1. Introduction

Multispectral imaging is an efficient technique for object identification, diagnosis, control or sorting, and therefore has an increasing number of applications in agriculture, machine vision, biomedicine, or defense and security [1]. Specific information can often be extracted from the spectral signature of objects in the visible (VIS) and near-infrared (NIR) ranges, where low-cost CMOS image sensors are available. Since a minimum of five to ten band-pass filters with different and arbitrary peak wavelengths are usually required, the standard red, green and blue (RGB) organic resists provided on CMOS sensors should be replaced by a new filter technology. A monolithic integration of the whole set of filters on the sensor is preferable, to benefit from snapshot operation, low form factor, and low cost fabrication, which are critical for fast changing scenes and tiny environments, for example in endoscopy or inspection with drones.

Several demonstrations of plasmonic multispectral filters have been achieved on CMOS image sensors in previous studies [2, 3]. Multiple filters corresponding with variable lateral dimensions may be realized with a single patterning step in one or a few metallic layers. However, in the VIS and NIR domain, plasmonic filter designs can hardly provide high peak transmission together with high spectral selectivity, ideally desired for high performance multispectral cameras. Another approach for fully integrated multispectral imaging has been shown with vertical silicon nanowires [4], spanning the VIS and NIR ranges by simply changing the nanowire diameter. The subtractive nature of the filters might induce some noise in the spectrum estimation as in three-channel color sensors [5].

Additive filters with selective spectral responses can be obtained with simple Fabry-Perot (FP) thin film stacks, where the variation of the optical thickness of one single layer allows to tune the wavelength across a wide spectral range without significant degradation of the spectral response. The monolithic integration of multispectral FP filters [6–8] relies on the variation of the physical thickness or the effective refractive index of the optical cavity.

The staircase filter architecture [7] obtained by varying the cavity physical thickness requires either numerous etching steps if they are performed filter by filter, or relations between filter wavelengths if some etching steps are common to different filters, as with the combinatorial etching technique [6, 9]. In addition, the process involves multiple partial etching steps of a layer. A tight control of both etching and monitoring equipments is required to avoid thickness errors and inaccuracies in peak wavelengths.

Alternatively, combining the advantages of both plasmonic and staircase FP filters, spectral tuning can be achieved on chip by spatial modulation of the refractive index in a FP cavity with constant thickness. Sub-wavelength patterns are formed in a first transparent material and subsequently gap-filled by a second transparent material with a different refractive index. The effective refractive index and peak wavelength are directly related to the volume ratio. One single patterning step is necessary whatever the number of filters, and the spectral responses may show high transmission and rejection. Initially proposed by Kaushik and Stallard [10], the concept was demonstrated for multispectral purposes in the VIS range with CMOS compatible materials, Al as FP mirrors, patterned SiN and PMMA for filling [11]. Integration of multispectral NIR filters was performed on a CMOS imager with Si and SiO₂ for both the nanostructured layer and the thin film Bragg reflectors of the FP cavity [8].

In the first part of the present study, we experimentally illustrate the ability of multispectral nanostructured FP filters to provide precisely constant peak wavelength across the whole surface of an image sensor in the focal plane of an optical system [12] (Fig. 1). This property has never been emphasized before, and may be particularly useful in the field of active 3D imaging based on structured light [13] or time-of-flight [14], where the band-pass filter is a thin film interference stack today. This filter is used to transmit the light emitted by a monochromatic source and back-reflected by the scene, and to reject ambient parasitic light. Whatever the position of the filter in the optics and sensor system, its band-pass has to be wide enough to anticipate the unavoidable spectral shift caused by oblique incidence in the field of view.

 

Fig. 1 Principle of blue-shift compensation by nanostructured layer in FP filters. 1) At the center of the sensor, light is normally incident and the filter is centered at λ₁. 2) At the sensor edge, the oblique incidence of light induces a blue-shift of the filter response: λ₂ < λ_1. 3) By changing the lateral dimensions of the nanostructures, the effective refractive index of the FP cavity is increased and the blue-shift is compensated: λ₃ = λ₁.

Download Full Size | PDF

For optical systems with a field of view significantly larger than the optical aperture, if the filters are on-chip FP, the spectral width can be narrowed in the design. This is because the spectral shift with respect to normal incidence can be completely compensated with a correction only depending on the incidence angle, variable over the sensor surface. With nanostructured FP filters, all corrections are simultaneously realized with a single patterning step, simply adjusting the lateral size of the nanostructures. Therefore, with narrower band-pass filters, and provided the wavelength of the source is stable enough, this technology can provide higher signal to noise ratio and higher range in distance measurements, or enable the use of a lower power source.

In the second part of the study, we try to address the issue of extending the spectral range of high-performance multispectral nanostructured FP filters, with a design compatible with simultaneous integration of all filters on a single chip. The targeted range is the complete VIS and NIR domain, to exploit the full detection bandwidth of Si detectors. The technology should be compatible with small pixels. The corresponding process flow should be as simple as possible, with the minimum number of technological steps, using a single set of materials for all the filters. The filter architecture has to minimize the possible errors caused by process inaccuracies, keeping in mind the wafer-level manufacturing issues in semiconductor foundries.

2. Compensation of blue-shift under oblique incidence with nanostructured filters

We first estimate the refractive index change required in the spacer layer inside the optical cavity to compensate the blue-shift of the resonance wavelength in interference FP filters illuminated in oblique incidence.

The blue-shift of the resonance peak essentially arises from the reduction of the phase shift experienced in the spacer layer under oblique incidence [15]. The usual phase Eq.:

First order derivation of (1) with respect to θ provides approximate expressions of the blue-shift δλ (Table 1) [15] in several typical cases of FP filters. For FP filters with a single dielectric layer between metallic reflectors, the derivation relies on the approximation of constant parameters φ_a and φ_b in small ranges of λ and θ_r, which is mainly correct in s polarization but tends to over-estimate the blue-shift in p polarization. For all-dielectric FP filters with Bragg mirrors, the spectral shift includes a significant contribution from φ_a and φ_b but remains a blue-shift in both cases of low- and high-index spacers. Whatever the FP type, the blue-shift follows a quadratic dependence with the incident angle θ_i. The blue-shift may be somewhat minimized by choosing a high-index spacer for all types of FP, and furthermore with a high cavity order in all-dielectric filters with high-index spacer. However, this also modifies the peak transmission and the spectral width of the filter.

Table 1. First order estimate of spacer index variation δn for compensation of blue-shift δλ in FP filters*

View Table | View all tables in this article

The technique for blue-shift compensation proposed in this paper may be used for all types of FP filters and lets the designer free to choose the spacer index and cavity order depending on the design specifications. If the incident angle is locally constant on the filter surface, and the spacer thickness is kept constant, the index variation δn required for the compensation may easily be derived from Eq. (1) again. Approximate expressions for δn are given in Table 1 in the above mentioned cases of FP filters with metallic and Bragg mirrors, neglecting the variations of reflection phase shifts φ_a and φ_b with the spacer index in the first case.

It can be shown from the formulas of Table 1, that the index variation δn required for the blue-shift compensation is always higher in all-dielectric band-pass filters with n_H and n_L alternating quarter-wave layers as Bragg mirrors, compared to metal-dielectric filters with any spacer index between n_H and n_L. In addition, for all-dielectric filters, δn is higher for high-index than for low-index spacers, whereas this is the inverse in metal-dielectric filters. The question arises, whether it is possible to provide the index variation in all filter types, with usual dielectric materials and standard patterning tools.

Trying to address that issue in the field of 3D active imaging consumer applications, we consider a typical optical system with ± 25° chief ray angle, f/2.5 aperture, and a working wavelength of 800nm. The resonance peak is blue-shifted by a few tens of nanometers at the edge of the field of view, considering refractive indices of materials commonly available in semiconductor foundries, ranging from 1.45 with SiO₂, 2 with SiN, 2.2-2.3 with TiO₂, up to 3.4-3.5 with Si at 800nm. Only the case of metallic FP filter with Si spacer may not require blue-shift compensation, since the shift is only a few nanometers, much less than the spectral width. For all-dielectric FP filters with low index contrast n_H/n_L of a few 10⁻², the compensation may not be achievable because δn tends to diverge, but this kind of design has no practical interest. For most of the other designs, the blue-shift can be completely compensated with a spacer index variation of a few 10⁻² to 10⁻¹. These variations are accessible with a suitable effective index medium composed of two transparent materials such as those mentioned above, with a refractive index difference of a few 10⁻¹.

The effective index of the nanostructured spacer may be derived from an effective medium theory like the Maxwell-Garnett model [16]. A simple linear interpolation between high and low index values seems to be sufficient for a first order estimation of the required fill factor.

The compensation of the blue-shift should ideally be gradually increasing from on axis to peripheral pixels. As the interest of the technology is to provide simultaneous compensation of all blue-shifts over the whole sensor or filter surface with a single lithography step, the minimum amount of blue-shift to be compensated has to be considered, and compared to the smallest achievable lateral size and variation of size in the nanostructured layer. Standard semiconductor technology generally allows fine tuning of the lateral size, e.g. a few nanometers, while the minimum size is a few tens of nanometers. It turns out that the best practical solution to manufacture filters with spatially varying compensation of the blue-shift, relies on the nanostructuration of the spacer for all the pixels, including at sensor center and edge.

Four typical examples with simple FP filters designed in the first cavity order on glass substrate are now presented to illustrate the compensation of the blue-shift with a nanostructured layer. The design wavelength is 800nm. The reflectors are either Ag layers or Si/SiO₂ quarter-wave Bragg mirrors (Fig. 2). High-index and low-index nanostructured spacers are considered, composed of different volume ratios of SiN and SiO₂ for the Ag mirrors, Si and SiO₂ in the case of the Bragg mirrors. Double cavity designs are chosen for the Ag filters for a higher selectivity. Because the resonance wavelength of double cavity is similar to single cavity filters, the blue-shift and the index variation for compensation are still given by Table 1 in first approximation.

 

Fig. 2 Left: Design (left) and simulated transmittance (right) of several simple FP filters on glass: with Ag reflectors (designs 1, 2), Bragg reflectors (designs 3, 4), low-index spacer (designs 1, 3), high-index spacer (designs 2, 4). For each filter, the normal incidence filter response (black dashed line) is blue-shifted at 25° incidence (grey dotted line). The blue-shift is compensated (red plain line) by modifying the effective index of one or two nanostructured layers (see Table 2 for the dimensions of the nanostructures). The effective index is shown in the designs, without blue-shift compensation (black), and with compensation (modifications in red).

Download Full Size | PDF

For low-index spacers (resp. high-index), the minimum width of the patterns (for 0°, resp. for 25°) is in the range of 50 to 100nm with a square geometry, accessible in technology. The period of the nanostructures, 200 to 300nm, is small enough to avoid light losses by diffraction and subsequent decrease of the transmittance in the zero order. For each filter, the variations of fill factor related to δn given by Table 1 are first estimated with the effective index theory. Then, the corresponding dimensions are refined by numerical simulation using a Rigorous Coupled Wave Analysis (RCWA) software, to compensate the blue-shift accurately. The resulting spectral responses are shown in Fig. 2 in each case. The difference between the exact results for un-polarized light shown in Table 2 and approximate values calculated from Table 1 is typically 10 to 20%, but the general trends are confirmed. The required size variations for compensation of the blue-shift are also in the scope of the patterning technology, ranging between 15nm and 40nm for the full 25° variation (Table 2), between 7nm and 20nm for gradual compensation with 5nm blue-shift increments. With cavity thicknesses of 120nm to 250nm, the aspect ratio of the patterns remains at moderate values between 1.8 and 3.5 for the patches, 1 and 1.7 for the gaps in between. This shows that the blue-shift compensation technique with nanostructured filters is realizable in principle for various architectures of FP filter designs.

Table 2. Characteristics of the nanostructured layer for blue shift compensation in interference filter designs (see Fig. 2)

View Table | View all tables in this article

In the case of the FP filter with Si/SiO₂ quarter-wave Bragg mirrors, the compensated spectral response has zero shift but significantly altered shape and half transmission compared to initial filter. However, all the characteristics of the initial spectral response can be maintained in the compensated filter by proper refinement of the filter design [17], targeting similar transmittances and similar blue-shifts in both s and p polarizations (Fig. 2, designs 3 and 4).

The simulations and analysis presented above try to investigate the applicability of the compensation technique in various filter cases with dimensional parameters accessible in standard technology. We now present a demonstrator example to further assess the manufacturability of the technology.

For our demonstrator, we choose CMOS compatible materials developed for interference filter integration in a previous study [18], namely Cu for the mirrors, SiN for the patterning, and SiO₂ for the gap-filling. The design is a simple single-cavity FP stack (Fig. 3(a), design 5), with a resonance set at 800nm. The nanostructured spacer for 0° incidence is of low-index type, with volume ratio SiO₂ 91% and SiN 9%, obtained by SiN square patches with 60nm lateral size and 200nm period (Table 2) arranged in a regular array. The spacer thickness is set at 120nm to limit the aspect ratio for the gap-filling around 1. Cu layers have to be sandwiched between SiN layers for a correct adherence. Below the spacer, an etching stop SiO₂ layer is required. Therefore, the filter is designed in the second order resonance, which in itself does not change the blue-shift in the case of metal-dielectric FP filters. However, these homogeneous layers are part of the reflecting stacks on each side of the spacer and the approximation of constant phase shift in the angular and spectral neighborhood does not stand. It is difficult to derive approximate expressions for the blue-shift and the index variation for compensation, because the reflections at dielectric/dielectric layer interfaces inside the FP cavity are not negligible. In practice, δn is determined by trial and error in numerical simulations. The exact values calculated by RCWA are δλ = −24nm at 25° and δn = 0.11 (Table 2). The volume ratio in the spacer for compensation at 25° is 75% SiO₂ and 25% SiN, which corresponds to SiN patches with 100nm lateral size and 200nm period. Both gratings for 0° and 25° are formed at a few millimeters distance with a single mask on the same wafer. The whole filter deposition is performed on a bare Si wafer covered by a SiO₂ and SiN anti-reflective coating and a 1µm thick SiO₂ layer simulating the back-end stack of a front-side CMOS image sensor, to optimize the filter process in conditions representative of a later integration. Therefore, spectral responses are measured in reflectance.

 

Fig. 3 (a) Design of Cu/SiN/SiO2 FP filter on Si substrate and back-end stack, the effective index for compensation of blue-shift at 25° is shown in red. (b) Measured reflectance in normal incidence (dark dashed line), at 25° incidence (grey dotted line), at 25° with blue-shift compensation (red line). (c) and (d) SEM cross sections of the filter. The nanostructured layer is realized by patterning a SiN layer and gap-filling with SiO₂. SiN patches arrays with different lateral dimensions, for example 80nm in (c) and 120nm in (d) are realized in a single patterning step on the wafer.

Download Full Size | PDF

From the measured spectral responses in Fig. 3(b)), the compensation of the blue-shift at 25° incidence appears almost complete, since the blue-shift is reduced from −24nm to −2nm with the 40nm lateral size increase of the SiN patches. This result shows that the developed patterning and filling process is accurate enough to provide the targeted volume ratios of the two dielectric materials. Figure 3(c) and 3(d) show the good morphological quality of the filter in cross section. It is noteworthy that the compensation is fully achieved, although the demonstrator has been realized without any specific control of layer optical constants and thicknesses, because this work does not aim the precise control of nominal spectral responses. Incidentally, the absolute peak wavelength and spectral shape somewhat differ from the simulated ones, which are not shown on Fig. 3(b) for clarity.

It is interesting to observe that the technique is not only applicable for simple case studies, but also for more complex filter designs. Let us consider for example a band-pass filter suitable for the detection of a light source with some dispersion on the emission wavelength specified in a ± 10nm range around 800nm. A design with 20nm spectral width and ~90% transmission may be proposed with Si and SiO₂ non quarter-wave layers, including a single nanostructured Si/SiO₂ layer with variable volume ratio for detection of the same spectral range between 0° and 25° incidence (Fig. 4 and Table 2, design 6). Without compensation, the spectral width of the filter should be 30nm to take into account the 10nm blue-shift at 25°. Here, the design enabling compensation is obtained in two steps. A first optimization is performed with a spectral target but without blue-shift compensation, using a thin film software [17]. The most sensitive layers regarding their refractive index are identified. Then, the design is refined by simultaneous optimization of the stacks for 0° and 25° in average polarization, with Matlab optimization functions. All the layer thicknesses are set equal between the designs at 0° and 25°, while their values are the variables of the optimization, together with the index variation of one of the most sensitive layers.

 

Fig. 4 Left: Designs (left) and simulated transmittances (right) of an IR band-pass and an IR-cut filter on glass. For each filter, the normal incidence filter response (black dashed line) is blue-shifted at 25° incidence (grey dotted line). The blue-shift is compensated (red plain line) by modifying the effective index of one single nanostructured layer (see Table 2 for the dimensions of the nanostructures). The effective index is shown in the designs, without blue-shift compensation (black), and with compensation (modifications in red).

Download Full Size | PDF

Another example is shown with an edge filter designed to transmit the VIS domain and reject the NIR wavelengths up to the cut-off of Si. This type of IR-cut filter is used in the optical system of most of color image sensors nowadays, in addition to RGB color resists which transmit NIR light. A typical design with 32 TiO₂ and SiO₂ layers is basically obtained with Optilayer [17], showing a 20nm blue-shift of the cut-off at 25° incidence. The narrowing of the resulting red filter due to the blue-shift of the IR-cut is difficult to tackle by basic signal processing because the signal is lost. The correction of the so-called color shading effect at the edge of the field of view may be achieved by specific algorithms [19, 20]. Our approach provides a solution at the device level and does not require any estimation of the illuminant. Here again, the blue-shift may be fully compensated by modifying the effective index of a single layer (Fig. 4 and Table 2, design 7). With 32 layers and a total thickness of 3.7µm, the filter is hardly integrable on a CMOS image sensor due to the limited thickness of standard photoresist in the lithography step for pad opening. The preferred integration is the hybridization of a glass wafer onto the sensor, with the coated side facing the sensor. The glass wafer may simply be the handling wafer used in the standard through-silicon-via process [21].

3. Multispectral Fabry-Perot filters with extended spectral range

In this section, we investigate how to extend the maximum tuning range accessible through the effective index variation of the spacer layer in a Fabry-Perot design. Five limitations prevent to span the whole VIS and NIR range: the limited extent of the high reflectivity domain of the reflectors, the refractive index difference between available transparent materials, absorption in dielectric layers especially in the low wavelengths, the presence of the second-order resonance in the low wavelengths when the first-order resonance is in the NIR, and the dispersion of phase shifts upon reflection on the mirrors.

Mirrors with high reflectivity over the VIS and NIR domain can be realized with Ag, taking advantage of the high k/n (extinction coefficient/refractive index) ratio over the whole range, but hardly with other metals, or with dielectric mirrors because the width of the stop-band is limited by the refractive index contrast between alternating layers.

Among the standard dielectric materials with high transparency over the VIS and NIR domains, SiO₂ and TiO₂ have one of the highest index difference (Δn ~0.8). Since the variations of the phase shift at reflection on Ag mirrors remain small in a broad spectral range and the free spectral range is large between the first and second order resonances, the range of the spectral tuning achievable from variable effective index in a nanostructured SiO₂/TiO₂ FP spacer is limited by the refractive index difference. If the minimum wavelength is set at λ_min, the maximum wavelength λ_max is solution of the Eq.:

In this respect, the staircase implementation of FP multispectral filters [7] potentially offers an extended spectral range, since it is not limited by any refractive index contrast and the spacer thickness can be increased as long as the second order does not appear. The VIS and/or NIR domains may be addressed depending on the choice of materials with sufficient transparency in the targeted domain. However, the practical wafer-level implementation has to deal with process inaccuracies, if partial etching is used, as outlined in the introduction. An alternative process to realize the staircase [22, 23] relies on alternating deposition and etching steps of a first dielectric material, over an etching stop layer common to all the etching steps and made of a second dielectric material. This is efficient for suppressing the etching non uniformity over the wafer surface. However, when more than three filters are targeted, the complete process requires numerous technological steps including possible inaccuracies on the various deposited thicknesses.

Trying to overcome these limitations, it is natural to think of both modifying the physical thickness and the effective index of the spacer to access wider ranges of the cavity optical thickness and spectral range of the resonance. This idea was already implemented, although on separate substrates, in a previous study [11] where the VIS and NIR domains are addressed through the nanostructuration of two spacers with different cavity lengths.

The approach we propose here tries to address the issue of spectral range extension, together with a technological process of limited complexity and compatible with the realization on a single substrate. The architecture of the filter array is shown on Fig. 5(a), in the case of single cavity FP filters.

 

Fig. 5 Cross section sketch of multispectral FP filter array with extended range in VIS and NIR range. For single cavity designs (left), the spectral tuning is achieved by one single nanostructured layer with variable lateral dimensions of the patterns. The extension to the NIR is obtained by an additional homogeneous layer inside the FP cavity. Second-order resonances are eliminated by absorption in a specific layer on top of the stack. The double cavity designs (right) provide higher spectral selectivity and requires a symmetric arrangement in the vertical axis to enable planarization after each gap-filling step.

Download Full Size | PDF

The architecture is based on three characteristics:

The thickness of the nanostructured layer is set by the minimum or the maximum wavelength of the VIS filters, considering that they are respectively obtained with n_L and n_H only. The nanostructuration provides the intermediate wavelengths. For the NIR filters, the thickness of the spacer may be chosen so that the NIR spanned range is adjacent to the VIS range. The maximum wavelength of the NIR filters is lower than in the configuration where the full thickness of the NIR FP cavity is nanostructured, but this is necessary for the realization on a single wafer. However, choosing a low-index spacer provides a higher spectral range for the NIR filters, compared to a high-index spacer. For the filters with the highest wavelengths, the second order resonance appears in the VIS domain. It may be strongly attenuated by adding a layer with high transparency in the NIR but absorption in the VIS, for example in Si. This layer can be deposited on top of the stack and etched on some of the filters.

An example of design and simulated transmittances on glass are shown on Fig. 6. The multispectral FP filters use Ag for the mirrors, TiO₂ and SiO₂ as high- and low-index dielectric materials, with a single and a double cavity architecture respectively on left and right parts of the figure. All the simulations are performed including the metal and dielectric losses. The effective indices are calculated with the Maxwell-Garnett formula [16]. The corresponding characteristics of the nanostructured spacers are displayed in Table 3. TiO₂ can be etched in plasma [24] and subsequently gap-filled with SiO₂. Alternative high-index and transparent materials in that spectral range are Ta₂O₅, ZrO₂ or Nb₂O₅.

 

Fig. 6 VIS designs (a, b), NIR designs (c, d), and simulated transmittances on glass (e, f) of multispectral Ag/TiO₂/SiO₂ FP filters with either single-cavity (a, c, e) or double-cavity (b, d, f) architectures. The designs rely on one single nanostructured layer per FP cavity with variable lateral dimensions of nanostructures for spectral tuning, an additional homogeneous spacer in the FP cavity to access the NIR domain, and an absorptive layer on top of NIR filters for the elimination of the shortwave sidebands.

Download Full Size | PDF

Table 3. Characteristics of the nanostructured layer in Ag / TiO₂ (H) / SiO₂ (B) multispectral FP filters

View Table | View all tables in this article

The spectral range of the single cavity filter set covers the VIS and NIR 450-950nm range, although the NIR FP cavity is not fully nanostructured. Only 2 lithography steps are required in the whole process, one for the patterning of the cavity, and the other for the etching of the Si layer on top of the stack. In this example, 11 filters with 30nm spectral width and 50nm regular spacing are designed. The number of filters can be increased without changing the process. In the 450-950nm range, any set of wavelengths can be realized, only limited by the smallest dimension and the smallest dimension variation provided by the lithography and patterning process. The spectral width of the filters can be reduced to decrease the correlation between filters. As usual in the design of metal/dielectric FP filters, this is at the expense of a lower transmission, except with transmission-induced designs [15], but they are not suitable to cover a broad spectral range with a high transmission because the properties of the multilayer dielectric matching stacks are not broadband. Alternatively, the spectral content of the scene may be reconstructed taking advantage of non-discrete spectral bands [25, 26]. Therefore, relatively broad Ag FP filters with high transmission should be suitable for multispectral detection in low light conditions.

The double cavity design provides a higher selectivity and lower variations of the peak transmission. It is constrained by the requirement of a flat surface below each of the two nanostructured layer, since the process includes chemical-mechanical polishing steps. Nevertheless, appropriate thicknesses can be found for the first dielectric layer in both the VIS and NIR stacks, resulting in relatively high transmission. The whole process for the double cavity multispectral FP filters requires 5 lithography steps, including 2 for the nano-patterning.

With a total stack thickness of 0.3µm and 0.7µm for the single- and double-cavity solutions, optical cross-talk effects should be avoided considering an integration on sensors with small pixels, of the order of 1µm. Edge effects are supposed to have minor impact on the spectral responses if the pixel size covers at least three periods of the nanostructures.

4. Conclusion

In this study, we tried to investigate the potential of the dielectric nanostructuration technology to address two major issues of interference filter arrays, focusing on the field of VIS and NIR image sensors.

The first issue is the blue-shift under oblique incidence, one major usual drawback of interference filters. Rather than suppressing the angular dependence, we proposed a simple technique to compensate for it when the incidence angle is variable over the filter surface, for example in the image plane of an imaging system with low aperture compared to the chief ray angle. The red-shift results from the increase of the effective refractive index of the layer within the filter stack which is the most sensitive to refractive index changes. This layer is realized by nanostructuring a first dielectric material and gap-filling with another dielectric material with a different refractive index. We realized a demonstrator with Cu filters, patterned SiN and gap-filling with SiO₂ to show that accurate compensation can effectively be achieved with standard semiconductor manufacturing tools, and independently of the fine tuning of the nominal spectral response. Therefore, this technique for blue-shift compensation is robust regarding layer optical constants and thicknesses. The main technological requirement is the control of lateral dimensions in the patterning process. The key point of the effective index method is its simple technological implementation with a single patterning step even when a number of intermediate corrections are desirable, providing gradual compensation of the blue-shift.

The technique is applicable for both high and low index spacers, for FP designs based on metallic or Bragg mirrors, and also for non Fabry-Perot interference filters as shown in a few examples of band-pass or edge filters. It is therefore possible to design narrower band-pass filters for 3D active imaging, allowing higher distance measurement range, or lower power consumption and lower cost of the light source. It is also possible to compensate the blue-shift of the cut-off wavelength in IR-blocking filters, and suppress the color shading in highly demanding imaging applications.

The second issue is the extent of the spectral range of multispectral filter arrays over the whole VIS and NIR range, using a single set of materials to minimize the process complexity. Our solution advantageously combines the variations of layer thickness and refractive index to increase the spectral range while minimizing the number of patterning steps. One single etching step with sub-wavelength patterns proves to be necessary whatever the number of filters, or two if a higher selectivity is desired. Here also, the robustness regarding layer thicknesses is a key advantage of the process. This technology may be suitable for manufacturing multispectral filter arrays with high selectivity in a monolithic or hybrid integration on a number of devices such as image sensors or displays. The technology is compatible with a small pixel size, typically 1µm, since the total thickness of the filters is 0.3 to 0.7µm. The simplicity of the process is particularly advantageous over a pure staircase approach as soon as more than three different filters are required, as in RGBIR imaging or multispectral imaging.

It turns out that for metal/dielectric filters, small index variations in a single nanostructured dielectric layer are generally sufficient to achieve blue-shift compensation, while a large index contrast is required for multispectral filtering in a large spectral range. Ultimately, in multispectral imaging systems with large fields of view, both functions may be simultaneously addressed by a single metal/dielectric filter array, to provide a filter set spanning the VIS and NIR domains with constant wavelength for each channel across the whole field of view, by simply varying the lateral dimensions of the patterns in a single lithography step.

Therefore, we believe that the traditional thin film interference filters, combined with modern and accurate patterning technologies, remain the most competitive filter architecture in many applications with demanding specifications about optical performance, technological complexity and manufacturing tolerances.