Mid-infrared ultrafast laser control system and method based on data-driven neural network

Abstract

The application discloses a kind of mid-infrared ultrafast laser control system and method based on data-driven neural network, system: physical model subsystem includes pump source, pump combiner, output coupler, gain fiber, polarization controller and few-mode fiber;Simulation model subsystem includes simulation pump source, simulation pump combiner, simulation output coupler, simulation gain fiber, simulation wave plate group and simulation few-mode fiber;Neural network subsystem includes data input module, LSTM module, depth training module, output module and data driving module;Data result verification subsystem includes comparison and inspection module.The method is: simulation model subsystem generates simulation data, neural network subsystem trains simulation data, and obtains output data;Data result verification subsystem obtains error analysis result and feeds back to neural network subsystem;Final result data is used to redesign and control physical model subsystem.The system and method can realize the accurate control to laser output system.

Description

Technical Field

[0001] This invention belongs to the fields of fiber optics and neural network technology, specifically relating to a mid-infrared ultrafast laser control system and method based on data-driven neural networks. Background Technology

[0002] Laser technology is an important branch of modern physics, widely used in communications, materials processing, medicine, and military fields. Laser state control is a crucial technology within laser technology, significantly impacting its output characteristics, stability, and reliability. With technological advancements, ultrafast laser state control is becoming increasingly important in many areas, such as satellites, spacecraft, and fighter jets. However, traditional laser state control methods typically employ experimental data to train models. This approach requires extensive experimental data and a complex training process, and it cannot quickly respond to sudden changes in state, presenting numerous technical challenges that need to be addressed.

[0003] To address the aforementioned issues, there is an urgent need to provide a mid-infrared ultrafast laser state control system and method based on data-driven neural networks. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a mid-infrared ultrafast laser control system and method based on a data-driven neural network. This system facilitates rapid state prediction through a priori methods and enables precise control of the laser output system. Furthermore, it significantly improves the response sensitivity of the laser output system. This method enhances the response time sensitivity, stability, and maintainability of laser and application systems.

[0005] To achieve the above objectives, the present invention provides a mid-infrared ultrafast laser control system based on a data-driven neural network, comprising a mid-infrared band physical model subsystem, a mid-infrared band numerical simulation model subsystem, a neural network subsystem, and a data result verification subsystem.

[0006] The mid-infrared band physical model subsystem includes a pump source, a pump combiner, an output coupler, a gain fiber, a polarization controller, and a few-mode fiber. The pump input of the pump combiner is connected to the pump output of the pump source, the input of the output coupler is connected to the signal output of the pump combiner, one end of the gain fiber is connected to the first output of the output coupler, the second output of the output coupler serves as the physical laser output, the input of the polarization controller is connected to the other end of the gain fiber, one end of the few-mode fiber is connected to the output of the polarization controller, and the other end of the few-mode fiber is connected to the signal input of the pump combiner.

[0007] The mid-infrared band numerical simulation model subsystem includes a simulation pump source, a simulation pump combiner, a simulation output coupler, a simulation gain fiber, a simulation waveplate group, and a simulation few-mode fiber. The pump input end of the simulation pump combiner is connected to the pump output end of the simulation pump source. The input end of the simulation output coupler is connected to the signal output end of the simulation pump combiner. One end of the simulation gain fiber is connected to the first output end of the simulation output coupler, and the second output end of the simulation output coupler serves as the simulated laser output end. The incident end of the simulation waveplate group is connected to the other end of the gain fiber. One end of the simulation few-mode fiber is connected to the exit end of the simulation waveplate group, and the other end of the simulation few-mode fiber is connected to the signal input end of the simulation pump combiner.

[0008] The neural network subsystem includes a data input module, an LSTM module, a deep training module, an output module, and a data driving module. The input end of the data input module is connected to the mid-infrared band numerical simulation model subsystem to obtain the simulation data of the mid-infrared band numerical simulation model subsystem. The input end of the LSTM module is connected to the output end of the data input module. The input end of the deep training module is connected to the output end of the LSTM module. The input end of the output module is connected to the output end of the deep training module. The input end of the data driving module is connected to the output end of the output module, and its output end is connected to the mid-infrared band physical model subsystem to control the mid-infrared band physical model subsystem through the output results.

[0009] The data result verification subsystem includes a comparison and verification module, a raw computation data link, a neural network output result feedback link, and an error feedback link. The comparison and verification module is used to compare and verify the output data of the neural network subsystem with the raw data of the mid-infrared band numerical simulation model subsystem, and calculate the error to obtain the error analysis result. The input end of the comparison and verification module is connected to the mid-infrared band numerical simulation model subsystem through the raw computation data link to obtain the raw data of the mid-infrared band numerical simulation model subsystem. The input end of the comparison and verification module is also connected to the mid-infrared band numerical simulation model subsystem through the neural network output result feedback link to obtain part of the output data of the neural network subsystem. The output end of the comparison and verification module is connected to the neural network subsystem through the error feedback link to feed back the obtained error analysis result to the neural network subsystem to improve the output accuracy of the neural network subsystem.

[0010] As a preferred embodiment, the pump source is a fiber laser or a solid-state laser, with a suitable wavelength in the 2-micron band.

[0011] As a preferred embodiment, the mid-infrared band numerical simulation model subsystem is modeled using a nonlinear Schrödinger equation, a dual-polarization coupled nonlinear Schrödinger equation, or a generalized Ginzberg-Landau equation.

[0012] As a preferred embodiment, the neural network subsystem employs a long short-term memory neural network.

[0013] In this invention, by setting up a mid-infrared band numerical simulation model subsystem, a ring resonator can be obtained through simulation. This allows for convenient modification of pump power data and laser emission patterns, as well as convenient alteration of the length, dispersion, nonlinearity, and gain coefficient of the gain fiber. It also facilitates convenient modification of the length, dispersion, and nonlinearity of the few-mode fiber, convenient control of the polarization controller to change the polarization state, convenient modification of the parameters of various optical components, and convenient acquisition of output laser data. Therefore, the mid-infrared band numerical simulation model subsystem can obtain massive amounts of simulation data at low cost, which is beneficial for training the neural network subsystem. By connecting the mid-infrared band numerical simulation model subsystem and the mid-infrared band physical model subsystem using a neural network subsystem, massive training can be performed on the vast amounts of data obtained from the simulation. This training allows for the acquisition of a neural network model for laser output prediction and control. The output of this neural network model can then be used to control the mid-infrared band physical model subsystem, thereby achieving the desired laser output. Furthermore, the acquired neural network model can be used to predict the output signal light of the mid-infrared band physical model subsystem with known parameters. Connecting the data result verification subsystem to the mid-infrared band physical model subsystem via the original computational data link facilitates the acquisition of the original simulation data. Connecting the data result verification subsystem to the output module of the neural network subsystem via the neural network output result feedback link facilitates the acquisition of the neural network subsystem's output data. Connecting the data result verification subsystem to the neural network subsystem via an error feedback link allows for the feedback of error analysis results obtained from comparing the original simulation data and the neural network subsystem's output data, further improving the output results. This system can obtain a high-precision laser control system by combining simulation and deep learning. It is beneficial to make rapid state prediction through prior knowledge and to achieve precise control of the laser output system. At the same time, it can significantly improve the response sensitivity of the laser output system.

[0014] This invention also provides a mid-infrared ultrafast laser control method based on a data-driven neural network, which employs a mid-infrared ultrafast laser control system based on a data-driven neural network, including the following steps:

[0015] Step 1: Pre-define the laser subsystem to be designed and controlled, i.e., the mid-infrared band physical model subsystem;

[0016] Step 2: Perform data simulation using NLSE / CNLSE / GL equations to obtain a numerical simulation model that includes a simulated pump source, a simulated pump combiner, a simulated output coupler, a simulated gain fiber, a simulated waveplate group, and a simulated few-mode fiber. This results in the simulation of the mid-infrared band numerical simulation model subsystem.

[0017] Step 3: Generate a large amount of simulation data by constructing a mid-infrared band numerical simulation model subsystem, and divide the obtained simulation data into two parts: 90% of the simulation data is used as training data for the neural network, and 10% of the simulation data is used as validation data for the neural network.

[0018] Step 4: Input the training data into the LSTM module through the data input module. Process the data using the input gate, storage unit, forget gate and output gate in the LSTM module. Then input the obtained data into the deep training module for deep training. Repeat this process multiple times to complete the data training. Finally, output the result data to the output module.

[0019] Step 5: Through the output module, a portion of the data generated by the numerical simulation model system is input to the comparison and verification module via the neural network output result feedback link. Simultaneously, the raw data of the mid-infrared band numerical simulation model subsystem is obtained through the raw calculation data link and input to the comparison and verification module. The comparison and verification module compares and verifies the training result data with the untrained raw data, performs calculation and analysis to obtain error result data, and then feeds the error analysis result data back to the neural network subsystem through the error feedback link. The cycle from Step 3 to Step 5 is repeated to continuously improve the training accuracy of the neural network subsystem until the set number of iterations is reached. Finally, the obtained result data is output to the data-driven module through the output module.

[0020] Step Six: Using the data-driven module, the final result data is used to inversely design and control the mid-infrared band physical model subsystem, and the pump power of the pump source, the length of the gain fiber, and the length of the few-mode fiber are determined. At the same time, the polarization state is adjusted by controlling the polarization controller.

[0021] As a preferred option, in step one, the pre-selected optical devices include a pump source, a pump combiner, an output coupler, a gain fiber, a polarization controller, and a few-mode fiber.

[0022] As an alternative, in step six, the neural network subsystem that has reached the training accuracy can also be used to predict the mid-infrared band physical model subsystem with known parameters in order to predict the output signal light.

[0023] This method trains a neural network using a physical model, significantly reducing the number of training iterations and improving state recognition efficiency, thus solving the challenge of high-speed, high-precision response in fast-moving laser systems. Simultaneously, through rapid prior state prediction, it quickly anticipates unknown abrupt changes, significantly improving the system's response time sensitivity. While this invention uses a fiber laser as an example, it can be widely applied to other laser systems. Compared to existing methods and technologies for controlling ultrafast laser systems using physical means, this invention has the following advantages:

[0024] 1. Compared with traditional neural network training methods that rely on real experimental data, training through physical models significantly reduces the number of training iterations and greatly improves state recognition efficiency. It also saves physical and data resources, overcomes difficulties such as the difficulty in generating experimental data, insufficient data volume, and high cost under extreme physical conditions, and solves the problems of high-speed and high-precision response in fast-moving laser systems or (such as satellites, spacecraft, fighter jets, etc.).

[0025] 2. This invention can rapidly predict the future state of the laser and its response data through prior data, and quickly predict unknown sudden changes or faults, which greatly improves the response time sensitivity, stability and maintainability of the laser and application system.

[0026] 3. The system of this invention has strong compatibility. By simply changing the number of parameters, model type and structure of the mathematical model used, and combining it with other intelligent algorithms, such as RNN, CNN, DNN, etc., it can generate data for training of various laser systems. Therefore, it can control more parameters with higher precision, and it is also conducive to interfacing with future intelligent systems, which greatly expands the range of systems that can be controlled. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the structure of a mid-infrared ultrafast laser control system based on a data-driven neural network in this invention;

[0028] Figure 2 This is a schematic diagram of the evolution of the time-domain signal and spectral signal of traditional soliton formation in the numerical simulation model subsystem of Embodiment 1 of the present invention;

[0029] Figure 3 This is a schematic diagram of the prediction results of the time evolution model of the evolution dynamics of the neural network subsystem in Embodiment 1 of the present invention;

[0030] Figure 4 This is a schematic diagram of the spectral evolution model prediction results of the evolution dynamics of the neural network subsystem in Embodiment 1 of the present invention;

[0031] Figure 5This is a schematic diagram of the evolution of the time-domain signal and spectral signal of dissipative soliton formation in the numerical simulation model subsystem of Embodiment 2 of the present invention;

[0032] Figure 6 This is a schematic diagram of the prediction results of the time evolution model of the evolution dynamics of the neural network subsystem in Embodiment 2 of the present invention;

[0033] Figure 7 This is a schematic diagram of the spectral evolution model prediction results of the neural network subsystem evolution dynamics in Embodiment 2 of the present invention.

[0034] In the diagram: 1-1, Pump source; 1-2, Pump combiner; 1-3, Gain fiber; 1-4, Polarization controller; 1-5, Few-mode fiber; 1-6, Output coupler; 2-1, Mathematical modeling; 2-2, Simulated pump source; 2-3, Simulated pump combiner; 2-4, Simulated gain fiber; 2-5, Simulated waveplate group; 2-6, Simulated few-mode fiber; 2-7, Simulated output coupler; 3-1, Data input module; 3-2, LSTM module; 3-3, Deep training module; 3-4, Output module; 3-5, Data-driven module; 4-1, Neural network output feedback link; 4-2, Raw computation data link; 4-3, Error feedback link; 4-4, Comparison and verification module. Detailed Implementation

[0035] The present invention will be further described below with reference to embodiments.

[0036] Example 1: Negative Dispersion System Data-Driven Neural Network Prediction of Traditional Soliton Dynamics

[0037] like Figure 1 As shown, the present invention provides a mid-infrared ultrafast laser control system based on a data-driven neural network, including a mid-infrared band physical model subsystem, a mid-infrared band numerical simulation model subsystem, a neural network subsystem, and a data result verification subsystem;

[0038] The mid-infrared band physical model subsystem includes a pump source 1-1, a pump combiner 1-2, an output coupler 1-6, a gain fiber 1-3, a polarization controller 1-4, and a few-mode fiber 1-5. Specifically, the pump source 1-1 has a wavelength of 793 nm and a maximum power of 12 W. The gain fiber 1-3 is a 3.5 m Tm-doped few-mode fiber with a core of 15 μm and a numerical aperture (NA) of 0.15, theoretically supporting both ring light and fundamental mode. The other undoped fibers used have the same combination structure as above. The coupling rate of the laser output is 10%, and the second-order group velocity dispersion of the entire system is negative. The pulse-activated switch is a waveplate group based on the nonlinear polarization rotation effect. The pump input of the pump combiner 1-2 is connected to the pump output of the pump source 1-1. The input of the output coupler 1-6 is connected to the signal output of the pump combiner 1-2. One end of the gain fiber 1-3 is connected to the first output of the output coupler 1-6, and the second output of the output coupler 1-6 serves as a physical laser output for laser signal detection. The input of the polarization controller 1-4 is connected to the other end of the gain fiber 1-3. One end of the few-mode fiber 1-5 is connected to the output of the polarization controller 1-4, and the other end of the few-mode fiber 1-5 is connected to the signal input of the pump combiner 1-2. Specifically, the pump input of the pump combiner 1-2 and the pump output of the pump source 1-1 are connected by fiber fusion splicing. The pump combiner 1-2, the output coupler 1-6, the gain fiber 1-3, the polarization controller 1-4, and the few-mode fiber 1-5 are sequentially connected in a loop via pigtails to form a closed optical path.

[0039] The mid-infrared band numerical simulation model subsystem includes a simulation pump source 2-2, a simulation pump combiner 2-3, a simulation output coupler 2-7, a simulation gain fiber 2-4, a simulation waveplate group 2-5, and a simulation few-mode fiber 2-6. The pump input of the simulation pump combiner 2-3 is connected to the pump output of the simulation pump source 2-2. The input of the simulation output coupler 2-7 is connected to the signal output of the simulation pump combiner 2-3. One end of the simulation gain fiber 2-4 is connected to the first output of the simulation output coupler 2-7, and the second output of the simulation output coupler 2-7 serves as the simulated laser output. The incident end of the simulation waveplate group 2-5 is connected to... The other end of the gain fiber 1-3 is connected to the simulated few-mode fiber 2-6, one end of which is connected to the output end of the simulated waveplate group 2-5, and the other end of which is connected to the signal input end of the simulated pump combiner 2-3. The simulated pump source 2-2 can emit pulsed laser or continuous laser. The length, dispersion, nonlinearity, and gain coefficient of the simulated gain fiber 2-4 can be changed. The length, dispersion, and nonlinearity of the simulated few-mode fiber 2-6 can also be changed. The simulated waveplate group 2-5 forms a saturable absorber-like structure, corresponding to the angles of the three waveplates (two quarter-waveplates and one half-waveplate) of the fiber polarization controller in the physical system, which can be changed. The optical devices in the mid-infrared band numerical simulation model subsystem are simulation models mathematically modeled using NLSE / CNLSE / GL equations.

[0040] The mid-infrared band numerical simulation model subsystem specifically comprises:

[0041] 1. Considering the polarization state, the numerical model of the coupled generalized nonlinear Schrödinger equation (CGL) is used. In the numerical simulation model system, for the ordinary undoped fiber transmission model, the dynamic simulation equation of the thulium-doped mode-locked fiber laser pulse after considering the birefringence effect is obtained, and the transmission model of the undoped fiber (simulated few-mode fiber 2-6) is shown in formulas (1) and (2):

[0042] (1);

[0043] (2);

[0044] 2. For the gain fiber transmission model, the dynamic simulation equation of the thulium-doped mode-locked fiber laser pulse after considering the gain filtering effect is obtained, and the transmission model of the thulium-doped fiber (simulated gain fiber 2-4) is shown in formulas (3) and (4):

[0045] (3); (4);

[0046] In the formula, ;

[0047] The distributed gain function is: ;

[0048] 3. The output coupler (simulated output coupler 2-7) model is shown in formulas (5) and (6):

[0049] (5);

[0050] (6);

[0051] 4. The waveplate model based on nonlinear polarization rotation (simulated waveplate group 2-5) is shown in formula (7):

[0052] (7);

[0053] In the formula, The phase delay bias introduced by the insertion polarization controller can be finely adjusted to achieve mode locking. In ring fiber lasers based on nonlinear polarization rotation effects, noisy laser light gradually evolves into a pulse signal after multiple intracavity cycles.

[0054] 5. The numerical models for the evolution of laser spatial transverse modes under nonlinear effects are shown in equations (8), (9), and (10):

[0055] (8); (9);

[0056] (10);

[0057] The optical components in this numerical simulation model are all simulation models mathematically modeled using NLSE / CNLSE / GL equations. Part of the data generated by the mid-infrared band numerical simulation model subsystem is used for training the neural network subsystem, and part is used to verify the accuracy of the neural network output results.

[0058] The neural network subsystem includes a data input module 3-1, an LSTM module 3-2, a deep training module 3-3, an output module 3-4, and a data-driven module 3-5. The input terminal of the data input module 3-1 is connected to the mid-infrared band numerical simulation model subsystem to obtain simulation data from the mid-infrared band numerical simulation model subsystem. The input terminal of the LSTM module 3-2 is connected to the output terminal of the data input module 3-1. The input terminal of the deep training module 3-3 is connected to the output terminal of the LSTM module 3-2. The input terminal of the output module 3-4 is connected to the output terminal of the deep training module 3-3. The input terminal of the data-driven module 3-5 is connected to the output terminal of the output module 3-4. The output module 3-1 connects to the mid-infrared band physical model subsystem and is used to control the mid-infrared band physical model subsystem through the output results. The data input module 3-1 inputs a portion of the data generated by the mid-infrared band numerical simulation model subsystem into the LSTM module 3-2 for processing, and then trains it through the deep training module 3-3. The LSTM module 3-2 includes an input gate, a storage unit, a forget gate, and an output gate. The output module 3-4 outputs data in two parts: one part is used to feed back to the data result verification subsystem for comparison and verification with the original data of the mid-infrared band numerical simulation model subsystem, and the other part is output to the data driving module 3-5 for driving and controlling the mid-infrared band physical model subsystem.

[0059] The data result verification subsystem includes a comparison and verification module 4-4, a raw computation data link 4-2, a neural network output result feedback link 4-1, and an error feedback link 4-3. The comparison and verification module 4-4 is used to compare and verify the output data of the neural network subsystem with the raw data of the mid-infrared band numerical simulation model subsystem, and calculate the error to obtain the error analysis result. The input end of the comparison and verification module 4-4 is connected to the mid-infrared band numerical simulation model subsystem through the raw computation data link 4-2 to obtain the raw data of the mid-infrared band numerical simulation model subsystem. This raw data is not used for neural network training. The input end of the comparison and verification module 4-4 is also connected to the mid-infrared band numerical simulation model subsystem through the neural network output result feedback link 4-1 to obtain the final part of the output data of the neural network subsystem. The output end of the comparison and verification module 4-4 is connected to the neural network subsystem through the error feedback link 4-3 to feed back the obtained error analysis result to the neural network subsystem to improve the output accuracy and precision of the neural network subsystem.

[0060] The pump source 1-1 may be, but is not limited to, a fiber laser or a solid-state laser. The applicable wavelength may be, but is not limited to, the 2-micrometer band, and may be extended to other mid-infrared bands of 2-5 micrometers.

[0061] The mid-infrared band numerical simulation model subsystem is obtained through data modeling method 2-1. The data modeling method can be, but is not limited to, the nonlinear Schrödinger equation (NLSE), the dual-polarization coupled nonlinear Schrödinger equation (CNLSE), or the generalized Ginzburg-Landau equation (GL). The output result can be a single-polarization system, a dual-polarization system, or a multi-transverse-mode system. The types of optical devices included in the mid-infrared band numerical simulation model subsystem can be modified through data modeling method 2-1, depending on the selection of the mid-infrared band physical model subsystem.

[0062] The neural network subsystem may be, but is not limited to, a long short-term memory neural network (LSTM).

[0063] This invention also provides a mid-infrared ultrafast laser control method based on a data-driven neural network, which employs a mid-infrared ultrafast laser control system based on a data-driven neural network, characterized by the following steps:

[0064] Step 1: Pre-determine the laser subsystem to be designed and controlled, namely the mid-infrared band physical model subsystem (2-micron conventional soliton fiber laser pulse). The pre-selected optical devices include pump source 1-1, pump combiner 1-2, output coupler 1-6, gain fiber 1-3, polarization controller 1-4, and few-mode fiber 1-5, etc.

[0065] Step 2: Perform data simulation using NLSE / CNLSE / GL equations, specifically prioritizing GL equations for data simulation. This will yield a numerical simulation model that includes a simulated pump source 2-2, a simulated pump combiner 2-3, a simulated output coupler 2-7, a simulated gain fiber 2-4, a simulated waveplate group 2-5 (polarization controller), and a simulated few-mode fiber 2-6. This will simulate the mid-infrared band numerical simulation model subsystem.

[0066] Step 3: Generate a large amount of simulation data by constructing a mid-infrared band numerical simulation model subsystem, and divide the obtained simulation data into two parts: 90% of the simulation data is used as training data for the neural network, and 10% of the simulation data is used as validation data for the neural network.

[0067] Step 4: Input the training data into the LSTM module 3-2 through the data input module 3-1. Process the data using the input gate, storage unit, forget gate and output gate in the LSTM module 3-2. Then input the obtained data into the deep training module 3-3 for deep training. Repeat this process multiple times to complete the data training. Finally, output the result data to the output module 3-4.

[0068] Step 5: Through output module 3-4, a portion of the data generated by the numerical simulation model simulation system 2 is input to the comparison and verification module 4-4 via the neural network output result feedback link 4-1. Simultaneously, the raw data of the mid-infrared band numerical simulation model subsystem is obtained through the raw calculation data link 4-2 and input to the comparison and verification module 4-4. The comparison and verification module 4-4 compares and verifies the training result data with the untrained raw data, performs calculation and analysis to obtain error result data, and then feeds the error analysis result data back to the neural network subsystem through the error feedback link 4-3. The cycle of steps 3 to 5 is repeated to continuously improve the training accuracy of the neural network subsystem until the set number of iterations is reached. The final result data is then output to the data driving module 3-5 through output module 3-4.

[0069] Step Six: Using the data-driven module 3-5, the final result data is used to inversely design and control the mid-infrared band physical model subsystem, determining the pump power of pump source 1-1, the length of gain fiber 1-3, and the fiber length of few-mode fiber 1-5. Simultaneously, the polarization state is adjusted by controlling the polarization controller 1-4. Furthermore, a neural network subsystem with trained accuracy can be used to predict the output signal light from the known parameters of the mid-infrared band physical model subsystem.

[0070] In this embodiment, the output laser center wavelength is set to 2000 nm, the output laser intensity accounts for 10% of the total intensity, the total cavity length is 16 m, and the total second-order dispersion is -1.12 ps in the numerical simulation. 2 The gain bandwidth of the thulium-doped fiber is 30 nm, and the nonlinear coefficient of the passive fiber is 3 W. -1 / km. In the experiment, the laser travels at the speed of light along the annular cavity, and within an extremely short time, the laser has already traversed the cavity tens of millions of times. In this embodiment... Figure 2 This is a schematic diagram of the evolution of the time-domain and spectral signals of traditional soliton formation in a numerical simulation model system based on CNLSE and GL equations. During the evolution of the time-domain signal, the peak signal intensity increases and the pulse width narrows, and then remains stable. During the evolution of the spectral signal, the signal widens, Kelly sidebands appear, and then remain stable.

[0071] A prediction model for the transmission characteristics of a 2 µm picosecond pulsed laser in optical fiber, based on LSTM, was built on the Keras platform. The transmission process of the 2 µm pulsed laser in a passive optical fiber was simulated using a numerical simulation program, with a set of data saved at regular intervals. 2000 sets of simulation data were generated, each set representing the light propagation data in the optical fiber. At regular intervals, 100 sets of data were recorded after transmission. The input data format used in the neural network was 2000256100, where 2000 is the number of data sets and 256 is the data length. Figure 3 This diagram illustrates the prediction results of the time-evolution model of the neural network system's evolutionary dynamics, with the initial state being continuous light. The data in the middle represents the evolution process of the soliton pulse predicted by the LSTM neural network, while the data on the right shows the difference between the numerical simulation results and the prediction results. The predicted dynamic process and the numerical simulation process show good consistency. The diagram clearly shows that during the evolution from noise to soliton, the pulse gradually forms three pulse peaks as it travels within the cavity. In subsequent evolution, only one pulse peak is effectively retained, with the pulse width narrowing, the peak intensity increasing, and finally stabilizing. Error analysis reveals a significant deviation in the process of preserving a single pulse during multi-pulse purification. Figure 4 This diagram illustrates the spectral evolution prediction results of the neural network system's evolutionary dynamics model. The spectral evolution predicted by the LSTM neural network model is clearly visible, starting as a line spectrum, gradually widening, and eventually exhibiting a clear Kelly spur, the most distinctive characteristic of soliton pulse spectra. Throughout the evolution process, the numerical simulation and the neural network prediction show good consistency. Even in up to 100 cavity cycles, the dynamics predicted by the network and the dynamics simulated by GNLSE are essentially consistent, with a relative difference of less than 4 dB throughout the entire evolution process.

[0072] Example 2: Using data from a positive dispersive system to drive a neural network to predict the dynamic behavior of dissipative solitons

[0073] In this embodiment, unless otherwise specified, all parameters are the same as in Embodiment 1. In this embodiment, the second-order group velocity dispersion in the control numerical model is set to positive. The total dispersion within the cavity is adjusted to the positive dispersion region, the total cavity length is 16m, and the total second-order dispersion is 1.12 ps. 2 The gain bandwidth of the thulium-doped fiber is 30 nm, and the nonlinear coefficient of the passive fiber is 3 W. -1 / km. In the numerical simulation, the waveplate angles were adjusted to 50, 115, and 70 degrees respectively, the gain saturation flux was 500 pJ, and the gain fiber gain coefficient was 4. In this embodiment... Figure 5This is a schematic diagram of the evolution of the time-domain signal and spectral signal of dissipative soliton formation in a numerical simulation model system. Continuous light is selected as the input signal. During the evolution process, the pulse signal is formed by multiple cycles, forming a wide pulse signal. The wide pulse signal gradually narrows to form an ultrashort pulse, while the spectral signal remains flat at the top during the widening process. Figure 6 The diagram shows the prediction results of the time evolution model of the neural network system evolution dynamics. The neural network prediction and results are in good agreement with the numerical simulation results. The final predicted spectrum shows a square wave shape, which is consistent with the characteristics of dissipative soliton spectrum, and the overall error is less than 5dB.

Claims

1. A mid-infrared ultrafast laser control system based on a data-driven neural network, comprising a mid-infrared band physical model subsystem, characterized in that, It also includes a mid-infrared band numerical simulation model subsystem, a neural network subsystem, and a data result verification subsystem; The mid-infrared band physical model subsystem includes a pump source (1-1), a pump combiner (1-2), an output coupler (1-6), a gain fiber (1-3), a polarization controller (1-4), and a few-mode fiber (1-5). The pump input of the pump combiner (1-2) is connected to the pump output of the pump source (1-1). The input of the output coupler (1-6) is connected to the signal output of the pump combiner (1-2). One end of the gain fiber (1-3) is connected to the first output of the output coupler (1-6), and the second output of the output coupler (1-6) serves as the physical laser output. The input of the polarization controller (1-4) is connected to the other end of the gain fiber (1-3). One end of the few-mode fiber (1-5) is connected to the output of the polarization controller (1-4), and the other end of the few-mode fiber (1-5) is connected to the signal input of the pump combiner (1-2). The mid-infrared band numerical simulation model subsystem includes a simulation pump source (2-2), a simulation pump combiner (2-3), a simulation output coupler (2-7), a simulation gain fiber (2-4), a simulation waveplate group (2-5), and a simulation few-mode fiber (2-6). The pump input of the simulation pump combiner (2-3) is connected to the pump output of the simulation pump source (2-2), and the input of the simulation output coupler (2-7) is connected to the signal output of the simulation pump combiner (2-3). One end of the simulated gain fiber (2-4) is connected to the first output end of the simulated output coupler (2-7), and the second output end of the simulated output coupler (2-7) serves as the simulated laser output end. The incident end of the simulated waveplate group (2-5) is connected to the other end of the gain fiber (1-3). One end of the simulated few-mode fiber (2-6) is connected to the output end of the simulated waveplate group (2-5), and the other end of the simulated few-mode fiber (2-6) is connected to the signal input end of the simulated pump combiner (2-3). The neural network subsystem includes a data input module (3-1), an LSTM module (3-2), a deep training module (3-3), an output module (3-4), and a data-driven module (3-5). The input end of the data input module (3-1) is connected to the mid-infrared band numerical simulation model subsystem to obtain the simulation data of the mid-infrared band numerical simulation model subsystem. The input end of the LSTM module (3-2) is connected to the output end of the data input module (3-1). The input end of the deep training module (3-3) is connected to the output end of the LSTM module (3-2). The input end of the output module (3-4) is connected to the output end of the deep training module (3-3). The input end of the data-driven module (3-5) is connected to the output end of the output module (3-4), and its output end is connected to the mid-infrared band physical model subsystem to control the mid-infrared band physical model subsystem through the output results. The data result verification subsystem includes a comparison and verification module (4-4), a raw calculation data link (4-2), a neural network output result feedback link (4-1), and an error feedback link (4-3). The comparison and verification module (4-4) is used to compare and verify the output data of the neural network subsystem with the raw data of the mid-infrared band numerical simulation model subsystem, and calculate the error to obtain the error analysis result. The input of the comparison and verification module (4-4) is connected to the mid-infrared band numerical simulation model subsystem through the raw computation data link (4-2) to obtain the raw data of the mid-infrared band numerical simulation model subsystem. The input of the comparison and verification module (4-4) is also connected to the mid-infrared band numerical simulation model subsystem through the neural network output result feedback link (4-1) to obtain part of the output data of the neural network subsystem. The output of the comparison and verification module (4-4) is connected to the neural network subsystem through the error feedback link (4-3) to feed back the obtained error analysis results to the neural network subsystem to improve the output accuracy of the neural network subsystem.

2. The mid-infrared ultrafast laser control system based on a data-driven neural network according to claim 1, characterized in that, The pump source (1-1) is a fiber laser or a solid-state laser, with a suitable wavelength in the 2-micrometer band.

3. A mid-infrared ultrafast laser control system based on a data-driven neural network according to claim 1 or 2, characterized in that, The mid-infrared band numerical simulation model subsystem is modeled using the nonlinear Schrödinger equation, the dual-polarization coupled nonlinear Schrödinger equation, or the generalized Ginzberg-Landau equation.

4. A mid-infrared ultrafast laser control system based on a data-driven neural network according to claim 3, characterized in that, The neural network subsystem employs a long short-term memory neural network.

5. A mid-infrared ultrafast laser control method based on a data-driven neural network, employing a mid-infrared ultrafast laser control system based on a data-driven neural network as described in any one of claims 1 to 4, characterized in that, Includes the following steps: Step 1: Pre-define the laser subsystem to be designed and controlled, i.e., the mid-infrared band physical model subsystem; Step 2: Perform data simulation using NLSE / CNLSE / GL equations to obtain a numerical simulation model that includes a simulated pump source (2-2), a simulated pump combiner (2-3), a simulated output coupler (2-7), a simulated gain fiber (2-4), a simulated waveplate group (2-5), and a simulated few-mode fiber (2-6). This results in the simulation of the mid-infrared band numerical simulation model subsystem. Step 3: Generate a large amount of simulation data by constructing a mid-infrared band numerical simulation model subsystem, and divide the obtained simulation data into two parts: 90% of the simulation data is used as training data for the neural network, and 10% of the simulation data is used as validation data for the neural network. Step 4: Input the training data into the LSTM module (3-2) through the data input module (3-1). Process the data using the input gate, storage unit, forget gate and output gate in the LSTM module (3-2). Then input the obtained data into the deep training module (3-3) for deep training. Repeat this process multiple times to complete the data training. Finally, output the result data to the output module (3-4). Step 5: Through the output module (3-4), a portion of the data generated by the numerical simulation model simulation system (2) is input to the comparison and verification module (4-4) through the neural network output result feedback link (4-1). At the same time, the original data of the mid-infrared band numerical simulation model subsystem is obtained through the original calculation data link (4-2) and input to the comparison and verification module (4-4). The comparison and verification module (4-4) compares and verifies the training result data with the untrained original data, and performs calculation analysis to obtain error result data. Then, the error analysis result data is fed back to the neural network subsystem through the error feedback link (4-3). The cycle from step 3 to step 5 is repeated to continuously improve the training accuracy of the neural network subsystem until the set number of iterations is reached. The final result data is then output to the data driving module (3-5) through the output module (3-4). Step 6: Using the data-driven module (3-5), the final result data is used to reverse design and control the physical model subsystem of the mid-infrared band, and the pump power of the pump source (1-1), the length of the gain fiber (1-3), and the fiber length of the few-mode fiber (1-5) are determined. At the same time, the polarization state is adjusted by controlling the polarization controller (1-4).

6. The mid-infrared ultrafast laser control method based on a data-driven neural network according to claim 5, characterized in that, In step one, the pre-selected optical devices include a pump source (1-1), a pump combiner (1-2), an output coupler (1-6), a gain fiber (1-3), a polarization controller (1-4), and a few-mode fiber (1-5).

7. The mid-infrared ultrafast laser control method based on a data-driven neural network according to claim 6, characterized in that, In step six, the neural network subsystem that has reached the training accuracy can also be used to predict the mid-infrared band physical model subsystem with known parameters in order to predict the output signal light.

Images