Machine learning enhanced quantum vacuum compression light source system and method

By introducing machine learning algorithms into the quantum vacuum compression light source system and intelligently adjusting key parameters, the problem of low stability and compression of quantum vacuum compression light sources in the existing technology is solved, and efficient quantum compression and long-term stability improvements are achieved.

CN118943878APending Publication Date: 2024-11-12EAST CHINA NORMAL UNIV
View PDF 0 Cites 6 Cited by

Patent Information

Application Number
CN202411004276.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2024-07-02
Filing Date
2024-07-25
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

The compression degree of existing quantum vacuum compressed light sources in the 795nm band is generally low and are susceptible to factors such as ambient temperature, magnetic field, and pump light intensity, resulting in insufficient system stability and it is difficult to maintain efficient operation for a long time in practical applications.

Method used

Using machine learning-enhanced quantum vacuum compression light source system, neural network algorithms intelligently search and adjust parameters such as pump light intensity, frequency, atomic gas chamber temperature, magnetic field, etc., to correct the compression degree in real time, improve the robustness of the system and Compression degree.

Benefits of technology

The long-term stability and compression degree of quantum compressed light sources have been significantly improved, and the compression degree has been increased to 4.2dB, and the stability time can be maintained for more than 1 hour, enhancing the system's resistance to environmental interference.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118943878A_ABST
    Figure CN118943878A_ABST
Patent Text Reader

Abstract

The invention discloses a machine learning enhanced quantum vacuum compression light source system. Comprising a coherent light source, an acousto-optic modulator, a single-mode fiber, a Glan laser prism, a first polarization beam splitter, a first focusing lens, a rubidium atom gas chamber, a Helmholtz coil, a magnetic shield, a second focusing lens, a quarter-wave plate, a first half-wave plate, a second polarization beam splitter, a first reflector, piezoelectric ceramics, an electro-optic intensity modulator and a third polarization beam splitter. The device comprises a first polarization beam splitter, a first photoelectric detector, a digital operation locking module, a second half-wave plate, a fourth polarization beam splitter, a second photoelectric detector, a third photoelectric detector, a spectrum analyzer, a computer, a signal generator and a temperature controller. The performance of the quantum vacuum compression light system is optimized by using the neural network technology in artificial intelligence, the compression degree and long-term stability of the system are further improved through dynamic real-time multi-parameter adjustment, and a quantum light source is helped to reduce light field noise.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the fields of artificial intelligence, engineering control, quantum optics and precision measurement technology, and mainly relates to a machine learning enhanced quantum vacuum compression light source system and method, which is used to prepare a long-term stable and reliable 795nm band quantum vacuum compression light source and further improve its compression degree. Background Art

[0002] In recent years, with the development of quantum science and technology such as quantum precision measurement, quantum communication and quantum computing, including gravitational wave detection, the launch of the "Mozi" and the "Jiuzhang" computer, experimental technology based on quantum physics has played an increasingly important role in future scientific and technological innovation and social development. The quantum industry based on quantum physics will become an important field for future scientific and technological development. Quantum precision measurement is to use the basic laws of quantum physics to seek higher-precision measurement methods. Squeezed state quantum light source, as a non-classical light field that exceeds the quantum noise limit, is widely used to improve the measurement accuracy of optical interferometers and has broad application prospects in high-precision measurements such as gravitational wave detection. So far, there have been a variety of methods to generate vacuum compressed quantum light sources in different bands. Generally speaking, the conditions for the generation of vacuum compressed light sources are very harsh, the equipment is complex and expensive, which makes it difficult to popularize the corresponding experimental technology, and the compression degree in the 795nm band is generally low.

[0003] In addition, a large number of studies [1-3] It shows that quantum vacuum compression light sources are easily affected by various factors such as ambient temperature, magnetic field, pump light intensity, detuning, and loss, which will lead to a decrease in the system's compression degree. This also leads to the lack of long-term stability of existing domestic vacuum compression light sources, which are generally limited in practical application scenarios. Therefore, how to solve the long-term stability of quantum vacuum compression light sources in practical applications and at the same time increase the compression degree of vacuum compression light fields as much as possible is a problem that needs to be solved urgently in the current scientific research and industrial communities. Summary of the invention

[0004] In order to solve the problems existing in the existing vacuum compression light source, the goal of the present invention is to provide a new type of machine learning enhanced quantum vacuum compression light source system and corresponding method, aiming to solve the problems that the traditional quantum compression light source is susceptible to multi-parameter interference, insufficient stability and low compression. Inspired by artificial intelligence technology, the present invention adopts a neural network algorithm of machine learning and is adapted to the 795nm band quantum vacuum compression light source preparation hardware system. It can intelligently search for multiple parameters such as the pump light intensity, frequency, atomic gas chamber temperature, and magnetic field of the vacuum compression light source to seek the best quantum compression performance, and can adaptively correct the compression of the vacuum compression light in real time to keep it stable for a long time. The system of the present invention can better solve the negative impact of fluctuations in multiple factors such as pump light intensity, frequency, atomic gas chamber temperature, and magnetic field on the performance of the compression light source from the interactive cycle of the machine learning program and the quantum light source experimental hardware. Compared with a single vacuum compression light source experimental system, the present invention significantly suppresses the noise drift of the compression light source and greatly improves the long-term robustness and quantum compression of the system. The invention has better expanded the prospects of vacuum compression light sources in precision measurements such as gravitational wave observation, geographical survey, and electromagnetic sensing.

[0005] In order to achieve the above object, the present invention provides a machine learning enhanced quantum vacuum squeezed light source system, comprising:

[0006] Coherent light source is 795nm linearly polarized light generated by semiconductor laser, which is used as pump light for quantum vacuum compression light. After interacting with atoms, it excites vacuum compression light.

[0007] Acousto-optic modulators have high conversion efficiency and can use the acousto-optic effect to produce multi-level diffraction to adjust the intensity of incident light;

[0008] Single-mode optical fiber plays the role of adjusting the spatial distribution of the light spot. It integrates the light spot after acousto-optic modulation into a Gaussian light spot to ensure the purity of the spatial mode.

[0009] Glan laser prism, with high extinction ratio, used to improve the linear polarization purity of incident laser, wherein the high extinction ratio refers to an extinction ratio greater than 1:10000;

[0010] A first polarization beam splitter is used as a polarizing device, and together with the aforementioned Glan laser prism, ensures that the incident light satisfies the vertical linear polarization;

[0011] The first focusing lens is mainly used to focus the light spot. The interaction between the focused light spot and the atoms will be stronger.

[0012] The rubidium atomic gas cell is a cylindrical glass bubble filled with rubidium atomic gas, which is used as a medium for the generation of quantum vacuum compressed light;

[0013] Helmholtz coils, used to apply magnetic fields to atoms and enhance the interaction between light and atoms;

[0014] Magnetic shielding, composed of three layers of Permalloy, is used to shield the crosstalk of the external environment magnetic field to ensure the stability of the compression performance of the quantum compression light source;

[0015] a second focusing lens for realigning the light beam into a parallel light beam;

[0016] A quarter-wave plate, which can adjust the polarization state of the outgoing light;

[0017] The first half-wave plate, in conjunction with the aforementioned quarter-wave plate, adjusts the polarization angle of light to ensure that the outgoing light maintains vertical polarization as much as possible and suppresses horizontal polarization light leakage;

[0018] The second polarization beam splitter plays a beam splitting role, separating the vacuum compression light and the pump light through orthogonal polarizations;

[0019] The first reflector is capable of changing the propagation direction of light so that the vacuum compressed light is directed toward the subsequent device;

[0020] Piezoelectric ceramics, used to scan the relative phase between the intense pump light and the vacuum compression light;

[0021] The electro-optic intensity modulator can control the intensity of the outgoing light through voltage, ensuring that the intensity of the outgoing light remains constant while other system parameters change;

[0022] A third polarization beam splitter, used for combining vacuum compression light and strong pump light;

[0023] A first photoelectric detector is used to monitor the light intensity of the transmission part of the pump light in real time and transmit the detection signal to the digital operation locking module;

[0024] The digital operation locking module contains a proportional integral differential PID operation circuit, which can operate the input signal and feedback lock the light intensity of the transmitted part of the pump light;

[0025] The second half-wave plate is used to rotate the polarization angle of linear polarized light by 45 degrees;

[0026] A fourth polarization beam splitter, used to split the light beam 50:50, and divide it equally and balancedly;

[0027] The second photodetector and the third photodetector are respectively used to receive light and convert it into electrical signals, and then subtract the two electrical signals to perform balanced zero-beat detection and subtract common mode noise;

[0028] Spectrum analyzer, used to analyze the noise power spectrum of the electrical signal after balanced subtraction, and then obtain the compression degree of vacuum compressed light;

[0029] The computer has a built-in neural network algorithm for machine learning, which can use the compression degree of vacuum compressed light as the optimization target, perform adaptive search and multi-parameter optimization, and the prediction results after calculation will be updated to subsequent instruments in real time through communication cables;

[0030] A signal generator, used to generate a stable analog electrical signal from the multi-channel digital signal transmitted by the computer machine learning algorithm, and quickly feed it back to the aforementioned coherent light source, acousto-optic modulator, and Helmholtz coil;

[0031] The temperature controller is used to receive the temperature control signal transmitted by the computer after calculation, so as to realize the function of correcting the temperature of the rubidium atom gas chamber in real time;

[0032] Specifically:

[0033] Coherent light source, line width of 200kHz, stable optical power, and linearly polarized light. Acousto-optic modulator, with up to 80% acousto-optic diffraction conversion efficiency, and device response bandwidth of 2MHz.

[0034] The rubidium atomic gas chamber needs to be thermostatically controlled, and the temperature generally needs to be controlled within the range of 40-75°C to ensure sufficient atomic density. In addition, the end faces of both sides of the glass gas chamber are coated with an anti-reflection film with a transmittance of more than 90%, reducing the loss of vacuum compressed light.

[0035] The Helmholtz coil is a highly symmetrical structure. When powered, a uniform parallel magnetic field is generated between the coils to enhance the interaction between light and atoms. The coil is made of 99.9% pure copper.

[0036] Driven by an external 0-10V voltage, the piezoelectric ceramic can change the relative phase difference between the strong pump light and the vacuum compression light, and scan out the optimal compression degree of the quantum vacuum compression light. The piezoelectric ceramic maintains a 1Hz triangle wave scanning frequency.

[0037] The second photodetector and the third photodetector perform signal differential subtraction to subtract common mode noise and reduce interference. In order to perform balanced zero-beat detection, the two light intensities entering the detector must maintain a 50:50 balance, and the detector photoelectric conversion quantum efficiency must reach more than 93% to reduce the detection loss of vacuum compressed light.

[0038] The machine learning program part receives the noise power spectrum from the spectrum analyzer in real time, and obtains the compression information of the quantum vacuum compression light source from it; when the compression of the experimental system is lower than the set threshold (usually set to 4dB), the artificial neural network will dynamically search the parameters such as the quantum system pump light power, frequency, magnetic field, temperature, etc. until the compression converges to near the stable maximum value; in this process, the neural network algorithm in machine learning begins to predict and construct the relationship between parameters and quantum compression, and updates the parameters through the built-in gradient descent algorithm, and judges whether the numerical convergence is achieved through high-order derivatives. The neural network predicts the current optimal compression and optimal parameters each time, thereby continuously shrinking the parameter range predicted and updated by the neural network around the optimal parameters, and finally the quantum compression of the system will converge to a steady state for a long time.

[0039] like Figure 2 As shown in the figure, the part in the box on the right is the specific cycle process of the machine learning algorithm. When the vacuum compression collected by the experimental system is lower than the set threshold, it indicates that the quantum compression light source has deviated from the maximum compression at this time. At this time, the neural network algorithm part (machine learning program) begins to search quickly within a given range, so that the quantum compression of the vacuum compression light field approaches the global optimal value. When the optimal target value is found (meeting the end condition of the neural network algorithm), at this time, the target quantity collected by the compression source hardware system is used for real-time calculation to continuously correct the pump light intensity, frequency, atomic gas chamber temperature, and magnetic field parameters; usually, the training samples collected in the initial few cycles are not enough for the neural network algorithm to predict the optimal compression. After 30 cycles, the parameter search results of the neural network will gradually begin to converge, and the parameters will be corrected in real time in a small range around the optimal value. In this way, the quantum vacuum compression light experimental system is locked at the maximum compression for a long time. During the whole process, the machine learning program part and the hardware part have been cyclically interacting to control the system compression degree from drifting with environmental factors. This is the specific process of machine learning enhancement.

[0040] The neural network algorithm is integrated in the python code and belongs to the program part of the present invention. Figure 2 As shown, in the present invention, the classification conditions are set as follows: (S s -S0)>0.1. Among them, the quantum compression degree S of the current experimental system s The difference S from the reference maximum compression Q s -S0. When S s -S0 is less than 0.1, the neural network algorithm will start to run and find the optimal value according to the loss function criterion. In the present invention, the loss function is set to t0 is the start time of the neural network, t1 is the current time, S s(t) represents the quantum compression degree of the vacuum compression light source at any time t. When the loss function begins to converge and reaches a stable extreme value, the neural network algorithm ends the search and can predict the pump light intensity, frequency, atomic gas chamber temperature, and magnetic field parameters corresponding to the optimal compression degree of the experimental system. Figure 2 The end condition in : the loss function L reaches a stable minimum (no smaller L value can be found in 30 consecutive iterations), which indicates that the compression degree of the experimental system is also stable.

[0041] The machine learning algorithm used in the present invention is the feedforward neural network FNN in the Python Tensorflow open source package. The hybrid machine learning algorithm of the present invention is based on the FNN neural network, uses its built-in differential evolution algorithm to generate initial training samples, uses the mean square error loss function as the optimization criterion, and Figure 2 The process shown interacts with the vacuum compression light source hardware system in real time. This unique artificial neural network control architecture plays a core role in the optimization process of quantum vacuum compression light sources. Neural network algorithms have practical applications in frontier fields of physics such as NV spin reading, optical pulse prediction, topological invariant search, quantum gas preparation, multi-parameter estimation, cold atom BEC, etc. [4-10] , which has been developed in the fields of biology, chemistry, and computer science. However, in the locking and optimization of vacuum compressed light, the use of similar frameworks is the first time. Figure 1 The experimental system for optimizing the vacuum compression light source is shown in the figure, and the neural network algorithm control program is designed according to the specific experimental parameters and optimization goals. The corresponding algorithm flow chart is shown in the attached figure. Figure 2 shown.

[0042] The present invention also proposes a quantum vacuum compressed light source method based on machine learning enhancement, which utilizes the above-mentioned quantum vacuum compressed light source optimization experimental system. In the method, the machine learning program is combined with the above-mentioned quantum vacuum compressed light source experimental system, and the noise and corresponding compression of the vacuum compressed light are corrected in real time in the order of parameter initialization, spectrum data acquisition, compression extraction, neural network training, parameter prediction update, feedback voltage transmission, and modulation device response, thereby reducing the long-term slow drift caused by factors such as optical path jitter, ambient temperature, and magnetic field interference, and improving the long-term stability of the 795nm vacuum compressed light source near the maximum compression working point.

[0043] The interaction between the machine learning algorithm and the vacuum compression light source implemented by the present invention is not a simple splicing-type series connection. Instead, the neural network algorithm is looped in parallel to meet the real-time feedback requirements of the quantum vacuum compression light source experiment. There are two main difficulties: First, it is necessary to select appropriate classification conditions. The classification conditions determine the prediction and parameter search effects of the neural network algorithm. If the classification conditions are set too broadly, the search and training time of the neural network algorithm may be too long, resulting in the inability to achieve real-time locking in the experiment. If the classification conditions are too precise, once the maximum compression locking point is jumped out due to sudden external disturbances, the algorithm's feedback will not be able to pull the working point back to the locked position. Therefore, the classification conditions (S s -S0)>0.1 is reasonably set according to the actual dynamic locking effect of the experimental device in the present invention, which can not only shorten the training steps of the neural network, but also resist the interference of multiple factors, and improve the robustness of the quantum compression light source when it is running near the working point. Second, the neural network algorithm needs to keep the optimized optimal parameters locked in real time. In order to ensure that the optimal parameters can reflect the optimal position of the experimental system, it is necessary to select the parameters for optimization that meet the experimental conditions according to the actual experimental situation. For example, there is a mapping relationship between the compression degree of the quantum vacuum compression light source and the pump light intensity, frequency, atomic gas chamber temperature, magnetic field parameters, etc. Therefore, the pump light intensity, frequency, atomic gas chamber temperature, and magnetic field parameters are selected as optimization parameters to ensure that the system can respond to parameter changes.

[0044] The specific method includes the following steps:

[0045] Step 1: Run the laser to heat the rubidium atom gas chamber to 50-75°C. Then adjust the optical path, scan the piezoelectric ceramic so that the interference signal can be observed on the second photodetector and the third photodetector, and then adjust the reflector to make the interference contrast as close to 100% as possible, and then turn the wave plate to reduce the light intensity of one path to vacuum. Then use the digital operation locking module to lock the electro-optical intensity modulator so that the intensity of the pump transmission light remains unchanged each time it is locked;

[0046] Step 2: After completing step 1, a balanced zero-beat detection is performed using the 50:50 light intensity detected by the second photodetector and the third photodetector. The detection signal enters a spectrum analyzer to obtain the compression degree of the quantum vacuum compressed light at a 1 MHz frequency band at different scanning phases.

[0047] Step 3: Initialize the machine learning parameters. By setting the laser frequency, pump light intensity, rubidium atomic gas chamber magnetic field, and temperature as optimization parameters, and setting the maximum number of iterations and parameter boundary values, the neural network algorithm is used to generate the initial training set data. Each time a round of parameters is updated, the laser frequency, pump light intensity, rubidium atomic gas chamber magnetic field, and temperature will be controlled by the signal generator and temperature controller respectively. The quantum vacuum compressed light system also responds and starts to collect the differential signal of the photodetector at the next moment and perform spectrum analysis;

[0048] Step 4: Based on the initial training set, the neural network algorithm in machine learning begins to predict and construct the relationship between parameters and quantum compression, and updates the parameters through the built-in gradient descent algorithm. It uses high-order derivatives to determine whether the numerical convergence has occurred, and predicts the optimal compression and optimal parameters. These optimal parameters and optimal parameters will continue to be updated and added to the training set.

[0049] Step 5: After machine learning predicts the optimal parameters, continue to obtain the next round of newly collected quantum compression data to determine whether the compression degree has deviated significantly from the previous optimal threshold, and decide whether a new round of parameter updates and function model predictions are needed;

[0050] Theoretical formula for the degree of compression of vacuum compressed light [1-2] as follows:

[0051] S=10lg(1-glsinφcosφ+g 2 l 2 cos 2 φ),

[0052] Among them, S is the compression degree of the compressed light, l is the action length of the rubidium atomic gas cell, g is the interaction coefficient between light and atoms, and φ is the scanning phase of the vacuum compressed light.

[0053] The beneficial effects of the present invention compared with the prior art are as follows:

[0054] The present invention uses machine learning to enhance the quantum compression of the traditional 795nm quantum vacuum compression light source and the long-term stability of the system. In the experiment, the compression of the vacuum compression light source reached a maximum of 4.2dB. Compared with the original 3.2dB, the compression increased by 1dB. And the compression of the quantum compression light source remained stable at 4.0dB±0.1dB for more than 1 hour.

[0055] By introducing machine learning technology, the ability of quantum vacuum compression light source to modify parameters and resist environmental interference is greatly enhanced. When the compression degree of the vacuum compression light source gradually deviates from the optimal position under external influence, it can still adaptively search and find the optimal locking parameters again. The robustness of the system is greatly improved; under the condition of long-term continuous acquisition, thanks to the above-mentioned adaptive neural network method, the noise compression degree working index of the light source is optimized. In the present invention, the data results of continuous acquisition for 1.1 hours are shown in the attached figure. Figure 4 As shown in the figure: By comparison, it can be found that the stabilization time has increased by 1 hour, which reflects that the noise slow drift has been significantly reduced, that is, the long-term stability of the experimental system has been significantly improved. Figure 3 It can be seen that the short-term noise spectrum at 1MHz shows that the quantum compression degree after dynamic optimization of multi-parameters using machine learning technology is at least 1dB higher than before optimization, and the performance of the compressed light source is further improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without paying any creative work.

[0057] Figure 1 This is a structural diagram of the machine learning enhanced quantum vacuum squeezed light source system in the present invention.

[0058] Figure 2 It is a schematic flow chart of the quantum vacuum squeezed light source method based on machine learning enhancement in the present invention.

[0059] Figure 3 It is a comparison diagram of the noise spectrum at 1 MHz of the vacuum compression light source before and after optimization in the present invention.

[0060] Figure 4 This is a comparison chart of the long-term stability of the compression degree of the vacuum compression light source before and after using machine learning technology.

[0061] Figure 5 This is a graph showing the changes in various parameters when the machine learning method is used to dynamically optimize the quantum vacuum compression light source in the present invention. DETAILED DESCRIPTION

[0062] The invention is further described in detail with reference to the following specific embodiments and drawings. The process, conditions, experimental methods, etc. for implementing the present invention, except for the contents specifically mentioned below, are all common knowledge and common common sense in the art and are not particularly limited by the present invention.

[0063] The present invention proposes a machine learning enhanced quantum vacuum compression light source system. By introducing an artificial neural network, the system can automatically search for various experimental parameters that affect quantum compression, and feed back the predicted optimal compression value to the hardware system for locking. After a certain period of iterative cycles, the system quantum compression will gradually converge to a stable maximum value, and each experiment will achieve dynamic stability. Compared with the traditional quantum compression light source, the present invention combines the technology of artificial intelligence to significantly suppress the noise slow drift of the quantum system, thereby improving the long-term robustness and compression index of the system. The specific experimental device includes a coherent light source, an acousto-optic modulator, a single-mode optical fiber, a Glan laser prism, a first polarization beam splitter, a first focusing lens, a rubidium atomic gas chamber, a Helmholtz coil, a magnetic shield, a second focusing lens, a quarter wave plate, a first half wave plate, a second polarization beam splitter, a first reflector, a piezoelectric ceramic, an electro-optic intensity modulator, a third polarization beam splitter, a first photodetector, a digital operation locking module, a second half wave plate, a fourth polarization beam splitter, a second photodetector, a third photodetector, a spectrum analyzer, a computer, a signal generator, and a temperature controller. The present invention has good application prospects in gravitational wave observation, geographical survey, electromagnetic sensing, distributed sensing, deformation detection, etc., and has important value in the fields of quantum precision measurement and industrial production.

[0064] Reference Figure 1 The machine learning enhanced quantum vacuum compressed light source system structure of the present invention includes: a coherent light source 1, an acousto-optic modulator 2, a single-mode optical fiber 3, a Glan laser prism 4, a first polarization beam splitter 5, a first focusing lens 6, a rubidium atomic gas chamber 7, a Helmholtz coil 8, a magnetic shield 9, a second focusing lens 10, a quarter wave plate 11, a first half wave plate 12, a second polarization beam splitter 13, a first reflector 14, a piezoelectric ceramic 15, an electro-optic intensity modulator 16, a third polarization beam splitter 17, a first photodetector 18, a digital operation locking module 19, a second half wave plate 20, a fourth polarization beam splitter 21, a second photodetector 22, a third photodetector 23, a spectrum analyzer 24, a computer 25, a signal generator 26, and a temperature controller 27;

[0065] Among them, the coherent light source 1 is a linearly polarized light with stable power. After passing through the single-mode optical fiber 3 and the Glan laser prism 4, the spatial mode and the purity of the linearly polarized light are further purified. After the incident pump light is focused and enters the rubidium atomic gas chamber, light and atoms interact to produce vacuum compression light. Subsequently, the vacuum compression light and the strong pump light are separated by the second polarization beam splitter 13. By scanning the relative phase between the two, balanced zero-beat detection can be performed on the second photodetector 22 and the third photodetector 23. The detection signal is transmitted to the spectrum analyzer 24 and read by the computer, and then passed to the neural network algorithm. From the attached Figure 3From the noise spectrum of the spectrometer shown in the figure, the computer can extract the compression degree of the vacuum compressed light. The neural network uses quantum compression as the optimization target and continuously and dynamically adjusts the parameters related to it, such as pump light intensity, frequency, and atomic gas chamber magnetic field, to find the optimal quantum compression degree and keep it stable. In theory, the formula for expressing the compression degree of vacuum compressed light is as follows: S = 10lg (1-glsinφcosφ+g 2 l 2 cos 2 φ), where S is the compression degree of vacuum compressed light, l is the action length of the rubidium atomic gas chamber, g is the interaction coefficient between light and atoms, and φ is the scanning phase of vacuum compressed light. Obviously, under the conditions of fixed atomic gas chamber length and global phase scanning, the compression degree is mainly related to the interaction coefficient between light and atoms. This coefficient is mainly affected by actual experimental factors such as light intensity, polarization, frequency, temperature, and magnetic field. Therefore, before the introduction of machine learning optimization technology, it is very susceptible to interference from external factors, is not very stable, and is prone to drift. After the introduction of neural networks in machine learning, the quantum vacuum compression source continuously updates and corrects various experimental parameters during the dynamic cycle to resist drift, so that it can remain stable for more than 1 hour. At the same time, the neural network has a global search function, which can jump out of the local maximum value and reach a stable global maximum compression value. This is also a new breakthrough that machine learning technology has brought to the traditional quantum compressed light source preparation technology.

[0066] When the system is in operation, if the quantum compression gradually begins to deviate from the set threshold of 4.0dB, the machine learning algorithm will use the pump light intensity, frequency, and magnetic field as parameters to search in a small range and correct the quantum experimental system back to the predicted maximum compression value, and then continue the cycle until it converges completely.

[0067] The complete process of the quantum vacuum squeezed light source method based on machine learning enhancement described in the present invention is as follows: Figure 2 As shown, first, run the laser to heat the rubidium atomic gas chamber 7 to 62°C. Then adjust the optical path, scan the piezoelectric ceramic 15 so that the interference signal can be observed on the second photodetector 22 and the third photodetector 23, and then adjust the reflector to make the interference contrast as close to 100% as possible, and then rotate the wave plate to reduce the light intensity of one path to vacuum. Then use the digital operation locking module 19 to lock the electro-optical intensity modulator 16 so that the intensity of its pump transmission light remains unchanged each time it is locked; during detection, the 50:50 light intensity detected by the second photodetector 22 and the third photodetector 23 is used for balanced zero-beat detection, and the detection signal enters the spectrum analyzer 24 to obtain the compression degree of the quantum vacuum compressed light at the 1MHz frequency band under different scanning phases, as shown in the attached figure. Figure 3 As shown;

[0068] After the hardware system is adjusted, the machine learning parameters are initialized. By setting the laser frequency, pump light intensity, rubidium atom gas chamber magnetic field, and temperature as optimization parameters, and setting the maximum number of iterations and parameter boundary values, the neural network algorithm is used to generate the initial training set data. Each time a round of parameters is updated, the laser frequency, pump light intensity, rubidium atom gas chamber magnetic field, and temperature will be controlled by the signal generator 26 and the temperature controller 27 respectively. The quantum vacuum compressed light system also responds and starts to collect the photodetector differential signal at the next moment and perform spectrum analysis;

[0069] Based on the initial training set, the neural network algorithm in machine learning begins to predict and construct the relationship between parameters and quantum compression, and updates the parameters through the built-in gradient descent algorithm. It uses high-order derivatives to determine whether the numerical convergence has occurred, and predicts the optimal compression and optimal parameters. These optimal parameters will continue to be updated and added to the training set.

[0070] Step 5: After machine learning predicts the optimal parameters, continue to obtain the next round of newly collected quantum compression data to determine whether the compression degree has deviated significantly from the previous optimal threshold, and decide whether a new round of parameter updates and function model predictions are needed;

[0071] In summary, the above-mentioned machine learning-based method will significantly improve the continuous operation stability and quantum compression degree of quantum vacuum squeezed light sources. This method is also universal and can be carried out in complex environments, which will help improve the application value of quantum light sources in actual measurement environments.

[0072] The neural network in the machine learning method uses a feedforward fully connected neural network, which iterates parameters according to the principle of back propagation BP and uses the mean square error loss function as the optimization criterion. According to the actual operation of the quantum vacuum squeezed light source, the specific end condition is set as follows: no smaller loss function value can be found after 100 consecutive iterations. The training set data of the neural network is provided by the quantum compression degree and pump light intensity, frequency, and x-direction magnetic field parameters collected at different times. The initial training set is generated by the neural network with random numbers within a given boundary.

[0073] The core part of the machine learning hybrid algorithm used in the present invention is provided by the following python computing package:

[0074] M-LOOP: Machine Learning Online Optimization Package, is an open source code package on Git-hub that provides a basic framework and interface for combining neural network algorithms with traditional instrument control.

[0075] TensorFlow: Google's machine learning package, used to build the neural network part of the hybrid algorithm in this invention;

[0076] Scikit-learn: Python open source scientific computing package used to perform multi-dimensional array operations on data collected by hardware systems.

[0077] Example

[0078] Press Attach Figure 1 The optical path structure shown in the figure, this embodiment specifically implements the method of the present invention in the quantum vacuum compression light source hardware system. During the experiment, first adjust the optical path, and place the rubidium atomic gas chamber in a magnetic shielding environment with a temperature of about 62°C. The entire experimental optical path is fixed on an optical platform. The incident pump light intensity is initially set at 16mW and the frequency is 377.1051THz. In order to prevent the detector from being oversaturated, the intensity of the pump transmission light is locked at 1mW during detection. Then, use two detector photoelectric probes to perform balanced zero-beat detection. As shown in the attached figure Figure 3 In this embodiment, the quantum compression degree of the noise spectrum of quantum vacuum compressed light at 1 MHz is tested before and after optimization by the machine learning method. The scanning time of the spectrometer is set to 1 s, and the resolution bandwidth RBW is 10 kHz.

[0079] In this embodiment, a slow scan voltage of 10V and a frequency of 1Hz is applied to the piezoelectric ceramic so that the lowest noise point of the vacuum compression light can be displayed on the spectrum. In order to avoid interference caused by changes in laser phase jitter, the phase scanning amplitude must be greater than one cycle during the entire experiment. However, due to the inevitable influence of optical path jitter, pump light intensity, frequency, magnetic field, ambient temperature, and loss changes, the actual output compression of quantum vacuum compression cannot reach 4dB, and the compression degree will drift slowly over time. The quantum vacuum compression light source must be locked when it is actually measured or injected into other systems. In this embodiment, a machine learning neural network algorithm is used to dynamically optimize the experimental system for multi-parameter adjustment. During the optimization process, a signal generator is used to feedback to the system. Figure 1 The voltage parameter range of the laser shown is 0-1.2V, the voltage parameter range of the acousto-optic modulator is 0-10V, and the constant voltage of the Helmholtz coil is 0-0.6V. Figure 5 As shown in the figure, during the machine learning optimization process, the parameters of pump light intensity, frequency, and magnetic field (normalized) are dynamically adjusted over time, constantly approaching the direction of maximum compression, narrowing the parameter range, and finally converging to a stable state. Compared with the original system, the compression of the quantum compression light source enhanced by machine learning reaches up to 4.2dB, an increase of about 1dB.

[0080] Comparative Example

[0081] In the above embodiments, the long-term stability change and compression degree of quantum vacuum squeezed light before and after the enhancement by machine learning technology were tested and compared. During the experiment, the noise compression degree at 1 MHz was mainly measured. Both cases were compared with the shot noise of the coherent light source.

[0082] As attached Figure 3 As shown in the figure, both with and without machine learning optimization show obvious noise compression compared with the shot noise of the classical coherent light field. Compared with the 3.2dB of the original system, the compression degree of the quantum compressed light source enhanced by machine learning reaches up to 4.2dB (MLL in the figure), an improvement of about 1dB. This shows that machine learning machines can overcome the shortcomings of traditional single parameter adjustment that easily fall into local maximum values, realize the search and prediction of global maximum compression, and help quantum multi-parameter nonlinear systems reduce noise. In addition, Figure 4 As shown, the changes in the compression of the quantum compressed light source system before and after the use of machine learning technology are monitored respectively. It can be found that the vacuum compressed light system before optimization is more likely to drift due to changes or instability in experimental parameters such as light intensity, frequency, magnetic field, and temperature. After the dynamic search of machine learning technology reaches convergence, the system can remain stable for a long time, the stabilization time can reach more than 1 hour, and there is no need to sacrifice the compression degree, and the compression degree can be stabilized at about 4.0dB. Therefore, the quantum vacuum compressed light source system and method based on machine learning in the present invention can effectively help some bottlenecks of quantum light sources under the existing technology, so that its long-term stability and noise compression degree in practical applications are improved.

[0083] References:

[0084] 1.Bao G,Wu S,Liu S,et al.Enhancement of the signal-to-noise ratio ofanatomic magnetometer by 10dB[J].Physical ReviewApplied,2019,11(5):054075.

[0085] 2.Yu Z,Liu S,Guo J,et al.Enhancing vacuum squeezing via magneticfield optimization[J].Optics Express,2022,30(10):17106-17114.

[0086] 3.Matsko A B,Novikova I,Welch G R,et al.Vacuum squeezing in atomicmedia via self-rotation[J].Physical ReviewA,2002,66(4):043815.

[0087] 4.Qian P,Lin X,Zhou F,et al.Machine-learning-assisted electron-spinreadout of nitrogen-vacancy center in diamond[J].Applied Physics Letters,2021,118(8):084001.

[0088] 5.Lohani S,Knutson E M,Zhang W,et al.Dispersion characterization andpulse prediction with machine learning[J].OSAContinuum,2019,2(12):3438-3445.

[0089] 6.Zhang P,Shen H,Zhai H.Machine learning topological invariants withneural networks[J].Physical review letters,2018,120(6):066401.

[0090] 7.Barker A J,Style H,Luksch K,et al.Applying machine learningoptimization methods to the production ofa quantum gas[J].Machine Learning:Science andTechnology,2020,1(1):015007.

[0091] 8. Tranter AD, Slatyer HJ, Hush MR, et al. Multiparameter optimization of a magneto-optical trap using deep learning [J]. Nature communications, 2018, 9(1): 4360.

[0092] 9. Carleo G, Cirac I, Cranmer K, et al. Machine learning and the physical sciences[J]. Reviews of Modern Physics, 2019, 91(4):045002.

[0093] 10.Wigley PB, Everitt PJ, van den HengelA, et al. Fast machine-learning online optimization of ultra-cold-atom experiments[J]. Scientific reports, 2016, 6(1):25890.

[0094] The protection content of the present invention is not limited to the above embodiments. Without departing from the spirit and scope of the inventive concept, changes and advantages that can be thought of by those skilled in the art are included in the present invention and are protected by the attached claims.

Claims

1. A machine learning enhanced quantum vacuum squeezed light source system, characterized in that: include: Coherent light source (1), acousto-optic modulator (2), single-mode optical fiber (3), Glan laser prism (4), first polarization beam splitter (5), first focusing lens (6), rubidium atom gas chamber (7), Helmholtz coil (8), magnetic shield (9), second focusing lens (10), quarter wave plate (11), first half wave plate (12), second polarization beam splitter (13), piezoelectric ceramic (15), electro-optic intensity modulator (16), third polarization beam splitter (17), first photodetector (18), digital operation locking module (19), second half wave plate (20), fourth polarization beam splitter (21), second photodetector (22), third photodetector (23), spectrum analyzer (24), computer (25), signal generator (26), temperature controller (27); wherein: The coherent light source (1) is a 795nm linearly polarized light generated by a semiconductor laser, which is used as pump light for quantum vacuum compression light. After interacting with atoms, it excites vacuum compression light. The acousto-optic modulator (2) has high conversion efficiency and can generate multi-order diffraction by using the acousto-optic effect to adjust the light intensity of the incident light; The single-mode optical fiber (3) plays a role in adjusting the spatial distribution of the light spot, integrating the light spot after acousto-optic modulation into a Gaussian light spot, thereby ensuring the purity of the spatial mode; The Glan laser prism (4) has a high extinction ratio and is used to improve the linear polarization purity of the incident laser. The high extinction ratio refers to an extinction ratio greater than 1:10000. The first polarization beam splitter (5) is used as a polarizing device, and together with the Glan laser prism (4) ensures that the incident light satisfies vertical linear polarization; The first focusing lens (6) is used to focus the light spot, and the interaction between the focused light spot and the atoms will be stronger; The rubidium atom gas chamber (7) is a cylindrical glass bubble filled with rubidium atom gas and used as a medium for generating quantum vacuum compressed light; The Helmholtz coil (8) is used to apply a magnetic field to the rubidium atoms to enhance the interaction between the light spot and the atoms; The magnetic shield (9) is made of Permalloy and is used to shield the rubidium atomic gas chamber (7) and the Helmholtz coil (8) from crosstalk from the external environmental magnetic field, thereby ensuring the stability of the compression performance of the quantum compression light source; The second focusing lens (10) is used to readjust the light beam into a parallel light beam; The quarter wave plate (11) is used to adjust the polarization state of the parallel light beam; The first half-wave plate (12) cooperates with the quarter-wave plate (11) to adjust the polarization angle of light, ensure that the emitted light maintains vertical polarization as much as possible, and suppress horizontal polarization light leakage; The second polarization beam splitter (13) plays a beam splitting role, separating the vacuum compression light and the pump light through orthogonal polarization; The piezoelectric ceramic (15) is used to scan the relative phase between the strong pump light and the vacuum compression light; The electro-optic intensity modulator (16) can control the intensity of the emitted light through voltage, thereby ensuring that the intensity of the emitted light remains constant while other system parameters change; The third polarization beam splitter (17) is used to combine the vacuum compressed light and the strong pump light; The first photodetector (18) is used to monitor the light intensity of the transmitted part of the pump light in real time and transmit the detection signal to the digital operation locking module (19); The digital operation locking module (19) internally contains a proportional integral differential PID operation circuit, which can operate the input signal and feedback-lock the light intensity of the transmission part of the pump light; The second half-wave plate (20) is used to rotate the polarization angle of the linearly polarized light by 45 degrees; The fourth polarization beam splitter (21) is used to split the light beam 50:50 to achieve balanced and equal division; The second photodetector (22) and the third photodetector (23) are respectively used to receive light and convert it into an electrical signal, and then subtract the two electrical signals to perform balanced zero-beat detection and subtract common mode noise; The spectrum analyzer (24) is used to analyze the noise power spectrum of the electrical signal after balanced subtraction, thereby obtaining the compression degree of the vacuum compressed light; The computer (25) has a built-in neural network algorithm for machine learning, which can take the compression degree of vacuum compressed light as the optimization target, perform adaptive search and multi-parameter optimization, and the prediction results after calculation will be updated to subsequent instruments in real time through communication cables; The signal generator (26) is used to generate a multi-channel digital signal transmitted by a computer machine learning algorithm into a stable analog electrical signal, and quickly feed it back to the aforementioned coherent light source (1), acousto-optic modulator (2), and Helmholtz coil (8); The temperature controller (27) is used to receive the temperature control signal transmitted after computer calculation, so as to achieve the function of adjusting the temperature in real time.

2. The system according to claim 1, characterized in that The coherent light source (1) has a line width of 200 kHz and is linearly polarized light; and / or, The acousto-optic modulator (2) has an acousto-optic diffraction conversion efficiency of 80%, and the device response bandwidth reaches 2 MHz; and / or, The rubidium atom gas chamber (7) needs to be thermostatically controlled within a range of 40-75° C. to ensure sufficient atomic density; preferably, both side surfaces of the rubidium atom gas chamber (7) are coated with an anti-reflection film with a transmittance of more than 90%, thereby reducing the loss of vacuum compressed light; and / or, The Helmholtz coil (8) is a highly symmetrical structure. After power is supplied, a uniform parallel magnetic field is generated between the coils to enhance the interaction between light and atoms. The coil is made of 99.9% pure copper. and / or, The magnetic shield is a three-layer Permalloy.

3. The system according to claim 1, characterized in that Further including: A first reflector (14); the first reflector (14) is arranged behind the second polarization beam splitter (13) and is used to change the propagation direction of light so that the vacuum compressed light is aligned with a subsequent device.

4. The system according to claim 1, characterized in that The piezoelectric ceramic (15) can change the relative phase difference between the strong pump light and the vacuum compression light when driven by an external voltage of 0-10V, and scan out the optimal compression degree of the quantum vacuum compression light; the piezoelectric ceramic (15) maintains a 1Hz triangle wave scanning frequency.

5. The system according to claim 1, wherein: The second photodetector (22) and the third photodetector (23) perform signal differential subtraction to subtract common mode noise and reduce interference; in order to perform balanced zero-beat detection, the two light intensities entering the detector are kept in a 50:50 balance, and the detector photoelectric conversion quantum efficiency needs to reach 93% to reduce the detection loss of vacuum compressed light.

6. The system according to claim 1, wherein: The machine learning program part of the computer (25) receives the noise power spectrum from the spectrometer (24) in real time, and obtains the compression information of the quantum vacuum compression light source therefrom; when the compression of the experimental system is lower than the set threshold, the artificial neural network will dynamically search the pump light power, frequency, magnetic field, and temperature parameters of the quantum system until the compression converges to near a stable maximum value; in this process, the parameter range predicted and updated by the neural network is continuously reduced, and finally the quantum compression of the system will remain in a steady state for a long time; preferably, the threshold is set to 4dB.

7. A machine learning enhanced quantum vacuum squeezed light source method, characterized in that: The following steps are involved: Step 1: operate a coherent light source (1) and an acousto-optic modulator (2) to heat a rubidium atom gas chamber (7) to 50-75° C.; then adjust the optical path, scan the piezoelectric ceramic (15) so that interference signals can be observed on both the second photodetector (22) and the third photodetector (23); then adjust the reflector so that the interference contrast is as close to 100% as possible; then rotate a quarter wave plate (11) and a first half wave plate (12) to reduce the light intensity of one path to vacuum; then use a digital operation locking module (19) to lock the electro-optic intensity modulator (16) so that the intensity of its pump transmission light remains unchanged each time it is locked; Step 2: using the 50:50 light intensity detected by the second photodetector (22) and the third photodetector (23) to perform balanced zero-beat detection, the detection signal enters the spectrum analyzer (24), and the compression degree of the quantum vacuum compressed light at the 1 MHz frequency band under different scanning phases is obtained; Step 3: Initialization of machine learning parameters. The frequency of the coherent light source (1), the intensity of the pump light, the magnetic field of the rubidium atom gas chamber, and the temperature are set as optimization parameters, and the maximum number of iterations and parameter boundary values ​​are set. The neural network algorithm is used to generate initial training set data. Each time a round of parameters is updated, the frequency of the coherent light source (1), the intensity of the pump light, the magnetic field of the rubidium atom gas chamber, and the temperature will be controlled by the signal generator (26) and the temperature controller (27) respectively. The quantum vacuum compressed light system also responds accordingly and starts to collect the differential signal of the photodetector at the next moment and perform spectrum analysis. Step 4: Based on the initial training set, the neural network algorithm in machine learning begins to predict and construct the relationship between parameters and quantum compression, and updates the parameters through the built-in gradient descent algorithm. It uses high-order derivatives to determine whether the numerical convergence has occurred, and predicts the optimal compression and optimal parameters. The optimal parameters and optimal parameters will continue to be updated and added to the training set. Step 5: After machine learning predicts the optimal parameters, continue to obtain the next round of newly collected quantum compression data to determine whether the compression degree has deviated significantly from the previous optimal threshold, and decide whether a new round of parameter updates and function model predictions are needed.

8. The method according to claim 7, characterized in that The theoretical formula for the compressibility of the vacuum compressed light is as follows: S=10lg(1-glsinφcosφ+g 2 the 2 what 2 (φ), Among them, S is the compression degree of the compressed light, l is the action length of the rubidium atom gas cell, g is the interaction coefficient between light and rubidium atoms, and φ is the scanning phase of the vacuum compressed light; and / or, The implementation of the method needs to rely on the machine learning enhanced quantum vacuum compression light source system described in any one of claims 1-6.

9. The method according to claim 7, characterized in that The specific cycle process of the machine learning algorithm includes: When the collected vacuum compression is lower than the set threshold, it indicates that the quantum compression light source has deviated from the maximum compression. At this time, the neural network algorithm begins to search quickly within a given range, so that the quantum compression of the vacuum compression light field approaches the global optimal value. When the optimal target value is found, that is, the end condition of the neural network algorithm is met, at this time, real-time calculations are performed through the collected target quantities to continuously correct the pump light intensity, frequency, atomic gas chamber temperature, and magnetic field parameters. After 30 cycles, the parameter search results of the neural network gradually begin to converge, and the parameters are corrected in real time in a small range around the optimal value, locking the quantum vacuum compression light experimental system at the maximum compression for a long time. The neural network algorithm and the hardware part interact in a loop all the time to control the system compression degree so that it does not drift with environmental factors.

10. The method according to claim 9, characterized in that The neural network algorithm is integrated in the python code, and the classification conditions are set: (S s -S0)>0.1; where the quantum compression degree of the current system S s The difference S from the reference maximum compression Q s -S0, when S s -S0 is less than 0.1, the neural network algorithm starts to run, and the optimal value is found according to the loss function criterion; the loss function is set to: Among them, t0 is the start time of the neural network, t1 is the current time, S s (t) represents the quantum compression degree of the vacuum compression light source at any time t; When the loss function begins to converge and reaches a stable extreme value, the neural network algorithm ends the search and predicts the pump light intensity, frequency, atomic gas chamber temperature, and magnetic field parameters corresponding to the system's optimal compression degree; the termination conditions include: the loss function L reaches a stable minimum value, which indicates that the compression degree of the experimental system is stable; the stable minimum value is that no smaller L value can be found after 30 consecutive iterations.

Citation Information

Cited By

  • Quantum random number generator based on co-fired ceramic integration

    CN119225690A

  • Miniaturized quantum random number generator based on phase noise

    CN119322603A

  • Rubidium optical pump magnetometer and measuring method thereof

    CN120143273A

  • Artificial intelligence-driven compressed light detection efficiency optimization method

    CN121189194A

  • Arbitrary parameter Gaussian state real-time generation system based on modulator

    CN122546527A