Optoelectronic device for multimode demultiplexing spectrometer and spectral reconstruction method

Abstract

The photoelectric device of multimode demultiplexing spectrometer and spectral reconstruction method relate to mode division (demultiplexing) spectroscopy. The photoelectric device is composed of a plurality of photoelectric detectors in cascade, each of which includes an SOI substrate, a silicon oxide layer, a multimode waveguide with two side branches, a two-dimensional material absorber and a metal electrode. The spectral reconstruction method includes a single-mode spectral reconstruction method, a single-sample multimode reconstruction algorithm and a multi-sample resolution enhancement reconstruction method. Different spectral information is modulated on different waveguide modes, then input into the multimode demultiplexing spectrometer, and a photocurrent is generated on each photoelectric detector. By using the spectral reconstruction method, single-mode spectral reconstruction under single sample, one-time multimode mixed spectral reconstruction and resolution enhancement spectral reconstruction under multi-sample are realized. The multimode demultiplexing spectrometer can simultaneously realize mode demultiplexing function and spectral detection capability, greatly reducing the complexity of the wavelength-division-mode-division hybrid multiplexing system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mode division (demultiplexing) spectroscopy, in particular to an optoelectronic device of a multimode demultiplexing spectrometer and a spectral reconstruction method. BACKGROUND

[0002] With the rapid development of information technology, the amount of network data is showing an explosive growth. In order to cope with the increasingly tense communication resources, various optical multiplexing technologies have emerged, such as time division multiplexing (TDM), wavelength division multiplexing (WDM), mode division multiplexing (MDM), etc. MDM technology modulates information onto mutually orthogonal mode channels, thereby achieving the purpose of synchronous transmission of signals. Spectroscopy is widely used in many scientific research and engineering fields. In recent years, many research results on the miniaturization of spectrometers have been published, which extend the spectral detection to those use scenarios that emphasize real-time and portability, and have relatively low requirements for spectral accuracy and stability, such as real-time biochemical sensing, mobile spectral imaging, and online monitoring of coherent optical communication quality, etc.

[0003] At present, there is no optoelectronic device and method that can simultaneously realize mode demultiplexing and spectral measurement. At the receiving end of the MDM system, a mode demultiplexer is needed to convert the high-order mode of each channel to the base mode, separate each channel, and then use a spectrometer to detect the information on each separated channel. In addition, although some spectral detection devices based on SOI (Silicon-On-Insulator) platform technology significantly improve the degree of integration, an additional photodetector needs to be used outside the device to capture the optical signal. The above shortcomings increase the number of optoelectronic devices and the complexity of the optical circuit at the receiving end, which restricts the miniaturization and integration development of the spectrometer technology in the MDM field.

[0004] In the MDM system, different modes of optical fields are superimposed on each other, showing a nonlinear relationship between the spectrum and the detected physical quantity, which causes great difficulty for the traditional spectral reconstruction algorithm based on linear algebra. In recent years, the deep integration of machine learning and photonics technology provides a new way of thinking and method for designing and solving nonlinear problems such as multimode spectral reconstruction. Benefiting from its end-to-end learning characteristics, the machine learning model usually does not need professional prior knowledge and can learn the nonlinear mapping relationship between the input and output physical quantities of a complex physical system. And machine learning is data-driven, which is a powerful technology that is transferable, updatable and iterative.

[0005] In conclusion, it is essential to invent a device and method that can simultaneously possess mode demultiplexing and spectral reconstruction capabilities, and introducing machine learning algorithms during multimode spectral reconstruction is undoubtedly an effective technical approach. Summary of the Invention

[0006] One object of the present invention is to provide an optoelectronic device for a multimode demultiplexing spectrometer that can simultaneously realize mode demultiplexing and spectral reconstruction functions, in order to address the above-mentioned shortcomings of the prior art.

[0007] Another object of the present invention is to provide an application of the optoelectronic device of the aforementioned multimode demultiplexing spectrometer, which can realize three types of on-chip integrated multimode demultiplexing spectral reconstruction, including single-mode spectral reconstruction, single-shot multimode reconstruction (SMR), and multi-shot resolution-enhanced reconstruction (MRR).

[0008] The optoelectronic devices of the multimode demultiplexing spectrometer consist of multiple photodetectors U1 to U2. m Cascaded configuration;

[0009] The photodetector is fabricated using SOI technology and, from bottom to top, comprises an SOI substrate, a silicon oxide layer, a multimode waveguide with branches on both sides, a two-dimensional material absorber, and a metal electrode.

[0010] The SOI substrate includes a silicon substrate and a buried oxide layer disposed on the silicon substrate;

[0011] The silicon oxide layer is located on the SOI substrate;

[0012] The multimode waveguide with branches on both sides is located on the SOI substrate and buried inside the silicon oxide layer;

[0013] The two-dimensional material absorber is located on the silicon oxide layer;

[0014] The metal electrode is located on the two-dimensional material absorber.

[0015] Furthermore, the two-dimensional photoelectric materials that can be selected for the two-dimensional material absorber include, but are not limited to, single-layer graphene, black phosphorus, and transition metal sulfides; the multimode waveguide with branches on both sides consists of a main multimode waveguide and a number of branch waveguides distributed on both sides.

[0016] The photodetector's dimensions include length a and width b; the silicon oxide layer's dimensions include thickness c; the multimode waveguide with side branches' design parameters include the number of branches, thickness t, and main multimode waveguide width w, where each branch has a length s, a width h, and a spacing g between adjacent branches; the spacing between adjacent photodetectors on the optoelectronic device of the multimode demultiplexing spectrometer is d.

[0017] The optical signal transmission sequence of the photoelectric devices in the multimode demultiplexing spectrometer is from photodetector U1 to U... m The input terminal is on the photodetector U1 side, and the output terminal is on the photodetector U... m On the other hand, the photocurrent signal is generated and measured on each photodetector, and then transmitted to the host computer after passing through a current amplifier and an analog-to-digital converter.

[0018] The optoelectronic devices of the multimode demultiplexing spectrometer can be used in spectral reconstruction. The spectral reconstruction includes single-mode spectral reconstruction, single-sample multimode reconstruction (SMR), and multi-sample resolution-enhanced reconstruction (MRR).

[0019] The specific steps for single-mode spectral reconstruction using the optoelectronic devices of the multimode demultiplexing spectrometer include:

[0020] (1) Initialization: Scan the normalized photocurrent response values ​​of each wavelength sampling point in different modes within the working bandwidth and construct the photocurrent response matrix; this step is only performed once during the initialization phase.

[0021] (2) Constructing a single-mode spectrum: Modulate the spectrum onto one of the waveguide modes allowed by the multimode demultiplexing spectrometer and couple it into the input of the multimode demultiplexing spectrometer;

[0022] (3) Detecting the photocurrent generated by the single-mode spectrum: The current value on each photodetector is sampled by a current amplifier and an analog-to-digital converter and transmitted to the host computer;

[0023] (4) Construct the objective function: The objective spectral function is approximately transformed into an orthogonal basis expansion, and the orthogonal basis coefficients, photocurrent response matrix and measured current value are used to construct the optimal objective function;

[0024] (5) Spectral reconstruction: Select an appropriate optimization method, calculate the orthogonal basis coefficients, and then substitute them into the orthogonal basis expansion to obtain the reconstructed spectrum.

[0025] The specific steps for single-sample multimode reconstruction (SMR) using the optoelectronic devices of the aforementioned multimode demultiplexing spectrometer include:

[0026] (1) Constructing a multimode mixed spectrum: Modulate target spectrum 1 to target spectrum k onto mode 1 to mode k respectively, superimpose k modes to form a multimode mixed spectrum, and couple it into the input terminal of the multimode demultiplexing spectrometer;

[0027] (2) Detecting the photocurrent generated by multimode mixed spectrum: The current value on each photodetector is sampled by a current amplifier and an analog-to-digital converter and transmitted to the host computer;

[0028] (3) Constructing a dataset: Repeat steps (1) and (2) to continuously change the target spectrum 1 to the target spectrum k to construct a dataset for machine learning, including a training set and a test set; in each dataset, the input of each sample is the current value corresponding to the target multimode mixed spectrum, and the label is the k target spectra before mixing, i.e., target spectrum 1 to target spectrum k;

[0029] (4) Training a machine learning model for single-sample multi-mode reconstruction: Using the dataset constructed in step (3), train and test a supervised machine learning model to obtain the trained machine learning model.

[0030] (5) Measured single-sample multi-mode reconstruction capability: Input any multi-mode mixed spectrum from the input end of the multi-mode demultiplexing spectrometer to obtain the reconstructed k target spectra.

[0031] The specific steps for performing multi-sample resolution enhancement reconstruction (MRR) using the optoelectronic devices of the aforementioned multimode demultiplexing spectrometer include:

[0032] (1) Modulate the spectrum to multiple modes in time division: Modulate the target spectrum to modes 1 to k in time slots 1 to k respectively, and input it from the input terminal of the multimode demultiplexing spectrometer in time slots 1 to k respectively;

[0033] (2) Time-division detection of photocurrent: In each time slot (i.e., time slot 1 to time slot k), the current value on each photodetector is sampled by a current amplifier and an analog-to-digital converter, and the current value of each detector in each time slot is transmitted to the host computer.

[0034] (3) Constructing a dataset: Repeat steps (1) and (2) to continuously change the target spectrum and construct a dataset for machine learning, including a training set and a test set; in each dataset, the input of each sample is the current value of each detector corresponding to the target spectrum in k time slots, and the label is the target spectrum;

[0035] (4) Training a machine learning model for single-sample multi-mode reconstruction: Using the dataset constructed in step (3), train and test a supervised machine learning model to obtain the trained machine learning model.

[0036] (5) Measured multi-sample resolution enhancement and reconstruction capability: Input any target spectrum from the input end of the multi-mode demultiplexing spectrometer to obtain the reconstructed target spectrum with enhanced resolution.

[0037] The multimode demultiplexing spectrometer described above utilizes the dispersion of the structure and the photoelectric conversion function of the detector to achieve integrated mode demultiplexing, spectral dispersion, and detection. This method enables single-mode spectral reconstruction, single-sample multimode reconstruction, and multi-sample resolution-enhanced reconstruction. It significantly reduces the complexity of wavelength division multiplexing (WDM)-mode division multiplexing hybrid multiplexing systems, and has positive implications for the miniaturization and on-chip integration of spectrometers in WDM systems. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the structure of the multimode demultiplexing spectrometer in Embodiment 1 of the present invention.

[0039] Figure 2 This is a schematic diagram of the photodetector in Embodiment 1 of the present invention.

[0040] Figure 3 This is a left view of the structure of the photodetector in Embodiment 1 of the present invention.

[0041] Figure 4 This is a front view of the photodetector structure of Embodiment 1 of the present invention.

[0042] Figure 5 This is a top view of the photodetector structure of Embodiment 1 of the present invention.

[0043] Figure 6 This is a schematic diagram of the single-mode spectral reconstruction method in Embodiment 2 of the present invention.

[0044] Figure 7 This is a comparison chart of the reconstruction results of the single-mode spectral reconstruction method in Embodiment 2 of the present invention.

[0045] Figure 8 This is a schematic diagram illustrating the training / testing single-sample multimodal reconstruction (SMR) method of Embodiment 3 of the present invention.

[0046] Figure 9 This is a comparison chart of the reconstruction results of the training / testing single-sample multimodal reconstruction (SMR) method in Embodiment 3 of the present invention.

[0047] Figure 10 This is a schematic diagram illustrating the training / testing process of the multi-sample resolution augmentation reconstruction (MRR) method in Embodiment 4 of the present invention.

[0048] Figure 11 This is a comparison chart of the reconstruction results of the training / testing multi-sample resolution augmentation reconstruction (MRR) method in Embodiment 4 of the present invention.

[0049] The markings in the diagram are as follows:

[0050] 1: SOI substrate; 2: Silicon oxide layer; 3: Multimode waveguide with branches on both sides; 4: Two-dimensional material absorber; 5: Metal electrode; U1-U m Photodetector 1 to photodetector m; I1-I m :U1 to U m The corresponding currents 1 to m are: a: the length of the photodetector; b: the width of the photodetector; c: the thickness of the silicon oxide layer 2; w, t: the width and thickness of the multimode waveguide 3 with branches on both sides; s, h, g: the length, width and spacing between each branch of the multimode waveguide 3 with branches on both sides and the adjacent branch; d: the spacing between adjacent photodetectors. Detailed Implementation

[0051] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. The drawings are not strictly to scale, and the same reference numerals denote the same components. Furthermore, to describe that A and B are directly adjacent and A is on top of B, the expression "A is on top of B" will be used. Directional terms in the description, such as "left" and "right," only represent the orientation or positional relationship in the drawings and do not imply that the device or element referred to must have a specific orientation or position.

[0052] The specific embodiments described are provided to better illustrate one way of the present invention, and do not represent the best implementation method, nor do they constitute a limitation on the content and scope of protection of the present invention.

[0053] Example 1

[0054] Reference Figure 1 The overall structure of the optoelectronic device of the multimode demultiplexing spectrometer described in Embodiment 1 of the present invention includes m cascaded photodetectors, namely U1-U2. m The number of photodetectors is m, and the spacing between adjacent photodetectors is d. The optical signal is input from the left side of the photoelectric device, transmitted along the x-direction, and output from the right side. U1-U m The induced photocurrent on the screen is measured and then uploaded to the host computer as data for subsequent spectral reconstruction.

[0055] Reference Figures 2-5The detailed structure of the photodetector described in Embodiment 1 of this invention includes an SOI substrate 1, a silicon oxide layer 2, a multimode waveguide 3 with branches on both sides, a two-dimensional material absorber 4, and a metal electrode 5. The SOI substrate 1 includes a silicon substrate and a buried oxide layer disposed on the silicon substrate; the silicon oxide layer 2 is located on the SOI substrate 1; the multimode waveguide 3 with branches on both sides, made of silicon, is located on the SOI substrate 1 and buried inside the silicon oxide layer 2; the two-dimensional material absorber 4, made of monolayer graphene, is located on the silicon oxide layer 2; and the metal electrode 5, made of gold, is located on the two-dimensional material absorber 4. For a single photodetector, the optical signal is input from one side of the multimode waveguide 3 with branches on both sides and output from the other side. Most of the signal is localized inside the multimode waveguide 3 with branches on both sides, forming an electric field. Through evanescent field coupling, the formed electric field is partially absorbed by the two-dimensional material absorber 4, generating a response current in the two-dimensional material absorber 4. The response current is led out through the metal electrode 5 and, after sampling, is used as measurement data for spectral reconstruction. Since different modes of light exhibit different characteristics in terms of effective refractive index and refractive index distribution in a medium, different response characteristics can be generated on different photodetectors by changing the input mode and the size of the branch waveguide.

[0056] This embodiment 1 provides a set of parameter design schemes. The length of the photodetector is a = 1.5 μm, the width of the photodetector is b = 12 μm, the thickness of the multimode waveguide 3 with two side branches is t = 220 nm, the thickness of the silicon oxide layer 2 is c = 230 nm, the width of the multimode waveguide 3 with two side branches is w = 3 μm, the number of photodetectors is m = 25, and the spacing between adjacent photodetectors is d = 1 μm. The branches of the multimode waveguide 3 with two side branches are symmetrically distributed along the multimode main waveguide, and the dimensions of each branch waveguide are randomly generated, randomly selected values ​​in the parameter space (s,h,g). Among them, the length s of each branch of the multimode waveguide 3 with two side branches is ∈ [180,390] ​​nm, the width h of each branch of the multimode waveguide 3 with two side branches is ∈ [180,4200] nm, and the spacing g between adjacent branches is ∈ [180,300] nm. Furthermore, the operating bandwidth of the spectrometer was set to 1500–1600 nm, and modes 1 to k were set to TE1–TE4. Through simulation and experimentation, the spectral response of the photodetector was obtained. To evaluate the diversity of the spectral response characteristics of the photodetector, the condition number was used as an evaluation metric. Generally, the smaller the condition number of a matrix, the higher its instability, which translates to greater diversity in the spectral response, and usually results in a higher reconstructed spectral resolution. Therefore, among the multiple sets of design parameters obtained using the above method, the set with the smallest condition number for the spectral response matrix should be prioritized. This parameter design scheme has been practically verified and is a feasible technical approach.

[0057] The optoelectronic devices of the multimode demultiplexing spectrometer described in Example 1 can be fabricated on a standard 220 nm thick SOI wafer, corresponding to SOI substrate 1. A multimode waveguide 3 with branches on both sides is fabricated using electron-beam lithography (EBL) and inductively coupled plasma (ICP) processes. A silicon oxide layer is then backfilled, and its top layer is polished. Subsequently, a silicon oxide layer of approximately 10 nm thickness is deposited on top to fabricate silicon oxide layer 2. A monolayer graphene sheet is then transferred onto silicon oxide layer 2 and etched using EBL and ICP processes to fabricate a two-dimensional material absorber 4. Finally, a 50 nm thick gold layer is deposited on the two-dimensional material absorber 4 to form a metal electrode 5.

[0058] Example 2

[0059] Reference Figure 6This embodiment 2 is an implementation of the single-mode spectral reconstruction method, consisting of five steps. S1: Initialization: Scan the normalized photocurrent response values ​​of each wavelength sampling point within the working bandwidth under different modes to construct the photocurrent response matrix R(λ). This step is executed only once during the initialization phase; S2: Constructing the single-mode spectrum: Modulate the target spectrum F(λ) onto one of the waveguide modes 1 to k allowed by the multimode demultiplexing spectrometer, and couple it into the input of the multimode demultiplexing spectrometer; S3: Detecting the photocurrent I1-I generated by the single-mode spectrum. m S4: Sample the current value on each photodetector and transmit it to the host computer; S5: Construct the objective function, approximate the target spectral function into the orthogonal basis expansion, and construct the optimal objective function with the orthogonal basis coefficients, photocurrent response matrix and measured current value; S6: Spectral reconstruction, select an appropriate optimization method, calculate the orthogonal basis coefficients, and then substitute them into the orthogonal basis expansion to obtain the reconstructed spectrum.

[0060] In this embodiment 2, the response matrix R of the i-th photodetector for the k-th mode i,k (λ)(i=1,2,...,m) is represented as:

[0061]

[0062] in, Let η be the electric field of the i-th photodetector corresponding to the k-th mode in the two-dimensional material absorber 4, and let η be the conversion coefficient from the light field to the photocurrent in the two-dimensional material absorber 4. The photocurrent of the i-th photodetector modulated with the target spectrum in the k-th mode is expressed as:

[0063]

[0064] Among them, F k (λ) represents the k-th target spectral function modulated onto the k-th mode, [λ min ,λ min [] indicates the operating wavelength range.

[0065] In this embodiment 2, the Gaussian expansion is used to approximate the target spectral function, F. k (λ) is derived from the expansion Approximate representation, where Let σ be the Gaussian width. The peak value of Gaussian peaks is uniformly distributed within the working bandwidth, α j,k Let Górski weights be the values ​​for the k-th target spectrum. An optimization algorithm based on Tikhonov regularization is used as the spectral reconstruction method. The objective function expression for the k-th target spectrum function is:

[0066]

[0067] Among them, A k It is the m×n matrix corresponding to the k-th target spectral function and α k =[α 1,k ,α 2,k ,…, n,k ] T Let c be the vector form of the Gaussian weights corresponding to the k-th target spectral function. k =[I 1,k ,I 2,k ,...,I m,k ] T I corresponding to the k-th target spectral function 1,k -I m,k The vector form of γ is the Tikhonov regularization factor. The objective function (3) is iteratively calculated until it converges to a minimum value, and then the solved α is... j,k Substituting into Gorsky expansion Finally, the target spectrum F reconstructed by single-mode is obtained. k (λ).

[0068] Furthermore, through practical testing, using the same design parameters as Example 1, Example 2 can reconstruct single-mode spectra modulated in any of the TE1-TE4 modes with a bandwidth of 1500–1600 nm. Its spectral reconstruction resolution (Full Width at Half-Maximum) is approximately 7 nm. The reconstruction results are shown in [Figure 1]. Figure 7 .

[0069] Example 3

[0070] Reference Figure 8 This embodiment 3 is an implementation of the single-sample multimodal reconstruction (SMR) method, which can be divided into five steps:

[0071] (1) Construct a multimode mixed spectrum, modulate the target spectrum 1 to the target spectrum k onto the modes 1 to k respectively, and then superimpose the k modes to form a multimode mixed spectrum and couple it into the input end of the spectrometer.

[0072] (2) Detect the photocurrent generated by the multimode mixed spectrum, sample the current value on each photodetector through a current amplifier and an analog-to-digital converter, and transmit it to the host computer.

[0073] (3) Construct a dataset by repeating steps (1) and (2) and continuously changing the target spectrum 1 to the target spectrum k to construct a dataset for machine learning, including a training set and a test set. In each dataset, the input of each sample is the current value corresponding to the target multimode mixed spectrum, and the label is the k target spectra before mixing, i.e., target spectrum 1 to target spectrum k.

[0074] (4) Train the machine learning model for single-sample multi-mode reconstruction. Using the dataset constructed in step (3), train and test the supervised machine learning model to obtain the trained machine learning model.

[0075] (5) Measured single-sample multi-mode reconstruction capability. Input any multi-mode mixed spectrum from the spectrometer input terminal to obtain the reconstructed k target spectra.

[0076] Furthermore, the machine learning model in Example 3 employs an architecture of cascaded fully-connected layers within a multi-layered convolutional neural network (CNN). Here, three CNN layers are used, with a ReLU non-linear activation function unit immediately following each CNN layer. After the three CNN layers, three fully-connected layers are cascaded, and the output of the last fully-connected layer is a vector of k reconstructed target spectra, denoted as [F1(λ1), F1(λ2), ..., F...]. k (λ n Finally, this vector is decomposed into k vectors [F1(λ1), F1(λ2), ..., F1(λ2)]. n )] to [F k (λ1),F k (λ2),...,F k (λ n )], which are the reconstructed spectra 1 to reconstructed spectra k.

[0077] Furthermore, through practical testing, when using the same design parameters as in Example 1, each input sample of the machine learning model in Example 3 is a 5×5 pixel image, where each pixel corresponds to I1-I m A value, where m = 25. TE1-TE4 are selected as the mode channels for the modulation spectrum. Samples corresponding to random spectra are generated, and 10,000 samples are split into training, validation, and test sets in an 8:1:1 ratio to train and test the machine learning model. Example 3 can simultaneously reconstruct a multimode mixed spectrum modulated at TE1-TE4 with a bandwidth of 1500–1600 nm, with a spectral reconstruction resolution of approximately 15 nm FWHM. The reconstruction results are shown in [reference needed]. Figure 9 .

[0078] Example 4

[0079] Reference Figure 10 This embodiment 4 is an implementation of the multi-sample resolution enhanced reconstruction (MRR) method, which includes the following five steps:

[0080] (1) Modulate the spectrum to multiple modes in time division. Modulate the target spectrum to modes 1 to k in time slots 1 to k respectively, and input it from the input terminal of the spectrometer in time slots 1 to k respectively.

[0081] (2) Time-division detection of photocurrent. In each time slot (i.e., time slot 1 to time slot k), the current value on each photodetector is sampled by a current amplifier and an analog-to-digital converter, and the current value of each detector in each time slot is transmitted to the host computer.

[0082] (3) Constructing the dataset. Repeat steps (1) and (2), continuously changing the target spectrum, to construct the dataset used for machine learning, including a training set and a test set. In each dataset, the input for each sample is the current value of each detector corresponding to the target spectrum in k time slots, and the label is the target spectrum.

[0083] (4) Train the machine learning model for single-sample multi-mode reconstruction. Using the dataset constructed in step (3), train and test the supervised machine learning model to obtain the trained machine learning model.

[0084] (5) Measured multi-sample resolution enhancement and reconstruction capability. Input any target spectrum from the spectrometer input terminal to obtain the reconstructed target spectrum with enhanced resolution.

[0085] Furthermore, the machine learning model in Example 4 employs an architecture of cascaded fully-connected layers within a multi-layered convolutional neural network (CNN). Here, two CNN layers are used, followed immediately by a ReLU non-linear activation function unit. After the two CNN layers, three fully-connected layers are cascaded, and the output of the last fully-connected layer is the vector form of the resolution-enhanced reconstructed spectrum, represented as [F(λ1), F(λ2), ..., F(λ...]. n )).

[0086] Using the same design parameters as in Example 1, the target spectrum is modulated to TE1-TE4 in time slots 1 to k, respectively, forming a k-channel, 5×5 pixel map, which serves as an input sample for the machine learning model in Example 4. The pixel of the k-th channel corresponds to I1-I of the k-th mode. mThe values ​​are given, where m = 25 and k = 4. Practical testing showed that using the same design parameters as in Example 1, and selecting TE1-TE4 as the mode channels for the modulation spectrum, generated samples corresponding to random spectra. These 10,000 samples were split into training, validation, and test sets in an 8:1:1 ratio to train and test the machine learning model. Example 4 can reconstruct a spectrum with a bandwidth of 1500–1600 nm, with a spectral reconstruction resolution of approximately 3 nm FWHM, representing a performance improvement of approximately 1.3 times compared to Example 3. The reconstruction results are shown in [Figure 4]. Figure 11 .

Claims

1. An optoelectronic device for a multimode demultiplexing spectrometer, characterized in that The plurality of photodetectors U1 to U m are connected in cascade. The photoelectric detector is prepared by an SOI process and comprises, from bottom to top, an SOI substrate, a silicon oxide layer, a multimode waveguide with two side branches, a two-dimensional material absorber, and a metal electrode. The SOI substrate comprises a silicon substrate and a buried oxygen layer arranged on the silicon substrate. The silicon oxide layer is located on the SOI substrate. The multimode waveguide with two side branches is located on the SOI substrate and is buried in the silicon oxide layer. The two-dimensional material absorber is located on the silicon oxide layer. The metal electrode is located on the two-dimensional material absorber. The optical signal transmission sequence of the optoelectronic device of the multi-mode demultiplexing spectrometer is from the photodetector U1 to U m , the input end is on the side of the photodetector U1, and the output end is on the side of the photodetector U m ; the photoelectric current signal is generated and measured on each photodetector, transmitted to the upper computer after passing through the current amplifier and the analog-to-digital converter.

2. The optoelectronic device of claim 1, wherein The two-dimensional photoelectric material selected for the two-dimensional material absorber includes monolayer graphene, black phosphorus, and transition metal sulfide.

3. Application of the photoelectric device of the multimode demultiplexing spectrometer of claim 1 in spectral reconstruction, wherein the spectral reconstruction comprises single-mode spectral reconstruction, single-sample multimode reconstruction, and multi-sample resolution enhancement reconstruction.

4. The use according to claim 3, characterized in that The specific steps of the single-mode spectral reconstruction include: (1) initialization: scanning the normalized photocurrent response value of each wavelength sampling point in the working bandwidth under different modes to construct a photocurrent response matrix; this step is only performed once in the initialization stage; (2) constructing a single-mode spectrum: modulating the spectrum to one of the waveguide modes allowed by the multimode demultiplexing spectrometer and coupling it into the input end of the multimode demultiplexing spectrometer; (3) detecting the photocurrent generated by the single-mode spectrum: sampling the current value on each photoelectric detector through a current amplifier and an analog-to-digital converter and transmitting it to the upper computer; (4) constructing an objective function: converting the target spectrum function into an orthogonal basis expansion, constructing an optimal objective function from the orthogonal basis coefficients, the photocurrent response matrix, and the measured current value; (5) spectral reconstruction: selecting a suitable optimization method, calculating the orthogonal basis coefficients, and substituting them into the orthogonal basis expansion to obtain the reconstructed spectrum.

5. The use according to claim 3, wherein The specific steps of the single-sample multimode reconstruction include: (1) Constructing multi-mode mixed spectrum: superimposing target spectrum 1 to target spectrum k modulated to mode 1 to mode k above into k mode to form a multi-mode mixed spectrum, and coupled into the input end of a multi-mode demultiplexing spectrometer; (2) detecting the photocurrent generated by the multimode mixed spectrum: sampling the current value on each photoelectric detector through a current amplifier and an analog-to-digital converter and transmitting it to the upper computer; (3) Constructing the dataset: Repeat steps (1) and (2), continuously changing the target spectrum 1 to the target spectrum 2. k The dataset used for machine learning is constructed, including a training set and a test set; in each dataset, the input of each sample is the current value corresponding to the target multimode mixed spectrum, and the label is the value before mixing. k The target spectra, namely target spectrum 1 to target spectrum 2. k ; (4) training a machine learning model for single-sample multimode reconstruction: using the data set constructed in step (3), training and testing a supervised machine learning model to obtain the trained machine learning model; (5) Real-time single sample multi-mode reconstruction capability: any multi-mode mixed spectrum is input from the input end of the multi-mode demultiplexing spectrometer, and the target spectrum after reconstruction is obtained k .

6. The use according to claim 3, characterized in that The specific steps of the multi-sample resolution enhancement reconstruction include: (1) Time division multiplexing of the spectrum into multiple modes: The target spectrum is input from the input end of the multi-mode demultiplexing spectrometer in time slot 1 to time slot k respectively, and is modulated into mode 1 to mode k upper, and is input from the input end of the multi-mode demultiplexing spectrometer in time slot 1 to time slot k respectively. (2) detecting the photocurrent in each time slot: sampling the current value on each photoelectric detector through a current amplifier and an analog-to-digital converter in each time slot and transmitting the current value of each detector in each time slot to the upper computer; (3) Constructing dataset: repeat step (1) and step (2), constantly change the target spectrum, construct the dataset used by machine learning, including training set and test set; in each dataset, the input of each sample is the current value corresponding to each detector under the target spectrum in each time slot, and the label is the target spectrum; k ​ (4) training a machine learning model for single-sample multimode reconstruction: using the data set constructed in step (3), training and testing a supervised machine learning model to obtain the trained machine learning model; (5) measuring the resolution enhancement reconstruction capability of the multi-sample: inputting any target spectrum from the input end of the multimode demultiplexing spectrometer to obtain the reconstructed resolution-enhanced target spectrum.

Images