Chip atomic clock based on deep learning

Abstract

The application discloses a chip atomic clock based on deep learning, which comprises a physical system, an electronic servo system and a deep learning system, adopts the deep learning system to improve the precision and stability of the chip atomic clock, and uses the deep learning system to learn and train a mathematical physical model of the deep learning chip atomic clock from a large number of samples through feature learning; the convolutional neural network is used to fit and identify the input-output mapping relationship between external influencing factors and CPT signals, realizes end-to-end signal detection, and provides real-time feedback for the chip atomic clock control system. The application automatically identifies and extracts the key features (such as temperature, magnetic field, etc.) of many external influencing factors of the chip atomic clock through the convolutional neural network model, provides accurate real-time feedback for the chip atomic clock control system, and guarantees the stability of the frequency of the atomic clock system.

Description

Technical Field

[0001] This invention relates to the field of atomic clock technology, and more particularly to a chip atomic clock based on deep learning. Background Technology

[0002] Atomic clocks are currently the most accurate time and frequency standards. Chip-based atomic clocks are atomic clocks with a size and weight comparable to a chip. Because they maintain high-precision time and frequency output while being small, lightweight, and low-power, chip-based atomic clocks have broad application prospects in scenarios with strict requirements on size and power consumption, and high frequency stability. Chip-based atomic clocks are a new generation of autonomous navigation and timing micro-terminals, widely used in cutting-edge technology fields such as positioning, navigation, surveying, communication, and radar.

[0003] The fundamental principle of a chip-based atomic clock is Coherent Population Trapping (CPT). The chip-based atomic clock achieves CPT resonance by using a microwave-modulated laser to obtain coherent two-color light that interacts with atoms in an atomic gas cell. This generates a stable microwave frequency, representing the radiative transition of a specific energy level in the atom, which serves as the spectral signal. Electronic devices then convert this microwave frequency into a standard frequency output convenient for practical applications. A chip-based atomic clock consists of a physical system and a control system. The physical system is the core of the entire chip-based atomic clock, with the laser and atomic gas cell being key components. The laser used in the chip-based atomic clock is a Vertical Cavity Surface Emitting Laser (VCSEL). The operating current, operating temperature, and microwave power of the VCSEL directly affect the CPT resonance frequency, thus directly impacting the stability of the chip-based atomic clock. Furthermore, intracavity losses during laser operation cause the laser cavity temperature to rise, resulting in changes in the laser frequency. Under ideal conditions (where atoms / ions are undisturbed and stationary), the energy level transition frequencies of atoms in an atomic chamber remain constant after laser irradiation, unchanged over time and location. However, in reality, it's impossible to create ideal conditions for atoms; they are subject to various external disturbances: temperature changes, magnetic fields, electric fields, blackbody radiation, etc. These non-ideal conditions all cause changes in the energy level transition frequencies within the atoms. The short-term frequency stability of a chip-based atomic clock depends on laser frequency fluctuations, power fluctuations, and polarization mode transitions, while its medium- to long-term stability is mainly affected by temperature changes, magnetic field changes, mechanical vibrations, and pressure changes. The long-term frequency stability and high environmental adaptability of a chip-based atomic clock directly affect the accuracy and robustness of the timekeeping system.

[0004] Currently, chip-based atomic clocks can only be passively corrected for temperature and external magnetic fields through physical system packaging, temperature-controlled electronic servo systems, and magnetic shielding. Although electronic servo systems can ensure the long-term frequency stability of chip-based atomic clocks under specific conditions, the long-term frequency drift problem remains difficult to overcome due to limitations in data sampling and analysis and passive servo capabilities. Furthermore, they are difficult to adapt to a wider range of environmental conditions, severely limiting their application scope.

[0005] Therefore, ensuring the high precision of atomic clocks and the stable performance of the entire system is a crucial research direction. Summary of the Invention

[0006] This invention proposes a deep learning-based chip atomic clock that integrates various external influencing factors (such as temperature changes, magnetic field changes, mechanical vibration, and pressure changes). It employs a precise and robust correction algorithm to improve the long-term stability and environmental adaptability of the chip atomic clock, breaking through key technologies such as intelligent servo systems. Details are described below:

[0007] A deep learning-based chip atomic clock, comprising: a physical system, an electronic servo system, and a deep learning system.

[0008] A narrow-linewidth frequency discrimination signal (CPT) is obtained through a physical system. An electronic servo system locks the output frequency to the stable atomic energy level within the atomic gas chamber. A deep learning system optimizes the long-term stability of the output frequency and provides real-time feedback. The deep learning system learns the frequency stability of the chip-based atomic clock under different operating conditions and external environments. A mathematical and physical model of the deep learning chip-based atomic clock is established by learning and training from a large number of samples using feature learning.

[0009] A convolutional neural network is used to fit and identify the input-output mapping relationship between external environmental influencing factors and CPT signals, realizing end-to-end signal detection and providing real-time feedback for the chip atomic clock control system.

[0010] The physical system includes: a vertical cavity surface-emitting laser (VCSEL), an atomic gas cell, a λ / 4 glass slide, and a photodetector;

[0011] The atomic gas chamber is fabricated using micro-electromechanical systems (MEMS) technology and needs to be filled with buffer gas to reduce collision frequency shift.

[0012] The electronic servo system includes: a servo loop;

[0013] The deep learning system adopts a convolutional neural network structure, including: an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer.

[0014] Furthermore, the servo loop includes the following components:

[0015] A frequency synthesizer, a loop filter, and a voltage-controlled oscillator together form an RF module.

[0016] The modulation and demodulation module consists of three components: attenuator, amplification and filtering, and microcontroller.

[0017] Specifically, the mathematical and physical model for establishing a deep learning chip atomic clock by learning and training from a large number of samples through feature learning is as follows:

[0018] A deep learning system is used to learn the frequency stability of a chip atomic clock under different operating states (mainly including different operating temperatures and currents of the VCSEL laser, different operating temperatures of the atomic gas chamber, and different operating currents of the C field) and under different external environments (mainly including different temperatures, different magnetic fields, mechanical vibrations, and pressure changes). A deep learning mathematical and physical model of the chip atomic clock is established by learning and training from a large number of samples through feature learning. Reinforcement learning using DQN is used to iterate repeatedly, and the value function is calculated according to its own ∈-greedy strategy. The mathematical and physical model is then updated based on the calculated value function.

[0019] Furthermore, the specific steps of using a convolutional neural network to fit and identify the input-output mapping relationship between external influencing factors and CPT signals to achieve end-to-end signal detection are as follows:

[0020] A servo loop is used to discriminate the crystal oscillator before the microwave chain using the CPT signal obtained after passing through the physical system, locking the crystal's output frequency at a stable atomic energy level difference, thus outputting a stable frequency. Simultaneously, a deep learning system is used to optimize the long-term stability of the chip-based atomic clock based on the real-time CPT signal. The specific optimization method is as follows:

[0021] The mathematical physics model obtained through training is used to analyze the optimal long-term stability under the current external environment in real time. Based on the model, the corresponding feedback quantity of the electronic servo system is obtained. The electronic servo system is then used to perform micro-control feedback adjustment of the working temperature and current of the VCSEL laser, the working temperature of the atomic gas chamber, and the working current of the C field.

[0022] The entire system is integrated using System In a Package (SIP) technology.

[0023] The beneficial effects of the technical solution provided by this invention are:

[0024] 1. This invention uses convolutional neural network deep learning to automatically identify and extract key features (such as temperature, magnetic field, mechanical vibration, pressure change, etc.) among many external influencing factors of chip atomic clock. It uses an electronic servo system to perform micro-control feedback adjustment on the working temperature and current of VCSEL laser, the working temperature of atomic gas chamber, and the working current of C field, thereby providing accurate real-time feedback for chip atomic clock and ensuring the long-term frequency stability of atomic clock system.

[0025] 2. This invention is the first to use deep learning to actively provide accurate real-time environmental feedback for the chip atomic clock, ensuring the long-term stability of the chip atomic clock and enabling it to adapt to a wider range of environments, thus greatly expanding its application scope.

[0026] 3. This invention breaks through the bottlenecks of traditional atomic clocks, such as unstable system frequency, slow adjustment, and large adjustment error caused by the inability to autonomously recognize changes in the external environment. It innovatively realizes a new generation of chip atomic clocks with more stable frequency and wider adaptability. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the structure of a chip-based atomic clock based on deep learning.

[0028] Figure 2 This is a schematic diagram of the structure of a deep learning system;

[0029] Figure 3 This is a schematic diagram of the servo loop structure. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below.

[0031] As an advanced branch of artificial intelligence, deep learning is based on an "end-to-end" learning paradigm and has outstanding advantages in processing multi-fidelity high-dimensional raw data. It provides a new approach to solving complex nonlinear problems and has achieved a series of successful applications in fields such as computer vision, health monitoring, and natural language processing. However, research on the application of deep learning methods in the field of chip atomic clocks has not yet been carried out.

[0032] Deep learning-based chip atomic clocks can expand the application scope of deep learning, build a new generation of autonomous intelligent chip atomic clocks, and fully combine data-driven and model-driven approaches. By training neural network models with good optimization and generalization performance, key features among the many external influencing factors of chip atomic clocks can be automatically identified and extracted. Stable and robust correction algorithms can be developed to provide accurate real-time feedback for chip atomic clock control systems. This is expected to significantly accelerate the optimization and stabilization process of chip atomic clock systems and solve the problems of inaccurate and incomplete data in traditional atomic clocks that rely on machine algorithms. This will overcome the bottlenecks of traditional atomic clocks, such as system frequency instability, slow adjustment, and large adjustment errors caused by the inability to automatically recognize changes in the external environment.

[0033] Therefore, deep learning is employed to improve the accuracy of atomic clocks and the stability of the overall system performance. Within the application scope of deep learning, a new generation of autonomous intelligent chip atomic clocks will be constructed. This will fully integrate data-driven and model-driven approaches, training a neural network model with excellent optimization and generalization performance. This model will automatically identify and extract key features from numerous external influencing factors affecting the chip atomic clock, developing stable and robust correction algorithms to provide precise real-time feedback for the chip atomic clock control system. This is expected to significantly accelerate the optimization and stabilization process of the chip atomic clock system, solving problems such as inaccurate and incomplete data in traditional atomic clocks relying on machine algorithms. This will overcome the bottlenecks of traditional atomic clocks, such as system frequency instability, slow adjustment, and large adjustment errors caused by the inability to automatically recognize changes in the external environment. Currently, no research on chip atomic clocks based on deep learning has been conducted domestically or internationally. This research is pioneering and leading, and the long-term stability and temperature adaptability of the achieved autonomous intelligent chip atomic clock are expected to reach world-leading levels.

[0034] Example 1

[0035] A deep learning-based chip atomic clock, see [link / reference] Figure 1 The structure of a chip atomic clock consists of a physical system, an electronic servo system, and a deep learning system.

[0036] The physical system includes a vertical cavity surface-emitting laser (VCSEL), an atomic gas cell, a λ / 4 glass slide, and a photodetector.

[0037] See Figure 2 The electronic servo system includes: a servo microcontroller circuit, namely: a servo loop. The servo loop contains the following components: a frequency synthesizer, a loop filter, and a voltage-controlled oscillator (the above three components form the RF module); an attenuator, an amplifier and filter, and a microcontroller (these three components form the modulation and demodulation module).

[0038] See Figure 3The deep learning system adopts a convolutional neural network structure, which includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer.

[0039] The working principle of a deep learning-based chip atomic clock is as follows: Figure 1 As shown: DC current and high-frequency microwaves output from the microwave chain are coupled and input into a vertical cavity surface-emitting laser (VCSEL), causing the laser to generate frequency-modulated multicolor light. This light is linearly polarized and becomes circularly polarized after passing through a λ / 4 glass plate, where it interacts with atoms in the MEMS atomic gas cell.

[0040] The energy levels of atoms in the atomic gas chamber are simplified to three levels (ground state 1>, ground state 2>, and excited state 3>). The positive and negative first-order sideband lights in the frequency-modulated multicolor light output from the laser resonate with atomic transitions, respectively exciting transitions from ground state 1> to excited state 3> and from ground state 2> to excited state 3>. When the frequency difference between the positive and negative first-order sideband lights is precisely equal to the frequency difference between ground state 1> and ground state 2> (i.e., the two frequency differences mentioned above are equal), coherent population trapping (CPT) occurs. CPT is an optical phenomenon well-known to those skilled in the art. During CPT resonance, atomic fluorescence radiation is in a dark state, and the corresponding transmitted light is in a bright state. Therefore, when CPT resonance occurs, the detector can detect a narrow-linewidth signal, which is the CPT signal. The narrow linewidth CPT signal obtained after passing through the physical system is used as a frequency discrimination signal by the servo loop to discriminate the crystal oscillator before the microwave chain, so that the frequency output by the crystal oscillator is strongly correlated with the stable atomic energy level frequency difference. The servo loop locks the standard frequency at the center frequency of the frequency discrimination signal, thereby making the crystal oscillator output a stable frequency.

[0041] When the external environment changes, deep learning methods are used to improve the long-term stability of the chip's atomic clock.

[0042] First, a mathematical and physical model of a deep learning chip atomic clock is established by learning and training from a large number of samples through feature learning. The deep learning system is used to learn the frequency stability of the chip atomic clock under different operating states (mainly including different operating temperatures and currents of the VCSEL laser, different operating temperatures of the atomic gas chamber, and different operating currents of the C field) and under different external environments (mainly including different temperatures, different magnetic fields, mechanical vibrations, and pressure changes). The mathematical and physical model of the deep learning chip atomic clock is established by learning and training from a large number of samples through feature learning. Then, reinforcement learning of DQN is used to iterate repeatedly, and the value function is calculated according to its own ∈-greedy strategy. The mathematical and physical model is updated according to the value function.

[0043] A convolutional neural network is used to fit and identify the input-output mapping relationship between external influencing factors and CPT signals, achieving end-to-end signal detection. A servo loop is used to discriminate the crystal oscillator before the microwave chain using the CPT signal obtained after passing through the physical system, locking the crystal oscillator's output frequency at a stable atomic energy level difference, thus outputting a stable frequency. Simultaneously, a deep learning system is used to optimize the long-term stability of the chip atomic clock based on the real-time obtained CPT signals. Specifically, the optimization method is as follows: the mathematical physics model obtained through training analyzes the optimal long-term stability under the current external environment in real time, and obtains the corresponding feedback quantity of the electronic servo system based on the model. The electronic servo system is used to perform micro-control feedback adjustment of the working temperature and current of the VCSEL laser, the working temperature of the atomic gas chamber, and the working current of the C field, thereby achieving end-to-end signal detection and providing accurate real-time feedback for the chip atomic clock control system.

[0044] See Figure 2 The hierarchical structure of the convolutional neural network (CNN) for a deep learning-based chip atomic clock includes an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer. The input layer receives and initially processes the CPT signal data. In the convolutional layers, effective features of the signal spectrum are extracted through convolution calculations. The pooling layers then compress the data to reduce overfitting. After flattening, the neurons in the fully connected layers are connected to the output layer. The CNN can fit and identify the input-output mapping relationship between external influencing factors and the CPT signal, thereby achieving end-to-end signal detection and providing accurate real-time feedback for the chip atomic clock control system.

[0045] Different modulation frequencies are used for different atoms. 87 Rb uses 3.4GHz and 6.8GHz. 85 Rb uses 1.5GHz and 3.0GHz, and cesium atoms use 4.6GHz and 9.2GHz. These values ​​are only used as examples in this embodiment of the invention. They can be set according to the needs of actual applications. This embodiment of the invention does not limit them.

[0046] Furthermore, in order to reduce the collision frequency shift, a buffer gas needs to be introduced into the atomic gas chamber, which is usually an inert gas such as helium, argon, xenon and radon. Different proportions of buffer gas can be selected according to the needs of actual applications, and the embodiments of the present invention do not limit this.

[0047] Preferably, deep learning is performed using convolutional neural networks, and the neural network model is generated using online learning. This means that a complete training dataset is not required at the beginning. As more real-time data arrives, the neural network-based CPT atomic clock environmental recognition model is continuously updated and optimized during operation.

[0048] In its implementation, DQN (deep Q-network) reinforcement learning is further employed. DQN is a method that combines neural networks and Q-learning. Through repeated reinforcement learning iterations, it continuously calculates the value function based on its own ∈-greedy policy and updates the policy network based on the obtained value function.

[0049] In the DQN method, the Q-value is a value function, where Q(s,a) represents the expected reward of taking action a (a∈A) in state s (s∈S) at a certain time. The environment s will provide a corresponding reward r based on the action a. The main idea of ​​the algorithm is to construct a Q-table to store the Q-values, and then select the action that maximizes the reward based on the Q-values. This algorithm is well known to those skilled in the art, and will not be described in detail in this embodiment.

[0050] Example 2

[0051] The following is combined Figure 1 and Figure 2 The solution in Example 1 will be further described below:

[0052] 101: Place the VCSEL laser tube, atomic gas cell, and photodetector on the substrate, ensuring that the optical centers are on the same horizontal line;

[0053] This means ensuring that the optical centers of the VCSEL laser tube, atomic gas cell, and photodetector are aligned at the same height and on the same horizontal line.

[0054] 102: The DC current and microwave output from the microwave chain are coupled (i.e., TEE) and input into the VCSEL laser tube, and output frequency-modulated polychromatic light; the frequency-modulated polychromatic light is converted into circularly polarized light after passing through the λ / 4 glass plate and reacts with the atoms in the atomic gas chamber;

[0055] 103: After passing through the atomic gas cell, when CPT resonance occurs, the detector can detect a narrow linewidth signal. This signal is the CPT signal. After the CPT optical signal is detected by the photodetector, the optical signal is converted into an electrical signal.

[0056] 104: Using a large number of CPT signals as samples for deep learning, a neural network-based CPT atomic clock environment recognition model was obtained through training a convolutional neural network;

[0057] In this embodiment of the invention, there are no restrictions on the parameters of each layer and the size and number of convolutional kernels in the CPT atomic clock environment recognition model based on neural networks. The training process of the model is as follows: obtain CPT signal dataset - data preprocessing - obtain training set - train set model - training ends.

[0058] 105: The servo loop is used to discriminate the frequency of the crystal oscillator before the microwave chain by using the CPT signal obtained after passing through the physical system, so that the output frequency of the crystal oscillator is locked at a stable atomic energy level difference, thereby outputting a stable frequency; at the same time, the crystal oscillator outputs two signals, one of which provides a standard frequency of 10MHz output, and the other of which is output to the microwave chain circuit.

[0059] 106: The CPT atomic clock environment recognition model based on a trained neural network is used to identify the real-time CPT signal, and the optimal long-term stability under the current external environment is analyzed in real time. The corresponding feedback quantity of the electronic servo system is obtained according to the model. The electronic servo system is used to perform micro-control feedback adjustment of the working temperature and current of the VCSEL laser, the working temperature of the atomic gas chamber, and the working current of the C field to improve the overall performance.

[0060] The system feedback quantity is the difference between the set target value and the actual value. The servo loop controls the laser based on the magnitude of the difference, thereby completing the locking.

[0061] 109: Using the DQN algorithm for deep reinforcement learning, the CPT atomic clock environment recognition model based on neural networks is continuously optimized, and finally the optimized chip atomic clock is obtained.

[0062] In this embodiment of the invention, a convolutional neural network is first used for deep learning, and then reinforcement learning is performed after a large amount of data is collected (the reinforcement learning algorithm is based on the DQN (Deep Q Neural Network) algorithm of the convolutional network).

[0063] Unless otherwise specified, the model numbers of the various devices in this embodiment of the invention are not limited, and any device that can perform the above functions is acceptable.

[0064] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0065] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A chip-based atomic clock based on deep learning, characterized in that, The chip atomic clock includes: a physical system, an electronic servo system, and a deep learning system. The narrow linewidth frequency discrimination signal (CPT) is obtained through a physical system. The output frequency is locked to the stable energy level of the atom in the atomic gas cell through an electronic servo system. The long-term stability of the output frequency is optimized and fed back in real time using a deep learning system. A deep learning system is used to learn the frequency stability of the chip atomic clock under different working conditions and different external environments. The mathematical and physical model of the deep learning chip atomic clock is established by learning and training from a large number of samples through feature learning. Based on the trained mathematical physics model, a convolutional neural network is used to fit and identify the input-output mapping relationship between external environmental influencing factors and CPT signals, realizing end-to-end signal detection and providing real-time feedback for the chip atomic clock control system.

2. The chip atomic clock based on deep learning according to claim 1, characterized in that, The physical system includes: a vertical-cavity surface-emitting laser, an atomic gas cell, a λ / 4 glass slide, and a photodetector; the atomic gas cell is fabricated using microelectromechanical systems (MEMS) technology and filled with a buffer gas to reduce collision frequency shift. The electronic servo system includes: a servo loop; The deep learning system adopts a convolutional neural network structure, including: an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer.

3. The chip atomic clock based on deep learning according to claim 1, characterized in that, The electronic servo system includes: A frequency synthesizer, a loop filter, and a voltage-controlled oscillator together form an RF module. The modulation and demodulation module consists of three components: attenuator, amplification and filtering, and microcontroller.

4. The chip atomic clock based on deep learning according to claim 1, characterized in that, The specific steps for establishing a mathematical and physical model for a deep learning chip atomic clock by learning and training from a large number of samples through feature learning are as follows: A deep learning system is used to learn the frequency stability of the chip atomic clock under different working conditions and under different external environments. The mathematical and physical model of the deep learning chip atomic clock is established by learning and training from a large number of samples through feature learning. The reinforcement learning of DQN is used to iterate repeatedly, and the value function is calculated according to its own ∈-greedy policy. The mathematical physics model is then updated based on the obtained value function.

5. A chip atomic clock based on deep learning according to claim 1, characterized in that, The specific steps for using a convolutional neural network to fit and identify the input-output mapping relationship between external environmental influencing factors and CPT signals to achieve end-to-end signal detection are as follows: The servo loop is used to discriminate the crystal oscillator before the microwave chain by using the CPT signal obtained after passing through the physical system, so that the output frequency of the crystal oscillator is locked at a stable atomic energy level difference, thereby outputting a stable frequency. The long-term stability of the chip-based atomic clock is optimized using a deep learning system based on real-time CPT signals. The specific optimization method is as follows: The mathematical physics model obtained through training is used to analyze the optimal long-term stability under the current external environment in real time. Based on the model, the corresponding feedback quantity of the electronic servo system is obtained. The electronic servo system is used to perform micro-control feedback adjustment of the working temperature and current of the VCSEL laser, the working temperature of the atomic gas cell, and the working current of the C field. The entire system is integrated using chip system-in-package technology.

Images