Spectrum envelope designable optical frequency comb generation method based on machine learning

Abstract

According to the optical frequency comb generation method with the designable spectrum envelope based on machine learning, interpolation is carried out on collected optical frequency comb spectrum data, so that more accurate peak value envelope and valley value envelope data of the optical frequency comb spectrum are extracted. The peak envelope serves as an input data set of a machine learning system, the valley envelope serves as an output data set of the machine learning system, the machine learning system is trained, and the corresponding valley envelope can be predicted through the peak envelope. According to the method, machine learning is applied to reconstruction from a frequency spectrum envelope to a complete optical frequency comb and determination of various modulation parameters of an initial optical frequency comb. And finally, performing specific modulation on the initial optical frequency comb and inputting the modulated initial optical frequency comb into a nonlinear optical fiber, and by utilizing the spectrum broadening characteristic of the nonlinear optical fiber, the obtained spectrum envelope type does not depend on a fixed mathematical function model, but supports multiple forms of envelope definition modes, including definition based on a control point interpolation mode.

Description

Technical Field

[0001] This invention relates to a technology in the field of optical frequency combs, specifically a method for using machine learning to reconstruct a complete optical frequency comb spectrum from a preset spectral envelope, and further generate an optical frequency comb with a desired spectral envelope. Background Technology

[0002] An optical frequency comb consists of a series of equidistant spectral lines with stable phase relationships. The main parameters defining an optical frequency comb are its center frequency, repetition rate, and bandwidth. It has wide applications in optical communication, spectroscopy, and other fields. The curve formed by connecting and normalizing the peak values ​​of each comb tooth in an optical frequency comb is called the spectral envelope. Summary of the Invention

[0003] This invention addresses the shortcomings of existing technologies, such as the reliance on structural parameter analysis of the nonlinear system itself in machine learning processes, which makes it difficult to extend to other nonlinear systems; the prediction of the generated optical frequency comb only includes a few characteristic values, or requires setting all spectral information of the desired optical frequency comb before prediction, and cannot predict and analyze only the spectral envelope of the generated optical frequency comb. This invention proposes a machine learning-based method for generating optical frequency combs with a designable spectral envelope. Machine learning is applied to the reconstruction from the spectral envelope to the complete optical frequency comb and the determination of various modulation parameters of the initial optical frequency comb. Finally, by specifically modulating the initial optical frequency comb and inputting it into a nonlinear optical fiber, the spectral broadening characteristics of the nonlinear optical fiber are utilized to obtain an optical frequency comb with the desired spectral envelope.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a machine learning-based method for generating optical frequency combs with spectral envelope designability, comprising:

[0006] Step 1, Data Processing Stage: An initial optical frequency comb is generated using a single-wavelength laser and a cascaded electro-optic modulator. This comb is then modulated and amplified by a programmable optical processor before being injected into a highly nonlinear fiber to produce a spectrally broadened output optical frequency comb. The output optical frequency comb spectral data is acquired by changing the modulation parameters of the programmable optical processor. Interpolation is performed on the output optical frequency comb spectral data to extract the peak envelope of each comb tooth and the valley envelope between comb teeth, which are then normalized and used as the sample set for subsequent stages. The interpolated output optical frequency comb spectral data is resampled using a larger frequency step size to obtain simplified output optical frequency comb spectral data, which serves as the sample set for subsequent stages.

[0007] The initial optical frequency comb is obtained as follows: a sinusoidal electrical signal generated by an RF source is applied as a driving voltage to two modulators. The voltage and phase of the electrical signal are adjusted by an electrical gain amplifier and a phase shifter, respectively. A DC bias voltage is also applied to the intensity modulator. When the electric field of a single wavelength of input light is... The electric field of the output light after passing through the cascaded modulator ,in: The angular frequency of the sinusoidal signal output by the radio frequency source. and The signal voltage is the value after being normalized by the half-wave voltage of the intensity modulator and phase modulator. The DC bias applied to the intensity modulator is normalized to the half-wave voltage of the intensity modulator. This refers to the chirp parameters of the system.

[0008] The aforementioned injected high nonlinear fiber refers to an initial optical frequency comb with high power generating self-phase modulation and four-wave mixing effects in a high nonlinear fiber, thereby causing spectral broadening and obtaining an output optical frequency comb with the same repetition rate as the initial optical frequency comb, but with a bandwidth much larger than that of the initial optical frequency comb.

[0009] Step 2, Training Phase: Firstly, the normalized peak and valley envelopes are used as a sample set to train the machine learning model, enabling the prediction of valley envelopes from peak envelopes. Then, optical frequency comb spectral data reconstruction can be achieved by interpolation between peak and valley values. Secondly, the modulation parameters of the initial optical frequency comb and simplified output optical frequency comb spectral data are used as a sample set to train the machine learning model, enabling the prediction of the modulation parameters of the initial optical frequency comb from the output optical frequency comb spectral data.

[0010] The machine learning model comprises two independent neural networks with different functions: a spectrum reconstruction neural network and a modulation parameter prediction neural network. These two neural network structures are invoked sequentially during operation. Specifically:

[0011] The aforementioned spectrum reconstruction neural network predicts the corresponding valley envelope information based on the peak envelope information of the optical frequency comb, and then reconstructs the complete optical frequency comb spectrum data through interpolation.

[0012] The described spectral reconstruction neural network comprises: a peak envelope input layer, a position encoding module, a convolutional feature extraction module, a spectral attention module, a valley envelope decoding module, and an output layer. Specifically: the peak envelope input layer receives the peak envelope vector from the output optical frequency comb; the position encoding module generates a position embedding vector based on the comb tooth index or frequency axis position, and concatenates or weightedly fuses it with the peak envelope; the convolutional feature extraction module contains multiple one-dimensional convolutional layers and activation functions to extract local variation features of the envelope; the spectral attention module consists of a multi-layer self-attention structure to mine non-local correlations between peak envelopes and enhance feature representation capabilities; the valley envelope decoding module upsamples and deconvolutions the fused features, outputting a data sequence matching the amount of peak envelope data; and the output layer uses a linear activation function to generate the final valley envelope prediction result.

[0013] The aforementioned spectrum reconstruction neural network is trained through supervised learning, using the valley envelope of the output optical frequency comb measured experimentally as the true label, constructing a mean square error loss function, and adding a high-order smoothness regularization term to avoid drastic jumps in the completed spectrum.

[0014] The modulation parameter prediction neural network predicts the modulation parameters required for the initial optical frequency comb based on the simplified output optical frequency comb spectrum data, thereby achieving designable control of the spectrum shape.

[0015] The modulation parameter prediction neural network includes: a spectral envelope input layer, a spectral feature extraction module, an attention enhancement module, a parameter regression module, and a multi-parameter output channel. Specifically: the spectral envelope input layer receives simplified complete spectral data; the spectral feature extraction module includes several one-dimensional convolutional layers and pooling operations to extract spectral shape distribution features; the attention enhancement module consists of a multi-layer self-attention structure to further model the nonlinear dependencies between different parts of the spectrum; the parameter regression module uses a densely connected network structure to improve feature representation depth and regression accuracy; and the multi-parameter output channel outputs the modulation parameters of the initial optical frequency comb, such as spectral phase and intensity distribution.

[0016] The modulation parameter prediction neural network is trained using supervised learning. It constructs a regression loss function using known modulation parameters and introduces a parameter boundary penalty term in combination with the input constraints of the physical system on the modulator to ensure that the prediction results are feasible and physically reasonable.

[0017] Step 3, Optical Frequency Comb Generation Stage: After reconstructing the optical frequency comb spectrum envelope by using a machine learning model and interpolation, the desired optical frequency comb spectrum envelope is input into the trained machine learning model to generate the modulation parameters required for the initial optical frequency comb. The modulation parameters are then input into a programmable optical processor to modulate the initial optical frequency comb, amplify its power, and inject it into a highly nonlinear optical fiber. The resulting optical frequency comb will have the desired spectrum envelope.

[0018] This invention relates to a spectral envelope-designable optical frequency comb generation system for implementing the above-described method, comprising: a data processing unit, a network training unit, a modulation parameter generation unit, and an optical frequency comb generation unit. The data processing unit extracts peak and valley envelope data by interpolating the output optical frequency comb spectral data acquired under different initial optical frequency comb modulations. It then trains a neural network to predict the valley envelope from the peak envelope. Finally, by interpolating between the two, it constructs a machine learning system capable of reconstructing the complete optical frequency comb spectrum from the spectral envelope. The network training unit simplifies the output optical frequency comb spectral data and combines it with the modulation parameters of the corresponding initial optical frequency comb. The modulation parameters are trained using a neural network to obtain a machine learning system that predicts the modulation parameters of the initial optical frequency comb from the output optical frequency comb spectrum. The modulation parameter generation unit, based on the set desired spectral envelope information, uses the machine learning system in the data processing unit to obtain complete optical frequency comb spectrum data and inputs it into the machine learning system in the network training unit for calculation to obtain the corresponding initial optical frequency comb modulation parameters. The optical frequency comb generation unit, based on the modulation parameters obtained in the modulation parameter generation unit, inputs them into the programmable optical processor to modulate the initial optical frequency comb, amplifies its power, and injects it into a highly nonlinear optical fiber, ultimately obtaining an optical frequency comb with the desired spectral envelope.

[0019] Technical effect

[0020] This invention extracts more accurate peak and valley envelope data of the optical frequency comb spectrum by interpolating the acquired optical frequency comb spectrum data. The peak envelope is used as the input dataset for a machine learning system, and the valley envelope as the output dataset. This allows the machine learning system to be trained to predict the corresponding valley envelope from the peak envelope. Compared with existing technologies, this invention overcomes the problem that the acquired spectrum data cannot accurately correspond to the frequency of each comb tooth in the actual optical frequency comb due to the resolution and step size of the spectrum acquisition equipment itself, thus more accurately reflecting the spectral shape of the optical frequency comb. Secondly, by decomposing the spectrum into peak envelope, valley envelope, and interpolation between peak and valley values, the complex mapping between the spectral envelope and the complete optical frequency comb spectrum is simplified to only needing to solve the mapping between the peak and valley envelopes to reconstruct the complete optical frequency comb spectrum through interpolation. By using machine learning, the function of predicting the valley envelope from the peak envelope is realized, and the effect of reconstructing the complete optical frequency comb spectrum from the spectral envelope is achieved. In addition, for the entire optical frequency comb generation system, the present invention no longer requires the desired complete spectrum as the input of the system, but only needs to set the desired spectral envelope to generate an optical frequency comb with the corresponding envelope. Attached Figure Description

[0021] Figure 1 This is a flowchart of the present invention;

[0022] Figure 2 A schematic diagram of the spectrum of the initial optical frequency comb generated for this embodiment;

[0023] Figure 3 This is a schematic diagram of interpolation and peak envelope and valley envelope extraction of the output optical frequency comb;

[0024] Figure 4 This is a comparative diagram showing the reconstruction operation performed on the acquired output optical frequency comb. Detailed Implementation

[0025] like Figure 1 As shown in this embodiment, a method for generating a spectral envelope-designable optical frequency comb based on machine learning is provided, comprising:

[0026] S1. An initial optical frequency comb is generated using a single-wavelength laser and a cascaded electro-optic modulator. Specifically, an input light with a wavelength of 1550nm and a power of 14dB is generated by a C-band laser and modulated by a cascaded intensity modulator and phase modulator to generate the initial optical frequency comb. The center frequency of the initial optical frequency comb is 193.4145THz, corresponding to 1550nm. A 10GHz sinusoidal electrical signal generated by an RF source is used as the driving voltage and applied to the two modulators. Therefore, the repetition rate of the initial optical frequency comb is 10GHz. The voltage and phase of the electrical signal are adjusted by an electrical gain amplifier and a phase shifter, respectively. A DC bias voltage is also applied to the intensity modulator. By adjusting the voltage and phase of the electrical signal, as well as the bias voltage, an initial optical frequency comb with a relatively flat spectrum and multiple combs is generated. Figure 2 As shown.

[0027] When the electric field of a single wavelength of input light is The electric field of the output light after passing through the cascaded modulator ,in: The angular frequency of the sinusoidal signal output by the radio frequency source. and The signal voltage is the value after being normalized by the half-wave voltage of the intensity modulator and phase modulator. The DC bias applied to the intensity modulator is normalized to the half-wave voltage of the intensity modulator. This refers to the chirp parameters of the system.

[0028] S2. Modulate the initial optical frequency comb using a programmable optical processor. Specifically, select the phase and intensity of five comb lines near the center frequency of the initial optical frequency comb as the modulation targets. Therefore, respectively, the phase... , , , , and intensity attenuation , , , , The modulation is applied to frequencies within the bands of 193.390 THz-193.399 THz, 193.400 THz-193.409 THz, 193.410 THz-193.419 THz, 193.420 THz-193.429 THz, and 193.430 THz-193.439 THz. The intensity of other frequencies is attenuated by 60 dB to avoid interference, which also allows the power after subsequent power amplification to be concentrated on the five modulated comb lines.

[0029] S3. The modulated initial optical frequency comb is amplified by a high-power erbium-doped fiber amplifier and then injected into a highly nonlinear fiber to generate a spectrally broadened output optical frequency comb. The spectral data of the output optical frequency comb is collected using a spectrometer. Specifically, the power of the initial optical frequency comb is amplified to 27 dBm using an EDFA and injected into a 2 km long highly nonlinear fiber. Due to self-phase modulation and four-wave mixing effects in the highly nonlinear fiber, the spectrum is broadened, resulting in an output optical frequency comb with the same repetition rate as the initial optical frequency comb but a much larger bandwidth.

[0030] S4. Change the modulation parameters of the initial optical frequency comb by the programmable optical processor, and repeat steps S1 to S3 to obtain multiple sets of corresponding modulation parameter sets and measured spectrum datasets. Specifically, randomly select 20,000 different sets of... Combination, in which Using a programmable optical processor, the initial optical frequency comb was modulated according to these 20,000 sets of modulation parameters. Simultaneously, 20,000 corresponding measured output optical frequency comb spectra were acquired.

[0031] S5, such as Figure 3 As shown, the measured spectrum data of the output optical frequency comb is processed, specifically including:

[0032] 5.1 The measured spectrum was interpolated using cubic spline interpolation with Lagrange boundaries to construct an interpolated output optical frequency comb spectrum from 193.1045THz to 193.7245THz with a frequency step of 10MHz.

[0033] 5.2 Numerical analysis was performed on the spectrum of the interpolated output optical frequency comb. The maximum intensity values ​​within every 10GHz interval in the frequency band of 193.1095THz-193.7195THz were taken, totaling 61 values, which were called peak power. Then, using the interpolation boundaries and 61 peak frequencies as intervals, the minimum intensity within each interval is calculated, resulting in 62 values, which are called valley power. .Will and The peak envelope is obtained by normalizing the combined data. Valley value envelope .

[0034] 5.3 The interpolated output optical frequency comb is resampled with a frequency step of 1.5 GHz to construct a simplified output optical frequency comb spectrum. Then, the above data processing was performed on 20,000 sets of measured output optical frequency comb spectra to obtain 20,000 sets of peak envelopes and corresponding valley envelopes, as well as simplified output optical frequency comb spectra.

[0035] S6. Construct a spectrum reconstruction neural network, using the peak envelope set as input data and the valley envelope set as output data for training. Optimize the machine learning model based on the accuracy of its judgments.

[0036] The aforementioned spectrum reconstruction neural network predicts valley envelopes from peak envelopes and then reconstructs the complete optical frequency comb spectrum through interpolation. The neural network structure includes an input layer, multiple feature extraction layers, an attention enhancement module, a decoding module, and an output layer. The input layer receives a normalized peak envelope vector and provides standardized envelope data to subsequent networks. The input dimension corresponds to the number of peak sampling points and is set to a 61-dimensional vector. The feature extraction layer includes multiple one-dimensional convolutional neural network (1D-CNN) modules, each consisting of one-dimensional convolution, a non-linear activation function, and batch normalization, extracting local morphological change features from the peak envelope and enhancing the model's ability to recognize intensity change trends. The attention enhancement module consists of multiple layers of self-attention mechanisms, employing multi-head attention to calculate the global correlation between different comb positions, improving the model's ability to model non-local features. The attention module introduces a positional encoding mechanism to preserve the positional information of the sequence in the frequency dimension; the decoding module maps the feature representation to the output dimension, and its structure includes upsampling operations and deconvolution layers, outputting a valley envelope vector (62-dimensional) that matches the input dimension; the output layer uses a linear activation function to generate the final predicted valley envelope result.

[0037] The spectrum reconstruction neural network is trained using supervised learning, employing 20,000 peak envelope sets and corresponding valley envelope sets obtained in step S5 as training samples. The loss function uses mean squared error (MSE), and a higher-order smoothing regularization term is introduced to constrain the continuity of the output envelope, preventing sharp abrupt changes in the completion result. During model training, the Adam optimizer is used for iterative parameter updates. The network structure is fine-tuned by evaluating the MSE loss, spectrum reconstruction accuracy, and envelope smoothness index on the validation set, improving overall prediction accuracy and stability. The optimized spectrum reconstruction neural network can then predict valley values ​​for any given peak envelope input in subsequent steps, and subsequently obtain complete spectrum data through interpolation.

[0038] S7. Construct a modulation parameter prediction neural network, using a simplified output optical frequency comb spectrum set as input data and the modulation parameter set as output data for training. Optimize the machine learning model based on the accuracy of its predictions.

[0039] The modulation parameter prediction neural network includes: a spectral envelope input layer, a spectral feature extraction module, an attention enhancement module, a parameter regression module, and a multi-parameter output channel. Specifically: the spectral envelope input layer receives the reconstructed complete spectral data; the spectral feature extraction module includes several one-dimensional convolutional layers and pooling operations to extract spectral shape distribution features; the attention enhancement module consists of a multi-layer self-attention structure to further model the nonlinear dependencies between different parts of the spectrum; the parameter regression module uses a densely connected network structure to improve feature representation depth and regression accuracy; and the multi-parameter output channel outputs the modulation parameters of the initial optical frequency comb, such as spectral phase and intensity distribution.

[0040] The modulation parameter prediction neural network is trained using supervised learning. It constructs a regression loss function using known modulation parameters and introduces a parameter boundary penalty term in combination with the input constraints of the physical system on the modulator to ensure that the prediction results are feasible and physically reasonable.

[0041] Due to the complex interactions between multiple parameters, including the length, nonlinear coefficient, and dispersion of nonlinear optical fibers, as well as the initial optical frequency comb and even the modulator bias conditions, it is difficult to determine how to modulate the initial optical frequency comb to obtain the desired optical frequency comb through simple mathematical calculations. Therefore, by introducing a deep neural network, [the desired optical frequency comb is achieved]. As input data, As output data, through continuous training, we obtain... and The correspondence.

[0042] S8. Input the desired optical frequency comb spectral envelope into the trained spectrum reconstruction neural network, and then reconstruct the complete optical frequency comb spectrum data through interpolation. Then input it into the trained modulation parameter prediction neural network to predict the modulation parameters required to generate an optical frequency comb with the desired spectral envelope. Specifically, this involves setting the desired spectral envelope. Then, the trained spectrum is used to reconstruct the neural network for prediction. Corresponding Finally, the peak and trough values ​​are combined in frequency order, and the complete optical frequency comb spectrum is reconstructed by segmented cubic Hermite interpolation with a frequency step of 1.5 GHz. The trained modulation parameter prediction neural network is then used to predict the optical frequency comb with the desired spectral envelope. The required modulation parameters of the initial optical frequency comb .

[0043] S9. Input the obtained modulation parameters into the programmable optical processor to modulate the initial optical frequency comb, and execute step S3 to obtain an optical frequency comb with the desired spectral envelope. Specifically, this involves: modulating the phase... , , , , and intensity attenuation , , , , The modulation is applied to frequencies within the bands of 193.390 THz-193.399 THz, 193.400 THz-193.409 THz, 193.410 THz-193.419 THz, 193.420 THz-193.429 THz, and 193.430 THz-193.439 THz, and the intensity modulation function of the programmable optical processor is used to attenuate frequencies outside these bands. The resulting output optical frequency comb, after power amplification and transmission through a highly nonlinear optical fiber, possesses the desired spectral envelope.

[0044] Based on specific practical experiments, users can set a set of target frequency points {f1, f2, ..., f k} and its corresponding target intensities {S1, S2, ..., S} k}, where fᵢ is the frequency coordinate and Sᵢ is the normalized envelope value at that frequency. The system reconstructs the continuous spectrum envelope function S(f) according to the following interpolation expression: S(f) = ∑ (i=1) (K) Sᵢ · φᵢ(f). Here, φᵢ(f) represents the interpolation basis function corresponding to the i-th control point, which can be a spline basis (such as cubic spline interpolation) or other interpolation functions with local support. The spectral envelope function S(f) can represent any physically realizable continuous or piecewise continuous function. The system uses the neural network model in this invention to perform complete spectral reconstruction of this envelope and output the modulation parameters of the initial optical frequency comb, thereby ultimately obtaining an optical frequency comb consistent with the target envelope through the spectral broadening system.

[0045] Compared with existing technologies, this invention does not require the complete desired spectrum as input to the machine learning system (in which case the physical rationality of the constructed desired spectrum is highly demanding, and repeated optimization is required to finally produce an effective output). It only needs to set the desired spectral envelope to generate an optical frequency comb with a corresponding envelope. At the same time, the frequency step size in the simplified output spectrum data can be controlled at around 1.5 GHz. Compared with the output spectrum data directly acquired by the spectrometer (which usually requires a frequency acquisition step size of around 500 MHz to ensure the accuracy and detail of the acquired spectrum data), this invention removes a large amount of redundancy in the data, which can improve the training efficiency and processing efficiency of the subsequent neural network.

[0046] In summary, the spectral envelope type implemented by this invention does not depend on a fixed mathematical function model, but supports multiple forms of envelope definition, including definition based on control point interpolation.

[0047] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A method for generating a spectral envelope-designable optical frequency comb based on machine learning, characterized in that, include: Step 1, Data Processing Stage: An initial optical frequency comb is generated using a single-wavelength laser and a cascaded electro-optic modulator. This comb is then modulated and amplified by a programmable optical processor before being injected into a highly nonlinear fiber to generate a spectrally broadened output optical frequency comb. The output optical frequency comb spectral data is acquired by changing the modulation parameters of the programmable optical processor. The output optical frequency comb spectral data is interpolated to extract the peak envelope of each comb tooth and the valley envelope between comb teeth, which are then normalized and used as the sample set for subsequent stages. The interpolated output optical frequency comb spectral data is resampled using a larger frequency step size to obtain simplified output optical frequency comb spectral data, which is used as the sample set for subsequent stages. Step 2, Training Phase: Firstly, the normalized peak envelope and valley envelope are used as a sample set to train the machine learning model, enabling the prediction of valley envelope from peak envelope. Then, the optical frequency comb spectrum data is reconstructed by interpolation between peak and valley values. Secondly, the modulation parameters of the initial optical frequency comb and the simplified output optical frequency comb spectrum data are used as a sample set to train the machine learning model, enabling the prediction of the modulation parameters of the initial optical frequency comb from the output optical frequency comb spectrum data. Step 3, Optical Frequency Comb Generation Stage: After reconstructing the optical frequency comb spectrum envelope by using a machine learning model and interpolation, the desired optical frequency comb spectrum envelope is input into the trained machine learning model to generate the modulation parameters required for the initial optical frequency comb. The modulation parameters are then input into a programmable optical processor to modulate the initial optical frequency comb, amplify its power, and inject it into a highly nonlinear optical fiber. The resulting optical frequency comb will have the desired spectrum envelope.

2. The method for generating a spectral envelope-designable optical frequency comb based on machine learning according to claim 1, characterized in that, The initial optical frequency comb is obtained as follows: a sinusoidal electrical signal generated by an RF source is applied as a driving voltage to two modulators. The voltage and phase of the electrical signal are adjusted by an electrical gain amplifier and a phase shifter, respectively. A DC bias voltage is also applied to the intensity modulator. When the electric field of a single wavelength of input light is... The electric field of the output light after passing through the cascaded modulator ,in: The angular frequency of the sinusoidal signal output by the radio frequency source. and The signal voltage is the value after being normalized by the half-wave voltage of the intensity modulator and phase modulator. The DC bias applied to the intensity modulator is normalized to the half-wave voltage of the intensity modulator. This refers to the chirp parameters of the system.

3. The method for generating a spectral envelope-designable optical frequency comb based on machine learning according to claim 1, characterized in that, The aforementioned injected high nonlinear fiber refers to an initial optical frequency comb with high power generating self-phase modulation and four-wave mixing effects in a high nonlinear fiber, thereby causing spectral broadening and obtaining an output optical frequency comb with the same repetition rate as the initial optical frequency comb, but with a bandwidth much larger than that of the initial optical frequency comb.

4. The method for generating a spectral envelope-designable optical frequency comb based on machine learning according to claim 1, characterized in that, The machine learning model includes two independent neural networks with different functions: a spectrum reconstruction neural network and a modulation parameter prediction neural network. The two neural network structures are called sequentially during use. Specifically, the spectrum reconstruction neural network predicts the corresponding valley envelope information based on the peak envelope information of the optical frequency comb, and then reconstructs the complete optical frequency comb spectrum data through interpolation. The modulation parameter prediction neural network predicts the various modulation parameters that need to be loaded into the initial optical frequency comb based on the simplified output optical frequency comb spectrum data, so as to achieve the designability and control of the spectrum shape.

5. The method for generating a spectral envelope-designable optical frequency comb based on machine learning according to claim 4, characterized in that, The described spectrum reconstruction neural network comprises: a peak envelope input layer, a position encoding module, a convolutional feature extraction module, a spectrum attention module, a valley envelope decoding module, and an output layer. Specifically: the peak envelope input layer receives the peak envelope vector from the output optical frequency comb; the position encoding module generates a position embedding vector based on the comb tooth index or frequency axis position, and concatenates or weights it with the peak envelope; the convolutional feature extraction module contains multiple one-dimensional convolutional layers and activation functions to extract local variation features of the envelope; the spectrum attention module consists of a multi-layer self-attention structure to mine non-local correlations between peak envelopes and enhance feature representation capabilities; the valley envelope decoding module upsamples and deconvolves the fused features, outputting a data sequence matching the peak envelope data volume; and the output layer uses a linear activation function to generate the final valley envelope prediction result.

6. The method for generating a spectral envelope-designable optical frequency comb based on machine learning according to claim 4 or 5, characterized in that, The aforementioned spectrum reconstruction neural network is trained through supervised learning, using the valley envelope of the output optical frequency comb measured experimentally as the true label, constructing a mean square error loss function, and adding a high-order smoothness regularization term to avoid drastic jumps in the completed spectrum.

7. The method for generating a spectral envelope-designable optical frequency comb based on machine learning according to claim 4, characterized in that, The modulation parameter prediction neural network includes: a spectral envelope input layer, a spectral feature extraction module, an attention enhancement module, a parameter regression module, and a multi-parameter output channel. Specifically: the spectral envelope input layer receives simplified complete spectral data; the spectral feature extraction module includes several one-dimensional convolutional layers and pooling operations to extract spectral shape distribution features; the attention enhancement module consists of a multi-layer self-attention structure to further model the nonlinear dependencies between different parts of the spectrum; the parameter regression module uses a densely connected network structure to improve feature representation depth and regression accuracy; and the multi-parameter output channel outputs the modulation parameters of the initial optical frequency comb, such as spectral phase and intensity distribution.

8. The method for generating a spectral envelope-designable optical frequency comb based on machine learning according to claim 4 or 7, characterized in that, The modulation parameter prediction neural network is trained using supervised learning. It constructs a regression loss function using known modulation parameters and introduces a parameter boundary penalty term in combination with the input constraints of the physical system on the modulator to ensure that the prediction results are feasible and physically reasonable.

9. A spectral envelope scalable optical frequency comb generation system for implementing the method of any one of claims 1-8, characterized in that, include: The system comprises a data processing unit, a network training unit, a modulation parameter generation unit, and an optical frequency comb generation unit. Specifically: The data processing unit extracts peak and valley envelope data by interpolating the output optical frequency comb spectrum data acquired under different initial optical frequency comb modulations. It then trains a neural network to predict the valley envelope from the peak envelope. Interpolation between the two forms a machine learning system capable of reconstructing the complete optical frequency comb spectrum from the spectral envelope. The network training unit simplifies the output optical frequency comb spectrum data and, combined with the modulation parameters of the corresponding initial optical frequency comb, trains a neural network to obtain a machine learning system that predicts the modulation parameters of the initial optical frequency comb from the output optical frequency comb spectrum. The modulation parameter generation unit, based on the set desired spectral envelope information, uses the machine learning system in the data processing unit to obtain the complete optical frequency comb spectrum data and inputs it into the machine learning system in the network training unit for computation, obtaining the corresponding initial optical frequency comb modulation parameters. The optical frequency comb generation unit, based on the modulation parameters obtained in the modulation parameter generation unit, inputs them into a programmable optical processor to modulate and amplify the initial optical frequency comb before injecting it into a highly nonlinear optical fiber, ultimately obtaining an optical frequency comb with the desired spectral envelope.

Images