Preparation method of terahertz computational spectrometer based on deep learning reconstruction algorithm

Abstract

The present invention discloses a preparation method of a terahertz computational spectrometer based on a deep learning reconstruction algorithm, comprising the following steps: S1: preparation of an optical filtering structure and a terahertz detector; S2: construction of a training data set; S3: construction of a network model for reconstructing the transmission spectral lines of substances in the terahertz band; S4: training of the network model for reconstructing the transmission spectral lines of substances; S5: transmission spectrum reconstruction. The present invention provides a preparation method of a terahertz computational spectrometer based on a deep learning reconstruction algorithm, which combines micro-nano processing technology and deep learning technology, proposes a nano-scale optical filtering structure, and combines it with a terahertz detector to detect photocurrent; proposes a deep learning reconstruction network algorithm to achieve high-precision measurement of the transmission spectrum in the terahertz band range of 70-200 cm ‑1 -1.

Description

Technical Field

[0001] The present invention relates to the technical field of terahertz spectrometers, and in particular to a preparation method of a terahertz computational spectrometer based on a deep learning reconstruction algorithm. Background Art

[0002] Terahertz spectroscopy technology has the advantages of non-destructive testing, being able to penetrate most non-metallic materials and having stronger transmittance compared to mid-infrared and near-infrared light. In recent years, it has developed rapidly. Currently, with the maturity of high-precision micro-nano processing technology, optical devices can be highly integrated and developed towards miniaturization. The size of the detection device can be reduced to the micron level. Based on this, the transmission spectrum line of the substance to be measured can be reconstructed through a spectral reconstruction algorithm. The computational spectrometer based on this has portability and real-time performance, enabling more practical scenarios for the analysis of terahertz spectra.

[0003] Traditional terahertz spectrometers require large optical systems and detectors to obtain a wider detection bandwidth and higher resolution. Nano-photonics technology refers to directly integrating nano-optical filtering structures on the pixels of photodetectors, which can reduce the physical size of the optical system. However, at the same time, it will inevitably sacrifice some device performance, and the resolution that can be obtained by the integrated nano-scale filtering structure will be reduced. Therefore, software algorithms undertake most of the performance compensation work. In the field of transmission spectrum reconstruction, traditional orthogonal matching pursuit algorithms (OMP), least squares fitting algorithms, BP basis pursuit algorithms, etc. have all been developed and applied relatively fully. However, such algorithms rely on a large number of filter units, require prior knowledge of the filtering response of the spectrometer, and are only applicable to a relatively narrow transmission spectrum range, with complex calculations. For spectrometers with a small number of spectra and a wide wavelength band, the reconstruction performance of these methods is not ideal, which greatly restricts the development of computational spectrometers based on spectral reconstruction.

[0004] Since the emergence of convolutional neural networks, they have been developed and matured in the fields of image and signal processing. Therefore, applying the reconstruction technology based on convolutional neural networks to transmission spectrum reconstruction will make up for the deficiencies of traditional reconstruction methods and is beneficial to the accuracy improvement and wide application of such computational spectrometers. Summary of the Invention

[0005] The purpose of the present invention is to provide a preparation method of a terahertz computational spectrometer based on a deep learning reconstruction algorithm. By combining micro-nano processing technology and deep learning technology, a nano-scale optical filtering structure is proposed, which is combined with a terahertz detector to detect photocurrent; a deep learning reconstruction network algorithm is proposed to achieve high-precision measurement of the transmission spectrum in the terahertz band range of 70 - 200 cm -1 for the transmission spectrum.

[0006] The technical solution adopted in the preparation method of a terahertz computational spectrometer based on a deep learning reconstruction algorithm disclosed by the present invention is as follows:

[0007] A preparation method of a terahertz computational spectrometer based on a deep learning reconstruction algorithm includes the following steps:

[0008] S1: Preparation of an optical filtering structure and a terahertz detector. A 100-nm-thick aluminum metal film is sputtered on high-purity silicon by photolithography, and the designed optical filtering structure is photolithographed on the aluminum film; a terahertz detector is prepared by micro-nano processing technologies such as ion implantation, photolithography, and lift-off.

[0009] S2: Construction of a training data set. Using the terahertz band transmission spectrum of a real substance actually measured by a Fourier transform spectrometer as a label and the measurement data of the computational spectrometer as training data, a network model data set is established.

[0010] S3: Construct a network model for reconstructing the transmission spectral line of a substance in the terahertz band. Based on the EDSR network of the super-resolution algorithm in the image field, the number of network residual blocks is reduced to 8 layers; a transposed convolution layer is introduced as an upsampling and network channel adjustment method, and a linear fully connected layer is added to adjust the final network output, making the final network more suitable for the reconstruction of the substance transmission spectral line and reducing network redundancy.

[0011] S4: Train the network model for reconstructing the transmission spectral line of a substance. The mean square error is used as the loss function of the network described in S3; the adaptive moment estimation method is used as the network optimizer to train and obtain the network weight model.

[0012] S5: Transmission spectrum reconstruction. Using a mercury lamp as a light source, after the light source irradiates the substance to be measured, the current sampling value can be obtained through the optical filtering structure and the detector prepared in S1. The measured data is input into the network trained in S4 for reconstruction to obtain the final substance transmission spectrum.

[0013] As a preferred solution, the specific steps of S1 include:

[0014] S1-1: Prepare the absorption region by ion implantation technology. Different doses of phosphorus ions are implanted on a high-resistance germanium substrate to form germanium-doped phosphorus absorption layers and electrode layers with different concentrations, and the blocking layer is naturally formed by the substrate.

[0015] S1-2: Photolithograph the shape of the electrode region by photolithography and deposit an aluminum film by magnetron sputtering technology, and then use acetone for lift-off to obtain a metal Al electrode.

[0016] S1-3: Sputter a 100-nm-thick aluminum metal film on high-purity silicon by photolithography, photolithograph the designed optical filtering structure on the aluminum film, and prepare a filter for filtering different wavelengths by etching and magnetron deposition coating processes.

[0017] As a preferred solution, the specific steps of S2 include:

[0018] S2-1: Use the actual transmission spectrum data measured by the Fourier transform spectrometer as the comparison label dataset; use the sparse sampling data measured by the computational spectrometer as the training dataset; divide the training set, validation set, and test set according to a ratio of 8:1:1;

[0019] S2-1: These two parts of data will jointly serve as the input of the network model.

[0020] As a preferred solution, the specific steps of S3 include:

[0021] S3-1, Depth expansion layer: The depth expansion layer performs convolution operations through 64 convolutional kernels to expand the 1 input channel to 64 to increase the model parameter capacity. The size of the network filter in the depth expansion layer is 5×1, and the number of network channels is 64;

[0022] S3-2, Residual network layer: The input of the residual network layer is the output data obtained from the depth expansion layer, which consists of a total of 8 cascaded residual sub-networks for extracting the residual features of the input data; each residual sub-network structure contains two branches. One branch includes a convolutional layer, a Relu activation layer, a convolutional layer, and a scaling layer, and the other branch directly connects the previous layer input to the output. The outputs of the two branches are added to obtain the output of the residual sub-network; 8 sub-networks are cascaded to obtain the output of this branch. The size of the convolutional kernel in the residual sub-network is 5×1, and the number of network channels is 64; the output of the residual network layer is the output of the 8 cascaded residual sub-networks;

[0023] S3-3, Upsampling layer: Use transposed convolution as the upsampling method to upsample the output obtained from the residual network layer by a specified multiple as needed; at the same time, adjust the number of channels from 64 to 1, and the size of the transposed convolution kernel is 7×1;

[0024] S3-4, Data adjustment layer: Since the length of the data after the upsampling layer is different from that of the comparison label data, a linear fully connected layer is introduced to adjust the data of the upsampling layer, aiming to make the final output of the network consistent with the length of the comparison label data for convenient comparison of the reconstruction effect.

[0025] As a preferred solution, the specific steps of S4 include:

[0026] S4-1, Define the network loss function: Use the mean squared error as the loss function of the network. The calculation formula of the mean squared error is:

[0027]

[0028] S4-2, where n is the length of the actual transmission spectrum line data, and [x1, x2, ……, x n are each data point of the actual transmission spectrum. are each data point of the network-reconstructed transmission spectrum. N is the number of samples participating in training each time. Define the network training method: set the initial learning rate of the network to 0.0001, and the batch data volume to 8; adopt the adaptive moment estimation method as the network optimizer.

[0029] As a preferred solution, the specific steps of S5 include: using a mercury lamp as the light source. After the light source irradiates the substance to be measured, the current sampling value can be obtained through the filtering structure prepared in S1 and the detector; use the measured and sampled data as the data to be reconstructed, and input it into the trained network in S4 for reconstruction to obtain a high-precision reconstructed substance transmission spectrum.

[0030] The beneficial effect of the preparation method of a terahertz computational spectrometer based on a deep learning reconstruction algorithm disclosed by the present invention is that: the terahertz waves selected by the filtering structure are then used to obtain the sampled response current value through the terahertz detector for algorithm reconstruction, which can greatly reduce the physical size of the spectrometer, providing an idea for the on-chip integration of micro-nano level optical filtering structures and detectors;

[0031] Introduce the residual network idea in deep learning into the field of substance transmission spectrum reconstruction, overcoming the drawback of traditional transmission spectrum methods that require prior knowledge of the response function of the spectrometer filtering structure, and at the same time reducing the requirement for the resolution of the hardware system, and achieving high-precision reconstruction of the transmission spectrum under the conditions of wide-band and small-number sampling. The present invention can achieve the reconstruction of the substance transmission spectrum in the terahertz band of 70 - 200 cm -1 The reconstruction of the substance transmission spectrum has a reconstruction mean square error of 0.00005. Description of the Drawings

[0032] Figure 1 is a flowchart of the preparation method of a terahertz computational spectrometer based on a deep learning reconstruction algorithm of the present invention.

[0033] Figure 2 is a schematic diagram of the optical filtering structure and detector structure of the preparation method of a terahertz computational spectrometer based on a deep learning reconstruction algorithm of the present invention.

[0034] Figure 3 is a schematic diagram of the reconstruction network structure of the preparation method of a terahertz computational spectrometer based on a deep learning reconstruction algorithm of the present invention.

[0035] Figure 4 is a schematic diagram of the peak transmission spectrum line of a partial optical filtering structure of the preparation method of a terahertz computational spectrometer based on a deep learning reconstruction algorithm of the present invention.

[0036] Figure 5 This is the reconstructed network transmission spectrum reconstruction result diagram of a preparation method of a terahertz computational spectrometer based on a deep learning reconstruction algorithm according to the present invention. Specific embodiments

[0037] The present invention will be further described and explained below in conjunction with specific embodiments and the accompanying drawings of the specification:

[0038] Please refer to Figure 1 , a preparation method of a terahertz computational spectrometer based on a deep learning reconstruction algorithm, comprising the following steps:

[0039] Please refer to Figure 2 , S1: Preparation of an optical filtering structure and a terahertz detector. A 100-nm-thick aluminum metal film is sputtered on high-purity silicon by photolithography, and the designed optical filtering structure is lithographed on the aluminum film; a terahertz detector is prepared by micro-nano processing technologies such as ion implantation, photolithography, and lift-off.

[0040] The specific steps of S1 include:

[0041] S1-1: An absorption region is prepared by ion implantation technology. Different doses of phosphorus (P) ions are implanted on a high-resistance germanium (Ge) substrate 1 to form germanium-doped phosphorus (Ge:P) absorption layers 2 and electrode layers 3 with different concentrations, and the blocking layer 4 is naturally formed by the substrate;

[0042] S1-2: The shape of the electrode region is lithographed by photolithography and an aluminum (Al) film is deposited by magnetron sputtering technology, and then acetone is used for lift-off to obtain a metal Al electrode 5;

[0043] S1-3: A 100-nm-thick aluminum (Al) film 6 is sputtered on high-purity silicon (Si) by photolithography, and the designed optical filtering structure 7 is lithographed on the aluminum film, and a filter 8 for filtering different wavelengths is prepared by etching and magnetron deposition coating processes.

[0044] S2: Construction of a training data set. The terahertz band transmission spectrum of a real substance measured by a Fourier transform spectrometer is used as a label, and the measurement data of the computational spectrometer is used as training data to establish a network model data set.

[0045] The specific steps of S2 include:

[0046] S2-1: The actual transmission spectrum data measured by the Fourier transform spectrometer is used as a comparison label data set; the sparse sampling data measured by the computational spectrometer is used as a training data set; the training set, validation set, and test set are divided according to a ratio of 8:1:1;

[0047] S2-1: These two parts of data will jointly serve as the input of the network model.

[0048] Please refer to Figure 3 , S3: Construct a network model for reconstructing the transmission spectrum of substances in the terahertz band. Based on the EDSR network, a super-resolution algorithm in the field of images, reduce the number of network residual blocks to 8 layers; introduce a deconvolution layer as an upsampling and network channel adjustment method, and add a linear fully connected layer to adjust the final network output, making the final network more suitable for the reconstruction of the substance transmission spectrum and reducing network redundancy.

[0049] S3-1, Depth expansion layer: In a convolutional neural network, the performance of the network model is usually improved by increasing the number of convolutional kernels or the number of network layers. The spatial complexity of the network model has a quadratic growth relationship with the number of convolutional kernels. That is to say, more convolutional kernels within the limit of computing resources can obtain more feature parameters. The depth expansion layer of the present invention performs convolution operations through 64 convolutional kernels, expanding 1 input channel to 64 channels to increase the model parameter capacity. The size of the network filter in the depth expansion layer is 5×1, and the number of network channels is 64.

[0050] S3-2, Residual network layer: The role of residual convolution is to learn the residual features of the input data by calculating the difference between the output and the input. During the calculation process, residual convolution only calculates the part with differences, which can avoid network bloat while increasing the number of network layers and speed up the learning speed of the network. The residual network layer consists of a total of 8 cascaded residual sub-networks for extracting residual features. The input of the residual network layer is the output data obtained from the depth expansion layer; each residual sub-network structure contains two branches. One branch includes a convolutional layer, a Relu activation layer, a convolutional layer, and a scaling layer. The outputs of the two branches are added to obtain the output of the residual sub-network; the size of the convolutional kernel in the residual sub-network is 5×1, and the number of network channels is 64. The output of the residual network layer is the output of the 8 cascaded branches of the residual sub-networks.

[0051] S3-3, Upsampling layer: The upsampling layer contains a convolutional layer, a normalization layer, and a deconvolution layer. The deconvolution is used as the upsampling method to upsample the output obtained from the residual network layer by a specified multiple (5 times) as needed; the deconvolution upsampling method has trainable parameters and can take into account the characteristics of the data during upsampling to improve the reconstruction effect. At the same time, the number of channels can be directly adjusted from 64 to 1 through a single convolutional kernel, which is consistent with the input channel of the reconstruction network. Here, the size of the deconvolution kernel is 7×1.

[0052] S3-4, Data adjustment layer: Since the data after passing through the upsampling layer has a different length from the comparison label data, it is not convenient for the calculation and evaluation of the mean square error. A linear fully connected layer is introduced to adjust the output data of the upsampling layer, aiming to make the length of the transmission spectrum data finally output by the reconstruction network consistent with the comparison label data, which is convenient for comparing the reconstruction effects.

[0053] S4: Train the network model for reconstructing the transmission spectrum of the substance, using the mean square error as the loss function of the network described in S3; use the adaptive moment estimation method as the network optimizer to train and obtain the network weight model.

[0054] The specific steps of S4 include:

[0055] S4-1. Define the network loss function: Use the mean square error as the loss function of the network. The calculation formula of the mean square error is:

[0056]

[0057] S4-2. Here, n is the length of the actual transmission spectrum data, [x1, x2, ……, x n is each data of the actual transmission spectrum, is each data of the network-reconstructed transmission spectrum, and N is the number of samples participating in training each time.

[0058] Define the network training method: Set the initial learning rate of the network to 0.0001, and the batch size to 8; use the adaptive moment estimation method (Adam) as the network optimizer. The Adam optimization algorithm adaptively adjusts the learning rate using the first-order moment estimation and maintains the gradient stability by combining the first-order moment and second-order moment estimations. Therefore, good performance can be obtained through a small amount of hyperparameter tuning, and the network convergence speed can be accelerated.

[0059] S5: Transmission spectrum reconstruction. Using a mercury lamp as the light source, after the light source irradiates the substance to be measured, the current sampling value can be obtained through the optical filtering structure and the detector prepared in S1. Input the measured data into the network trained in S4 for reconstruction to obtain the final transmission spectrum of the substance.

[0060] The specific steps of S5 include: Using a mercury lamp as the light source, a schematic diagram of a part of the transmission spectrum filtered by different wavelengths by the optical filtering structure prepared in S1 is as Figure 4 shown. After the mercury lamp irradiates the substance to be measured, the spectrum transmitted through the filtering structure irradiates the detector to obtain the current sampling value. Use the measured weak current signal data as the data to be reconstructed and input it into the network trained in S4 for reconstruction to obtain a high-precision reconstructed transmission spectrum of the substance.

[0061] As Figure 5The figure shows the reconstructed transmission spectrum of the substance to be measured (dark blue paint). By comparison, it can be seen that the two spectral lines basically coincide. After calculation, the mean square error between the reconstructed transmission spectral line of the present invention and the transmission spectral line measured by the Fourier spectrometer is 0.00005. The computational spectrometer proposed by the present invention can achieve the same accuracy as the traditional Fourier spectrometer in measuring the transmission spectrum of unknown substances in the terahertz band under the condition of reducing the detector to the micro-nano scale.

[0062] The present invention provides a preparation method of a terahertz computational spectrometer based on a deep learning reconstruction algorithm. The terahertz wave selected by the filtering structure is then used to obtain the sampled response current value through the terahertz detector for algorithm reconstruction, which can greatly reduce the physical size of the spectrometer and provide an idea for the on-chip integration of micro-nano optical filtering structures and detectors.

[0063] Introducing the residual network idea in deep learning into the field of substance transmission spectrum reconstruction overcomes the drawback that the traditional transmission spectrum method requires prior knowledge of the response function of the spectrometer filtering structure, and at the same time reduces the requirement for the resolution of the hardware system, and realizes high-precision reconstruction of the transmission spectrum under the conditions of wide band and small number of samplings. The present invention can realize the reconstruction of the substance transmission spectrum in the terahertz band of 70 - 200 cm -1 The reconstruction of the substance transmission spectrum, and the mean square error of reconstruction reaches 0.00005.

[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than limiting the protection scope of the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A preparation method of a terahertz computational spectrometer based on a deep learning reconstruction algorithm, characterized in that, It includes the following steps: S1: Preparation of the optical filtering structure and the terahertz detector. A 100-nm-thick aluminum metal film is sputtered on high-purity silicon by photolithography, and the designed optical filtering structure is lithographed on the aluminum film; the terahertz detector is prepared by micro-nano processing technologies such as ion implantation, photolithography, and lift-off. The specific steps of S1 include: S1-1: Preparation of the absorption region by ion implantation technology. Different doses of phosphorus ions are implanted on a high-resistance germanium substrate to form germanium-doped phosphorus absorption layers and electrode layers with different concentrations, and the blocking layer is naturally formed by the substrate. S1-2: Lithograph the shape of the electrode region by photolithography and deposit an aluminum film by magnetron sputtering technology, and then use acetone for lift-off to obtain the metal Al electrode. S1-3: Sputter a 100-nm-thick aluminum metal film on high-purity silicon by photolithography, lithograph the designed optical filtering structure on the aluminum film, and prepare a filter for filtering different wavelengths by etching and magnetron deposition coating processes. S2: Construction of the training data set. The terahertz band transmission spectrum of a real substance actually measured by a Fourier transform spectrometer is used as a label, and the measured data of the spectrometer is calculated as training data to establish a network model data set. S3: Construct a network model for reconstructing the terahertz band substance transmission spectrum line. Based on the EDSR network, a super-resolution algorithm in the image field, reduce the number of network residual blocks to 8 layers; introduce a deconvolution layer as an upsampling and network channel adjustment method, and add a linear fully connected layer to adjust the final network output, making the final network more suitable for the reconstruction of the substance transmission spectrum line and reducing network redundancy. The specific steps of S3 include: S3-1, Depth expansion layer: The depth expansion layer performs convolution operations through 64 convolution kernels to expand the number of input channels from 1 to 64 to increase the model parameter capacity. The size of the network filter in the depth expansion layer is 5×1, and the number of network channels is 64. S3-2, Residual network layer: The input of the residual network layer is the output data obtained from the depth expansion layer, which contains a total of 8 cascaded residual sub-networks for extracting the residual features of the input data. Each residual sub-network structure contains two branches. One branch includes a convolution layer, a Relu activation layer, a convolution layer, and a scaling layer, and the other branch directly connects the input of the previous layer to the output. The outputs of the two branches are added to obtain the output of the residual sub-network. The 8 sub-networks are cascaded to obtain the output of this branch. The size of the convolution kernel of the residual sub-network is 5×1, and the number of network channels is 64. The output of the residual network layer is the output of the 8 cascaded residual sub-networks. S3-3, Upsampling layer: Use deconvolution as the upsampling method to upsample the output obtained from the residual network layer by a specified multiple as needed; at the same time, adjust the number of channels from 64 to 1, and the size of the deconvolution kernel is 7×1. S3-4, Data adjustment layer: Since the length of the data after passing through the upsampling layer is different from that of the comparison label data, a linear fully connected layer is introduced to adjust the data of the upsampling layer, aiming to make the final network output consistent with the length of the comparison label data for facilitating the comparison of the reconstruction effect. S4: Train the network model for reconstructing the transmission spectrum of the substance, using the mean square error as the loss function of the network described in S3; use the adaptive moment estimation method as the network optimizer to train and obtain the network weight model; S5: Transmission spectrum reconstruction. Using a mercury lamp as the light source, after the light source irradiates the substance to be measured, the current sampling value can be obtained through the optical filtering structure and detector prepared in S1. Input the measured data into the network trained in S4 for reconstruction to obtain the final transmission spectrum of the substance.

2. The preparation method of a terahertz computational spectrometer based on a deep learning reconstruction algorithm according to claim 1, characterized in that, The specific steps of S2 include: S2-1: Use the actual transmission spectrum data measured by the Fourier transform spectrometer as the comparison label data set; use the sparse sampling data measured by the computational spectrometer as the training data set; divide the training set, validation set, and test set according to the ratio of 8:1:1; S2-1: These two parts of data will jointly serve as the input to the network model.

3. The preparation method of a terahertz computational spectrometer based on a deep learning reconstruction algorithm according to claim 1, wherein The specific steps of S4 include: S4-1, Define the network loss function: Use the mean square error as the loss function of the network. The calculation formula for the mean square error is: S4-2, where n is the length of the actual transmission spectrum data line, is each data of the actual transmission spectrum, is each data of the network reconstructed transmission spectrum, N is the number of samples participating in training each time, and the network training method is defined as follows: set the initial learning rate of the network to 0.0001, and the batch data volume to 8; adopt the adaptive moment estimation method as the network optimizer.

4. The preparation method of a terahertz computational spectrometer based on a deep learning reconstruction algorithm according to claim 1, characterized in that, The specific steps of S5 include: Using a mercury lamp as the light source, after the light source irradiates the substance to be measured, the current sampling value can be obtained through the filtering structure and detector prepared in S1; use the measured and sampled data as the data to be reconstructed, input it into the network trained in S4 for reconstruction to obtain the high-precision reconstructed transmission spectrum of the substance.

Images