Molecular sensing based on surface-enhanced Raman scattering (SERS) has shown promise as a powerful analytical tool for biomedical diagnostics, identification of chemical compounds, food safety, and environmental monitoring. (1−4) A frontier for SERS is represented by the ability to perform label-free and chemical-specific sensing in applications where sampling is difficult, including (i) probing biological tissue, (5,6) (ii) site-specific study for living cells, (7,8) and (iii) on-site environmental monitoring, such as detection for hazardous heavy metal and toxic gas. (9−11) Integrating a SERS substrate on a fiber optic probe can create a compact and multifunctional sensor that can be used as a point-of-care SERS amplifier or remote sensing probe where the excitation laser and the resulting Raman signal are guided and collected by the same waveguide (the through-fiber detection scheme). Additionally, the ability to implement SERS in small-diameter optical fibers (12,13) has the potential to extend SERS applications to environments where sample perturbations need to be minimized, such as brain structures. In the context of studying neurochemical dynamics, vibrational spectroscopy enhanced by plasmonic nanostructures is of particular interest by its ability to detect and identify neurotransmitters even at attomolar concentrations, (14) with the release of these molecules linked to both physiological and pathological states, including Alzheimer’s disease, depression, schizophrenia, and Parkinson’s disease. (15−18)

SERS-active fiber optic probes have been mainly fabricated by covering the termination of an optical fiber with metallic nanoparticles. (19−31) Of specific significance is the case of tapered optical fibers (TFs), the plasmonic structures on the TF surface can engage with both radiative and guided modes over a customizable segment ranging from a few hundred micrometers to several millimeters. (32) Some guided modes can radiate out by thinning the fiber waveguide, while the remaining guided modes in the waveguide create evanescent fields at the surface. (12,33,34) Both types of fields can interact with plasmonic structures covering the entire tapered surface, expanding the sensing areas that can interact with the surrounding environment. Additionally, the wide and uniform interaction surface of the NIs-covered taper can be used for depth-selective SERS measurement if taking advantage of its mode-division demultiplexing. (35) Moreover, the tapered shape reduces invasiveness when interfacing with deep brain structures, making it more suitable and effective for neurological studies. Fabricating plasmonic structures on the TFs requires the structuring of a curved surface with a nonconstant curvature radius. Only bottom-up approaches have been demonstrated to successfully enable SERS sensing capabilities on a TF probe, including electrostatic self-assembly, (10,27,36−43) direct nucleation reaction, (44−47) dip coating, and laser-assisted evaporation. (23,48−51) However, plasmonic nanoparticle distributions on the surface fabricated by these approaches are largely random. The morphologies show an uneven statistical distribution across the large surface, cluster formation, and multilayer deposition, inevitably introducing uncertainty and ambiguity on the correlation between the device performance and the morphological status, limiting the understanding and design of the best-performance remote sensing tapered SERS probe.

Here, we aim at uncovering the correlation between particle size and density versus the probe’s SERS performance, to provide design principles for particle-decorated TF remote sensing probes. A tunable solid-state dewetting technology was employed to fabricate nanoislands (NIs) with varied geometric parameters. The monolayer densely packed NIs are surfactant-free and in direct contact with the silica surface of TFs. The formed patterns have analyzable nanoparticle sizes, densities, and statistically uniform distributions through the entire extent of the taper’s surface. This allows us to systematically study, for the first time, the effect of the particle size, density, and analyte on the overall SERS performance of the through-fiber detection system.

Figure 1 shows schematic steps for fabricating NIs patterns with different geometrical parameters on the nonplanar surface of the fiber taper. TFs were obtained by a heat-and-pull method from commercially available core/cladding step-index silica fibers. By incrementally increasing the deposited gold film’s thickness, after dewetting, NIs’ patterns can be tuned (for details, see section S1).

For the experimental demonstration, we fabricated four sets of NIs-TFs with initial film thicknesses of Thk_i = 1.7, 3.3, 5, and 6.7 nm. Under an optical microscope, the surface of the four sets of NIs-TFs at growing thickness shows a gradually increasing pink color (Figure 2a), generated by an increase in NIs’ size as identified by scanning electron microscope (SEM) (Figure 2b). To extract the geometric parameters of the NIs patterns quantitatively, SEM image batches were taken along the entire extent of the fiber taper (a representative high magnification image at the very tip of TF is shown in Figure 2c). Image recognition methods (described in Methods) were used to compute the coverage rate, particle number and particle density for each image. For the different Au film thicknesses, 2 TFs were included in the analysis. For each fiber, the number of sampled images with different magnifications along the taper was no less than 10. The measured average diameters at different taper positions are reported in Figure 2d, showing that the different NIs patterns distribute uniformly from the very tip to the taper end. The final geometric parameters of NIs’ patterns were then extracted by averaging between two fibers, and each statistic is an averaging of all the data points at different positions along the whole taper. Figure 2e shows the average diameters and NIs’ densities calculated, the diameters are D₁ = 14 ± 2, D₂ = 16 ± 2, D₃ = 33 ± 3, and D₄ = 49 ± 3 nm, and the NIs densities are ρ₁ = 2895 ± 7, ρ₂ = 1872 ± 57, ρ₃ = 507 ± 93, and ρ₄ = 236 ± 27 NIs/μm², for the ith Thk_i (i = 1, 2, 3, 4). Figure 2f displays the coverage rates (C) for the four different patterns, found to be about 40%, independent of initial film thickness, and the corresponding average heights (H_i) of the NIs are H_i = 4, 8, 12, and 16 nm (computed with H_i = Thk_i/C_i).

To understand the electromagnetic performances of the different NIs patterns, we implemented a model with Au nanodisks arranged in a hexagonal periodic manner (Figure 3a). One nanodisk of diameter D occupies the center of a hexagonal unit cell. The diameters of the nanodisks are the average diameters for each pattern in Figure 2e. The heights of the nanodisks are taken from Figure 2f as H_i = Thk_i/C_i. The ratio between nanodisk’s occupied surface and the overall unit cell area in 2D plane defines the coverage rates, which equal the coverage rates obtained from SEM image analysis (Figure 2f). Given the experimental average values for D and C as a function of Thk, the resulting effective gaps (G) in a hexagonal periodic arrangement can be computed to be 7, 8, 16, and 22 nm (Figure 3b; details in Methods). These geometric parameters define the NIs patterns and, hereinafter, we mark the patterns fabricated by initial film coating of Thk_i = 1.7, 3.3, 5, and 6.7 nm as D14H4G7, D16H8G8, D33H12G16, and D49H16G22 respectively.

The electromagnetic behavior of the structures was simulated by a finite element method (commercial implementation by COMSOL Multiphysics) with an unpolarized plane wave excitation source. Representative field enhancement distributions at 785 nm in one unit cell are displayed in Figure 3c. The plasmonic resonances are shown as the complement of transmittance in Figure 3d: as each pattern has different geometric parameters, the plasmonic resonances of the systems are slightly different in terms of position, intensity, and spectral width. The maximum field enhancements as a function of wavelength are reported in Figure 3e, showing that at above 595 nm the smallest NIs D14H4G7 give the strongest field enhancement, the cross-section of the field enhancement is reported in Figure S2. Although the field enhancement localized by the NIs is distributed on all edges, the interaction between the molecules and the particles requires spatial overlap. This excludes the region below the NIs since there is the silica fiber substrate. Therefore, the regions where the excitation light interacts with the target molecules include the spaces close to both the top and side surfaces of the Au nanoparticles. Assuming that analytes are distributed uniformly in the surrounding environment, the absolute amount of the SERS signal for each unit cell correlates with the field enhancement integral in the space with nonzero field distribution (i.e., everywhere in the unit cell apart from the nanodisk). To better quantify the effect of electric field enhancement on the SERS performance, we calculated the VFE by integrating the local filed enhancement E/E_in over the volume of one unit cell, taken as an extruded hexagon around the nanodisk rising 10% over the nanodisks’ height (inset in Figure 3f, cross-section in Figure S2, VFE normalized to unit volume in Figure S3). The calculated VFE shows that at all analyzed wavelengths, D49H16G22 has the strongest VFE, followed by D33H12G16, D16H8G8, and D14H4G7 (Figure 3f). Assuming that such patterns cover the entire fiber surface, we have estimated the number of unit cells N that covers the conical taper surface (see Methods). For a fixed surface area, the smaller NIs system will contain more unit cells than the bigger particles. Figure 3g reports the unit cell VFE multiplied by N for each pattern, showing that for most of the wavelength regions on which the calculations were performed the bigger particle patterns experience the higher absolute field enhancement (except for 570 to 620 nm). These results suggest that if the electromagnetic effect is the main contribution to the overall SERS signal, the larger particle pattern will generate the highest SERS scattered signal.

To verify the correlation of the NIs-TFs’s performances with the 4 different NIs patterns on the TF surface, we conducted through-fiber SERS measurements in solution. We first used R6G as a chemically stable analyte to test the LOD of the probes. Different concentrations (from 10^–9 to 10^–3 M) of R6G aqueous solutions were prepared, and then through-TF SERS measurements were conducted with a custom-built Raman microscope (details in Methods). For each NIs-TF, the measurements started from the lowest concentration solution and increased toward the highest concentration in steps, using a single fiber for each concentration ramp. Each spectrum was taken immediately after the fiber tip was immersed entirely into the aqueous solution, and the fiber tip was kept in the solution only during the exposure time. Four different patterns (D14H4G7, D16H8G8, D33H12G16 and D49H16G22) of the NIs-TFs set were tested, with n = 2 fibers for each pattern. The measured spectra at different concentrations are reported in Figure 4a–d, in the original spectra, they are composed of a contribution from the analyte and a Raman background generated by the probe itself. This latter is not the same for the four NIs patterns (Figure S7). To improve the visibility and contrast of the analyte’s signal, a reference spectrum method was employed to subtract the silica background. This process began by acquiring the fiber Raman background from clean samples in water without any analytes present. Then, the analyte’s signal was obtained by subtracting the fiber background from the original SERS spectrum, as shown in the bottom panels of Figure 4a–d. For all types of NI-TFs, the R6G Raman signature can be identified from 10^–6 M, and the peak intensities increase as the NIs grow. Among all the subtracted spectra, 12 R6G Raman peaks were identified (1018, 1048, 1092, 1128, 1188, 1275, 1314, 1364, 1513, 1579, 1607, 1653 cm^–1). In Figure 4e–h, a more detailed data analysis is presented with the bubble charts, allowing a better comparison between the different samples. The raw spectra were first normalized to the silica peak at 1055 cm^–1, then the silica background measured in water was subtracted from the normalized spectra. Working on the subtracted normalized spectra set, the 12 peaks were analyzed individually. A local linear baseline was removed from each peak, then we integrated the area of each peak. The summation of the corresponding peak intensities of the two data sets was used as a marker for the bubble sizes in the bubble chart (see Figure S5 for more details). For all 4 types of patterns, R6G signature peaks appear at the concentration of 10^–6 M. Most of the peaks increase in intensity concurrently as concentration increases, and the patterns comprised of larger NI’s, result in a more intense R6G SERS signal with a direct correlation to the total VFE increase.

To broaden our findings beyond R6G, we employed NI-TFs to detect serotonin. This monoamine neurotransmitter, smaller than R6G and showing a stronger affinity for gold, influences numerous neurological functions and is implicated in conditions like depression and anxiety. (52,53) Thus, tunable NIs-TFs sets with n = 2 fibers per pattern were measured. The representative spectra data sets and the corresponding subtracted spectra sets are shown in Figure 5a–d. The same data analysis approach as R6G was adopted (details in Figure S5). There are 8 peaks identifiable across all the spectra data sets, at 908, 1036, 1115, 1339, 1420, 1482, 1538, and 1648 cm^–1. The detailed analysis results are shown in the bubble chart in Figure 5e–h. Interestingly, for serotonin, more peaks and lower LOD (10^–7 M) are obtained with smaller NIs patterns (D14H4G7 and D16H8G8). Additionally, from all the data sets there is an obvious degradation of the detection capability (or even vanishing) at high concentrations. Further tests on the probe that degraded at high serotonin concentrations show that it is still able to detect R6G molecules but with less peak intensity and one order of decreased detection sensitivity at 10^–5 M (detailed data presented in Figure S6). Considering the lowest LOD, the peak numbers, and the spectra prominence, D16H4G8 NIs patterns show the best performance.

In a through-fiber detection setup with plasmonic NIs-TFs, various factors impact SERS signal detection of a target analyte. These include: (1) NIs patterns’ near-field enhancement, (2) relative strength between molecular SERS signal and fiber-generated Raman background, (3) analyte interaction with gold NIs, and (4) spatial overlap between analytes and near-field enhancement. Our study shows distinct influences of these factors on R6G and serotonin, a neurologically relevant neurotransmitter.

For R6G, our experiments show that the total VFE of the tunable NIs patterns on the taper surface mainly dominates the spectral response. The higher SERS signals can be obtained along with gradually increased total VFE. The modulation on the background reflection does not affect the intensity of characteristic peaks significantly, since R6G has a large Raman cross-section which can provide high relative strength when mixing with the silica Raman background from the fiber. R6G does not possess functional groups with a strong affinity to gold (therefore it is unlikely that R6G will bond on NIs and generate surface chemical reactions), and the spectral response is stable with well-recognized molecular vibrations. (54,55) Additionally, the molecular size of R6G (∼1 nm (56,57)) is well below the narrowest investigated gap (∼7 nm). Given the low gold affinity and this factor, we hypothesize a uniform spatial overlap between R6G molecules and near-field enhancement, aligning with assumptions in our simulations.

For serotonin, its small Raman cross-section results in a weak signal (relative to fiber background), which makes the modulation of the background have more impact on the detection of serotonin peaks. Thus, the increasing background observed as a function of the NI’s size is more likely to diminish the peaks’ prominence at low concentrations. This may explain the lower LOD obtained for samples D14H4G7 and D16H8G8. Smaller serotonin molecules, roughly 60% the size of R6G, with amino functional groups, can bond to gold surfaces more effectively, (58−62) especially in narrow gap regions and near the surface, even at low concentrations. This results in better spatial overlap between molecules and more intense ’hot spots’ generated by narrower gaps in smaller NIs patterns like D14H4G7 and D16H8G8. Along with concentration increase, the increased density of serotonin bonded on the gold surface might induce a polymerization reaction (63) forming a Raman inactive dielectric layer dampening the SERS sensitivity at high concentrations (≥10^–4 M). The interplay of all those factors results in the serotonin spectra response shown in Figure 5. The serotonin SERS bands vary in the literature due to factors like excitation wavelength, dissolving environments, and nanometal materials used. (14,64−66) Peak disappearance and shifts have been noted in indole ring modes. (67) Studies often focus on thiol molecules due to their strong bonding with Au surfaces. (68−71) For instance, 4-nitrobenzenethiol (NBT) can dimerize into 4,4′-dimercaptoazobenzene (DMAB) under laser radiation, showing new peaks in the SERS spectrum. (68) Au nanoparticles on microsphere systems have shown similar reactions. (71)

In pursuit of improved sensitivity for SERS probes, another important aspect we should consider is the collection properties of the SERS signal for nanoparticle-decorated fiber probes. Different from the volumetric collection scheme of the blank TF, (35) the collection of nanoparticle-decorated TF mainly happens at the interface. The NIs on the tapered fiber surface are first excited by the fiber photonic-guided and radiative modes, forming the local field enhancement around the NIs’ surfaces, which are so-called “hot spots”. When molecules are in those hot spots, a Raman dipole can be generated. The NIs then work as antennas to convert that evanescent molecular source field to propagating far-field radiation, and the plasmonic mode dominates the far-field radiation patterns. The radiation patterns can be studied by Fourier imaging. As shown by Shegai et al., (72) the SERS emission is distributed in all the measured directions in a trimer system. Also, Zhu et al. (73) show in a dimer antenna on a glass and dielectric interface system that the simulated emission patterns radiate power into the substrate, with the emission angle distribution exceeding the critical angle for total internal reflection at the air-dielectric interface. Thus, the nanoislands on the tapered fiber surface facilitate Raman emission into the fiber with a wide angular distribution. The portion of light with angles within the fiber’s collection capability, limited by the fiber’s numerical aperture (NA), is collected and forms guided modes within the fiber, which then propagate back for detection. For a more comprehensive analysis of plasmonic fiber, Kim et al. (74) simulated dimer systems on top of the flat fiber. Factors including the particle geometry, excitation laser wavelength, refractive index, and NA of the fiber have been discussed, elaborating the importance of engineering the plasmonic “hot spot” close to optical fibers, deepening the understanding of particle-decorated fiber detection systems.