Van der waals homojunction micro-spectrometer based on material thickness gradient change and method

Abstract

The application belongs to the technical field of semiconductor devices, and particularly relates to a van der Waals homojunction microspectrometer based on material thickness gradient change, which comprises a Si substrate, a SiO2 substrate, a gate, a BN substrate, a source, a drain, and a van der Waals homojunction. The SiO2 substrate is arranged at the top end of the Si substrate, the gate is arranged at the top end of the SiO2 substrate, and the BN substrate is arranged at the top end of the gate. The van der Waals homojunction is arranged at the top end of the BN substrate, and the thickness of the van der Waals homojunction changes from thick to thin in a gradient manner from one side to the other side. The source is arranged at the thicker side of the van der Waals homojunction, and the drain is arranged at the thinner side of the van der Waals homojunction. By controlling the thickness change of the homojunction, the energy band structure of the homojunction can be effectively regulated, and the efficient separation and transmission of interlayer carriers can be induced. In combination with an optimized spectrum reconstruction algorithm, high-dimensional photoelectric response data collected based on the homojunction can be quickly analyzed, and high-precision reconstruction of incident spectrum can be realized.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor device technology, and specifically relates to a van der Waals homojunction micro spectrometer and method based on material thickness gradient variation. Background Technology

[0002] Spectrometers, as key analytical instruments, are widely used in chemical and biological detection, industrial material characterization, and imaging sensing. However, traditional spectrometers rely on dispersive elements such as gratings or Michelson interferometers, which have complex internal mechanical structures and large sizes, limiting their flexibility and integration in practical applications. Therefore, miniaturization of spectrometers while maintaining high resolution and low cost has become a research focus. Currently, common miniaturization technologies mainly include micro-dispersive elements, narrowband filters, and Fourier transform interferometry systems.

[0003] In recent years, photodetectors based on two-dimensional materials have provided a new approach to the miniaturization of spectrometers. These materials possess unique photoelectric response behavior, and their spectral absorption characteristics can be tuned by bias voltage, enabling a single device to selectively detect multiple wavelengths under different voltages. This effectively replaces the dispersive function of traditional dispersive elements at the electrical level. Combined with computational spectral reconstruction techniques, the distribution of the incident spectrum can be analyzed from the device's response data without the need for an optical dispersive structure. Summary of the Invention

[0004] In view of this, this application provides a van der Waals homojunction micro spectrometer based on material thickness gradient variation, which has the same precision and accuracy but is smaller in size and has a wider range of applications.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] In a first aspect, the present invention provides a van der Waals homojunction micro-spectrometer based on a material thickness gradient, comprising a Si substrate, a SiO2 substrate, a gate electrode, a BN substrate, a source electrode, a drain electrode, and a van der Waals homojunction, wherein the SiO2 substrate is disposed at the top of the Si substrate, the gate electrode is disposed at the top of the SiO2 substrate, and the BN substrate is disposed at the top of the gate electrode; the van der Waals homojunction is disposed at the top of the BN substrate, and the thickness of the van der Waals homojunction varies in a gradient from thick to thin from one side to the other; the source electrode is disposed on the thicker side of the van der Waals homojunction, and the drain electrode is disposed on the thinner side of the van der Waals homojunction.

[0007] Preferably, the van der Waals homojunction is a WSe2 van der Waals homojunction, a MoTe2 van der Waals homojunction, or a MoS2 van der Waals homojunction.

[0008] Preferably, the van der Waals homojunction transitions from a multilayer to a single layer, or from a thicker multilayer to a thinner multilayer.

[0009] Preferably, the thickness of the van der Waals homojunction varies in the range of 0.7 nm to 20 nm.

[0010] Preferably, the source and drain materials are at least one of gold, platinum-gold, and chromium-gold.

[0011] A second aspect of the present invention provides a method for reconstructing spectral information of the aforementioned miniature spectrometer, comprising the steps of:

[0012] S1 Simulation data generation and noise enhancement based on physical model;

[0013] S2 constructs a deep neural network model;

[0014] S3 Model pre-training based on multi-scale loss function;

[0015] S4 is a migration fine-tuning method based on small sample real data.

[0016] After completing the above training and fine-tuning, the fixed model parameters are deployed to the computing terminal. Based on the measured current vectors under different gate voltages and bias voltages, the unknown incident spectrum can be recovered.

[0017] Preferably, in step S1, a large number of hybrid-driven training datasets are generated by constructing a deep neural network and employing Monte Carlo data augmentation strategies and noise models, including:

[0018] K sets of virtual spectral vectors are randomly generated by computer. The virtual spectrum is composed of the superposition of the base spectrum and characteristic peaks, and its mathematical expression is: ,in, The base spectrum is used to simulate background noise or broadband continuous spectral components in the detection environment, and its value is usually set to [0, ... A constant that is randomly distributed within a certain range is used to enhance the model's ability to suppress background interference. The center wavelength is randomly selected. The half-width and height are random. Let M represent the random peak intensity, and M be the number of peaks contained in a single spectrum. Combining the spectral response matrix R with an injected noise model, a simulated input current is generated. : ,in, Represents the ideal photocurrent. Gaussian white noise representing the thermal noise of the analog readout circuit , Shot noise related to light intensity ; and the input current vector and tag spectrum Normalization is performed separately to accelerate neural network convergence.

[0019] Preferably, the deep neural network model in step S2 is a fully connected deep neural network with residual connections, comprising: an input layer, a feature extraction layer, a feature compression layer, and a spectral reconstruction layer;

[0020] The number of nodes in the input layer N in This corresponds to the number of times the bias voltage and gate voltage are changed in the micro spectrometer.

[0021] The feature extraction layer is set to a 2-layer structure, and a residual connection mechanism is introduced between adjacent feature extraction layers. A Dropout regularization layer is added after each feature extraction layer to force the network to learn robust global features.

[0022] The feature compression layer compresses and refines high-dimensional features, removes redundant information, and focuses on the contour features of the spectrum.

[0023] The number of nodes in the spectral reconstruction layer N out , which corresponds to the number of discrete wavelength points in the spectrum to be reconstructed.

[0024] Preferably, in step S3, the hybrid dataset generated in step S1 is divided into a training set and a validation set, which are then input into the network from step S2 for training; a multi-scale combined loss function is used. ,in, This is the mean square error loss; For spectral angle mapping loss, , To balance the hyperparameters of the weights, the AdamW optimizer is used during training, with a preset initial learning rate and a cosine annealing learning rate decay strategy, until the model converges.

[0025] Preferably, step S4 includes: keeping the weight parameters of the feature extraction layer of the training model unchanged, and only fine-tuning the feature compression layer and the spectral reconstruction layer; then using a small amount of collected real data to fine-tune and update the end of the model.

[0026] Compared with the prior art, the beneficial effects of the present invention are:

[0027] This invention provides a miniature spectrometer whose core is a two-dimensional van der Waals homojunction with varying thickness. By controlling the thickness variation, its band structure can be effectively modulated, inducing efficient separation and transport of interlayer carriers. Furthermore, combined with an optimized spectral reconstruction algorithm, high-dimensional photoelectric response data collected based on this homojunction can be rapidly reconstructed with high precision from the incident spectrum. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the structure of a van der Waals homojunction micro spectrometer based on material thickness gradient variation provided in an embodiment of the present invention;

[0029] Figure 2 This is a flowchart of spectral information reconstruction of a van der Waals homojunction micro-spectrometer based on material thickness gradient provided in an embodiment of the present invention;

[0030] Figure 3 This is a response curve of a van der Waals homojunction micro-spectrometer based on material thickness gradient variation provided in an embodiment of the present invention;

[0031] Figure 4 This is a single-peak reconstruction comparison diagram of a van der Waals homojunction micro-spectrometer based on material thickness gradient provided in an embodiment of the present invention;

[0032] Figure 5 This is a broadband reconstruction comparison image of a van der Waals homojunction micro-spectrometer based on material thickness gradient variation provided in an embodiment of the present invention. Detailed Implementation

[0033] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0035] This application provides a van der Waals homojunction micro-spectrometer based on material thickness gradient variation, which will be described in detail in the following embodiments.

[0036] Figure 1This diagram illustrates a van der Waals homojunction micro-spectrometer based on a material thickness gradient according to an embodiment of this application. Specifically, it includes a Si substrate 1, a SiO2 substrate 2, a gate electrode 3, a BN substrate 4, a source electrode 5, a drain electrode 6, and a van der Waals homojunction 7. The SiO2 substrate 2 is disposed at the top of the Si substrate 1, the gate electrode 3 is disposed at the top of the SiO2 substrate 2, and the BN substrate 4 is disposed at the top of the gate electrode 3. The van der Waals homojunction 7 is disposed at the top of the BN substrate 4, and its thickness gradients from thick to thin from one side to the other. The source electrode 5 is disposed on the thicker side of the van der Waals homojunction 7, and the drain electrode 6 is disposed on the thinner side of the van der Waals homojunction 7.

[0037] The source 5 and drain 6 can be made of various metal materials such as gold, platinum / gold, and chromium / gold. It should be noted that a wide variety of metals can be selected, and the metal materials used for the source and drain can be the same or different to meet different device design and performance requirements.

[0038] The thickness of the van der Waals homojunction 7 gradually varies, from multilayer to single layer, or from thicker multilayer to thinner multilayer. The range of material thickness variation can be selected from 0.7nm to 20nm, and can be either evenly spaced or varying at different distances to meet different device design and performance requirements.

[0039] Homojunctions, as a type of epitaxial structure based on interlayer stacking, exhibit highly tunable bandgap characteristics with varying layer number and stacking configuration. By controlling the number of layers and allowing for gradient thickness variations, a continuous transition of the bandgap from a direct bandgap to an indirect bandgap can be achieved, forming a tunable band alignment at the K-valley. This provides an ideal platform for interlayer separation and transport of photogenerated carriers. This structure combines the synergistic advantages of intrinsic material properties and interfacial coupling effects, laying the material foundation for constructing high-performance, low-dimensional optoelectronic devices.

[0040] Furthermore, because the material thickness of the homojunction structure gradually varies in the transverse direction of the device, the band gap of each region changes with its position, thus forming an inherent band gap step within the same two-dimensional material. Under the control of the applied bias voltage and gate voltage, the width of the depletion region of the electrode-material junction region can be controllably shifted along the thickness gradient direction, allowing different band gap regions to participate in carrier generation and separation sequentially under different bias voltages. This adjustable depletion region position effectively changes the device's absorption of photons of different wavelengths and the probability of carrier collection, thereby increasing the nonlinearity of the spectral response under different applied bias voltages and gate voltages. Combined with a reconstruction algorithm, the unknown spectrum can be reconstructed using the nonlinear spectral response curves under different bias voltages and gate voltages. Figure 3The figure illustrates a van der Waals homojunction micro-spectrometer based on material thickness gradient variation in different V values, according to an embodiment of the present invention. ds The response curve below.

[0041] The van der Waals homojunction 7 can be grown using chemical vapor deposition and positioned on top of the BN substrate 4 to adjust the photoelectric response output. For example... Figure 1 As shown, the thickness changes in a gradient from one side to the other, with the thicker side connected to the source 5 on the left and the thinner side connected to the drain 6 on the right. ds V is the voltage difference between the source and drain. g This is the voltage between the gate and the source.

[0042] Van der Waals homojunctions can be made of various two-dimensional materials such as WSe2, MoTe2, and MoS2.

[0043] Further reference Figure 2 The process of spectral information reconstruction of the miniature spectrometer based on van der Waals homojunction with varying material thickness according to the present invention is specifically implemented as follows:

[0044] S1 Simulation data generation and noise enhancement based on physical model;

[0045] S2 constructs a deep neural network model;

[0046] S3 Model pre-training based on multi-scale loss function;

[0047] S4 is a migration fine-tuning method based on small sample real data.

[0048] Specifically, step S1 is based on the physical response matrix. Monte Carlo data augmentation strategies are used to obtain a large training dataset. In a specific embodiment, the specific implementation steps of this process are as follows:

[0049] K sets of virtual spectral vectors (e.g., K=50,000) are randomly generated by computer. To simulate various complex light sources in nature and in the laboratory, a hybrid model of "substrate + multi-peak Gaussian" is used. The virtual spectrum is composed of the substrate spectrum and characteristic peaks superimposed, and its mathematical expression is as follows: ,in, The base spectrum is used to simulate background noise or broadband continuous spectral components in the detection environment, and its value is usually set to [0, ... A constant that is randomly distributed within a certain range is used to enhance the model's ability to suppress background interference. The center wavelength is randomly selected. It is a random half-width (covering narrowband and broadband features). Let M represent the random peak intensity, and M be the number of peaks contained in a single spectrum. Combining the spectral response matrix R with an injected noise model, a simulated input current is generated. : ,in, Representing the ideal photocurrent, its mathematical expression is: , Gaussian white noise representing the thermal noise of the analog readout circuit Johnson noise in the simulated readout circuit; Shot noise related to light intensity This simulates the randomness of photon arrival, with its variance proportional to the light intensity. It also considers the input current vector and the tag spectrum. Normalization is performed separately to accelerate neural network convergence.

[0050] Specifically, in step S2, the constructed deep neural learning network is a fully connected network with residual connections and Dropout mechanism. Its overall network architecture includes an input layer, a feature extraction layer, a feature compression layer, and a spectral reconstruction layer.

[0051] In some embodiments, the number of nodes in the input layer This corresponds precisely to the number of times the bias and gate voltage are changed during scanning in the micro-spectrometer. The feature extraction layer is set to two layers, and a residual connection mechanism is introduced between adjacent feature extraction layers, i.e., the output of the l-th layer... It depends not only on the nonlinear transformation of the previous layer, but also directly on the input of the previous layer: A Dropout regularization layer is added after each hidden layer to randomly disconnect some neurons during training, simulating the failure or instability of individual pixels in the detector and forcing the network to learn robust global features. The feature compression layer compresses and refines high-dimensional features, removing redundant information and focusing on the contour features of the spectrum. The number of nodes in the spectral reconstruction layer... This corresponds to the number of discrete wavelength points in the spectrum to be reconstructed. The spectral reconstruction layer uses the Sigmoid activation function, and the output spectral intensity values ​​are limited to the range [0, 1].

[0052] Specifically, in step S3, the pre-training of the pre-trained model based on the multi-scale loss function is carried out by dividing the above-generated mixed dataset into a training set and a validation set, and inputting them into the fully connected network constructed in step S2 for training.

[0053] In some embodiments, the hybrid dataset generated in step S1 is divided into a training set (90%) and a validation set (10%) in a 90:10 ratio.

[0054] In some embodiments, to balance the overall profile fitting of the spectrum with the accuracy of peak positions, a multi-scale combined loss function is employed: ,in, The mean squared error loss is used to constrain the Euclidean distance in intensity between the reconstructed spectrum and the true spectrum, ensuring accurate energy distribution. For the spectral angle mapping loss, the spectrum is treated as a high-dimensional vector, and the cosine of its included angle is calculated. This term is insensitive to absolute intensity but extremely sensitive to peak position and spectral shape. , To balance the hyperparameters of the weights, in some preferred embodiments, the following settings are made: , This configuration significantly improves the reconstruction accuracy of weaker bands while maintaining the signal-to-noise ratio. During training, the AdamW optimizer is used with a preset initial learning rate, coupled with a cosine annealing learning rate decay strategy, until the model converges. In some specific embodiments, the AdamW optimizer is used during training, with an initial learning rate set to... Combined with a cosine annealing learning rate decay strategy, the learning rate gradually decays to the minimum according to a cosine curve during training. The total training parameters were set to 200, and the model was trained until it converged.

[0055] Specifically, in step S4, although the amount of simulation data is large, it cannot fully simulate the uneven distribution of interface defect state density, differences in contact resistance, and stray light interference caused by the CVD growth process in real devices. In this embodiment of the invention, transfer learning is introduced to perform fine-tuning based on a small sample of real data. The weight parameters of the feature extraction layer of the training model remain unchanged, and only the final spectral reconstruction layer of the network is fine-tuned. Using a small amount of collected real data, the model's final layer is fine-tuned and updated with a small learning rate.

[0056] In some specific embodiments, the fine-tuning parameter settings are performed using 50 sets of real data for secondary training, and the learning rate is lowered to... To prevent the destruction of learned features; the number of iterations is set to 50, and only the terminal parameters are adjusted for fast convergence. Figure 5 The image shows a comparison of broadband reconstruction before and after transfer learning. It can be seen that after transfer learning, the broadband reconstruction result is closer to the true value.

[0057] Specifically, in step S5, the online end-to-end spectral reconstruction and output are achieved by deploying the fixed model parameters to the computing terminal after completing the aforementioned training and fine-tuning. Then, based on the measured current vectors under different gate voltages and bias voltages, the unknown incident spectrum can be recovered.

[0058] The above scheme constructs a spectral mapping model based on a deep neural network to generate a large amount of mixed training data. Subsequently, it uses a Monte Carlo enhancement strategy and a multi-level noise injection mechanism to simulate the real photoelectric response output. Next, it uses a fully connected network with residual connections and dropout mechanisms, combined with a multi-scale loss function, to pre-train and fine-tune the model, thereby establishing a nonlinear mapping from the photocurrent space to the wavelength space, achieving end-to-end high-precision reconstruction of the original spectrum. Figure 4 The figure shows a comparison between the single-peak reconstruction of the van der Waals homojunction micro spectrometer based on material thickness gradient variation of the present invention and existing commercial spectrometers. Compared with existing commercial spectrometers, the micro spectrometer of the present invention is smaller in size and has a wider range of applications while having the same precision and accuracy.

[0059] In summary, this invention provides a miniature spectrometer based on a two-dimensional van der Waals homojunction structure with varying thickness. By controlling the thickness variation, its band structure can be effectively modulated, inducing efficient separation and transport of interlayer carriers. Combined with an optimized spectral reconstruction algorithm, a large-scale mixed training dataset is generated by constructing a deep neural network and employing a Monte Carlo data augmentation strategy. Subsequently, a fully connected network structure with residual connections and Dropout mechanisms is designed, and the model is pre-trained using a multi-scale loss function to establish a nonlinear mapping relationship from photocurrent to spectrum. Finally, a transfer learning mechanism is introduced to fine-tune the network terminals using a small number of real samples, effectively eliminating domain bias between simulation and measured data. This method exhibits excellent nonlinear fitting ability and system adaptability under complex photoelectric response environments, significantly improving the spectral reconstruction accuracy and real-time performance of the miniature spectrometer under noise, drift, and other interference conditions. It enables rapid and high-precision reconstruction of the incident spectrum from high-dimensional photoelectric response data collected based on the homojunction of this invention.

[0060] It should be noted that the above-described embodiments of the methods are presented in a sequential manner for ease of understanding. However, those skilled in the art should understand that this application does not limit the order of execution of the steps. Within the scope of protection of this application, the execution order of some steps can be adjusted according to specific needs, or they can be performed in parallel when conditions permit. Furthermore, those skilled in the art should also understand that the embodiments presented in the specification are preferred examples, and the steps, functional units, or modules included therein are not all the contents that this application must possess.

[0061] In the above embodiments, the content emphasized by each embodiment is different. For parts that are not described in detail in some embodiments, please refer to the relevant content of other embodiments in this specification.

[0062] The preferred embodiments disclosed above are intended to help explain the principles and implementation of the present invention. Optional embodiments do not exhaustively describe all technical details and should not be construed as limiting the scope of protection of the present invention. Obviously, those skilled in the art can make various modifications or equivalent substitutions based on the content of this application without departing from the overall concept of this application. The embodiments cited or described in this application are only used to more clearly illustrate the present invention, and the scope of protection of the present invention should be determined by the claims and their equivalents.

Claims

1. A van der Waals homojunction micro-spectrometer based on material thickness gradient variation, characterized in that, The spectrometer comprises a Si substrate (1), a SiO2 substrate (2), a gate electrode (3), a BN substrate (4), a source electrode (5), a drain electrode (6), and a van der Waals homojunction (7), wherein, The SiO2 substrate (2) is disposed at the top of the Si substrate (1), the gate (3) is disposed at the top of the SiO2 substrate (2), and the BN substrate (4) is disposed at the top of the gate (3); The van der Waals homojunction (7) is disposed on the top of the BN substrate (4), and the thickness of the van der Waals homojunction (7) varies from thick to thin from one side to the other. The source electrode (5) is disposed on the thicker side of the van der Waals homojunction (7), and the drain electrode (6) is disposed on the thinner side of the van der Waals homojunction (7).

2. The miniature spectrometer according to claim 1, characterized in that, The van der Waals homojunction (7) is a WSe2 van der Waals homojunction, a MoTe2 van der Waals homojunction, or a MoS2 van der Waals homojunction.

3. The miniature spectrometer according to claim 1, characterized in that, The van der Waals homojunction (7) is gradually transitioned from a multilayer to a single layer, or from a thicker multilayer to a thinner multilayer.

4. The miniature spectrometer according to claim 1, characterized in that, The thickness of the van der Waals homojunction (7) varies from 0.7 nm to 20 nm.

5. The miniature spectrometer according to claim 1, characterized in that, The source electrode (5) or drain electrode (6) material is at least one of gold, platinum-gold, and chromium-gold.

6. The method for reconstructing spectral information of a miniature spectrometer according to any one of claims 1-5, characterized in that, Including the following steps: S1 Simulation data generation and noise enhancement based on physical model; S2 constructs a deep neural network model; S3 Model pre-training based on multi-scale loss function; S4 is a migration fine-tuning method based on small sample real data.

7. The method according to claim 6, characterized in that, In step S1, a large number of hybrid-driven training datasets are generated by constructing a deep neural network and employing Monte Carlo data augmentation strategies and noise models, including: K sets of virtual spectral vectors are randomly generated by computer. The virtual spectrum is composed of the superposition of the base spectrum and characteristic peaks, and its mathematical expression is: ,in, The base spectrum is used to simulate background noise or broadband continuous spectral components in the detection environment, and its value is usually set to [0, ... A constant randomly distributed within a certain range is used to enhance the model's ability to suppress background disturbances; The center wavelength is randomly selected. The half-width and height are random. For random peak intensity, M is the number of peaks contained in a single spectrum; Then, by combining the spectral response matrix R, a noise model is injected to generate the simulated input current. : ,in, Represents the ideal photocurrent. Gaussian white noise representing the thermal noise of the analog readout circuit , Shot noise related to light intensity ; And the input current vector and tag spectrum Normalization is performed separately to accelerate neural network convergence.

8. The method according to claim 6, characterized in that, The deep neural network in step S2 is a fully connected neural network, comprising: an input layer, a feature extraction layer, a feature compression layer, and a spectral reconstruction layer. The number of nodes in the input layer This corresponds to the number of times the bias voltage and gate voltage are changed in the micro spectrometer. The feature extraction layer is set to 2 layers, and a residual connection mechanism is introduced between adjacent feature extraction layers. A Dropout regularization layer is added after each feature extraction layer to force the network to learn robust global features. The feature compression layer compresses and refines high-dimensional features, removes redundant information, and focuses on the contour features of the spectrum. The number of nodes in the spectral reconstruction layer , which corresponds to the number of discrete wavelength points in the spectrum to be reconstructed.

9. The method according to claim 6, characterized in that, In step S3, the mixed dataset generated in step S1 is divided into a training set and a validation set, which are then input into the network in step S2 for training. Employing a multi-scale combined loss function: ,in, The mean squared error loss is used to constrain the Euclidean distance in intensity between the reconstructed spectrum and the true spectrum. For the spectral angle mapping loss, the spectrum is treated as a high-dimensional vector, and the cosine of its included angle is calculated; , The hyperparameters for balancing the weights, The AdamW optimizer was used during training, with a preset initial learning rate and a cosine annealing learning rate decay strategy, until the model converged.

10. The method according to claim 6, characterized in that, In step S4, the weight parameters of the feature extraction layer of the training model are kept unchanged, and only the feature compression layer and the spectral reconstruction layer are fine-tuned; then, the end of the model is fine-tuned and updated using a small amount of collected real data.

Images