The increasing degree of automation in industrial production processes and agriculture, together with the rising quality standards for food products, cosmetics, and pharmaceutics, is driving the need for fast, reliable, and non-destructive sensing methods. Optical methods are in the spotlight in this development, as reflection and transmission spectra can be measured non-invasively and rapidly. Molecular vibrations of typical organic bonds, such as O–H and C–H, create a chemical fingerprint in the near-infrared and mid-infrared range that allows retrieving material information from the optical measurements.11. C. Pasquini, Anal. Chim. Acta 1026, 8 (2018). https://doi.org/10.1016/j.aca.2018.04.004 In particular, near-infrared (NIR) spectroscopy has been successfully investigated for several decades for a large variety of application cases.22. M. Manley, Chem. Soc. Rev. 43, 8200 (2014). https://doi.org/10.1039/c4cs00062e Currently, the main limitation is the need for a spectrometer, which is mostly restricted to a laboratory environment due to its typical size, cost, and lack of robustness. Portable and handheld spectrometers are also reaching the market,33. K. Beć, J. Grabska, and C. W. Huck, Chem. -Eur. J. 27, 1514 (2021). https://doi.org/10.1002/chem.202002838 but their cost remains limited by the use of discrete components and large InGaAs arrays. Chip-based solutions have the potential to enable the pervasive application of spectroscopy in industrial and consumer markets.4,54. A. McGonigle, T. Wilkes, T. Pering, J. Willmott, J. Cook, F. Mims, and A. Parisi, Sensors 18, 223 (2018). https://doi.org/10.3390/s180102235. J. Kulakowski and B. d’Humières, Miniature, micro and chip-size spectrometers: technologies, market trends and customers’ needs (Tematys, 2020). Much research effort has been dedicated to the integration of spectrometers,6,76. A. Li, C. Yao, J. Xia, H. Wang, Q. Cheng, R. Penty, Y. Fainman, and S. Pan, Light: Sci. Appl. 11, 174 (2022). https://doi.org/10.1038/s41377-022-00853-17. Z. Yang, T. Albrow-Owen, W. Cai, and T. Hasan, Science 371, eabe0722 (2021). https://doi.org/10.1126/science.abe0722 but so far, integrated NIR solutions have not reached commercial maturity. One key design consideration is the fact that in these applications, input light originates from a large area and is spatially incoherent, preventing efficient coupling to a waveguide.An integrated NIR spectral sensing technology based on a detector array was recently introduced by some of the authors.88. K. D. Hakkel, M. Petruzzella, F. Ou, A. van Klinken, F. Pagliano, T. Liu, R. P. J. van Veldhoven, and A. Fiore, Nat. Commun. 13, 103 (2022). https://doi.org/10.1038/s41467-021-27662-1 The spectral sensor consists of a multi-pixel array of 16 resonant-cavity-enhanced (RCE) photodetectors, which each have a unique complex spectral response curve with several peaks. The planar cavity formed by two metal mirrors leads to an enhancement of the optical fields for specific wavelengths and higher absorptance.99. M. S. Ünlü and S. Strite, J. Appl. Phys. 78, 607–639 (1995). https://doi.org/10.1063/1.360322 Thereby, high responsivities can be achieved even with thin absorber layers,99. M. S. Ünlü and S. Strite, J. Appl. Phys. 78, 607–639 (1995). https://doi.org/10.1063/1.360322 which can enable low dark currents.10,1110. B. N. Sverdlov, A. E. Botchkarev, N. Teraguchi, A. A. Salvador, and H. Morkoç, Electron. Lett. 29, 1019 (1993). https://doi.org/10.1049/el:1993068011. M. Verdun, G. Beaudoin, B. Portier, N. Bardou, C. Dupuis, I. Sagnes, R. Haïdar, F. Pardo, and J.-L. Pelouard, J. Appl. Phys. 120, 084501 (2016). https://doi.org/10.1063/1.4961327 By varying the optical path length of the cavities, the absorption is enhanced for different wavelengths in each photodetector pixel. The response peaks of the 16 pixels cover a broad wavelength range and allow accurately measuring material properties based on their diffuse reflectance,8,12–148. K. D. Hakkel, M. Petruzzella, F. Ou, A. van Klinken, F. Pagliano, T. Liu, R. P. J. van Veldhoven, and A. Fiore, Nat. Commun. 13, 103 (2022). https://doi.org/10.1038/s41467-021-27662-112. R. F. Kranenburg, F. Ou, P. Ševo, M. Petruzzella, R. de Ridder, A. van Klinken, K. D. Hakkel, D. M. J. van Elst, R. van Veldhoven, F. Pagliano, A. C. van Asten, and A. Fiore, Talanta 245, 123441 (2022). https://doi.org/10.1016/j.talanta.2022.12344113. F. Ou, A. van Klinken, P. Ševo, M. Petruzzella, C. Li, D. M. J. van Elst, K. D. Hakkel, F. Pagliano, R. P. J. van Veldhoven, and A. Fiore, Sensors 22, 7027 (2022). https://doi.org/10.3390/s2218702714. F. Ou, A. van Klinken, K. D. Hakkel, M. Petruzzella, D. M. J. van Elst, P. Ševo, C. Li, F. Pagliano, R. P. J. van Veldhoven, and A. Fiore, Spectroscopy Outside the Laboratory for Today’s Spectroscopists (Spectroscopy Online, 2022), Vol. 37, p. 34. without the need to reconstruct a spectrum. In contrast to a spectrometer, determining the incoming spectrum is, therefore, not the goal for the spectral sensor presented here. In the previous work,88. K. D. Hakkel, M. Petruzzella, F. Ou, A. van Klinken, F. Pagliano, T. Liu, R. P. J. van Veldhoven, and A. Fiore, Nat. Commun. 13, 103 (2022). https://doi.org/10.1038/s41467-021-27662-1 the detectors’ peaks had linewidths between 100 and 200 nm and showed considerable non-resonant background (14%–25% of the peak responsivity). The large overlap between the spectral response functions of the individual photodetectors led to a significant correlation between the obtained photocurrent values, reducing the ability to discriminate narrow spectral features in the spectrum under test based on the photocurrent values. In our previous work, the significant non-resonant background and the large linewidth were identified as the limiting factors for the sensing performance. The spectral linewidth, in turn, was attributed to the surface roughness produced during the definition of the tuning layers by gray-scale electron-beam lithography. This fabrication approach is, at the same time, difficult to scale to larger fabrication volumes, which hampers the production of such photodetector arrays at an industrial scale. In this letter, we describe a largely improved and scalable fabrication method based on optical lithography that, together with improvements in the sensor design, leads to a substantial reduction of the peak linewidth (to 42–67 nm) and of the non-resonant background (to 3%–9% of the peak value). We also report a detailed electro-optical characterization of the devices, showing good matching with optical simulations. The robustness and small footprint of these sensors, together with the improved device characteristics and the use of a simple fabrication process based on optical lithography, make them suitable for use as embedded sensors in industrial and consumer applications.With respect to the multi-pixel array in our previous work,88. K. D. Hakkel, M. Petruzzella, F. Ou, A. van Klinken, F. Pagliano, T. Liu, R. P. J. van Veldhoven, and A. Fiore, Nat. Commun. 13, 103 (2022). https://doi.org/10.1038/s41467-021-27662-1 we aim here at reducing the linewidth of the responsivity peaks to well below 100 nm while still covering the complete wavelength range. To cover the free-spectral range of the resonant cavities (1150–1650 nm) with 16 response peaks, a linewidth of at least 30 nm is required. Simulations of the optical fields in our RCE photodetectors show that the intrinsic optical losses due to absorption and scattering during a roundtrip dominate over the outcoupling loss of the semi-transparent top mirror. Consequently, the linewidth of the response peaks can be decreased by reducing the intrinsic optical losses. This can be achieved by minimizing the roughness of interfaces in the layer stack , which cause scattering losses. Additionally, the intrinsic optical losses may be reduced by a smaller thickness of the absorber layer; however, when other losses start to dominate, the peak responsivity may be compromised. The optical response of the RCE photodetector is simulated with a 1D model using the transfer matrix approach to obtain the position-dependent intensity of the optical fields inside the cavity and the resulting absorption in the InGaAs absorber layer [see the layer stack in Fig. 1(b)]. While literature data were used for the dispersion of the complex refractive index for the metal mirrors (gold,1515. P. B. Johnson and R. W. Christy, Phys. Rev. B 6, 4370 (1972). https://doi.org/10.1103/physrevb.6.4370 silver, and germanium1616. E. Palik, Handbook of Optical Constants of Solids (Academic Press, 1998).) and the InP layers,1616. E. Palik, Handbook of Optical Constants of Solids (Academic Press, 1998). the dispersion for all other materials (SiO2, SiN, InGaAs) used in the simulation was determined by ellipsometer measurements on a set of reference samples with various layer thicknesses of these materials. For silicon nitride (SiN), we assumed the material to be purely transparent (k = 0), as this gave the best match to experimental data. The model is used to find the optimal trade-off between linewidth and peak responsivity and determine the appropriate thicknesses of the tuning layers for the 16 pixels. The result is the responsivity curves for the 16 pixels as shown in Fig. 1(a) for SiN tuning layer thicknesses from 50 to 325 nm in evenly spaced steps. According to simulations, this choice of tuning layers results in full coverage of the wavelength range between 890 and 1650 nm with an average linewidth of 52 nm [see Fig. 1(a)]. The peak responsivity of the pixels varies throughout the spectrum, as the spatial overlap of the field intensity with the absorbing InGaAs layers is different for every mode and tuning layer thickness. For a tuning layer thickness of 197 nm (pixel 9), the normalized field intensity for two resonant modes is plotted in Fig. 1(c) as an example. For the mode with three field maxima inside the cavity (mode 3 at λ = 1412 nm), a larger part of the electric field is overlapping within the absorbing layer and, therefore, generating photocurrent more efficiently compared to the mode with four field maxima (mode 4 at λ = 1070 nm), leading to a higher responsivity at 1412 nm than at 1070 nm.For the fabrication of the spectral sensor, differently from our previous work,88. K. D. Hakkel, M. Petruzzella, F. Ou, A. van Klinken, F. Pagliano, T. Liu, R. P. J. van Veldhoven, and A. Fiore, Nat. Commun. 13, 103 (2022). https://doi.org/10.1038/s41467-021-27662-1 a process based on optical lithography and the stepwise deposition of a dielectric tuning layer was used. Indeed, while the electron-beam gray-scale lithography process described earlier88. K. D. Hakkel, M. Petruzzella, F. Ou, A. van Klinken, F. Pagliano, T. Liu, R. P. J. van Veldhoven, and A. Fiore, Nat. Commun. 13, 103 (2022). https://doi.org/10.1038/s41467-021-27662-1 allowed the definition of all pixel heights in a single step using a resist as a tuning layer, it also produced a rough top surface, resulting in an increase in linewidth. In this work, by using SiN as a tuning layer, we achieve a flat surface and much improved linewidth.

A stack of indium phosphide (InP) and lattice-matched indium gallium arsenide (InGaAs) layers are grown on a 2′′ InP non-intentionally doped wafer using metal-oxide vapor phase epitaxy (MOVPE). To form a p-i-n-diode, the contact layers are doped with Zn on the p-side (1 × 1018 cm3 in InP, 1 × 1019 cm3 in InGaAs) and with Si on the n-side (5 × 1018 cm3 in InP, 1 × 1019 cm3 in InGaAs), while the absorber layer is non-intentionally doped. On top of the diode layer stack, a layer of silica (SiO2) is deposited. The surface is cleaned with oxygen plasma before depositing the bottom silver mirror by evaporation with a 2 nm germanium layer on both sides for improved adhesion. A layer of silica is deposited on the bottom mirror as well as on a 3′′ silicon wafer, onto which the InP wafer with the diode layer stack will be bonded. For the wafer bonding process, benzocyclobutene (BCB) is spin-coated on the silicon wafer, onto which the InP wafer is bonded under vacuum conditions. Afterward, the InP substrate and the sacrificial layers are removed by wet-chemical etching, exposing the stack of grown diode layers.

In order to create separate photodiodes with a common p-contact, a layer of patterned optical photoresist serves as an etching mask while the top four diode layers (n-InGaAs contact, n-InP layer, InGaAs absorber, and p-InP layer) are partially removed in a series of wet-chemical etching steps, leaving arrays of rectangular mesas, each with a size of 130 × 605 μm2. The next two optical lithography steps are used to create the contact pads on the p-side (Ti/Pt/Au 25 nm/75 nm/200 nm) and n-side (Ni/Ge/Au 30 nm/50 nm/250 nm) by evaporation and lift-off, which create ohmic contacts on the doped InGaAs surfaces. The p-contact around and in between the pixels is shared by all the pixels, while the n-contact on top of each mesa allows individually measuring the current of each pixel [see Fig. 2(a)]. The n-contact has a ring-shape to ensure efficient carrier collection while the blocking of incoming light from the top is minimized.For obtaining an individual wavelength response from each pixel, the optical path length is varied by depositing tuning layers of silicon nitride with different thicknesses. As determined from profilometer measurements on the fabricated arrays, this provides surfaces with significantly reduced roughness (standard deviation of ∼5 nm) compared to the ∼25 nm roughness observed in the grey-scale EBL process.88. K. D. Hakkel, M. Petruzzella, F. Ou, A. van Klinken, F. Pagliano, T. Liu, R. P. J. van Veldhoven, and A. Fiore, Nat. Commun. 13, 103 (2022). https://doi.org/10.1038/s41467-021-27662-1 The roughness obtained previously using a grey-scale EBL process is mainly caused by the three-dimensional proximity-effect correction, which is required to prevent thickness deviations toward the edges, and the re-exposure with high-energy electrons during the lithographic definition of the top mirror on top of the tuning layers. The transition to optical lithography eliminates both of these sources of roughness.In order to reduce the number of lithography/deposition steps, we use a combinatorial deposition technique1717. S.-W. Wang, M. Li, C.-S. Xia, H.-Q. Wang, X.-S. Chen, and W. Lu, Appl. Phys. B 88, 281 (2007). https://doi.org/10.1007/s00340-007-2726-3 of dielectric material in n steps to obtain 2n different layer thicknesses. In the first lithography step, the optically active area of all pixels is covered with uniform layers of silica (48 nm) and SiN (51 nm). In each of the following four lithography steps, a thin layer of SiN is deposited on 8 out of the 16 pixels, leading to 24 = 16 different total thicknesses of the tuning layer. The deposited thicknesses are measured as 146.3 nm (target 140 nm), 72.7 nm (target 70 nm), 36.2 nm (target 35 nm), and 17.8 nm (target 17.5 nm), respectively. Consequently, the fabricated SiN tuning layers have total thicknesses ranging from 51 to 324 nm in approximately even steps. The tuning layers extend on top of and between the diode mesas and leave only the area of the contact pads uncovered. This ensures a uniform thickness within the optically active area of the pixels. Additionally, the sidewall of the mesas is covered and thereby shielded from oxidation or other influences of the surrounding. In the last lithography and lift-off step, a thin layer of gold (17.5 nm) is evaporated on top of the tuning layers as a top mirror, forming a planar cavity together with the silver bottom mirror underneath the diode layer stack. In total, the multi-pixel array has a footprint of 1.41 × 1.41 mm2 [see Fig. 2(a)]. After the fabrication of the spectral sensors is finished, the detector arrays are characterized electrically and optically.The electrical properties of the individual photodiodes were analyzed by measuring their IV characteristics using a 2 kΩ resistor and a multimeter. The obtained results are shown in Fig. 2(b) on a semi-logarithmic scale. At a forward bias of 1 V, currents up to 5 mA were observed. At −1 V reverse bias, dark currents in the range of a few μA are observed for most pixels, corresponding to current densities of 5 × 10−3 A/cm2. At zero bias, the typical resistance of the pixels is determined to be 1 MΩ; however, there are substantial differences between the values for individual pixels, which could be attributed to different properties of the wet-etched side surface, which may influence the reverse current. Comparing the slope of the IV curve to lateral transport measurements shows that the resistance of the pixels at zero bias is dominated by the vertical conduction through the semiconductor layers. The other main contributions to the measured resistance are the two ohmic contacts between the doped InGaAs contact layers and metal contact pads, as well as the resistance caused by the lateral conduction through the p-doped contact layer. Overall, the electrical characteristics indicate a functional diode allowing efficient generation and transport of the photocurrent.The wavelength-dependent responsivity of each pixel is measured using a broadband-light source with an 850 nm long-pass filter and a monochromator (output linewidth 8 nm) for the illumination. The incoming light from a multi-mode fiber (550 μm core size) is focused onto the optically active area of a pixel using a microscope. The resulting photocurrent is read out via a resistor of 10 kΩ using a multimeter. The experimentally obtained responsivity curves for each pixel are shown in Fig. 3(a). In agreement with the simulation [see Fig. 1(a)], the responsivity peaks cover the complete wavelength range between 890 and 1650 nm. Measurements of different arrays across the wafer revealed a typical variation of the peak wavelengths in the range of 5 nm. This is mainly caused by the epitaxially grown diode layers and the plasma-deposited tuning layers becoming thinner toward the edge of the sample, which causes a blue shift of the resonance wavelengths. The use of industrial-grade deposition equipment would strongly improve this uniformity. Using a Lorentzian fit of the peaks, their linewidth is determined as (55 ± 7) nm (minimum: 41.8 nm and maximum: 67.2 nm), which is in very good agreement with the simulation results [see Fig. 3(b)]. As the linewidth of the peaks is related to the optical losses in the cavity, the agreement between experimental and simulation results shows that the optical losses are dominated by absorption and out-coupling losses , which are considered in the simulation. Consequently, scattering losses due to the roughness of the fabricated surfaces only have a minor contribution to the total optical losses. With respect to our previous work,88. K. D. Hakkel, M. Petruzzella, F. Ou, A. van Klinken, F. Pagliano, T. Liu, R. P. J. van Veldhoven, and A. Fiore, Nat. Commun. 13, 103 (2022). https://doi.org/10.1038/s41467-021-27662-1 the improved design and fabrication process result in a linewidth reduction by a factor of 2–3, depending on the pixel and the mode considered. Additionally, the non-resonant background is reduced by a factor of 3–4 to about 3%–9% of the peak responsivity, making the responsivity peaks much more pronounced in the spectrum. Figure 3(a) shows that the peak responsivities vary throughout the wavelength range between 0.11 and 0.30 A/W, which is significantly lower compared to the simulation [between 0.2 and 0.55 A/W, see Fig. 1(a)]. Differences between measurements and simulations can be explained by deviations of the real and imaginary parts of the refractive index of the deposited and grown materials from the experimentally determined values that are used in the simulations. For the metal mirrors (gold, silver, and germanium), literature data for the complex refractive index were used in the simulation, which might not correctly represent the properties of the evaporated material layers in our device.The overall envelope of the responsivity curves, with a maximum around 1400 nm and a minimum around 1100 nm, also shows very good agreement with the simulation data shown in Fig. 1(a). This proves that our 1D model describes the field distribution and expected response of the individual pixels convincingly.In order to determine the dynamic range of the sensors and the smallest possible light power that can be detected, the photocurrent was read out using a transimpedance amplifier (transimpedance: 5 × 105 V/A) in combination with a lock-in amplifier (time constant: 3 s, DC filter slope: 24 dB, and noise-equivalent bandwidth: 0.026 Hz). The incoming laser light is tuned to the responsivity maximum of the illuminated pixel (pixel 13, 1540 nm) and passes through an attenuator and a chopper modulating the incoming light at 417 Hz, which also serves as a reference input to the lock-in amplifier. For four different settings of the attenuator, the measured photocurrent is determined for the different output powers of the laser, leading to illumination powers ranging from 4 pW to 6 μW. This leads to photocurrents between a few pA and 1.5 μA (see Fig. 4). The maximum photocurrent is hereby limited by the lock-in amplifier and not by the maximum photocurrent of the photodetectors. For light powers between 60 pW and 6 μW, i.e., over five orders of magnitude, an approximately linear relation between light power and photocurrent is observed: a power fit of the data points results in an exponent close to 1 (see power fit in Fig. 4). A linear fit of the same data results in a slope of 0.249 A/W, which is in good agreement with the responsivity value measured for this pixel and wavelength [0.257 A/W, see Fig. 3(a)]. For light powers below 20 pW, the measured photocurrents approach a constant value of about 4 pA, likely related to cross-talk in the measurement system. As the observed fluctuation of this value is of the same order as its average, we consider it as the noise level, from which we determine the smallest detectable illumination power as ∼20 pW in the present experimental configuration. This minimum power value is not limited by the photodetector itself but is subject to the choice of the read-out system. Lower values for the minimum detectable power have been observed when reading out the sensor with the lock-in amplifier directly (larger load resistance); however, the achievable dynamic range is limited in that case.We finally assess the applicability of the sensor to practical sensing problems and the expected performance improvement with respect to the previous generation. First, an experimental study was performed on determining the concentration of a solution of methanol in isopropanol, as described in the supplementary material, Sec. A. Using a partial least squares (PLS) analysis, a strong prediction model is obtained featuring an average ratio of performance to deviation (RPD) of 25. Second, a multiple linear regression model (MLR) was applied to a set of measured sample spectra combined with the measured response curves of the photodetector arrays presented here and in our previous work88. K. D. Hakkel, M. Petruzzella, F. Ou, A. van Klinken, F. Pagliano, T. Liu, R. P. J. van Veldhoven, and A. Fiore, Nat. Commun. 13, 103 (2022). https://doi.org/10.1038/s41467-021-27662-1 to compare the performance. The determination of the fat content in raw cow milk from its reflection spectrum was chosen in view of its practical relevance and as an exemplary application of near-infrared spectroscopy.18,1918. H. G. Yakubu, Z. Kovacs, T. Toth, and G. Bazar, Crit. Rev. Food Sci. Nutr. 62, 810 (2022). https://doi.org/10.1080/10408398.2020.182954019. W. Wrzeszcz, P. Tomza, M. Kwaśniewicz, S. Mazurek, R. Szostak, and M. A. Czarnecki, RSC Adv. 6, 37195 (2016). https://doi.org/10.1039/c6ra04595b Based on the dataset reported by Aernouts et al.,2020. B. Aernouts, E. Polshin, J. Lammertyn, and W. Saeys, J. Dairy Sci. 94, 5315 (2011). https://doi.org/10.3168/jds.2011-4354 an improvement in the RPD value of up to 84% was obtained for low noise levels (details on the method and analysis are provided in the supplementary material, Sec. B). For higher noise levels, the sensing performance of both arrays is expected to be similar according to the simulations, as the noise level becomes the limiting factor in the accuracy of the prediction model rather than the correlation between the photocurrent values of different pixels. It must be noted that the expected sensing performance of the photodetector array presented in this work and the improvement with respect to our previous work are highly dependent on the spectral characteristics of the considered sensing problem. Nevertheless, the simulations for this application case show clearly that the reported improved spectral properties can lead to a significant improvement in the sensing performance for real application cases. More details on the simulation procedure and the datasets for the sample spectra can be found in the supplementary information.In conclusion, we demonstrated an integrated spectral sensor consisting of an array of photodiodes with individual spectral responses. We reported a fabrication approach based on a series of optical lithography steps, through which SiN tuning layers with 16 different thicknesses are deposited stepwise. Leading to largely decreased surface roughness and uniform cavity heights across the entire optically active area, this novel fabrication approach results in a significantly decreased linewidth of the response peaks and reduced non-resonant background as compared to the previous demonstration of InGaAs NIR spectral sensors.88. K. D. Hakkel, M. Petruzzella, F. Ou, A. van Klinken, F. Pagliano, T. Liu, R. P. J. van Veldhoven, and A. Fiore, Nat. Commun. 13, 103 (2022). https://doi.org/10.1038/s41467-021-27662-1 Additionally, the use of optical lithography opens the way to high-volume manufacturing of such spectral sensors. The experimental results for the optical response are in good agreement with optical simulations, which can be used to find suitable design parameters. In previous studies,8,12–148. K. D. Hakkel, M. Petruzzella, F. Ou, A. van Klinken, F. Pagliano, T. Liu, R. P. J. van Veldhoven, and A. Fiore, Nat. Commun. 13, 103 (2022). https://doi.org/10.1038/s41467-021-27662-112. R. F. Kranenburg, F. Ou, P. Ševo, M. Petruzzella, R. de Ridder, A. van Klinken, K. D. Hakkel, D. M. J. van Elst, R. van Veldhoven, F. Pagliano, A. C. van Asten, and A. Fiore, Talanta 245, 123441 (2022). https://doi.org/10.1016/j.talanta.2022.12344113. F. Ou, A. van Klinken, P. Ševo, M. Petruzzella, C. Li, D. M. J. van Elst, K. D. Hakkel, F. Pagliano, R. P. J. van Veldhoven, and A. Fiore, Sensors 22, 7027 (2022). https://doi.org/10.3390/s2218702714. F. Ou, A. van Klinken, K. D. Hakkel, M. Petruzzella, D. M. J. van Elst, P. Ševo, C. Li, F. Pagliano, R. P. J. van Veldhoven, and A. Fiore, Spectroscopy Outside the Laboratory for Today’s Spectroscopists (Spectroscopy Online, 2022), Vol. 37, p. 34. we demonstrated that chemical information can be directly retrieved by building a regression model from the measured photocurrent values. This sensing approach can be applied in relevant application cases ranging from plastic sorting to the measurement of the nutritional properties of milk. The reduction of the linewidth demonstrated here leads to less overlap between the response functions and less correlation between the photocurrent values, which will lead to even higher prediction performance. For two application cases, the sensing performance was assessed experimentally and using simulations, demonstrating that a significant improvement in the sensing accuracy can be obtained. Overall, this work contributes to making spectral sensors suitable for cost-efficient, handheld devices for material analysis applications outside of the laboratory.

DATA AVAILABILITY

Section:

ChooseTop of pageABSTRACTDATA AVAILABILITY <<REFERENCES

The data that support the findings of this study are available from the corresponding author upon reasonable request.

See the supplementary material for a detailed description of the experimentally determined sensing performance of the methanol concentration in an isopropanol solution (Sec. A) and of the simulations to compare the expected performance in determining the fat content in raw cow milk with the photodetector array presented in this work and in our previous work88. K. D. Hakkel, M. Petruzzella, F. Ou, A. van Klinken, F. Pagliano, T. Liu, R. P. J. van Veldhoven, and A. Fiore, Nat. Commun. 13, 103 (2022). https://doi.org/10.1038/s41467-021-27662-1 (Sec. B).We thank B. Aernouts (KU Leuven) for the use of the reflection data of raw cow milk samples as presented in the supplementary material. We acknowledge J. J. P. M. Schulpen for his support with the ellipsometer measurements. This research was partially funded under NWO TTW Project No. 16670 (A.v.K., C.L. F.O., and R.P.J.v.V.), NWO Project No. 17626 (D.J.v.E.), the Penta Call 2 project Environmental Sensors for AIR Quality (ESAIRQ) Grant No. 16113 (K.D.H., M.P., and A.F.), the High Tech Systems and Materials (HTSM) Project No. 14658 (F.P.), and the NWO Zwaartekracht Research Center for Integrated Nanophotonics.

Conflict of Interest

M.P., K.D.H., F.O., and F.P. are employees, and M.P., F.P., and A.F. are shareholders and co-founders of MantiSpectra B.V., the company commercializing the sensor. The remaining authors declare no competing interest.

Author Contributions

A.F., K.D.H., F.P., M.P., and A.v.K. developed the device concept. A.v.K, C.L., D.M.J.v.E., and M.P. optimized the fabrication recipe. C.L., A.v.K., D.M.J.v.E., and M.P. fabricated the devices. R.v.V. performed the sample growth of the active layers. A.v.K. performed the experiments. M.P. and D.M.J.v.E. developed the simulation code. A.F. supervised the project. A.v.K. and A.F. wrote the article. All authors discussed the work and contributed to writing and revising the article.

Anne van Klinken: Conceptualization (equal); Data curation (equal); Formal analysis (equal); Investigation (equal); Methodology (equal); Resources (equal); Visualization (equal); Writing – original draft (equal); Writing – review & editing (equal). Don M. J. van Elst: Methodology (equal); Resources (equal); Software (equal); Writing – review & editing (equal). Chenhui Li: Methodology (equal); Resources (equal); Writing – review & editing (equal). Maurangelo Petruzzella: Conceptualization (equal); Methodology (equal); Resources (equal); Software (equal); Writing – review & editing (equal). Kaylee D. Hakkel: Conceptualization (equal); Writing – review & editing (equal). Fang Ou: Writing – review & editing (equal). Francesco Pagliano: Conceptualization (equal). René van Veldhoven: Resources (equal). Andrea Fiore: Conceptualization (equal); Funding acquisition (equal); Methodology (equal); Project administration (equal); Supervision (equal); Writing – original draft (equal); Writing – review & editing (equal).

REFERENCES

Section:

ChooseTop of pageABSTRACTDATA AVAILABILITYREFERENCES <<1. C. Pasquini, Anal. Chim. Acta 1026, 8 (2018). https://doi.org/10.1016/j.aca.2018.04.004, Google ScholarCrossref2. M. Manley, Chem. Soc. Rev. 43, 8200 (2014). https://doi.org/10.1039/c4cs00062e, Google ScholarCrossref3. K. Beć, J. Grabska, and C. W. Huck, Chem. -Eur. J. 27, 1514 (2021). https://doi.org/10.1002/chem.202002838, Google ScholarCrossref4. A. McGonigle, T. Wilkes, T. Pering, J. Willmott, J. Cook, F. Mims, and A. Parisi, Sensors 18, 223 (2018). https://doi.org/10.3390/s18010223, Google ScholarCrossref5. J. Kulakowski and B. d’Humières, Miniature, micro and chip-size spectrometers: technologies, market trends and customers’ needs (Tematys, 2020). Google Scholar6. A. Li, C. Yao, J. Xia, H. Wang, Q. Cheng, R. Penty, Y. Fainman, and S. Pan, Light: Sci. Appl. 11, 174 (2022). https://doi.org/10.1038/s41377-022-00853-1, Google ScholarCrossref7. Z. Yang, T. Albrow-Owen, W. Cai, and T. Hasan, Science 371, eabe0722 (2021). https://doi.org/10.1126/science.abe0722, Google ScholarCrossref8. K. D. Hakkel, M. Petruzzella, F. Ou, A. van Klinken, F. Pagliano, T. Liu, R. P. J. van Veldhoven, and A. Fiore, Nat. Commun. 13, 103 (2022). https://doi.org/10.1038/s41467-021-27662-1, Google ScholarCrossref9. M. S. Ünlü and S. Strite, J. Appl. Phys. 78, 607–639 (1995). https://doi.org/10.1063/1.360322, Google ScholarScitation, ISI10. B. N. Sverdlov, A. E. Botchkarev, N. Teraguchi, A. A. Salvador, and H. Morkoç, Electron. Lett. 29, 1019 (1993). https://doi.org/10.1049/el:19930680, Google ScholarCrossref11. M. Verdun, G. Beaudoin, B. Portier, N. Bardou, C. Dupuis, I. Sagnes, R. Haïdar, F. Pardo, and J.-L. Pelouard, J. Appl. Phys. 120, 084501 (2016). https://doi.org/10.1063/1.4961327, Google ScholarScitation, ISI12. R. F. Kranenburg, F. Ou, P. Ševo, M. Petruzzella, R. de Ridder, A. van Klinken, K. D. Hakkel, D. M. J. van Elst, R. van Veldhoven, F. Pagliano, A. C. van Asten, and A. Fiore, Talanta 245, 123441 (2022). https://doi.org/10.1016/j.talanta.2022.123441, Google ScholarCrossref13. F. Ou, A. van Klinken, P. Ševo, M. Petruzzella, C. Li, D. M. J. van Elst, K. D. Hakkel, F. Pagliano, R. P. J. van Veldhoven, and A. Fiore, Sensors 22, 7027 (2022). https://doi.org/10.3390/s22187027, Google ScholarCrossref14. F. Ou, A. van Klinken, K. D. Hakkel, M. Petruzzella, D. M. J. van Elst, P. Ševo, C. Li, F. Pagliano, R. P. J. van Veldhoven, and A. Fiore, Spectroscopy Outside the Laboratory for Today’s Spectroscopists (Spectroscopy Online, 2022), Vol. 37, p. 34. Google Scholar15. P. B. Johnson and R. W. Christy, Phys. Rev. B 6, 4370 (1972). https://doi.org/10.1103/physrevb.6.4370, Google ScholarCrossref, ISI16. E. Palik, Handbook of Optical Constants of Solids (Academic Press, 1998). Google ScholarCrossref17. S.-W. Wang, M. Li, C.-S. Xia, H.-Q. Wang, X.-S. Chen, and W. Lu, Appl. Phys. B 88, 281 (2007). https://doi.org/10.1007/s00340-007-2726-3, Google ScholarCrossref18. H. G. Yakubu, Z. Kovacs, T. Toth, and G. Bazar, Crit. Rev. Food Sci. Nutr. 62, 810 (2022). https://doi.org/10.1080/10408398.2020.1829540, Google ScholarCrossref19. W. Wrzeszcz, P. Tomza, M. Kwaśniewicz, S. Mazurek, R. Szostak, and M. A. Czarnecki, RSC Adv. 6, 37195 (2016). https://doi.org/10.1039/c6ra04595b, Google ScholarCrossref20. B. Aernouts, E. Polshin, J. Lammertyn, and W. Saeys, J. Dairy Sci. 94, 5315 (2011). https://doi.org/10.3168/jds.2011-4354, Google ScholarCrossref© 2023 Author(s). All article content, except where otherwise noted, is licensed under a Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).