Mid-infrared Nanoantennas as Ultrasensitive Vibrational Probes Assisted by Machine Learning and Hyperspectral Imaging


 Infrared (IR) Spectroscopy has been developed for centuries and has been widely used to identify molecular structure from the massive information provided by IR fingerprint absorption, reflecting the vibration energy of the chemical bond. Due to the intrinsically weak light-matter interaction, IR spectroscopy serves low sensitivity and sizeable optical interaction length (~mm to ~cm) compared with other optical probes like Raman, florescent, and refractometry technology, which hinder the applications for ultra-sensitive biomolecular screening. Here, we report a new type of IR spectroscopy by wavelength gradient hook nanoantenna integrated with the microfluidic channel, enhancing the IR molecular absorption and bringing in refractometry function with ultrathin (~100 nm) optical interaction length. With the proof-of-concept demonstration of molecular recognition of mixed alcoholic liquids by machine learning and molecular fingerprint retrieving by hyperspectral images in one-time data acquisition, our work paves the way to advance, small-volume, real-time, ultra-sensitive, in-vitro biomolecular dynamic analysis in the aqueous environment.


Introduction
system for global molecular recognition. Thus, the performance of the sensing system needs to 92 improve by engineering the PNA structure. The intuitive way is to increase the intensity of the 93 electric field by squeezing the gap between adjacent PNA into nanometer scale. 25 In addition to sensitivity, the detection range is another critical FOM of PNA sensors, reflecting 102 the number of fingerprint absorption peaks that can be captured. Thanks to the sharp resonance 103 peaks of PNA, the enhancement becomes the maximum only when molecule fingerprint absorption 104 peaks match with the PNA resonance, which is a very narrow bandwidth. Therefore, to detect more 105 absorption peaks in the MIR region, multi-resonant PNA sensors are proposed to collect broadband 106 spectrum data to recognize lipids and proteins from separate absorption wavelengths. 34 107 Nevertheless, the individual resonances of PNA by a different order of resonance modes also fails 108 to cover the whole spectrum of IR fingerprint wavelength region from 5.5 μm to 10 μm because 109 of the gaps between two resonance peaks. 35,36 Therefore, to collect continuous spectral fingerprint 110 absorption, pixelated all-dielectric nanoantenna array and tunable antenna by incident angle were 111 proposed for ultra-broadband spectroscopic analysis for molecular barcode imaging and 112 fingerprint absorption retrieving. 37,38 However, the data from the different pixels are collected 113 separately, which cannot achieve dynamic monitoring. Hyperspectral IR imaging could be an 114 6 advanced solution to solve the problems, and PNA has been used to enhance the vibrational 115 molecular IR imaging captured by the focal plane array (FPA). 39 Furthermore, spectrum 116 reconstruction by IR imaging technology is reported to retrieve wavelength shift of nanoantenna 117 for molecular identification as refractometry microscopy with one-time data acquisition, but the 118 performance is still limited due to the use of a monochromatic light source. 40 119 We proposed a novel molecular identification platform by wavelength gradient hook isopropyl alcohol (IPA), and their mixture. We also integrated the WGHNA-SEIRAS platform 131 with a microfluidic channel, solving the water absorption issue by traditional IR spectroscopy for 132 aqueous measurement, which can be easily compatible with the biomolecular and cellular systems. 133 Our work paves the way to achieve fast global molecular recognition by rich spectrum information 134 from HNA-SEIRAS in non-contact, non-destructive, label-free, and miniaturized methods. Working Principles of WGHNA-SEIRAS 136 The concept of WGHNA-SEIRAS platform is shown in Fig.1. The MIR light shines from an IR 137 microscope and excites the plasmonic resonance of HNA on calcium difluoride (CaF2) substrate 138 with desired polarization state and perpendicular incidence. The transmitted or reflected light is 139 routed to IR FPA to capture the far-field spectral response from HNA. With the plasmon-phonon 140 coupling illustrated in Fig.1 a, where ω0 and ωm represent the angular frequency of resonance for HNA and molecular 146 vibration, respectively. γa and γr denote the radiative and absorptive losses of HNA, while γm is the 147 absorptive loss of molecules. μ is the coupling strength between HNA and molecular vibration. processing and ML for the raw data, the fingerprint barcoding and molecule identification are 160 achieved for array and supercell, respectively. In Fig. 1  the difference of molecular spectra with massive raw data collected. As demonstrated in Fig. 1  The design concept and experimental results of HNAs are shown in Fig.2. From equation 5, 171 we observe that the T and R are related to the resonance properties of HNA, which are radiative 172 and absorptive losses (γa and γr). The γa is related to the ohmic loss of plasmonic material (e.g., Au 173 in this work) and is almost robust among different antenna structures. Therefore, the philosophy 174 to use hook shape in PNA design is to engineer γr to tune the radiation from electron oscillation by 175 inducing inverse current from short arm (L3) of HNA. The method to control radiation capability 176 from the ratio of inverse current is merely adjusting the geometric difference (∆L) between long 177 arm (L1) and short arm (L3) of HNA as illustrated in Fig.2 a. The HNA performs a dipole 178 resonance at resonance frequency by enhancing localized electric field intensity, thus inducing the 179 current from one end to the other end ( Fig.2 b,c). The connection of two arms of HNA (L2) only 180 affects the resonance wavelength and is defined as a fixed value (400 nm) to fit the fabrication 181 9 resolution as shown in the SEM photo in Fig.2 where μ and f denote coupling efficiency between HNA and molecular vibration as well as the 194 ratio (γr / γa) between radiative (γa) and absorptive (γr) damping rate of the HNA, respectively. As for T and f=2 for R). In Fig. 3 g,h, a transition of line shape from Fano-like to EIT-like is observed 203 when the resonance wavelength of HNA matching with molecular absorption wavelengths.

204
Besides, the highest sensitivity is achieved when the resonance wavelengths of HNA and 205 molecules are well-matched (Fig.3 i). After optimization, the arm length ratio is fixed at L3:L1=1:3 206 to achieve the highest sensitivity at reflection mode. Therefore, to simultaneously achieve the best 207 sensitivity and broad bandwidth, the wavelength-gradient structures are designed by gradually 208 increasing the total length with the fixed folding degree of HNA.

210
The spectrum of the 16-element HNA supercell is shown in Fig.4 a measured by a Fourier-  In addition to the proof-of-concept characteristic of thin-film, we also integrated WGHNA into a 225 microfluidic system ( Fig. 4 d I), which is easily compatible with biomolecular systems for achieved by all-dielectric nanoantenna with high quality factors 37,38 , our approach behaves smaller 300 footprint and spatial tunability so that the whole enhanced spectra can be captured in one testing, 301 which dramatically reduced the time for broadband fingerprint retrieving, paving the way to 302 ultrasensitive and ultrafast molecules screening in ultra-broadband wavelength range with ultra-303 small volume. Although the spatial resolution is limited by the pixel of FPA (32*32 pixels) in our 304 demonstration (4*4 HNA pixels) to avoid mutual coupling, it is easy to improve by replacing the 305 FPA with more pixel numbers, smaller pixel area, and better detectivity (D*).

307
In this work, we propose a novel WGHNA-SEIRAS with high sensitivity and broad bandwidth The EIT-like line shape and Fano-like line shape can be expressed from Eq.8 when ω0=ωm and 546 ω0≠ωm, respectively. The plasmonic resonance can be easily obtained when there is no coupling 547 effect from molecules (μ=0).
Eq. 9-11 is used to extract absorptive and radiative loss of HNA by fitting the resonance spectrum 552 in the frequency domain from simulation (Extended Data Fig.1). By engineering the HNA 553 27 structure by changing ΔL with the constant L, the γr and γa can be tuned continuously, and ω0 554 remains unchanged. To explore the sensing performance, we have made some assumptions to 555 simplify Eq.8 in order to perform the analytical operation. First, we make ω0=ωm to match the 556 frequency of HNA and molecular vibration since the WGHNA is only designed for the molecular 557 absorption wavelength near the HNA resonance wavelength to have the best enhancement. Second, 558 we treat μ as a much smaller parameter compared with γm, γr, and γa. Therefore, we apply the 559 difference between Eq.9 and Eq.8 when ω = ω0 = ωm.
Since μ<<γm, 2 is a small real number close to 0. Therefore, we cancel the high order term and too much among different HNA devices, so that μ is also a constant. Additionally, γm is also 571 unchanged since we fix the absorption peaks of the "C=O" bond from PMMA in sensitivity 572 characterization. By applying the first derivative of f for Eq.13 and Eq.14. We further calculate 573 the maximum enhancement of the T and R spectrum and get the optimal condition that occurs 574 when f equals 0.5 and 2, respectively.   Fig.4 b). For pixel area of 80*80 μm 2 , we achieve the data acquisitions in 624 four times (Extended Data Fig.4 c). Compared with these two figures, a clearer square-shape 625 imaging of HNA array is shown in the picture by 4 pixels per acquisition. This result shows that 626 the spatial resolution of HNA array is mainly limited by pixels of FPA, which can be improved by 627 replacing the FPA with more pixel numbers, smaller pixel area, and better detectivity (D*).      sensing spectrum of PMMA and silk between HNA and HNA supercell. The HNA from P1, P8, and P16 is selected as a reference with the response to short, medium, and long-wavelength resonance, respectively. It shows HNA supercell has a good response over a broad range of wavelengths from 5.5 μm to 9 μm which HNA only covers a narrow bandwidth near resonance wavelength for enhancement of ngerprint absorption. (d)I. Schematic drawing of an integrated micro uidic HNA supercell system for liquid sensing. The HNAS on the CaF2 carrier chip is ip bonded to the PDMS surface with the alignment of HNAS into the micro uidic channel. The micro uidic channel is formed by a 3D printed mold and is xed on a microscope slide. The IR light is shining from the backside of the CaF2 chip, and re ected light is   IR ngerprint retrieval and molecules identi cation by hyperspectral IR imaging for HNA array. (a) Schematic illustration of wavelength gradient HNA array for hyperspectral imaging. Each pixel response to different IR wavelengths. (b) By pixelating the wavelength gradient HNA into a four by four arrangement, the hyperspectral imaging is captured by the FPA, representing the different spectrum response of each HNA pixels. P1 response shortest wavelengths (~4.67 μm) and P16 response the longest wavelengths (~7.46 μm). The wavelength difference is designed to be ~200 nm to construct a linear gradient in the wavelength domain. (c) Zoom-in picture for the HNA array at four selected wavelengths (i. 4.67 μm, ii. 5.66 μm, iii. 6.88 μm, iv. 7.46 μm). The expected pixel is illuminated at a resonant wavelength while other pixels are dark. The pixel is illuminated at HNA resonance, which are 4.67 μm for P1, 5.66 μm for P6, 6.88 μm for P12, and 7.46 μm for P16.