Cen et al.

eLight (2023) 3:9 Official Journal of CIOMP

elight.springeropen.com
https://doi.org/10.1186/s43593-023-00041-7

RESEARCH ARTICLE Open Access

Microtaper leaky‑mode spectrometer

with picometer resolution
Qingqing Cen1,2†, Sijie Pian1,2†, Xinhang Liu1,2, Yuwei Tang1,2, Xinying He1,2 and Yaoguang Ma1,2,3*   

Abstract
The wide application of optical spectroscopy makes miniaturized spectrometers with fundamental importance. The
scalability, high-performance, low-cost, and small footprint are still contradicting each other and limiting the applica-
bility of miniaturized spectrometer for practical application. Here we propose a compact spectrometer that satisfies
the four advantages. The device uses a fiber taper tip to generate complex leaky mode patterns within 1 mm length.
The unique correspondence between the pattern and wavelength operates effectively for hundreds of nanometers
spectral range while providing a spectral resolution around ~ 1 pm. The integration of multiple taper tips enables
hyperspectral imaging applications. The working range of our device can be further extended using different materi-
als and detectors while keeping the similar architecture.
Keywords Spectrometer, Microfiber, Deep learning, Hyperspectral

1 Introduction the existence of an inherent tradeoff between the above

The precision and elegance of obtaining abundant light- aspects are constraining this long-term theme of minia-
matter-interact information with a snapshot measure- turization from advancing. Generally, spectrometers with
ment makes optical spectroscopy indispensable elements dispersive elements require extra spatial separations tend
for modern industries and scientific research [1]. The to leave a large footprint [9, 10]. Filters (including nar-
miniaturization of traditionally bulky spectrometers rowband and reconstructive types) based designs suffer
has been strongly motivated by the vast applications, from power loss caused by either absorption or reflection
including biomedical sensing [2, 3], material analysis [4], as well as limited resolution and bandwidth due to lim-
optical communication [5, 6] and light source characteri- ited channel numbers [11, 12]. On-chip spectrometers
zation [7]. Researchers have been engineering spectrome- relying on nanofabrication usually have very low coupling
ters for lower cost, higher flexibilities, smaller size, better efficiency for wideband operation [13, 14]. Multimode
stabilities, and performances [8] for quite some time. But interferences could generate random speckles which
are associated with spectra information [13]. However,
†
Qingqing Cen and Sijie Pian contributed equally to this work most spectrometer designs based on it rely on random
medium such as rough surface [15], multimode fiber [16,
*Correspondence:
Yaoguang Ma 17], integrating sphere [18], photonic crystals [19], etc.,
[email protected] which usually require the companion of bulky or expen-
1
State Key Laboratory of Extreme Photonics and Instrumentation, sive equipment, such as a high-performance camera or
College of Optical Science and Engineering; International Research
Center for Advanced Photonics, Zhejiang University, Hangzhou 310027, even a microscope, to complete the measurement [13,
China 20]. Usually, optical fibers can generate random speckles,
2
ZJU-Hangzhou Global Scientific and Technological Innovation Center, but the length of the fiber used are on the scale of meters.
Hangzhou 311215, China
3
Jiaxing Key Laboratory of Photonic Sensing & Intelligent Imaging, Long photonics crystal fibers or multimode fibers are
Intelligent Optics & Photonics Research Center, Jiaxing Research Institute, vulnerable to environmental fluctuations thus weaken-
Zhejiang University, Jiaxing 314000, China ing the stability of spectrometers based on them [21, 22].

© The Author(s) 2023. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which
permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the
original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or
other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line
to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory
regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this
licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by/4.​0/.
Cen et al. eLight (2023) 3:9 Page 2 of 11

The rest design requires structures with fancy character- the core structure of our spectrometer consists of a
istics that are difficult or expensive to fabricate [13, 23, curved microfiber taper tip mounted on a CIS with
24]. Above all, a flexible low-cost tiny spectrometer with both sides of the taper glued to the substrate became
stable high performance is still elusive. insensitive to common environmental vibrations and
On the other hand, microfibers are ideal tools for manip- impacts. The reliability of our device enables us to cap-
ulating light field for its tailorable dispersion and small ture incoming signals over a long time period (usually,
footprint. But experiments usually use microfibers for con- the collection time for massive experimental data could
fining light inside the fiber to propagate as long as possible. exceed 8 h continuously).
If a microfiber taper is drawn under non-adiabatic condi- The core-size-mismatch between PMF and MMF ena-
tion, the fiber geometry will force the guiding modes to bles the excitation of multiple transverse modes. These
evolve from single-mode to multimode and then couple to modes will again couple to abundant transverse mode
single-mode/few-mode again in a short distance. The lat- propagating inside the sharp upper taper region first [29,
ter process will generate leaky modes that normally being 30], then leaked at the predesigned lower taper region
undesired for microfiber applications [20, 25–28]. But, by and project onto the CIS with a complex light field pat-
engineering the drawing conditions, we can maximize the tern as shown in Fig. 1b, c. The leakage only happens at
leaky modes generation within a 1 mm taper region and certain position and orientation around the taper with
make use of the leaked random interference speckles. The sub-nanometer-scale surface roughness (Fig. 1e, f ). The
sharp angle of the taper makes it less susceptible to envi- mode numbers of the taper, the taper length and the elon-
ronmental disturbance compared with adiabatic fiber gation show significant effect to the performance of the
tapers. Here we demonstrate a low-cost, scalable spectrom- spectrometer (see Additional file 1: S9 for details). In gen-
eter that has a picometer resolution and sub-millimeter eral, a larger core diameter can help excite more modes to
footprint. The spectrometer utilizes complex leaky modes produce more complex pattern. Appropriately extending
speckles projected from a curved microfiber taper tip that the taper length will also help us obtain more details of
uniquely determines the wavelength of the input signal. the patterns, which is expected to increase the accuracy
By solid packaging with a complementary metal-oxide of spectrum reconstruction. Besides, a larger elongation
semiconductor (CMOS) imaging sensor (see Materials and would produce a larger leaking area for leaky modes, thus
Methods for details), the data acquisition of our spectrom- more pixels can be used to record the spectral informa-
eter can be finished using a single snapshot with no exter- tion, which may increase the bandwidth of the device
nal equipment needed. A light-weight vision transformer [13]. Therefore, in the experiment, we selected a multi-
(ViT) network was used for analyzing the complex frames mode fiber with a core diameter of 105 μm to excite more
recorded by the CMOS image sensor (CIS). The correlation modes. The taper length and the curvature are fine tuned
between the spectral information and leaky mode images to reach a balance between small footprint and enough
can be easily constructed after training. In addition, this details of the patterns captured by the CIS. Using a home-
high-performance tiny device is fabricated with low-cost made fiber tapering machine (Additional file 1), the size
elements (the core components of the spectrometer cost of the captured pattern can be shrunk down to an area
less than $15) and could be used over long intervals while within 300 μm * 600 μm while maintaining enough infor-
keeping the accuracy and reliability. mation. Furthermore, to enable a maximum detection
efficiency, we also made sure the taper will leak almost all
2 Results the power coupled-in by shape parameter tuning of the
Figure 1a shows the schematic of our spectrometer. The taper (Fig. 1g, h). The output of this ultracompact high-
spectral signals lens-coupled into an in-line fiber polar- performance spectrometer is a complex interference pat-
izer (ILP) and a polarization maintaining fiber (PMF) to tern with ample fine structures that can be captured with
maintain the signal stability against coupling fluctua- a single shot within tens of milliseconds even when the
tions and polarization drifts [20], and are then guided power of the projected image is tens of microwatts.
through a multimode fiber (MMF). The fusion spliced Accurate reconstruction of the projected image to
region between the PMF and the MMF are glued to a spectroscopy information relies heavily on the algo-
rigid glass substrate to ensure a stable coupling. After rithms. To enable a fast and mobile friendly identi-
tapering a short part of the MMF using a homemade fication of the projected light pattern, we adopted a
tapering machine, we use a three-dimensional stage transformer-based model called MobileViT [31] for the
to bring the taper into contact with the CIS (Sony image classification as shown in Fig. 2. The combination
IMX219). Ultraviolet sensitive glue was used to fix the of convolutional neural networks (CNNs) and ViTs can
position and gesture of the fiber taper tip (Fig. 1d, also enable a lightweight and low latency network that could
see Additional file 1: Method for details). After fixation, be applicable for portable and field applications.
Cen et al. eLight (2023) 3:9 Page 3 of 11

Fig. 1 Anatomy of microtaper spectrometers. a Schematics of microtaper spectrometer. b, c Side-view and top-view of the leaky mode patterns
from a microtaper shines on the substrate. The wavelength of the incident laser used here is 600 nm. d Photo of a fabricated microtaper
spectrometer. e, f Scanning electron microscope images of the microtaper on the CIS. g Fiber transmission curves versus the taper parameter. The 4
fibers are tapered with elongation (x) of 2.72 mm, 3.20 mm, 4.64 mm and 7.76 mm respectively. By carefully control the shape of the microtaper, the
incident power can be transported to the leaky mode pattern with near 100% efficiency. h Fiber transmission as a function of interaction length.
The interaction length is the contact length of the tapered fiber waist region and the CIS. The 4 fibers are tapered with elongation (x) of 5 mm,
6 mm, 7 mm and 8 mm respectively. The power projected from the microtaper can also be tuned by changing the interaction length

The architecture of the network is shown in Fig. 2a. alternatively. The 4 groups of images in the mid-
The images collected by the CIS is first transmitted to dle shows some examples of the outputs after pass-
a 3 × 3 convolution layer (CNN) and 4 inverted resid- ing the 3 layers respectively. As the network’s depth
ual blocks respectively to downscale the dimensions. increases, the information inside the images is distil-
Then the outputs are passed into 3 layers of inverted lated gradually. We can visualize in the activation maps
residual blocks and mobileViT blocks (see Method) of our trained network (with Network Dissection; see
Cen et al. eLight (2023) 3:9 Page 4 of 11

Fig. 2 Identifying the spectral information from captured leaky mode pattern images. a The transformer architecture used for spectra
reconstruction. MobileViT units are used to introduce the benefits of both CNN and transformer. The 4 sets of images in the center are the
activations of the first convolutional layer and the three mobileViT blocks, respectively. b Clustering of leaky mode patterns visualizes the high
accuracy of the reconstructed results. Here 10 different clusters are used to show the identification results of 1000 patterns of different wavelengths.
c Reconstruction of tested peak spectra. The black dashed lines indicate the ground truths, while the solid lines are the corresponding predicted
spectra. The working bandwidth is limited by the performance of our silicon CIS used

Methods for details) and observed that, almost all the by sigmoid function and retrieve the final reconstructed
original fine features are preserved initially. But the fol- spectra.
lowing layers separate the main image characters into The ViT network works excellently for image identifi-
different channels, such as the overexposed area in the cations which are easily capable of associating light pat-
middle or the speckle area at the edges. As the images tern images with wavelength information. We performed
are processed during this process, the vision informa- a 2D multidimensional scaling (MDS) projection of the
tion is transformed into spectral information gradually. output of the last mobileViT layer. If leaky mode pat-
Finally, the data was sent to Dense layer to be activated terns are identified with similarities in the network, then
Cen et al. eLight (2023) 3:9 Page 5 of 11

the distance between their activation vectors will be less. pre-feed with suitable training data. (See Additional
Thus, the categorization process of subsequent network file 1: Figure S5e, S5f ).
layers separating out the different images can be visual- The spectral resolution of a spectrometer can be esti-
ized clearly. Figure 2b shows the clustering of 1000 dif- mated by identifying two nearby input wavelengths.
ferent leaky mode patterns (corresponding to different We acquired a series of images while scanning a nar-
spectra varies in both intensity and wavelength) that row linewidth laser (Matisse 2 TX-Light) from 753.01
are distinctly clustered by their spectral information to 753.15 nm in steps of around 0.95 pm to the trained
extracted, which demonstrates the high accuracy of the MobileViT network. We get test speckle pattern using
reconstructed method. method in [16]. We gradually change the separation dis-
Transfer-matrix-based reconstruction algorithms are tance between wavelengths to find the reconstruction
susceptible to experimental noises in practice. But for limit of our device. As shown in Fig. 3b, two peaks with
neural networks, noises can be considered and reduced wavelength difference around 1.53 pm can be identified
with proper training, since the training process can be in the NIR regime. The other cases with wavelength dif-
fed by thousands of data that carries noise with differ- ferences larger than that can be easily reconstructed. The
ent types and levels. Furthermore, the mobileViT net- recovery process still possesses high fidelity although the
work utilized here combines the advantages of both CNN addition process will introduce higher readout noise and
and ViT. The convolution and pooling operation in CNN making the reconstruction process less stable, leading to
make the system architecturally invariant to translations possible jittering of the single peak FWHM.
thus immune to light intensity induced leaky mode pat- The fiber taper tip parameter will also affect the per-
tern variation. And ViT could obtain the long-range formance of our spectrometer. Figure 3c shows an
information of the data, thus having better receptive field. estimation of the resolution ability of three spectrom-
Utilizing these unique advantages, we tested the wave- eters with different fibers (differs in the core diam-
length resolving ability of our spectrometer across a eter) as the tapering fiber. The left, middle, right parts
spectral range of 450–1000 nm (limited by the work- represent results for the standard single-mode fiber
ing bandwidth of our silicon CIS adopted in the experi- (core diameter = 8.2 μm) case, the standard multi-
ments). As Fig. 2c shows, the spectrometer accurately mode fiber (core diameter = 62.5 μm) case, the spe-
recovers the positions of the tested spectra (generated cial multimode fiber (core diameter = 105 μm) case,
by a supercontinuum light source (YSL Photonics SC- respectively. The estimation was done by calculat-
OEM) with an acousto-optic tunable filter (AOTF)) with ing spectral correlation function of speckle intensity,
an average peak signal-to-noise ratio (PSNR) of 46.7 dB. C(�) = �I(, x)I( + �, x)�/�I(, x)��I( + �, x)� − 1.
The reconstructed narrow-band lights from the proposed The 105 μm-core-diameter multimode fiber case shows
device (color solid line) and ground truth spectra from clearly better resolvability as a thicker fiber would yield
a commercial grating spectrometer (black dashed line, more supporting modes and enriching the projected
Ocean Optics, LEDPRO-50) shows great agreements. interference pattern for finer detail reconstructions.
The center-wavelength error of the monochromatic Similar trends can also be observed for fiber tapers with
light was approximately 0.0223%. The linewidth error is different shapes. As shown in Fig. 3d, three fiber taper
approximately 7.37%. spectrometers were fabricated using tapers with differ-
To evaluate the performance of our device, we fur- ent heated region, which resulting in different tapering
ther tested the bandwidth of spectra that can be identi- angles (see Additional file 1: Figure S6) [32]. From left to
fied with a single shot. Figure 3a shows some examples the right, the heated region on the same fibers is 1.5 mm,
of continuous spectra that were reconstructed via our 3 mm, 5 mm. A larger taper angle will introduce more
spectrometer. The full-width at half-maximum (FWHM) efficient mode coupling in the taper region and thus mak-
are around 30, 60, 90 nm respectively for the input sig- ing the final pattern projected more complex and contain
nals generated by synthesizing spectrum. The black more information for the recovery process. On the other
dashed lines show ground truths, which matches closely hand, using coupling fibers with larger diameter differ-
with the reconstructed spectra from our device. We find ences [33] or displace the fusion splice position [34] can
good agreement with the measured spectra at tens of nm further increase the mode numbers excited (see Addi-
bandwidth and qualitative agreement at near-hundred- tional file 1: Figure S7) to fully exploit the resolution abil-
nm bandwidth. For very broadband inputs, the super- ity of our device.
position of the interference pattern will level out the fine Due to the small footprint of our spectrometer,
structure and weaken the fine frequency reconstruction we can implement multiple fiber taper tips on one
fidelity. But the coarse patterns left can still be used for CIS and use it for hyperspectral imaging as shown in
the wide frequency range signal reconstruction when Fig. 4a. Here we use 20 tapers integrated on the CIS
Cen et al. eLight (2023) 3:9 Page 6 of 11

Fig. 3 Characterization of microtaper spectrometers. a Reconstruction of continuous spectra with FWHM 30, 60, 90 nm. The black dashed line is
ground truth. b Reconstruction of two narrow peaks with separation 1.53 pm, 6.89 pm, 18.95 pm. c Evaluation of resolution of three spectrometers
made from tapers with different core diameters. The core diameter from left to right is 8.2 mµ, 62.5 mµ, 105 mµ. d Evaluation of resolution of three
spectrometers made from tapers with different heater regions. The heater region length from left to right is 1.5 mm, 3.0 mm, 5.0 mm. Each column
of data in the heatmap represents the estimated result of a set of measurement data (see method). The black dashed lines represent estimated
resolution of each measurement data. The scale bar in c, d is shared by respective subfigures

and arrange the fibers to be aligned in a row for scan- spectra and recovered spectra). A pseudo-colored
ning. As a proof-of-principle demonstration, we first image shows the hyperspectral imager can easily differ-
use this hyperspectral imager to acquire an image for entiate broadband light signals in an image even when
a projected target. Our target is formed by projecting they are almost identical to human eyes. We then test
the logo of Zhejiang University using a supercontinuum the applicability for real world samples. We use some
laser and an AOTF to generate spectra within 590– color plastic sheets to test the color-resolving accuracy
630 nm, as shown in Fig. 4b. The illuminated light spec- of our device. The actual look of the final hyperspec-
tra can be identified clearly using our device (Dashed tral image and the photo of the plastic sheet sample are
lines and solid lines in Fig. 4c represent ground truth not identical (Fig. 4d, e) because the spectrum below
Cen et al. eLight (2023) 3:9 Page 7 of 11

Fig. 4 Hyperspectral imaging demonstration using our spectrometer. a 20-channel hyperspectral imager. b Reconstruction of a projected
image. The RGB colors of the input and prediction are calculated from their spectra with color matching functions. A pseudo-colored image
was constructed using the prediction color data to show the detail of the input image. c Spectra of the 11 colors used to assemble the image.
The black dashed lines indicate the ground truths, while the solid lines are the corresponding predicted spectra. d Photo of a color film patch. e
Reconstructed RGB image of d. The colors are calculated from the transmittance spectra which is predicted by our hyperspectral imager. f Spectra
of the center point of each patch in e denoted by the markers. The black dashed lines are the ground truths, while the blue solid lines with markers
are the corresponding predicted spectra. The purple shaded area indicates unreconstructed wavelength range (400–450 nm) due to the limitation
of light source spectral range. g Color coordinates in CIE 1931 color space of the center point of each patch in e denoted by the markers

450 nm is not recorded in our spectrometer (the AOTF imager and data from the conventional spectrometer
we used does not work for light below 450 nm), while shows great potential of the ability of our design. The
the photo is captured under direct sunlight. We recon- number of pixels available simultaneously (20 in this
struct the actual spectra of the center point of each case) for our hyperspectral imager is limited by the pat-
plastic sheets (Fig. 4f ) and compares them with the tern image size, resulting in the low-resolution of the
results from a conventional spectrometer (Agilent Cary recovered image (Fig. 4e). In practice, more pixels could
5000). The results have very good agreements indicat- be added efficiently by using more CIS with integrated
ing the high accuracy of our spectrometer for real- tapers or a CIS with larger area. Although each spec-
world scenarios. The CIE color maps are also employed trometer needs to be pre-calibrated before used effec-
to compare the color consistency between the ground- tively, it is realistic to construct a hyperspectral imager
truth color and our reconstruction results (Fig. 4g). The with thousands of “hyper-pixels” in each dimension by
high agreements between data from our hyperspectral updating the current design.
Cen et al. eLight (2023) 3:9 Page 8 of 11

3 Discussion 4 Methods
Previous miniaturized spectrometers require high-preci- 4.1 Micro taper spectrometer fabrication
sion fabrication or alignment of dispersive elements that The microfiber tapers are fabricated with a home-made
are difficult to synthesize or maintain. However, this is fiber drawing machine (see Additional file 1: Figure S1).
not the case for the microtaper spectrometer. The fiber A commercial optical fiber was cleaved and put inside a
taper structure is very easy to fabricate [35] and can be 5-mm-long ceramic tube. The tube was heated to 1200℃
integrated on the CIS [36] even with bare human hands. using a C ­ O2 laser. The two ends of the fiber were pulled
We can easily draw high quality fiber tapers with an alco- by two linear servo motor stages (Yaskawa SGLF Series)
hol burner (see Additional file 1: Figures S2, S3) and the fixed on a sliding rail. The flame-brushing technique was
integration process requires no alignment as long as the used to control the movement of the stages with the posi-
taper shed light on the CIS. This makes the microtaper tion of the ceramic tube fixed.
spectrometer special from other congeners: A spectrom- After tapering, the micro taper sample is fixed on a
eter with ultrahigh accuracy (~ 1 pm) and tiny footprint micro positioner. A CIS (Raspberry Pi Camera Module
(~ 1 ­mm2) can be made easily at home for core com- 2, Sony IM219) module is fixed directly under the taper
ponents less than $15 with some simple tools. Further after removing the lens and cover glasses. The micro
improvement of reconstruction quality can be achieved positioner is adjusted slowly and gradually to move the
by using high performance image sensors (smaller pixel taper into close contact with the CIS. A 650 nm laser was
size, larger bit depth, lower dark current level, etc.), coupled into the taper to assist the process as an indi-
enlarging numbers of leaky modes and improving the cator for the contact condition. The taper is fixed with
device stability. Since the algorithms are trained on a UV sensitive glue after the projected patterns observed
computer with commercial graphic card with ease in from the CIS is optimized. The tilting angle of the taper
our case, the scalability of this design is quite satisfac- is also finely tuned to make sure the leaking efficiency of
tory. Ideally, since the silica fiber is transparent from the taper. Finally, we cover the device with a cover glass
400 nm to almost 2000 nm, our spectrometer could be and seal the device with UV glue to ensure an air-tight
operational in this range with perfect performance with capsulation.
suitable CIS modules. Furthermore, if we consider the The hyperspectral imaging sample was fabricated with
actual propagation length of the fiber taper is quite short very similar method. We fix two wedge structures on the
(several centimeters long, the polarization maintaining substrate to elevate all the 20 fiber tapers. The relative
part can be discarded when the whole device is used in position and direction of the 20 tapers are adjusted indi-
a fixed scenario, including the input fiber for coupling.), vidually during the contact process mentioned above and
the working range could be even wider (less absorption) fixed using UV glues to make sure the maximized utili-
to cover from 0.3 nm all the way to 3 µm. Such spectrom- zation of the CIS chip area while preventing crosstalk of
eters may be applied to food inspection, drug identifica- neighboring microtaper spectrometers.
tion, and to personalized health diagnostics with very low
cost. Besides, if we consider polarization state in our cali-
bration procedure, we could acquire both the spectrum 4.2 MobileViT network training
and polarization state of unknown light. We use a supercontinuum source and an AOTF to obtain
Some issues need to be considered to further improve monochromatic spectral signals at the wavelength range
the performance of our spectrometer. For a broadband of 450–1000 nm of different intensities. The lasers are
light input, speckle overlap will result in a decrease in separated using a beam splitter before reaching the
image contrast, which leads to a trade-off between band- coupling lens of our spectrometer to be monitored by a
width and resolution of the spectrometer. Our spectrom- conventional spectrometer (Ocean Optics LEDPRO-50)
eter exhibits good results when analyzing coherent light. at the same time so that the spectrum of each signal is
For incoherent light with different wavelengths, differ- obtained. The leaky mode patterns corresponding to
ent mode distributions can be used to reconstruct the the signals were captured with the CIS respectively as
spectrum with a lower resolution, since reduced coher- training sets. We obtain 10-bits raw data from the CIS
ence will decrease the pattern complexity. In addition, directly and crop the region of interest (ROI) with the
although the spectrometer is easy to fabricate for bare size of 640 × 1400. Our dataset consists of images cor-
hands, more efforts in the automation of the manufacture responding to random spectra that can be obtained by
could lead to mass production of the spectrometers. randomly synthesizing images of different intensities of
different wavelengths. In order to obtain better train-
ing result, our training set usually has 10,000 images.
We train our MobileViT network on a server with a
Cen et al. eLight (2023) 3:9 Page 9 of 11

NVIDIA 3090Ti graphic card. All the images are normal- minimum resolution can be obtained by the spectral cor-
ized to a 0 to 1 scale. The exposure time, gain of the CIS relation function.
are kept constant during all the data collecting process. After acquiring a series of images to the trained Mobi-
For the spectra reconstruction tasks, the Dense output leViT network, we characterize the spectral resolution
layer is activated by the sigmoid function. The RMSprop of our microtaper spectrometer by testing its ability to
method is used to optimize the MSE loss. The learning distinguish two peaks ultra-narrow spectra. The pat-
rate exponentially decays over time with the initial value tern image for two peak ultra-narrow spectra were
of 5 × ­10–4. For the spectra classification tasks, the out- constructed by summing the images of the desired wave-
put layer is activated by the softmax function. The Adam lengths sequentially.
method is used to optimize the categorical crossentropy
with 0.1 label smoothing. 4.4 Hyperspectral imaging data acquisition
Inside the MobileViT, the input feature map is first The ZJU-logo images are projected using a supercon-
convoluted with 3 × 3 kernel to learn the local represen- tinuum source and an AOTF. The lasers are separated
tations by image-specific inductive biases of CNN, then using a beam splitter before reaching the coupling lens
projected to a high-dimensional space by the point-wise of our spectrometer to be monitored by a conventional
convolution. The output is further unfolded into a series spectrometer (Ocean Optics LEDPRO-50) at the same
of non-overlapping 2 × 2 patches, then fed to the L trans- time. The arbitrary colors of the ZJU logo are generated
formers. The self-attention-based transformers encode by combining the output of the 8 channels of the AOTF.
the inter-patch relationships to learn the global repre- All the spectrometers on the hyperspectral imager are
sentations. After the extractions of local and long-range pre-trained before the data collection using the method
information, the output is folded back and fused with described above. The plastic sheet samples are also tested
the original input. We train our MobileViT network on a with the same equipment.
server with a NVIDIA 3090Ti graphic card.
The clustering process shows the inherent connec- Supplementary Information
tions between the leaky mode patterns. Since the pat- The online version contains supplementary material available at https://​doi.​
terns are correlated with the wavelength information, org/​10.​1186/​s43593-​023-​00041-7.
patterns associated with similar spectra will be more
Additional file 1: Figure S1. Home-made fiber tapering system (HFTS).
resembling. We divide the spectral range into multiple Figure S2. Tapers drawn with varying heat-zone lengths by the HFTS.
parts (In Fig. 2b we used 10 parts). And adjust the output The heat-zone lengths used in the 4 tapers from top to bottom are 1
of our network to classify the data. The activation func- mm, 2 mm, 3 mm and 4 mm, respectively. Figure S3. Comparison of
hand-drawn tapered fibers and machine-drawn fibers. (a), Manually
tion is switched to softmax and optimized with Adam drawing a fiber by an alcohol lamp. (b), (c), Taper tips drawn by hands
optimizer. The distance between two different patterns is (b) and HFTS (c). d,e, Full tapered fibers drawn by hands (d) and HFTS (e).
generated by calculating the pairwise distance between The scale bars in b-e are 500 µm. Figure S4. Photo of a mini microtaper
spectrometer after fixing the microtaper on the CIS using UV sensitive
the activations of the two patterns. After we obtain the glue. Figure S5. Reconstruction of various spectra. (a) The green, orange,
distance matrix, we use MDS to project the matrix into a and red curves are the reconstruction results of narrow linewidth spectra
two-dimensional distribution. at three different central wavelengths λ. (b) The colored solid lines are the
reconstruction results of narrow linewidth spectra with the same central
wavelength and different relative light intensities P. The black dashed
line is the actual spectrum measured by a conventional spectrometer.
4.3 Resolution test (c-d) Reconstruction results of discrete multi-peak spectra by leaky-mode
We use Matisse 2 TX-Light tunable laser to accomplish spectrometer. (e) Reconstruction results of spectra filtered by filters with
resolution test. The laser is coupled into a standard bandwidths of 10nm. (f ) Reconstruction results of spectra filtered by filters
with bandwidths of 10nm. Figure S6. Spectrometer samples made from
multimode fiber using a fiber collimator (THORLABS different tapers and comparison of microfibers of different heated region.
F810APC-850). After passing through the multimode (a), spectrometer samples made from tapers with different parameters.
fiber, the light is propagated to the standard single mode From left to right, the first to the third cases are spectrometers made
from tapers with different core diameters: 8.2μm, 62.5μm, 105μm and the
fiber or SMA fiber to input the light to the spectrom- heated region of these tapers is 1.5mm. Their estimated resolutions are
eter. The fibers behind the in-line polarizer are fixed to displayed in Fig. 3C. The third to fifth cases are spectrometers made from
stabilize the spatial polarization of the light. During the tapers with different heated region: 1.5mm ,3mm ,5mm and the core
diameters are 105μm. Their estimated resolutions are displayed in Fig. 3D.
test, we tune the laser to change the wavelength gradu- The ROIs used to reconstruct the spectrum are marked in (b), Tapers
ally, at the same time, the Raspberry Pi stores the cap- drawn by HFTS with different heated regions. From top to bottom, the
tured pictures. This data collection lasts for about 280 s heated regions are 1.5mm, 3mm, 5mm corresponding to the situation in
the third to fifth spectrometer samples in (a). The scale bars are 500μm.
and we can get about 2200 raw pictures. Matching the Figure S7. Simulated number of modes excited when coupling with a
time of pictures captured with the recorded data of the vertical offset for the multimode fiber. The input fiber is a PANDA polariza-
wavelength meter, we can relate the wavelength of laser tion maintaining fiber (PMF). The output fiber is a multimode fiber (core
diameter = 105 µm). Figure S8. All the predicted spectra in figure 4e. The
with its correspondence pictures captured. The estimated
Cen et al. eLight (2023) 3:9 Page 10 of 11

and a conventional near infrared reflectance (NIR) spectrometer for

y-axis of all subplots is the normalized intensity, and the horizontal axis characterizing soil carbon and nitrogen. Geoderma 338, 422–429 (2019)
represents the wavelength range from 450nm to 770nm. Figure S9. The 5. B.I. Akca, C.R. Doerr, Interleaved silicon nitride AWG spectrometers. IEEE
simulation results of different tapers with different core diameter. The core Photonics Technol. Lett. 31(1), 90–93 (2019)
diameters of tapers are (a) 8.2 μm, (b) 62.5 μm and (c) 105 μm. Figure S10. 6. A. van Wijk, C.R. Doerr, Z. Ali, M. Karabiyik, B. Imran Akca, Compact ultra-
Simulated speckle pattern of different taper lengths. (a) 6000 μm (b) 3000 broad-bandwidth cascaded arrayed waveguide gratings. Opt. Express
μm (c) 1500 μm (d) 750 μm. In qualitative simulations of the effect of fiber 28(10), 14618–14626 (2020)
taper length on spectrometer performance, we set the fiber’s core diam- 7. Q.F. Yang et al., Vernier spectrometer using counterpropagating soliton
eter to 105 μm. The smallest diameter of fiber taper was fixed so that the microcombs. Science 363(6430), 965–968 (2019)
change of taper length would change the taper angle of the fiber taper. 8. Z. Yang et al., Single-nanowire spectrometers. Science 365(6457),
We then calculated speckle patterns with different taper lengths. Figure 1017–1020 (2019)
S11. Simulated speckle pattern of tapers with different elongations. (a) 9. S. Grabarnik, R. Wolffenbuttel, A. Emadi, M. Loktev, E. Sokolova, G. Vdovin,
3000mm (b) 4000mm (c) 5000mm. The taper angle of the fiber taper is Planar double-grating microspectrometer. Opt. Express 15(6), 3581–3588
fixed in the simulation, which means that changing the elongation will (2007)
change the diameter of the taper as it transitions into the waist region. 10. D. Thomae, T. Hönle, M. Kraus, V. Bagusat, A. Deparnay, R. Brüning, R. Brun-
Figure S12. The average correlation between obtained at time interval of ner, Compact echelle spectrometer employing a cross-grating. Appl. Opt.
each 5min in 12h. 57(25), 7109–7116 (2018)
11. X. Gan, N. Pervez, I. Kymissis, F. Hatami, D. Englund, A high-resolution
spectrometer based on a compact planar two dimensional photonic
Acknowledgements
crystal cavity array. Appl. Phys. Lett. 100(23), 231104 (2012)
The authors thank Prof Qing Yang for helpful discussion. Y.G. Ma acknowledges
12. S. Zhang, Y. Dong, H. Fu, S.-L. Huang, L. Zhang, A spectral reconstruction
the financial support from the National Natural Science Foundation of China
algorithm of miniature spectrometer based on sparse optimization and
(NSFC) (62222511, 61905213) and Natural Science Foundation of Zhejiang
dictionary learning. Sensors 18(2), 644 (2018)
Province China (LR22F050006). The authors thank Liying Chen from the State
13. B. Redding, S.F. Liew, R. Sarma, H. Cao, Compact spectrometer based on a
Key Laboratory of Modern Optical Instruments, Zhejiang University for help in
disordered photonic chip. Nat. Photonics 7(9), 746–751 (2013)
experiments.
14. W. Hartmann, P. Varytis, H. Gehring, N. Walter, F. Beutel, K. Busch, W. Per-
nice, Broadband spectrometer with single-photon sensitivity exploiting
Author contributions
tailored disorder. Nano Lett. 20(4), 2625–2631 (2020)
YM conceived the idea, designed the experiments, and supervised the
15. M. Chakrabarti, M.L. Jakobsen, S.G. Hanson, Speckle-based spectrometer.
research project. QC, XL and SP conducted the experiments, performed the
Opt. Lett. 40(14), 3264–3267 (2015)
manufacturing and characterization of the fibers and spectrometers. QC, SP,
16. B. Redding, S.M. Popoff, H. Cao, All-fiber spectrometer based on speckle
and XL performed the modeling and simulation. YM, QC, SP, and XL analyzed
pattern reconstruction. Opt. Express 21(5), 6584–6600 (2013)
the data and wrote the paper. All authors contributed to the experiments and
17. B. Redding, H. Cao, Using a multimode fiber as a high-resolution, low-loss
discussion of the manuscript.
spectrometer. Opt. Lett. 37(16), 3384–3386 (2012)
18. N.K. Metzger, R. Spesyvtsev, G.D. Bruce, B. Miller, G.T. Maker, G. Malcolm,
Funding
M. Mazilu, K. Dholakia, Harnessing speckle for a sub-femtometre resolved
National Natural Science Foundation of China (NSFC) (62222511, 61905213),
broadband wavemeter and laser stabilization. Nat. Commun. 8(1), 1–8
Natural Science Foundation of Zhejiang Province China (LR22F050006).
(2017)
19. Z. Xu, Z. Wang, M. Sullivan, D. Brady, S. Foulger, A. Adibi, Multimodal
Availability of data and materials
multiplex spectroscopy using photonic crystals. Opt. Express 11(18),
All data needed to evaluate the conclusions in the paper are present in the
2126–2133 (2003)
paper and supporting information. Additional data related to this paper may
20. N.H. Wan, F. Meng, T. Schröder, R.-J. Shiue, E.H. Chen, D. Englund, High-
be requested from the authors.
resolution optical spectroscopy using multimode interference in a
compact tapered fibre. Nat. Commun. 6(1), 1–6 (2015)
Declarations 21. H. Imani, S. Golmohammadi, A. Rostami, K. Abbasian. Resolution improve-
ment in high-speed fiber-optic spectrometers using photonic crystal
Competing interests fibers. Paper presented at international conference on photonics 2010
The authors declare that they have no conflict of interest. (Langkawi, Malaysia, 05–07 July 2010).
22. B. Redding, M. Alam, M. Seifert, H. Cao, High-resolution and broadband
all-fiber spectrometers. Optica 1(3), 175–180 (2014)
Received: 9 October 2022 Revised: 14 January 2023 Accepted: 14 Febru- 23. B. Redding, S.F. Liew, Y. Bromberg, R. Sarma, H. Cao, Evanescently coupled
ary 2023 multimode spiral spectrometer. Optica 3(9), 956–962 (2016)
24. Y. Kwak, S.M. Park, Z. Ku, A. Urbas, Y.L. Kim, A pearl spectrometer. Nano
Lett. 21(2), 921–930 (2021)
25. L. Zhang, F. Gu, J. Lou, X. Yin, L. Tong, Fast detection of humidity with
a subwavelength-diameter fiber taper coated with gelatin film. Opt.
References Express 16(17), 13349–13353 (2008)
1. A. Li, C. Yao, J. Xia, H. Wang, Q. Cheng, R. Penty, Y. Fainman, S. Pan, 26. X. Guo, L. Tong, Supported microfiber loops for optical sensing. Opt.
Advances in cost-effective integrated spectrometers. Light Sci. Appl. Express 16(19), 14429–14434 (2008)
11(1), 1–18 (2022) 27. C.-L. Hou, Y. Wu, X. Zeng, S. Zhao, Q. Zhou, G. Yang, Novel high sensitivity
2. G.Y. Belay, W. Hoving, A. van der Put, M. Vervaeke, J. Van Erps, H. Thien- accelerometer based on a microfiber loop resonator. Opt. Eng. 49(1),
pont, H. Ottevaere, Miniaturized broadband spectrometer based on a 014402 (2010)
three-segment diffraction grating for spectral tissue sensing. Opt. Lasers 28. M.Z.A. Razak, S.A. Reduan, A.S. Sharbirin, N. Jamaludin, M.Z. Zulkifli, H.
Eng. 134, 106157 (2020) Ahmad, Noncontact optical displacement sensor using an adiabatic
3. H. Yuan, C. Liu, H. Wang, L. Wang, L. Dai, Early pregnancy diagnosis of rab- u-shaped tapered fiber. IEEE Sens. J. 15(10), 5388–5392 (2015)
bits: a non-invasive approach using Vis-NIR spatially resolved spectros- 29. M.C. Frawley, A. Petcu-Colan, V.G. Truong, S. Nic Chormaic, Higher order
copy. Spectroc. Acta Pt. A-Molec. Biomolec. Spectr. 264, 120251 (2022) mode propagation in an optical nanofiber. Opt. Commun. 285(3),
4. B.G. Barthès, E. Kouakoua, M. Clairotte, J. Lallemand, L. Chapuis-Lardy, M. 4648–4654 (2012)
Rabenarivo, S. Roussel, Performance Comparison between a miniaturized 30. S. Ravets, J.E. Hoffman, P.R. Kordell, J.D. Wong-Campos, S.L. Rolston, L.A.
Orozco, Intermodal energy transfer in a tapered optical fiber: optimizing
Cen et al. eLight (2023) 3:9 Page 11 of 11

transmission. J. Opt. Soc. Am. A-Opt. Image Sci. Vis. 30(11), 2361–2371
(2013)
31. S. Mehta and M. Rastegari. MobileViT: light-weight, general-purpose, and
mobile-friendly vision transformer. arXiv preprint arXiv:​2110.​02178.
32. T.A. Birks, Y.W. Li, The shape of fiber tapers. J. Lightwave Technol. 10(4),
432–438 (1992)
33. L.V. Nguyen, D. Hwang, S. Moon, D.S. Moon, Y. Chung, High temperature
fiber sensor with high sensitivity based on core diameter mismatch. Opt.
Express 16(15), 11369–11375 (2008)
34. T. Wang, Y. Li, Y. Meng, Y. Qiu, B. Mao, Study of a fiber spectrometer based
on offset fusion. Appl. Opt. 59(15), 4697–4702 (2020)
35. L. Tong, R.R. Gattass, J.B. Ashcom, S. He, J. Lou, M. Shen, I. Maxwell, E.
Mazur, Subwavelength-diameter silica wires for low-loss optical wave
guiding. Nature 426(6968), 816–819 (2003)
36. Y. Ma, X. Li, H. Yu, L. Tong, Y. Gu, Q. Gong, Direct measurement of propaga-
tion losses in silver nanowires. Opt. Lett. 35(8), 1160–1162 (2010)