Visible to long-wave infrared chip-scale spectrometers based on photodetectors with tailored responsivities and multispectral filters

2.1 Background

Nanophotonic FADA microspectrometers are attractive for researchers as they are able to leverage current imaging technologies such as CMOS or CCD camera sensors and adapt them into microspectrometers by adding their own custom filter arrays. These commercial sensor arrays provide benefits such as a large number of pixels, on-board electronics to digitize and read-out intensity values, high dynamic range and sensitivities. Early versions of microspectrometers were commercial imaging sensors equipped with linear variable filters (LVF) [34], [35], [36], or a set of Fabry–Perot etalon bandpass filters [32]. These filter arrays were straightforward to fabricate, requiring only thin film deposition and no lithographic fabrication steps in the case of LVFs or photolithography to define the etalon pixels. These types of optical filters produce narrowband transmission spectra and as such do not usually require the implementation of a reconstruction algorithm, as a direct pixel read-out method is applicable.

The development of spectral reconstruction algorithms for broadband filter arrays [39], [52] allowed for filter arrays with more sophisticated transmission spectra to be used in FADA microspectrometer configurations. These include using chemically synthesized colloidal QDs as absorptive filters [45], PhC slabs [20], [42] and plasmonic nanoantennas [38] or apertures [40]. The FADA approach to measuring test sample transmission spectra is outlined in [40] and reproduced in Figure 1(d). Light from a source, with its own characteristic spectrum is passed through a test sample, where the source spectrum is modified by the wavelength dependent absorption and scattering of the sample. Light from the sample is then filtered by the nanophotonic filter array and collected by a set of suitable detectors. The measured photoresponse is then used with the known transmission spectra of each filter and that of the source to find the sample’s transmission spectrum.

2.2 Quantum dot absorptive filter array

Figure 2(a) shows a color photograph of a set of chemically synthesized core–shell colloidal Cadmium sulfide (CdS)/Cadmium selenide (CdSe) QDs of varying composition and radii [45]. Each circular aggregate of QDs forms an absorptive spectral filter, where all the QDs in the filter have the same radius and composition. By varying the growth conditions, it was possible for the authors to create 195 different QD filters with distinct, broadband transmission spectra, a selection of which can be seen in Figure 2(b). The QDs are then encapsulated in a polyvinyl butyral film and each of the 195 filters transmission characterized. The QD filter film is then attached to a commercial monochrome silicon CCD camera shown in Figure 2(c).

Figure 2:

Colloidal quantum dot (QD) FADA microspectrometer. (a) 195 absorption filters formed from uniquely sized QDs embedded in a polyvinyl butyral film. Transmission spectra of nine of the QD filters in the array (b). Filter array integrated into a commercial charge-coupled device (CCD) camera (c). Measured (crosses) and reference (solid line) spectra using the QD microspectrometer (d)-(g). Reprint permission obtained from [45].

Broadband test spectra were generated, and images were captured with the QD-FADA camera. Each pixel in the captured image was then either registered to a particular QD filter or not used as the light passes through the membrane unfiltered. By using least-squares L2 regularization the authors were able to reconstruct visible spectra ranging from 400 to 600 nm in wavelength as shown in Figure 2(d–g). This QD microspectrometer is also capable of reconstructing narrowband spectra, down to linewidths of 3 nm. The stability of the QD filter array was characterized by performing measurements of the same narrowband test spectra 6 months apart. The authors found no degradation in the function of the QD microspectrometer after this time period.

One of the limitations of using QDs as absorptive filters in a FADA microspectrometer include the heavy metals used to make these QDs are toxic and may in future be phased out of use because of this. Furthermore, the material choice limits the spectral range to around the band gap energies of the QDs, so that the microspectrometer range is limited by the filter array design, not by the intrinsic responsivity of the detector array. Finally, the authors evaluate the viability of a QD-FADA microspectrometer with 147 filters instead of 195 filters and found the test spectral reconstructions to be less accurate. This suggests limited viability for mass production as in this design each spectrometer will need 195 different QD synthesis steps. Nevertheless, filters based on QD absorption have natural advantages over the “top-down” fabricated filters as their transmission spectra have no angular or polarization dependence, whereas thin film, PhC slab, plasmonic nanoantenna and grating-based spectrometers often do.

2.3 Plasmonic and photonic crystal slab filter arrays

PhC slab filters consist of regular arrays of air holes in a dielectric film, where the transmission through the slab is determined by the array period, lattice constant and diameter of the holes. Figure 3(a) shows a schematic of a PhC-FADA microspectrometer, where the filter array consists of 36 distinct PhC designs etched into 500 nm of silicon-on sapphire and attached to a commercial CMOS sensor [42]. Simulations carried out suggested that increasing the number of filters beyond 30 had little effect on reducing the mean-squared-error of the simulated spectrometer chip’s spectral reconstructions. Each filter pixel is 32 × 32 μm and the total filter array is just 200  × 200 μm. Figure 3(b) shows the transmission spectra for three of the PhC filter arrays. The complex nature of the transmission spectra is exploited here to reduce the cross-correlation of each filter’s transmission, which has been shown to improve microspectrometer accuracy [53]. The PhC filter array has less than 70% transmission, perhaps due to absorption in the silicon used as the filter material, this reduces the overall sensitivity of the spectrometer chip, compared to a grating spectrometer. Here L2 regularization is used to estimate the incident light spectrum. Figure 3(c) shows the PhC-FADA spectrometer chip measured spectra as well as reference spectra of light from two different LED light sources. The results are in good agreement. The spectrometer chip was also used to measure monochromatic light with wavelengths from 550 to 750 nm. The spectral resolution for such measurements is claimed to be about 1 nm across this wavelength range. An interesting application of this spectrometer is shown in Figure 3(d), (e). Here the authors generate two beams with distinct spectra (see 3(d)) but appear to have the same color and indeed have the same Red-green-blue (RGB) color values. The PhC-FADA spectrometer is able to easily distinguish between these metameric beams, something a color CMOS camera could not do. The PhC slab filter array does, however, have a clear drawback: the transmission spectra are highly sensitive to the angle of incidence of the incoming light. The authors evaluate this and find that the reconstructed spectrum is accurate only up to a difference of 3° angle of incidence between the transmission spectra measurements of each filter (calibration) and the beam under test.

Figure 3:

PhC filters integrated onto a CMOS sensor array (a). Transmission spectra of three of the PhC filters (b). Reconstructed (blue) and reference (red) LED spectra (c). PhC slab microspectrometer can be used to distinguish two metameric colored light sources (reconstructed and reference spectra (d) and appearance and Red-green-blue (RGB) color values (e)). scanning electron microscope (SEM) and measured transmission spectra of LWIR filter array microspectrometer based on plasmonic apertures in a gold film. Reconstructed (red) and reference transmission spectra (blue) of various plastics: cellophane (g), polyvinyl chloride (PVC) (h) and Polyethylene (PE) (i). Reprint permission obtained from [40], [42].

Unlike PhC slab filters, many plasmonic filter arrays designs with transmission spectra insensitive to polarization angle and robust to angle of incidence have been developed. Figure 3(f) shows a scanning electron microscope (SEM) image of such a design. The filter design here consists of an array of coaxial nanoapertures in a 140 nm gold film on an undoped silicon substrate. The measured transmission spectra of a set of coaxial nanoapertures designed to filter LWIR light is also shown in Figure 3(f). The transmission peaks are tuned from 6 to 14 μm by varying the array period and aperture dimensions, with a peak transmission around 50%. The full-width-half-maximum of the transmission peaks are broad, around 1–2 μm and greater. Nevertheless researchers were able to use a set of 101 of these plasmonic filters to create a LWIR FADA microspectrometer [40]. The transmission of each filter array characterized using an FTIR microscope and spectrometer. In this case a detector array is synthesized by using the detector in the FTIR spectrometer and integrating the signal to give a scalar “photocurrent” reading for each filter and for each sample spectrum. A recursive least squares method is used to recover first the spectrum of the blackbody source within the FTIR system and then subsequently used to reconstruct the transmission spectra through several plastic films, shown in Figure 3(g–i). The measured and reference spectra are in good agreement, even for the fine absorption lines of the bonds in the polymer films, which is important for material identification in remote sensing applications. The angular sensitivity of the filter array is investigated through simulation. The transmission is shown to have a small reduction in magnitude up to 10° angle of incidence, but the transmission peak wavelengths are stable. This work and other demonstrations of MWIR and LWIR microspectrometers [38], [41], hold great promise for remote sensing due to the 104-fold reduction in mass and dimensions of these designs compared to conventional FTIR spectrometer systems and the importance of this spectral region in materials identification including noxious gases, polymers and explosives.

3.1 Background

Recently a new class of microspectrometer has emerged, where instead of replacing a dispersive element such as a Michelson interferometer or diffraction grating with a planar nanophotonic filter, the dispersion (and hence wavelength discrimination) occurs within the photodetectors themselves. Two methods of tailoring detector responsivities without the use of external filters have been put forward: (i) structural coloration of PIN photodiodes and (ii) material composition engineering to spatial tailor detector band gap energy. These approaches to microspectrometers again rely upon reconstruction algorithms to estimate incident spectra from measured photocurrents and known (calibrated) responsivities.

3.2 Structurally colored nanophotonic photodiode arrays

Silicon can be structurally colored by reactive ion etching nanophotonic elements such as arrays of vertical nanowires or sub-diffractive high contrast gratings. In both instances particular wavelengths of light will be strongly absorbed in the silicon, while other wavelengths are reflected or transmitted through the structure. Which wavelengths will be absorbed is determined by the leaky waveguide modes supported by the nanophotonic nanostructures, which in turn is determined by the array period and waveguide or nanowire cross-section. A range of these geometric parameters can be set within the one lithographic step. By creating a set of these arrays with various structural color in a PIN doped silicon substrate it is possible to create a filter-free microspectrometer chip.

Figure 4(a) shows a nanowire array based microspectrometer developed by researchers at the University of Melbourne [48]. Each colored square is a uniform array of nanowires of a specific radius, ranging from 60 to 175 nm, here the period is fixed and the radius of the nanowires controls the absorption of light at each pixel. The structural color provided by the silicon nanowires has been exploited to create a color printed image of the University’s emblem. Figure 4(b) shows an SEM image of one of the nanowire pixels after reactive ion etching and before planarization and metallization has been carried out. Each nanowire in an array is a vertical PIN photodiode, with 200 nm p + layer on top of 2 μm of lightly doped silicon region and then 500 nm of n + doped silicon. To make electrical contact with the top (p + layer) of each nanowire an epoxy planarization layer is used, then etched back to reveal just the nanowire tips and a thin transparent conductive oxide (TCO) deposited. Metal contacts, leads and pads are also deposited. Each nanowire array also sits on a mesa, which is itself a planar PIN photodiode connected in series with the nanowire diodes. Current measured with a positive bias is mostly photocurrent from the nanowire array, and when this bias is reversed, photocurrent from the mesa is measured. This allows for two distinct detectors within the one on-chip footprint. The responsivities of the nanowire and mesa detectors are complementary since any light that is not absorbed by the nanowire array is absorbed by the mesa detector below. Figure 4(c) shows the external quantum efficiency (EQE) of each of the 24 nanowire pixels in the microspectrometer chip. Peak EQE values of 30% occur at spectral positions corresponding to leaky hybrid (HE) waveguide modes. Corresponding reduction in EQE occurs in the mesa diodes. After calibration, the spectrometer chip was used to measure narrowband light sources across the visible spectrum as shown in Figure 4(d). The spectral reconstructions in this case employed L1 regularization, which is well suited to sparse spectra [51]. From these measurements the authors estimate the spectrometers resolution to be around 5 nm. Broad spectra were also measured using this design, but instead employed the RLS method, as it was found to outperform L1 regularization for these broader sources.

Figure 4:

Optical (a) and SEM (b) images of a silicon nanowire array microspectrometer (scale bars are: (clockwise from top left) 5 mm, 200 and 100 µm in (a) and 10 and 1 µm (inset) in (b)). Each pixel consists of an array of nanowires of a particular radius, varying the radius changes the structural coloration. Measured nanowire detector responsivities (c) and reconstructed and reference (dashed line) sample spectra (d). SEM images of a silicon fishnet pixel (e) (scale bars: 50 and 2 μm inset). Pixel responsivity and external quantum efficiency (EQE) spectra from both the fishnet and mesa detectors (f). Fishnet microspectrometer measured (red) and reference (blue) broad- and narrow-band sample spectra (g) and (h). Reprint permission obtained from [47], [48].

Figure 4(e) shows an SEM image of a fishnet pixel, another type of structurally colored silicon detector [48]. In this case each diode is comprised of a vertically oriented set of two interleaved orthogonal waveguide arrays. Unlike its nanowire counterpart, the interleaved design of the fishnet pixels means there is no need for planarization, etch-back, and TCO deposition. A combination of waveguide theory calculations and finite element method simulations suggest this structure can have near-unity absorption for wavelengths of light near each HCG mode cut-off wavelength, which are determined by the array period and width of each waveguide. The fishnet pixels are also etched into a PIN doped silicon substrate and sits upon its own mesa photodetector. Figure 4(f) shows the responsivity and EQE spectra of one of the 20 pixels in the fishnet microspectrometer chip for both fishnet and mesa detector. The fishnet detectors have peak sensitivities ranging from 400 to 620 nm. Here the mesa detectors are used to extend the operating range of the microspectrometer chip up to 800 nm, covering the entirety of the visible spectrum. After measuring the responsivities of each pixel and then the photocurrents generated by test spectra in each pixel a supervised machine learning algorithm is used to estimate the incident spectrum. The first stage consists of L2 regularization, the result of which is then fed into a simulated annealing stage to improve the estimated spectrum. Spectral measurements of a white LED and the same LED filtered with a bandpass filter are shown in Figure 4(g), (h), the results are in agreement with the reference spectra.

3.3 Single nanowire spectrometers via material composition engineering

The filter-free microspectrometer designs discussed thus far have all been based upon sets of individual photosensitive pixels, each with their own distinct responsivity spectra, which also can have varied response to illumination conditions such as the angle of incidence. Recently, however, it has been demonstrated that a filter-free microspectrometer can be formed from a single semiconductor nanowire, without the need for structural coloring [46]. In this design the nanowire is grown via a chemical vapor deposition process and is composed of compositionally graded CdS_xSe_1−x. One end of the 100 μm long nanowire is mostly CdS, the other end is mostly CdSe with a smooth transition of material composition in between. This means the band gap energy spatially varies from 1.74 eV (CdSe) to 2.42 eV (CdS) along the nanowire’s length and hence the absorption spectrum of the semiconductor nanowire also varies spatially. Figure 5(a) shows a photoluminescence (PL) color image and emission spectra from different points along the nanowire, where the change in emission color and wavelength indicate a change in band gap energy. After growth the nanowires are placed on a silica substrate and an array of electrical contacts are deposited to form a linear array of detectors with unique responsivities. This is shown in Figure 5(b), where each region of nanowire between to metal electrodes forms one of the 38 microspectrometer detector units. Unlike the previously discussed examples of photodiode-based filter-free microspectrometers, here each unit acts as a photoconductor, but judicious choice of the metal contacts could allow for Schottky photodiodes to be formed. A thin Al₂O₃ passivation layer is then deposited to encapsulate and protect the nanowire. Figure 5(c) shows the normalized responsivity spectra from each detector unit measured with a 0.5 V bias. The spectrometer operational range is just over 100 nm wide, ranging from 510 to 620 nm, this is restricted by the band gap energies of the bulk semiconductors chosen to grow the nanowire from. This limitation is not easy to overcome as the growth of compositionally graded nanowires is not possible with just any combination of semiconductors, however, the authors do suggest several alternative combinations. Spectra measured with the nanowire spectrometer is shown in Figure 5(d), where the results agree well with the reference spectra. Here L2 regularization is used to reconstruct incident spectra but the algorithm also has a method for detecting erroneous photocurrent measurements, by comparing the mean photocurrent to its standard of deviation. If a detector fault occurs or a particular photocurrent value is far outside what is expected the algorithm will exclude that detector from the reconstruction. Even though the total detector number is then reduced, by excluding data from suspicious detectors the reconstruction algorithm has improved accuracy and robustness to measurement error. The stated spectral resolution of this microspectrometer is 1 nm.

Figure 5:

Photoluminescence image and position-dependent spectra (a) from a single CdS_xSe_1−x nanowire (scalebar 20 µm). Optical image and schematic of the single nanowire microspectrometer (b) (scalebar 10 µm). Measured position-dependent responsivity spectra (c) and (d) measured (solid) and reference (dashed) sample spectra. Reconstructed absorption scanning spectra (e) from different locations within an onion cell, measured with a single nanowire microspectrometer. Reprint permission obtained from [46].

This nanowire microspectrometer chip can also operate as a scanning hyperspectral imaging sensor and provide in situ spectral imaging. Since each detector in the spectrometer is only 2.5 µm wide, scanning microscopy techniques can be used without the need of a pinhole aperture or single-mode optic fiber. Figure 5(e) shows a micrograph of a few red onion cells and the locations of absorption measurements. Below that are four distinct absorption spectra measured with the single nanowire microspectrometer chip. This application could be very useful for cytobiology and biomedicine research, where detailed spectral data registered to a location within a microscope image of each cell will provide important information about bio-composition and chemical concentrations and could be used to augment machine vision in artificial neural network powered experiments [54], [55].