as 1.66. Considering the limited optimization done so far and no fundamental restrictions, we believe the homogeneity of our CQD imager could be further improved to be in par with InGaAs imager. Fig. 4d-f by imaging the ST-52 step chart with 12 gray levels (density range from 0.1 to 3.7). All 12 gray levels are clearly distinguished by our PbS CQD imager (the inset of Fig. 4d). The average density of grayscale in Fig. 4d is normalized to 255 by the equation: gray level=255×(10 -density /1.01) (1/2.2)31 . The characteristic curve of our PbS CQD imager is shown in Fig. 4e where x-axis represents exposure in dB, which equals to -20log(target density/0.1) referred to the ISO-14524. The signal-to-noise ratio (SNR) curve in Fig. 4f demonstrates that SNR drops from 10 to 1 with a decreasing light exposure. The DR of our PbS CQD imager for SNR = 1 is 31 dB, which is limited by the maximum range (31.3 dB) of the ST-52 step chart.

The dynamic range (DR) of our PbS CQD imager is demonstrated in
The slant edge test for spatial frequency response is shown in Fig. 4g-i to measure spatial resolution and modulation transfer function (MTF). The high-resolution photograph of the ISO-12233 test chart captured by our 640×512 PbS CQD imager (pixel size 15 μm) in Fig. 4g demonstrates all kinds of shapes more clearly than that in supplementary Fig. S12a captured by the commercial 320×256 InGaAs imager (pixel size 30 μm). The horizontal and vertical spatial resolution of our PbS CQD imager is extracted from the edge spread function (ESF) as shown in Fig. 4h. The 10-90% edge rise distance of the horizontal slant edge is slightly better than the vertical (0.93 vs. 0.98 pixels), which are both better than 1.64(H) and 1.84(V) pixels of the InGaAs imager (supplementary Fig. S12b). The modulation transfer function (MTF) in Fig. 4i shows that the MTF50 (50% contrast spatial frequency) of the horizontal and vertical slant edge is 40 lp/mm and 37 lp/mm respectively. The resolution of our PbS CQD imager is about 4 times higher than the InGaAs imager (10 lp/mm) with nearly the same-area focal plane (supplementary Fig. S12c). Fig. 5| Applications of PbS CQD imager. Photographs of apple and water captured by the smartphone silicon imager (a) and our PbS CQD imager (d) illuminated with natural light. Photographs of hand captured by our PbS CQD imager (b) and the InGaAs imager (e) illuminated with 940 nm LED. c and f, Gray level along the red dotted line 1 and 2 in panel b and e. g, Photographs of water and ethanol captured by our PbS CQD imager and the InGaAs imager illuminated with 940 nm LED. Solution S1, S3: water, Solution S2, S4: ethanol. h, Normalized grayscale histogram of solution S1-4 in panel g. i, Normalized grayscale histogram of alcohol with various concentration associated to supplementary Fig.  S13. Fig. 5 shows several types of images captured by our PbS CQD imager. Under natural light, the photograph in Fig. 5a captured by the smartphone silicon imager demonstrates a well-looking apple and a shadowy level of water in the visible range. In comparison, the same scene captured by the PbS CQD imager in 400-1300 nm range gives more information, such as a hidden scar in the apple and a clear level of water as shown in Fig. 5d. The capability of our PbS CQD imager to capture infrared images provide additional valuable information that are hardly discernable by the silicon imager.

Applications of PbS CQD imager
Another application is demonstrated on vein imaging under 940 nm plane light source in transmission mode. In this band, penetration depth of tissue is over 10 mm 32 and our PbS CQD imager (EQE>60% @ 940 nm, J d <17.8 nA cm -2 ) is much more sensitive than the commercial InGaAs imager (EQE<15% @ 940 nm, J d~1 43 nA cm -2 ) 33 considering its higher EQE and lower J d . The photograph in Fig. 5b by our PbS CQD imager shows much clearer blood veins in the hand than that in Fig. 5e taken by the commercial InGaAs imager. The grayscale values along the red dotted line 1 and 2 are respectively shown in Fig. 5c and 5f. Our PbS CQD imager obtains more obvious and sharper change of grayscale (9.6 and 13.5) in the edge of vein compared with the commercial InGaAs camera (7.7 and 5.2), demonstrating greater application potential in vein imaging.
Our PbS CQD imager further provides the possibility for material identi cation as the commercial InGaAs imager. The images of water and ethanol, illuminated by 940 nm plane light source, are captured by our PbS CQD and the InGaAs imager as shown in Fig. 5g. The water bottles (S1, S3) are darker than the ethanol bottles (S2, S4) due to stronger 940 nm absorption of water than ethanol (supplementary Fig.  S13d). The normalized grayscale histogram in Fig. 5h demonstrates that our PbS CQD imager obtains broader and higher grayscale values than the InGaAs imager. The grayscale deviation between water (median 110) and ethanol (median 80) captured by our PbS CQD imager is greater than that captured by the InGaAs imager (median 60 to 40). The contrasts illustrate that our PbS CQD imager is better at water and ethanol identi cation than the InGaAs imager. Supplementary Fig. S13 shows the images of alcohol with various concentration by using our PbS CQD imager. The corresponding grayscale histogram in Fig.  5i demonstrates how our PbS CQD imager distinguishes different concentration of alcohol in this azeotropic mixture through comparing gray values. Our PbS CQD imager also performs very well in organic dye discrimination ( Figure S14), thanks to its wide responsive spectrum and tunable band gap via quantum con nement. All these examples highlight the broad application potential of our CQD imagers.

Conclusions
In summary, we demonstrate a low-cost and high-performance imager by monolithically integrating welldesigned CQD photodiodes and CMOS ROIC. Via optimization of charge transport layer and protective layer, our top illuminated PbS CQD photodiodes demonstrate a balanced D * of 2.1×10 12 Jones and -3dB bandwidth of 140 kHz. Monolithic integration of PbS CQD photodiodes and ROIC produced a nearinfrared CQD imager with the-state-of-the-art performance including a large pixel number of 640×512, a >60% EQE @ 940 nm, a spatial resolution of 40 lp/mm at MTF50, and comparable image quality with commercial InGaAs camera. We further showcase our CQD imager for apple scar detection, vein imaging, water/ethanol discrimination and dye identi cation, as examples to supplement silicon based visible camera. The success of CQD imager highlights the great potential of monolithic integration of functional devices onto ROIC, such as arrays of sensors, integrated photonics and even integration of sensing, memory and computing onto a single chip.