Filter technologies implemented on CMOS image sensors for spectrally selective applications often use a combination of on-chip organic resists and an external substrate with multilayer dielectric coatings. The photopic-like and near-infrared bandpass filtering functions respectively required by ambient light sensing and user proximity detection through time-of-flight can be fully integrated on chip with multilayer metal–dielectric filters. Copper, silicon nitride, and silicon oxide are the materials selected for a technological proof-of-concept on functional wafers, due to their immediate availability in front-end semiconductor fabs. Filter optical designs are optimized with respect to specific performance criteria, and the robustness of the designs regarding process errors are evaluated for industrialization purposes.
© 2014 Optical Society of America
The widespread use of mobile devices in recent years is partly driven by increasing customer demands for multiple functionalities on the same device. Due to their sensitivity in the visible and near-infrared (NIR) ranges, CMOS image sensors can be used for different applications than traditional RGB color imaging. The green channel can simply provide signals for monitoring ambient light and adjusting the brightness or the contrast of a display. Time-resolved applications, such as time-of-flight (TOF) for user proximity detection, can also be addressed with the NIR signal provided by a CMOS single-photon avalanche diode (SPAD) [1,2]. The manufacturing of multifunctional low-cost sensors is possible, because CMOS image sensors are mass-produced in silicon foundries, the different functions can be implemented on the same sensor, and these applications do not necessarily require any imaging optics.
Green light for ambient light sensing (ALS) is basically filtered with dye or a pigmented organic resist whose spectral transmittance closely approaches the photopic response of the human eye in the visible range. An additional filter is required to block NIR light, since the organic resists are transparent in this spectral range. The most obvious choice for the cut-infrared filter is the same as in all RGB microcamera modules. High rejection of infrared, together with high transmittance in the whole visible range, is achieved from interference effects in several tens of alternating high- and low-index dielectric layers deposited on a glass substrate.
Long-range TOF with Si sensors is currently performed with a monochromatic light source, such as a pulsed or modulated vertical cavity surface emitting laser (VCSEL) emitting in the NIR domain, for example at 850 nm, because it is invisible to the eye and still in the detection range of silicon. The distance from the sensor to a reflecting object in the scene is deduced from the measured delay between the signal detected by the SPAD capturing photons reflected from the object, and the pulse emitted by the source. A highly selective bandpass interference filter, including many dielectric layers deposited on an external glass substrate, is placed in front of the SPAD sensor. This filter improves the signal-to-noise ratio by blocking the whole visible and NIR domain thus rejecting most of parasitic ambient light, except for a small bandpass around the wavelength of the TOF source to take into account the typical 20 nm variations of the VCSEL center wavelength due to heat or manufacturing dispersion. However, optical design of filters with only dielectric layers does not allow strong rejection of visible light. A black organic resist, transparent in the NIR, is used on chip to enhance the visible rejection.
When both ALS and NIR functions are to be implemented together, a difficulty arises as the transmittances of the two filters on glass are incompatible. A single full sheet filter on glass cannot be used, as in micromodule cameras for color imaging, to suppress unwanted spectral domain for both functionalities. For ALS and NIR detection by the same chip, both infrared cutoff and infrared bandpass filters can be formed side by side on the same glass substrate to be placed in close vicinity to the CMOS sensor, just above the respective areas of the chip dedicated to ALS and NIR detection. This solution is effective if the sizes of respective areas are at least comparable to the distance between the filter and the detection matrix, i.e., often several hundreds of micrometers. Also, glass substrates with several different patterned filters are much more expensive than the mass-produced full sheet infrared cutoff filters. Another solution may be considered with fully integrated filters on chip for both ALS and NIR detection. In addition to a potential implementation pixel by pixel and cost reduction, such a solution would allow reducing the thickness of the module through the saving of the glass substrate. This is especially true in sensors without optics.
The integration of all dielectric filters on chip has already been shown in previous works [3,4]. All dielectric filters can have high transmission, but do not fulfill the spectral requirements of strong IR-cut and highly selective NIR filtering functions if the number of layers and total thickness are limited. In this study, we investigate the suitability of metal–dielectric interference stacks as fully integrated filters for ALS and NIR detection on a single CMOS chip. Silver–dielectric multilayer filters provide both efficient NIR cutoff filtering and color filtering [5,6]. However, Ag is not commonly used in CMOS fabs due to contamination issues. We focus here on copper–dielectric filters, as Cu is a CMOS-compatible material and has appropriate optical constants above 600 nm with a high extinction coefficient over refractive index ratio, k/n. Despite a central wavelength in the yellowish-green at 550 nm, the ALS filtering can be realized on chip with a single copper–dielectric filter. Transmission in the bandpass and rejection can be enhanced by introducing dielectric layers with different refractive indices in the filter, as in transmission-induced filters . The dielectric materials chosen here are SiN and as they are widespread in the semiconductor industry. The same materials can be used to design NIR bandpass filters in the prospect of simultaneous integration on a CMOS chip.
The paper is organized as follows. In Section 2, integration schemes for integration on SPAD wafers are identified, and filter designs are optimized. Simulated performances including robustness to process errors and angular behavior are compared to reference filters, trying to estimate the relevance for manufacturing this technology in semiconductor foundry. Section 3 addresses the key point of Cu adherence on SiN without any optically absorbing adherence layer, and describes the realization of proof-of-concept demonstrators on CMOS wafers.
2. Filter Design and Evaluation for Large-Scale Manufacturing
Special care was taken to consider filter architectures compatible with a potential double integration of ALS and NIR filters on CMOS SPAD wafers. Specific criteria were defined for the optimization of filter designs for ALS and NIR applications. The robustness of the stacks to process errors under normal and oblique illumination was then evaluated in the prospect of large-scale manufacturing.
A. Integration Schemes for a Double Integration
Two integration schemes were considered for the design investigation.
The first one favored the absence of modification of the back-end process, with filters formed above the passivation layers (Fig. 1, top). The ALS and IR bandpass filters could be independent, but configurations with some layers in common could be preferred for technological reasons. The process flow planned the deposition of the first filter (say IR bandpass) on the SiN passivation layer, IR bandpass filter patterning, ALS filter deposition and patterning, then top coat encapsulation for protection of filter edges against atmospheric moisture, and pad opening.
In the second integration scheme, at least one of the filters was deposited in two stages, respectively below and above the passivation layers (Fig. 1, bottom). The passivation layers were 250 nm thick covered by 500 nm thick standard SiN. Including these thick layers inside the filter cavities was a way of increasing the total thickness of the filter, without exceeding a maximum 1.2 μm height of patterned layers above the passivation layers, due to the limited thickness of the photoresist for the lithography process. Large optical thickness allows more freedom in multilayer design and helps to match targeted spectral response [8,9]. The integration of the first filter layers inside the top oxide layer of the back-end did not require any change of pad height, because they were etched in the vicinity of pads. The oxide layer only needed to be deposited in two stages, and did not require any planarization because the height of the bump generated by the buried filter layers on top of this oxide layer remained limited to a few hundred nanometers, without impact on a later lithography step.
In both integration schemes, the deposition of the second filter had to start with a Cu layer, to benefit from the high etching selectivity with dielectrics and therefore to avoid etching the top dielectric layer of the first filter at the end of the patterning of the second filter. Since a low-H adherence layer had to be deposited first, before the Cu layer (as detailed in Subsection 3.A), this low-H SiN layer was also the last layer of the first filter, not to mention the final top-coat encapsulation.
The minimum thicknesses of Cu, low-H SiN and layers were respectively 20, 20, and 100 nm, to avoid any additional deposition process developments. The thickness of the top coat was fixed to 105 nm as in the standard imager process. No more than two Cu layers were expected within the filter to limit the lateral overetching during the patterning process (Subsection 3.A). layers could be sandwiched between SiN layers, and only Cu/low-H SiN interfaces were considered in the design because oxidized Cu and induced delamination under Cu layers.
B. Definition of Performance Criteria
A specific request of customers concerned the packaging of the sensors, which had to include a black housing in order to make the sensors invisible to the user. The black housing could be a black window with high transmission in the NIR range and low transmission in the visible range (Fig. 2). This requirement complicated the conception of ALS sensors but made the design of IR bandpass filters easier.
Quantitative criteria were introduced to optimize the design of ALS and IR bandpass filters in the system also including the black housing, CMOS sensor quantum efficiency (QE) for both filters, and the black resist in the case of the IR bandpass filter (Fig. 2). The criteria were derived from system requirements and did not summarize to transmittance targets such as the photopic response.
The first criterion for the ALS sensor performance was related to the stability of the signal provided by the ALS sensor, considering various illuminants, typically the 3200 K blackbody close to the emission of quartz-halogen light bulbs, or fluorescent light sources and LEDs for indoor conditions, and D65  for outdoor conditions. The sensitivity under an illuminant , in , is proportional to the spectrally integrated product of the relative spectral power distribution of the illuminant (denoted ), black housing transmittance (denoted Black), filter transmittance (denoted Filter), and Si QE. It was computed as follows:
The integrated signal (in ) measured on the ambient light sensor over the integration time should be multiplied by a calibration factor, denoted , [in ]. This factor is inversely proportional to a sensitivity, in order to have a measurement in lux. From a system point of view, this is a system calibration (conversion /s to lux) performed, for example, under two reference light sources , typically an incandescent and a LED source. In this case, the calibration factor is the geometric mean of the inverse of the two sensitivities computed under each illuminant.
The ALS error was defined as the relative variation of the product of sensitivity and calibration factor within a set of illuminants, with respect to the two reference illuminants. For ,
In our specifications, the maximum ALS error had to be lower than 2.2 dB over the whole set of illuminants. ALS error was very sensitive to the IR cutoff. The value of 2.2 dB was the maximum ALS error for the reference filter, which was a green resist associated with an external IR cutoff filter.
The second criterion for the ALS sensor was related to the sensor sensitivity in low light conditions. ALS dark (in lux) was defined as the signal given by the sensor in the dark at a given temperature . It represented the dark current, DC (in /s), converted in lux through the calibration factor of the sensor:
In this study, the absolute value of ALS dark was not meaningful because the test vehicle used for the calibration was not fully representative of the sensors to be used in the ALS application. In the following, all ALS dark values are normalized with respect to the ALS dark of the reference filter. The filter design optimization had to minimize ALS dark.
The criteria for the evaluation of IR bandpass filters were closely related to transmission and rejection characteristics of the filters. The first criterion concerned the transmission of the filter in the bandpass:
The same specifications were applied for filters under oblique incidence, up to 30° on the ALS filters and 15° on the IR filters, limited by the geometry of the packaging. They were also expected for filters manufactured with process errors arising from deposition machines with time monitoring of layer thickness, and from characterization tools. In the framework of large-scale manufacturing, both random and systematic errors were supposed to result in random statistics with a Gaussian distribution centered on nominal values of layer thicknesses and optical constants. Three different sets of process dispersions were considered in this specific study:
C. ALS Filter Design Optimization and Evaluation of Robustness
ALS filters were designed with manually adjusted target transmittances and the random optimization tool in Optilayer software  to determine appropriate layer orderings. Further optimization was then performed with a proprietary software based on Matlab optimization functions, able to directly optimize the multilayer stacks in one dimension with respect to the specific criteria ALS error and dark. The complete stack down to the Si substrate was simulated to take into account all the interference effects during the optimization process.
The design optimization had competing requirements to simultaneously optimize ALS error and ALS dark (Fig. 3). Both criteria could be optimized with values close to the reference filter. With multilayer Cu/dielectric filters, the ALS error could be improved by a stronger rejection of the NIR (Fig. 4) at the expense of filter transmission in the visible range, degrading the ALS dark. This was obtained with thicker Cu layers. Also, optimized ALS filters were redshifted compared to the reference filter (Fig. 4). This was due to the sharp rise of Cu refractive index below 600 nm, which hindered the centering of a Cu filter at 550 nm. The redshift did not significantly impact the ALS error, but the low transmission below 600 nm impacted the ALS dark. However, Cu/dielectric filters do withstand the final forming gas annealing at 400°C contrary to organic resists used in the reference filter. Compared to the standard process with organic resist for the ALS filtering on SPAD, the forming gas annealing may provide a reduction of dark current , leading to improved ALS dark, due to deactivation of defects generated in Si during the filter process and the pad opening. Considering a potential threefold improvement of the ALS dark after the forming gas annealing, designed filters with normalized ALS dark up to 3, three times the value of the reference filter, could be considered as acceptable. Despite the limitations of filter performances caused by Cu refractive index and the need for relatively thick Cu layers to achieve sufficient IR cutoff, a set of optimized Cu ALS filters was found to respect both specifications on ALS error and ALS dark.
Filter robustness was then evaluated by statistical tests, simulating a sample of 5000 spectral responses among a Gaussian random distribution around the nominal thickness and refractive index values of a given stack. The input thickness and refractive index were chosen within a range of typically considered in statistical process control. All the layers down to the Si were taken into account in the robustness evaluation. Among the whole set of nominal filters shown in Fig. 3, the filter with ALS error of 1.36 and normalized ALS dark of 2.06, shown in Fig. 4, was selected because the dispersion of ALS errors simulated with the largest process errors (set 1) remained below the 2.2 dB limit, while the normalized ALS dark was most often below 3 (Fig. 5). The dispersion on performances could be further reduced by simultaneously improving the standard dispersions of all three materials (Fig. 5). An effort on a single material did not show significant improvement. Robustness simulated with process errors of set 1 under oblique incidence was acceptable for incidence angles up to 30°, only inducing a minor degradation of the ALS dark (Fig. 6).
D. IR Bandpass Filter Design Optimization and Evaluation of Robustness
Optilayer software proved to be sufficient to optimize IR bandpass filters, as the evaluation criteria were more or less equivalent to spectral transmittance targets. As expected, the best designs were obtained with the integration scheme including the passivation layers within the filters. The maximum transmission was 5% to 10% higher than for the integration scheme with all the filter layers above the passivation layers. Optimized Cu/dielectric filters had worse values of A parameters than the reference commercial filter, due to the lower peak transmission (Fig. 7). In some designs, this was somewhat compensated by better B parameters with a better centering of central wavelength, while C and D parameters were similar to the reference.
The weak point of the IR bandpass interference filters was the robustness with layer thickness and refractive index. The envelope of simulated A, B, C, D parameters clearly exceeded the acceptable limits of nominal specifications (Fig. 8), even for the third set of process errors, the most demanding for hardware monitoring. Nevertheless, significant differences were observed between the sensitivities of the individual layers of the stacks. With process errors specified in the three sets, the Cu layers were not very critical, and filter performance dispersion could be reduced by lower process error dispersion on dielectric materials only. Furthermore, only a few dielectric layers were really critical. Depending on the designs, the most critical layers were thick dielectric layers, such as the 500 nm SiN passivation layer, and/or the dielectric layers inside the Fabry–Perot cavities, between Cu layers. The stability of IR bandpass filter performance with process errors could be significantly improved in simulations (Fig. 8, yellow curves) by applying a drastic control of the only two passivation layers with 0.5% standard dispersion on their thickness, keeping the process error values of the set 2 for the other layers. Under 15° oblique incidence, the dispersion of filter performance did not degrade significantly, but the average C parameter was increased due to the blueshift of interference filters.
Several reasons explain why IR bandpass filters are more sensitive to process errors than ALS filters. The spectral shift induced by a given process error increases with the central wavelength of Fabry–Perot filters. Also, evaluation criteria are very sensitive to a given spectral shift of the IR bandpass filters with a small spectral width, whereas a blueshift or widening of ALS filters toward low wavelengths negligibly impacts ALS error and dark criteria. In addition, the robustness behavior markedly differs between NIR and visible ranges due to the much lower refractive index of Cu in the NIR. Optical constants of Cu are more favorable for the robustness of ALS filters, but this is at the expense of peak transmission and sensor sensitivity.
IR bandpass filter designs with a single metallic layer may be considered to improve the robustness, but do not provide sufficient performances in terms of rejection, with a limited number of layers. Other design techniques may also be tried, specifically targeting enhanced robustness  with monitoring of layer-by-layer sensitivity. Improved robustness may be obtained at the expense of nominal performance, which is not far beyond the acceptable values in the designs optimized in Subsection 2.D. In all cases, the technological limitations concerning the number, thickness, and ordering of the layers are strong constraints and may hinder the achievement of an acceptable trade-off between performance and robustness.
Alternative multilayer technologies may be investigated for IR bandpass functionality on CMOS chips, especially nonmetallic multilayers, such as stacks. Both the peak transmission and visible rejection can be improved with respect to Cu/dielectric filters, as aSi has very high index contrast with , absorbs visible wavelengths, and transmits NIR wavelengths. The robustness to process errors remains a critical question, because the filters are also interferential.
It should be noted, however, that the definition of criterion 1 for the IR bandpass filter robustness could be refined, to be more representative of a yield of the whole system including the light source. The present definition of criterion 1 corresponds to the specific case where the wavelength of the VCSEL is at the limits of its specifications (845 or 865 nm), shifted with respect to the filter central wavelength. Considering the wavelength of the VCSEL randomly chosen in the 845–865 nm interval would provide a better idea of the final yield. Incidentally, the dispersion of filter performances under process errors would more easily satisfy this specification.
The whole study assumed random Gaussian statistics for the distribution of process errors, which corresponds to large-scale production using several deposition machines and several characterization tools for calibration. The case of small-scale production with a single deposition machine and a single calibration tool is different. For a given material, the intrawafer thickness error is mainly systematic and its spatial distribution is often reproducible from layer to layer, and from wafer to wafer. With a tight control of the calibration tool, the dispersion of filter performances may be reduced.
3. First Demonstration on Functional CMOS Wafers
A. Process Developments
Technological developments were required to provide stacks including both Cu and SiN layers without any absorbing adherence layers, suitable for the ALS and NIR filtering.
All the experiments were carried out on p type 300 mm (100) wafers. The SiNx films were deposited by plasma-enhanced chemical vapor deposition (PECVD) and the Cu layers by physical vapor deposition (PVD) respectively in single-wafer CENTURA and ENDURA systems of applied materials. A reactive ion etch Aviza system was used for the dry etch processes of the SiNx films and wet etch treatments were done in a Semitool automated wet bench. Cross-sectional scanning transmission electron microscopy observations were performed with an FEI TECNAI OSIRIS system that operated at 120 kV and equipped with an energy dispersive x-ray detector.
The Cu films were deposited at ambient temperature. The working pressure was set between 0.1 and 1 Pa with a constant Ar debit of several tens of sccm. As in the back-end process of CMOS imagers on 300 mm wafers, Cu layers were only available in large thickness (typically 1 μm), and it was necessary to optimize the deposition conditions in order to provide semi-transparent thin Cu films in accord with the typical designs of metal/dielectric filters. By lowering the power of the DC plasma in the chamber down to 20 kW, the deposition rate decreased to few nanometers per second, which set a lower limit of about 20 nm for the thinnest Cu layers with acceptable reproducibility and uniformity.
One of the major challenges was to select and deposit dielectric layers able to form correct interfaces with either underlying or overlying Cu layers. Indeed well-known Cu barrier layers such as Ta, TaN, Ti, and TiN, commonly used in interconnection process, also serve as an adhesion promoter but were prohibited here, due to their high absorption detrimental for filter transmission. PECVD SiN films were first deposited with a standard process (Table 1) using a gas mixture at a pressure of few Torrs and a temperature of 380°C. The RF power was kept constant at around 800 W and the gas flow ratio set between 3 and 4.
In multilayer Cu/SiN stacks, to remove the Cu oxide layer formed during the air breaks, in situ thermal treatments with hydrogen plasma ( or ) were performed before starting the deposition of SiN . However, delamination often occurred at the lower Cu interface, with underlying SiN. As the failure mechanism was only observed after the last SiN deposition in SiN/Cu/SiN stacks, it was supposed to originate from desorption from the first SiN layer induced by the thermal budget during the deposition of the second SiN film. Previous studies [15–17] showed that out-diffusion from PECVD SiN appears when the temperature is close or higher than during the deposition due to multiple Si–H bonds breaking in the film. Different degassing conditions in a atmosphere before Cu deposition were tested without success. Moreover, the morphology aspect of the SiN surface was degraded with the degassing leading to a rougher surface. Therefore, the only way of providing a better thermal stability of the stacks was to modify the SiN deposition conditions (Table 1) in order to decrease the Si–H bonds concentration in the SiN layers . The development of SiN films with a gas flow ratio about 10 times lower than for standard SiN films (Table 1) allowed the complete elimination of delamination in SiN/Cu/SiN and SiN/Cu/SiN/Cu/SiN stacks, even after an annealing at 400°C during 2 h in atmosphere. The interfaces appeared flat, without any local adherence failure of the Cu films (Fig. 9). As expected, the Si–H and N–H concentrations deduced from Fourier transform infrared spectroscopy measurements were respectively lower and higher in the so-called low-H SiN compared to the standard SiN (Table 2). The refractive index was significantly lower (Table 2) while the extinction coefficient remained negligible in the whole visible and NIR range, below .
For the integration of filters on CMOS, filter patterning was required not only to access the electrical contacts but also to provide the filtering function to a restricted area corresponding to selected pixels.
The patterning was performed using a dry etch process with an plasma for SiNx films and a double chemistry at ambient temperature for the wet etching of Cu, with HF diluted at 0.5% to remove the residual interface layer above Cu and a DSP solution, mixture of dilute sulfuric acid () and hydrogen peroxide (), to remove Cu. Both and SiN layers could be etched by the same dry process with somewhat different etching speed. In SiN/Cu/SiN/Cu/SiN stacks, relatively vertical sidewalls were obtained on SiN patterns due to the good anisotropic properties of the dry etch process and a high selectivity of the two kinds of etch treatments used to pattern the whole stack. As the Cu layers were very thin (), the lateral attack of Cu during the wet etching was limited to below 1 μm, a value that may be reduced by refinement of the process for future developments requiring higher spatial resolution. Strong lateral overetching is undesirable as it may break the edges of the filter patterns, due to the top Cu layers exposed several times to the wet etching in a multilayer stack.
B. Filter Integration on CMOS Wafer
The feasibility of process integration was then investigated on 300 mm CMOS front-side imager wafers with arrays of pinned photodiodes  with 2.8 μm pixel pitch, available for a first demonstration. The major issue was the demonstration of ALS and IR bandpass filtering on functional wafers. Several separate wafers were independently processed for the ALS and the IR filters, with the prospect of a joint integration on a single wafer in a future stage. The most immediate integration scheme (Fig. 10), which avoided significant modifications of the back-end process flow, planned (i) the deposition of filters on top of the back-end stack, above the oxide layers needed for the interconnections; (ii) filter etching everywhere except over pixel matrices; (iii) Al contact pad opening with the etching of SiN and passivation layers, which had already been etched on the pixel area as in the usual imager process to reduce the height of the back-end stack for small pixels ; and (iv) SiON top coat deposition. No microlens was realized on top of the stack for this first demonstration, because wafers with potential broken filter fragments including contaminating Cu metal were not allowed to enter the lithography area of the clean room.
To secure the process for this first demonstration, filter designs only included five layers, except for one of the IR bandpass designs which contained 11 layers. At this stage, the simulations used optical constants for the three materials deduced from single layer ellipsometric measurements.
Filters were first deposited on bare Si wafers and the characterized reflectances were found to be in correct agreement with the simulations. The demonstrator was then realized according to the planned process flow. No degradation of dark current was observed compared to usual wafers. QE measurements were performed at normal incidence with the characterization bench described in . The spectral signature of ALS and IR filters was clearly present in QE measurements (Fig. 11). The measured QE was then compared to simulated QE obtained by the product of simulated filter transmittance and Si QE previously measured on wafers without filters. The oscillations in the spectral responses originated from the residual reflection of a few percent on the antireflective coating on Si. Measured and simulated QE could not be meaningfully compared in magnitude, because microlenses were not realized above the filters, inducing significant light losses in the areas of the pixels that are not photosensitive. The losses were not taken into account in the 1D simulations and could not be deduced from the ratio of maximum simulated versus measured transmissions, because simulated filter transmission depends on many other unknown parameters, such as Cu refractive index. The blueshift of the measured versus simulated responses may be explained by overestimated thicknesses of the SiN layers. A deeper understanding of the mismatches would require both the formation of microlenses, and a dedicated multisample characterization as previously demonstrated  but was not the objective of this first technological proof-of-concept. However, the differences in filter spectral width and the dispersion of IR bandpass filter spectral responses between center and edge of the 300 mm wafers motivated the part of the study concerning the potential performances and robustness of the filter technology (Subsections 2.C–2.E).
In this study, we realized a first demonstrator of fully integrated spectral filters on CMOS separate wafers with materials commonly used in foundries and standard patterning process. The filters were interference multilayer stacks with Cu, SiN, and layers. Specific developments were necessary to enhance the adherence of Cu layers over SiN layers and determine an appropriate etching process, but the technology remained simple and immediately accessible for usual equipments of the semiconductor industry.
The Cu/dielectric filters were candidates for the replacement of present filtering solutions, which combine on-chip organic resists and interference dielectric multilayer filters on glass. The major advantages of on-chip metal/dielectric filters were the reduced thickness of the module, the potential for implementation of different filters on pixels, and the elimination of costly external glass substrates with multiple filters. Although the design was constrained by technological limitations derived from compatibility with CMOS manufacturing and cost considerations, the performances of optimized filters were close to the reference filters. Enhanced performance could be obtained by integrating some of the back-end layers within the filter. The high rejection of metal/dielectric filters enabled to integrate the ALS function on chip without the NIR cutoff filter on glass. The same technology was found to be particularly well suited for the NIR bandpass filtering required in TOF applications, with both high transmission and high rejection. To our knowledge, thin film multilayers have no competing solution for on chip integration in the NIR range. However, the tough specifications in our specific targeted applications still required the use of a black resist to enhance rejection in the visible range. The technology was compatible with a double integration of ALS and IR bandpass filters on the same CMOS chip.
The manufacturability of this multilayer thin film technology was investigated in terms of robustness to process errors with assumed Gaussian statistics. The process errors on layer thickness and refractive index were considered in a range of . ALS filter designs were found to be robust enough with process errors typically observed in fab, provided that the dark current is reduced after the final back-end annealing. This has to be verified in the near future. In that case, the technology would be suitable for industrial implementation of standalone ALS filters. The bottleneck of the technology was the robustness to process errors for narrow IR bandpass designs. However, significant improvement may be achievable with a reduction of process errors in the deposition of only two dielectric layers in the filters. Smaller dispersions may result from a more drastic control of process, a limited number of machines for the deposition of these two layers in production, and deposition machines providing very good intrawafer thickness uniformity, with standard deviation typically less than 1%. Also, the introduction of in situ optical monitoring techniques [23,24] or trimming in imager foundries should help to reduce the dispersion of filter performances.
The framework of the study was a partnership between ST Microelectronics and CEA Leti. It was supported by the Minalogic competitiveness pole and partially funded by the European Union FEDER together with the DGCIS national board for industry renewal. The authors would like to thank François Leverd and Pascal Besson for etching developments, Katia Haxaire for Cu deposition developments, Marie-Lyne Charles for help in process development monitoring, Laurent Clément for TEM characterization, and Jérôme Hazart for the supply of stack optimization routine.
1. E. A. G. Webster, L. A. Grant, and R. K. Henderson, “A high-performance single-photon avalanche diode in 130-nm CMOS imaging technology,” IEEE Electron Device Lett. 33, 1589–1591 (2012). [CrossRef]
2. S. Cova, M. Ghioni, M. A. Itzler, J. C. Bienfang, and A. Restelli, “Semi-conductor based detectors,” Exp. Methods Phys. Sci. 45, 83–146 (2013). [CrossRef]
3. S. Kawada, S. Sakai, N. Akahane, R. Kuroda, and S. Sugawa, “A wide dynamic range checkered-color CMOS image sensor with IR-cut RGB and visible-to-near-IR pixels,” in Proceedings of IEEE Sensors Conference (IEEE, 2009), pp. 1648–1651.
4. S. Koyama, Y. Inaba, M. Kasano, and T. Murata, “A day and night vision MOS imager with robust photonic-crystal-based RGB-and-IR,” IEEE Trans. Electron Devices 55, 754–759 (2008). [CrossRef]
5. Y. T. Yoon and S. S. Lee, “Transmission type color filter incorporating a silver film etalon,” Opt. Express 18, 5344–5349 (2010). [CrossRef]
6. L. Frey, P. Parrein, J. Raby, C. Pellé, D. Hérault, M. Marty, and J. Michailos, “Color filters including infrared cut-off integrated on CMOS image sensors,” Opt. Express 19, 13073–13080 (2011). [CrossRef]
7. H. A. Macleod, Thin-Film Optical Filters III (Institute of Physics, 2001).
8. J. A. Dobrowolski, “Comparison of the Fourier-transform and flip-flop thin-film synthesis methods,” Appl. Opt. 25, 1966–1972 (1986). [CrossRef]
9. A. V. Tikhonravov, M. K. Trubetskov, T. V. Amotchkina, and M. A. Kokarev, “Key role of the coating total optical thickness in solving design problems,” Proc. SPIE 5250, 312–321 (2004). [CrossRef]
10. D65 illuminant, http://www.cie.co.at/publ/abst/datatables15_2004/std65.txt.
11. A. V. Tikhonravov and M. K. Trubetskov, OptiLayer Thin Film Software, http://www.optilayer.com.
12. J. L. Regolini, D. Benoit, and P. Morin, “Passivation issues in active pixel CMOS image sensors,” Microelectron. Reliab. 47, 739–742 (2007).
13. P. Verly, “Design of a robust thin-film interference filter for erbium-doped fiber amplifier gain equalization,” Appl. Opt. 41, 3092–3096 (2002). [CrossRef]
14. Y. L. Cheng, T. J. Chiu, B. J. Wei, H. J. Wang, J. Wu, and Y. L. Wang, “Effect of copper barrier dielectric deposition process on characterization of copper interconnect,” J. Vac. Sci. Technol. B 28, 567–572 (2010). [CrossRef]
15. C. Boehme and G. Lucovsky, “Dissociation reactions of hydrogen in remote plasma-enhanced chemical vapor-deposition silicon nitride,” J. Vac. Sci. Technol. A 19, 2622–2628 (2001). [CrossRef]
16. D. Benoit, P. Morin, and J. Regolini, “Determination of silicon nitride film chemical composition to study hydrogen desorption mechanisms,” Thin Solid Films 519, 6550–6553 (2011). [CrossRef]
17. S. C. Mao, S. H. Tao, Y. L. Xu, X. W. Sun, M. B. Yu, G. Q. Lo, and D. L. Kwong, “Low propagation loss SiN optical waveguide prepared by optimal low-hydrogen module,” Opt. Express 16, 20809–20816 (2008). [CrossRef]
18. F. Karouta, K. Vora, J. Tian, and C. Jagadish, “Structural, compositional and optical properties of PECVD silicon nitride layers,” J. Phys. D 45, 445301 (2012). [CrossRef]
19. E. R. Fossum and D. B. Hondongwa, “A review of the pinned photodiode for CCD and CMOS image sensors,” IEEE J. Electron Devices Soc. 2, 33–43 (2014).
20. M. Cohen, F. Roy, D. Herault, Y. Cazaux, A. Gandolfi, J. P. Reynard, C. Cowache, E. Bruno, T. Girault, J. Vaillant, F. Barbier, Y. Sanchez, N. Hotellier, O. LeBorgne, C. Augier, A. Inard, T. Jagueneau, C. Zinck, J. Michailos, and E. Mazaleyrat, “Fully optimized Cu based process with dedicated cavity etch for 1.75 μm and 1.45 μm,” in Proceedings of the International Electron Devices Meeting (IEEE, 2006), paper 4154411.
21. C. Mornet, J. Vaillant, T. Decroux, N. Virollet, D. Herault, and I. Schanen, “An image quality evaluation tool simulating image sensors including quantum efficiency off-axis effect,” Proc. SPIE 7876, 78760M (2011). [CrossRef]
22. L. Frey, P. Parrein, L. Virot, C. Pellé, and J. Raby, “Thin film characterization for modeling and optimization of silver-dielectric color filters,” Appl. Opt. 53, 1663–1673 (2014). [CrossRef]
23. B. Badoil, F. Lemarchand, M. Cathelinaud, and M. Lequime, “Manufacturing of an absorbing filter controlled by a broadband optical monitoring,” Opt. Express 16, 12008–12017 (2008). [CrossRef]
24. C. C. Lee, K. Wu, and M. Y. Ho, “Reflection coefficient monitoring for optical interference coating depositions,” Opt. Lett. 38, 1325–1327 (2013). [CrossRef]