A few years ago, we demonstrated the first wafer-scale integration of metal-dielectric Fabry–Perot interference filters on complementary metal oxide semiconductor (CMOS) image sensors for red-green-blue (RGB) color imaging [1]. The major advantage of these filters is their low chromatic crosstalk, enabling correct color rendering without the infrared (IR)-cut multilayer external filter which has to be used with the pigment-based organic color filters in the traditional technology. Despite the fact that the IR-cut filter requires the deposition of tens of layers with high thickness accuracy for an efficient rejection of the IR, the mass production for billions of sensors has considerably reduced the production cost. In addition, although the Ag filter designs could provide low color errors and high signal-to-noise ratios (SNRs) similar to the organic color filters with IR-cuts, the transmissions of the proposed double cavity Fabry–Perot Ag filters were only 45–65% [1], significantly lower than the organic color filters (80–90%). It was shown in another study [2] that small variations of the central wavelengths of the RGB filters have much more impact on the usual SNR indicator [3] than variations of the transmissions. The sensor sensitivity to incident light is not sufficiently taken into account in filter design optimization using this noise metric. In practice, the limited transmissions of Ag filters would be detrimental, especially in low light level conditions below 10 lux.

In the literature, other alternative technologies for RGB filters have been investigated with potential integration on CMOS image sensors, especially purely dielectric interference multilayers [4], nanostructured metallic [5,6], or dielectric [7] filters, respectively, based on plasmonic and grating mode resonances, or semiconductor nanowires [8]. However, the filters were generally designed considering one or two evaluation criteria among good color rendering, good similarity with color-matching functions, high transmission, wide color gamut, and low angular sensitivity, but often lacked one or several important criteria regarding image quality to compete with the organic color filters and justify a technological transfer to the image sensor semiconductor industry.

In this Letter, we first propose updating the previous assessment of Ag interference filters regarding suitability for high-performance color image sensing. We try to further optimize the performance of these filters with respect to revised evaluation criteria. Two directions are identified to maximize the transmission, namely, the introduction of designs derived from the induced transmission filter (ITF) theory, and the reduction of the Ag effective refractive index through the engineering of interfaces with dielectric layers.

The theory of ITF relies on the concept of potential transmittance of a metallic layer with a given thickness, which can be approached by adding dielectric quarter-wave stacks with a high and low index on each side of the metallic layer for a proper matching of the optical admittance of the external media [9]. Generally, this results in spectral responses with very high transmissions and narrow bands. Here, the challenge is to reach good colorimetric performances, SNRs, and high transmissions, together with a realistic process integration, based on a limited number of technological steps and a limited total thickness of the filter stacks. Since we target an integration compatible with small pixels (less than 5 μm), independent filter stacks such as used with the lift-off patterning process [10,11] are not allowed. Our patterned filter designs rely on architectures with a number of layers shared between all three RGB filters and patterning steps on a few layers with a common etching stop layer to generate a staircase with highly accurate thickness control [1]. This integration constraint translates into thickness equality constraints in the design of filter arrays. This is likely to broaden the resonance of the ITF designs, because the best matching stacks cannot be applied for each filter, but may finally help because the spectral responses should preferably look like the color-matching functions which are relatively broad. Therefore, we retain the idea of using two dielectric materials with an index contrast but, possibly, non-quarter-wave, as a new degree of freedom in the design of RGB Ag-dielectric filters. ${\mathrm{TiO}}_2$ and ${\mathrm{SiO}}_2$ are good candidates for this Letter, because they are available in most semiconductor clean rooms and provide a significant index contrast, together with high transparency. All the designs assume the Palik [12] optical constants for Ag layers. Several designs of patterned filters are investigated, with one or two metallic layers as in double-stage ITF designs, high- or low-index spacers adjacent to the metallic layers enabling different typical spectral widths, and several different matching layer stacks (Fig. 1). Assuming a filter integration at the end of the CMOS back-end stack under micro-lenses, the incident and output media are respectively defined as a transparent organic material and ${\mathrm{SiO}}_2$, neglecting the residual reflection of the AR coating usually present over the Si surface.

 

Fig. 1. Filter design types investigated. (a) 1M_H, (b) 2M_B, (c) 2M_H, and (d) 3M_H. The filter type name includes the number of metallic layers; H or B indicates a high- or low-index spacer layer. Layers with variable thicknesses between RGB filters are shown with arrows.

Download Full Size | PDF

The designs of the RGB filters should ideally be optimized with the best evaluation criteria representative of image quality perceived by the human eye, noise metrics, in particular [2], based on test images with numerical noise, and should be co-optimized with a suitable interpolation algorithm for a given spatial arrangement such as the Bayer color filter array [13]. Here, we only use three analytic criteria derived from filter transmittance curves, more compatible with the iterative algorithms of multilayer and multi-stack optimization for a faster, though more approximate ranking of the designs. We first wish to minimize the mean color error induced by the deviations of the filter spectral responses after a color correction matrix (CCM) from the color-matching functions and calculated with the $\mathrm{\Delta }E00$ color difference [14] root-mean-square value over the 24 patches of the Macbeth ColorChecker. The second objective is the minimization of the scene illuminance required to get an SNR of 10 defined by the traditional noise metric indicator YSNR10 [3]. The third criterion is the maximization of the sensor sensitivity which is proportional to the luminance defined from RGB signals in the ITU Rec. 601, $Y=0.299\mathrm{R}+0.587\mathrm{G}+0.114\mathrm{B}$, and normalized with respect to the value Y_REF obtained with reference organic color filters [15] combined with an IR-cut filter with a cutoff at 650 nm. Typical conditions of a daylight source, imaging system of a micro-camera embedded on a mobile device, and image sensor with 1.4 μm pixels are considered: D65 illuminant, scene reflectance 18%, optics transmission 80%, optics aperture 2.8, monochrome response of a 3 μm thick Si layer, frame rate 15 fps, noise floor 3e- and photo-response nonuniformity 1%. Due to a smoother cutoff at 650–700 nm and small, but nonzero, transmission in the IR for some of the Ag filter designs, the evaluation is performed over the whole 380–1000 nm range by extending in the IR the spectral responses of the color patches and the D65 curve, respectively, using measurements of a previous study [16] and the 6504 K blackbody law. In every optimization, the initial design includes quarter-wave matching stacks and spacer layer thicknesses given by the ITF theory for each of the RGB filters, or averaged between the RGB filters depending on the equality constraints. At each iteration on layer thicknesses, a new CCM is calculated by linear least squares to minimize $\mathrm{\Delta }E00$. The solutions of this multi-objective optimization problem are determined using Matlab optimization functions by varying the relative weights of the three evaluation criteria. More specifically, following the recommendation of a previous study [17], the designs should be compared at a fixed mean color error, for example, given by the reference filters. First, the designs are successively optimized with respect to each of the three criteria to identify the achievable extremes. Then, the Pareto front [18] is determined in the ($\mathrm{\Delta }E00$, YSNR10) space, regardless of the sensitivity, enabling the identification of a design with the best possible YSNR10 and the targeted $\mathrm{\Delta }E00$. Finally, another Pareto front is calculated in the (Y/Y_REF, YSNR10) space with $\mathrm{\Delta }E00$ fixed at the reference value 2.2, to investigate the compromise between these two criteria. The results for the various Ag RGB filter architectures are represented in Fig. 2. The evolution of the performances of the reference filters with the thickness of the organic filters is also plotted for a fair comparison, because thinner organic filters, in principle, may be considered to provide higher sensitivities.

 

Fig. 2. Simulated Pareto fronts [18] in the (a) ($\mathrm{\Delta }E00$, YSNR10) space, and (b) (Y/Y_REF, YSNR10) space for the four studied filter types, 1M_H (red), 2M_B (dark blue), 2M_H (cyan), and 3M_H (gray). The simulated performances of the organic color filters with external IR-cuts are plotted for reference (black) with decreasing thickness from 0.6 μm ($\mathrm{\Delta }E00=2.15$, $\mathrm{YSNR}10=66\mathrm{lux}$, $\mathrm{Y}/\text{Y\_REF}=1$) to 0.2 μm.

Download Full Size | PDF

It turns out that for all the Ag filter architectures, the derivative at the initial point of the Pareto front in the (Y/Y_REF, YSNR10) space is weak, enabling a significant improvement of filter transmission at the expense of a minor degradation of YSNR10. For example, the 2M_H designs with $\mathrm{\Delta }E00$ fixed at 2.2 increase their sensitivity by 16% with only 4 lux more. This is easily achieved by a reduction of the Ag layer thicknesses from 37 to 34 nm. The 2M_H ITF designs clearly have the best performances of all studied Ag filters, with all three criteria possibly quite similar to the reference filters at normal incidence and better at oblique incidence if a single CCM is used for all the pixels in the data processing. Interestingly, the performances of 2M_H designs are quite close to the optimized designs of RGB Ag filters with fully independent thicknesses achievable with the lift-off process. The 2M_B filters have a larger angular dependency. The 1M_H filters can have 22% improved sensitivity versus the organic color filters thinned down to 0.2 μm, with the same color error and YSNR10. However, they cannot reach the low YSNR10 value of the reference filters with standard 0.6 μm thicknesses, simultaneously with a small color error. The 3M filters designed with a single dielectric material definitely have lower sensitivity or higher YSNR10 than the reference filters at fixed $\mathrm{\Delta }E00$. The flexibility of the Ag interference filters, arising from the property of spectral tuning simply controlled by dielectric thickness, also clearly appears on Fig. 2. Very low values of $\mathrm{\Delta }E00$ (0.4) are possible, together with a still correct YSNR10 (80 lux), but reduced relative sensitivity (0.78). Conversely, some designs have better YSNR10 values (53 lux) than the reference filters, with higher, but possibly acceptable, color error (4.1) and 0.94 relative sensitivity. Samples of interesting spectral responses are shown in Fig. 3.

 

Fig. 3. Simulated transmittances of example optimized designs of (a) type 1M_H, (b) type 2M_B, (c) type 2M_H, and (d) type 3M_H, simulated at 0° (thick solid lines) and 25° incidence in air (thin solid lines). Typical transmittances of RGB color filters with external IR-cuts at 0° are shown in dashed lines. Performances at normal incidence are calculated for each design.

Download Full Size | PDF

The next key to improving the metal/dielectric filter performances is to increase the $k/n$ ratio of the metal [19], which enables transmission enhancement, while the rejection is maintained. Hereafter, $n+ik$ is the effective complex refractive index of the Ag layers, determined from the best fit of simulated versus measured transmittances [20]. Ag is the best choice among metals compatible with optical coatings, due to the low $n$ values all over the visible and near IR ranges. However, optical losses around the Ag/dielectric interface were reported in a number of studies [20,21] and interpreted by a plasmon peak generated by the interface roughness and increased by a high dielectric constant of the dielectric materials. This is in contradiction with the requirement on the high-index spacer layer mentioned above for high-performance RGB Ag filter designs. We first tried to evaluate the effective optical constants in our $\mathrm{Ag}/{\mathrm{TiO}}_2$ multilayer stacks deposited on 200 mm diameter glass wafers by magnetron sputtering at room temperature. The Ag layers were encapsulated by a thin capping of ${\mathrm{TiO}}_2$ deposited without ${\mathrm{O}}_2$ plasma to prevent Ag oxidation during the reactive sputtering process normally used for the ${\mathrm{TiO}}_2$ layers. A simple modeling of the stacks without roughness or thin interface layers was used to determine the effective optical constants of the Ag layers which include the optical contributions of both bulk and interfaces, from retrofit analysis of spectrophotometric measurements on a set of multilayer stacks specifically designed for the characterization [22]. At first glance, the Ag $n$ and $k$ were found, respectively, close to Palik and Johnson–Christy [23] data (Fig. 4), thus leading to a higher $k/n$ ratio than in the design study presented above. However, a plasmon peak was evidenced in the blue region of the $n$ spectrum, consistent with a significantly lower transmission in measurements of blue filters compared with simulations using Palik optical constants. We tried to enhance the filter transmission by introducing lower index interlayers as in previous studies [24]. Replacing ${\mathrm{TiO}}_2$ with SiN at both Ag interfaces shifted the plasmon peak in the UV and enhanced the blue filter transmission but, unexpectedly also degraded the transmission of the green and red filters. This was correlated with a rise of the Ag effective refractive index in the Drude domain, compared to interfaces with ${\mathrm{TiO}}_2$, but the interpretation at the microscopic level is still unclear. In contrast, with the ${\mathrm{TiO}}_2$ capping replaced by a SiN capping at the upper Ag interface only, the plasmon peak was both blue shifted and reduced, and the Ag effective refractive index was maintained to low values in the Drude region. This configuration provided the highest transmissions for the three filters, about 80%, and a full width at half-maximum of 100 nm for 1M_H filters.

 

Fig. 4. Measured (a) effective refractive index and (b) extinction coefficient of Ag layers with different bottom and top interfaces, ${\mathrm{TiO}}_2/\mathrm{Ag}/{\mathrm{TiO}}_2$ (dotted green), ${\mathrm{TiO}}_2/\mathrm{Ag}/\mathrm{SiN}$ (solid green), and SiN/Ag/SiN (dashed green). The modeling of Ag includes one Tauc–Lorentz and three Lorentz oscillators. Literature data are plotted for comparison.

Download Full Size | PDF

Finally, we fabricated periodic patterned arrays of RGBIR Ag filters with various pixel sizes from 100 to 1.4 μm, on 200 mm glass wafers for transmittance characterization. The designs were of the 1M_H type. A single 33 nm thick Ag layer with the thin interface layers addressed above was sandwiched between two high-index layers, each one including a staircase to tune the thickness from B to IR, formed by three cycles of SiN deposition and complete etching with a stop on the underlying ${\mathrm{TiO}}_2$ and three different masks. The external ${\mathrm{TiO}}_2$ and ${\mathrm{SiO}}_2$ layers of the matching stacks were common to all four filters. An additional 145 nm thick Si layer was deposited on the top and etched everywhere, except over the IR pixels, to cut the second-order resonance in the blue otherwise transmitted by the IR filter. The complete process included the deposition of 16 layers and seven patterning steps, with total thicknesses of 362, 417, 459, and 673 nm for the B, G, R, and IR stacks, respectively. We measured the transmittances of the patterned filters with a micro-spectrophotometer designed in the lab by coupling a microscope with a spectro-photometer. Figure 5(a) shows the measured spectral responses of the filters, in good agreement with the simulations using the Ag optical constants previously determined on the calibration stacks. The peak transmittances are around 80% for RGB and 65% for the IR filter. For a correct measurement of the patterned filters, the field diaphragm in the Koehler illumination system of the microscope had to be closed to illuminate a single pixel only and avoid the parasitic detection of light incident through the neighboring pixels and propagating up to the spectrometer entrance slit via reflections on the tube inner surface. This explains the very small residual parasitic bump visible in the RGB spectral responses at 800 nm. No significant change was found between the spectral responses of pixels with different sizes down to 5 μm. The pixels below 5 μm were not spectrally characterized because the high numerical aperture required in the illumination would blue shift and widen the transmittances. However, the visual aspect of the transmitted colors is similar to large pixels, at a fixed numerical aperture of 0.30, as illustrated in Figs. 5(b)–5(d).

 

Fig. 5. (a) Measured transmittances of RGBIR Ag-dielectric 1M_H filters patterned on a glass wafer with a 50 and 5 μm pixel pitch (i.e., thick/thin solid lines), compared with simulated transmittances (dashed lines). (b)–(d) Transmission microscope images of RGBIR filter mosaics with 50, 5, and 2 μm pixels.

Download Full Size | PDF

In summary, we have proposed new designs of RGB Ag/dielectric interference multilayers based on the ITF concept to optimize the sensor sensitivity, together with the high performance of the other parameters for color imaging, i.e., color restitution and SNR at normal and oblique incidence. We also improved the filter transmission by reducing the parasitic plasmonic effects at Ag interfaces with a careful choice of thin dielectric interlayers. To the best of our knowledge, this is the only alternative filtering technology able to compete with the organic color filters in terms of imaging performance. The technology is more complex than the organic color filters due to a higher number of deposition, lithography, and etching steps. The real advantage of these filters is the ability to integrate other channels, in addition to RGB, such as panchromatic or IR, which is not achievable with the organic color filters alone. Despite the 80% transmission measured in our RGBIR filter arrays, we still have not experimentally shown the full potential of the technology yet, because a 30% sensitivity enhancement compared to the organic color filters is expected from simulations with the optical constants of Johnson and Christy for the Ag layers. This may be achievable with higher quality interfaces of the Ag layers or other deposition techniques. With a thickness of a few hundreds of nanometers and a patterning process with high spatial resolution, this may be the unique solution for RGBW or RGBIR high-performance imaging with pixels smaller than 3 μm.