Cross-environment self-adaptive laser coherent combination method, equipment and medium

Abstract

The invention relates to the technical field of laser synthesis, and discloses a cross-environment self-adaptive laser coherent synthesis method, equipment and a medium. The method comprises the following steps: constructing environment distribution containing various parameter combinations, and randomly sampling an environment from the environment distribution at the beginning of each training period; training an adaptive strategy network based on a sampled environment; deploying the trained adaptive strategy network to a laser coherent combination system, and fixing network parameters of the adaptive strategy network; and updating the internal memory state through initial interaction between the self-adaptive strategy network and the current environment, and outputting a phase correction instruction to perform phase control of laser synthesis. According to the method, the environment overfitting problem existing in laser coherent synthesis in traditional reinforcement learning is solved, rapid self-adaption can be achieved only through a small amount of initial interaction, the debugging cost of system deployment and scene switching is greatly reduced, different noise, turbulence intensity and laser array scales can be adapted, and the stable and efficient phase locking control effect is always kept.

Description

Technical Field

[0001] This invention relates to the field of laser synthesis technology, specifically to a cross-environment adaptive laser coherent synthesis method, equipment, and medium. Background Technology

[0002] Laser coherent combining systems are used to precisely and rapidly control the phase of multiple laser beams. Existing active phase control technologies can be mainly divided into two categories: traditional optimization algorithms and machine learning-based intelligent algorithms. However, both of them reveal fundamental flaws when facing complex and ever-changing real-world environments.

[0003] Traditional optimization algorithms are typically represented by Stochastic Parallel Gradient Descent (SPGD). SPGD iteratively optimizes the system by applying physical random perturbations and estimating the gradient direction based on changes in the evaluation function. Its advantages include simple structure and no need for precise phase measurement. However, its inherent limitations restrict its application in high-performance systems: First, the number of convergence steps increases exponentially with the number of laser channels, leading to a rapid decrease in the control bandwidth of large-scale arrays, making it difficult to meet real-time requirements. Second, SPGD relies on physical perturbations for gradient estimation, which not only introduces additional noise, causing the system to continuously jitter near the optimal solution, but also suffers from significant control lag and slow convergence speed in dynamic environments.

[0004] Deep reinforcement learning (DRL), particularly methods based on the Actor-Critic framework, while introducing a decision-making process, generally suffers from a deeper "environment overfitting" problem. Agents are typically trained in a single, fixed simulation environment, resulting in fixed neural network weights after training. This is equivalent to learning an "expert policy" that can only cope with specific noise models, turbulence intensities, and array configurations seen during training. In real-world applications, environments are constantly changing: moving a system from indoors to outdoors causes drastic changes in atmospheric turbulence; prolonged operation alters thermal drift characteristics; and array sizes may be adjusted during system upgrades or maintenance. In these new environments, the "train once, use for life" approach of conventional DRL strategies quickly becomes ineffective, with performance plummeting. The only solution is time-consuming and expensive complete retraining.

[0005] In summary, existing technologies lack an intelligent control paradigm that can autonomously and quickly adapt to new environments. Summary of the Invention

[0006] In view of this, the present invention provides a cross-environment adaptive laser coherent synthesis method, device and medium to solve the above problems.

[0007] To address the above technical problems, this invention provides a cross-environment adaptive laser coherent synthesis method, comprising: Construct an environment distribution containing multiple parameter combinations, and randomly sample the environment from the environment distribution at the beginning of each training cycle; An adaptive policy network with memory capabilities is trained based on sampled environments; the adaptive policy network is used to learn a general laser phase control strategy across environments. The trained adaptive policy network is deployed to the laser coherent synthesis system, and the network parameters of the adaptive policy network are fixed. The adaptive policy network updates its internal memory state through initial interaction with the current environment and outputs phase correction commands to achieve phase control of laser coherent synthesis.

[0008] As an optional approach, constructing an environment distribution that includes multiple parameter combinations involves setting variable ranges for noise parameters, atmospheric turbulence intensity parameters, and laser array configuration parameters, and generating multiple simulation environments by combining these variable ranges to form an environment distribution.

[0009] In practical applications of laser coherent synthesis, noise interference, atmospheric turbulence intensity, and laser array configuration are all environmental variables that affect the phase control effect. These variables can change in various ways depending on the application scenario. If the construction of the environmental distribution does not specifically cover these variables, the trained adaptive policy network will not be able to access the typical environmental characteristics of the real scene and will find it difficult to learn general control capabilities across environments. Setting the variable range of this parameter and combining it to generate a simulation environment can enable the constructed environmental distribution to better replicate the actual environmental changes of laser coherent synthesis.

[0010] As an alternative approach, training adaptive policy networks with memory capabilities includes: At the beginning of each training cycle, the internal memory state of the adaptive policy network is reset, enabling the adaptive policy network to perform a preset number of interactions in the sampled environment, recording the observation data, output actions and corresponding rewards for each step; The policy parameters of the adaptive policy network are updated based on observation data, output actions, and rewards. Repeat the above training cycles until the policy parameters of the adaptive policy network converge and / or the training objective of each training cycle is to maximize the cumulative reward within that cycle.

[0011] Through the above technical solution, since the adaptive policy network has memory capabilities, if the internal memory state is not reset during training, the memory traces of the past environment will interfere with the training process of the current environment. Furthermore, without fixed requirements for interactive data recording and parameter update logic, network training will lack a standardized process, making it difficult to form a stable and universal control strategy. By using maximizing cumulative reward as the training objective, the network's learning direction can align with the beam quality optimization requirements of laser coherent synthesis, ensuring that after the policy parameters converge, the network can stably learn a universal phase control strategy across environments.

[0012] As an optional method, the interaction includes: Receive the far-field diffraction pattern of coherent laser synthesis under the current environment as observation data, and output phase correction command based on the observation data; Obtain the immediate reward after the phase correction command is applied to the environment, and store the observation data, phase correction command, and immediate reward together.

[0013] Through the above technical solution, since the far-field diffraction pattern obtained from laser coherent synthesis is a direct representation of the phase synchronization state, and the phase correction command is the action executed by the network for control, if there is no clear data type of observation and action output form during the interaction process, and no associated storage of interaction-related data, the network will find it difficult to optimize its strategy based on the interaction process. Using the far-field diffraction pattern as observation data allows the network to clearly understand the phase state characteristics of the current environment, and associate and store the observation data, phase correction command, and real-time rewards, thus more clearly perceiving the control effects of different control actions in the corresponding environment.

[0014] As an optional approach, it also includes: after completing the interactive data collection for multiple training cycles, updating the policy parameters of the adaptive policy network based on the observation data, output actions, and rewards of all training cycles, and repeating this step until the policy parameters converge.

[0015] By using the above technical solution, the strategy parameters are updated based on the integrated data from multiple training cycles. This allows the optimization of parameters to cover various typical scenarios in the environment distribution, avoiding overfitting of the network to a single environment and making the optimized strategy parameters more consistent with the characteristics of the overall environment distribution.

[0016] As an alternative approach, deploying the trained adaptive policy network to the laser coherent synthesis system involves fixing the network parameter information of the adaptive policy network and resetting the internal memory state of the adaptive policy network.

[0017] Resetting the internal memory state can completely eliminate interference from historical environmental memories during the training phase, allowing the network to interact with the real environment in its initial state and ensuring that the network can recognize the characteristics of the real environment.

[0018] As an alternative approach, updating the internal memory state through the initial interaction between the adaptive policy network and the current environment includes: Output initial phase correction command, receive observation data and real-time rewards from the laser coherent combining system, update internal memory state based on initial phase correction command, observation data and real-time rewards, and repeat the above operation until the initial interaction is completed a preset number of times.

[0019] Since the adaptation of the network to the real environment after deployment requires a gradual process of identifying environmental characteristics, without standardized initial interaction steps, the network will lack a systematic environmental exploration process and will be unable to quickly capture the core characteristics of the real environment. By performing the initial interaction a preset number of times according to the steps, the internal memory state can be quickly adapted to the characteristics of the real environment. After completing the preset number of initial interactions, the network can accurately perceive information such as noise, turbulence, and array configuration of the real environment.

[0020] As an optional approach, it also includes: After the initial interaction is completed, the adaptive policy network continuously outputs phase correction instructions based on the updated internal memory state; if the environment changes, it updates the internal memory state and adjusts the phase correction instructions through new observation data, output actions and rewards.

[0021] Because factors such as atmospheric turbulence intensity and thermal drift of equipment can cause the environment to change slowly, if only fixed control is completed after the initial interaction, the network will not be able to cope with the dynamic changes in the environment, and the phase control effect will gradually decrease with the environmental changes. After the initial interaction, the continuous output of phase correction commands based on the adapted memory state can enable the system to enter a stable state and ensure the basic effect of laser coherent synthesis. When the environment changes, the memory state is updated by new interaction data and the phase correction commands are dynamically adjusted, so that the network's control strategy can adapt to the environmental changes in real time without offline retraining, thus always ensuring the stability and effectiveness of phase control.

[0022] On the other hand, the present invention also provides an electronic device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the cross-environment adaptive laser coherent synthesis method.

[0023] On the other hand, the present invention also provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of the cross-environment adaptive laser coherent synthesis method.

[0024] The beneficial effects of this invention are as follows: This invention solves the overfitting problem in traditional reinforcement learning for laser coherent synthesis. It eliminates the need for retraining in new environments, enabling rapid adaptation with only a few initial interactions, significantly reducing the debugging costs of system deployment and scene switching. It also significantly improves system robustness and scalability, adapting to different noise levels, turbulence intensities, and laser array sizes, maintaining stable and efficient phase-locked control performance at all times. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the cross-environment adaptive laser coherent synthesis method of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been presented in the various embodiments of the present invention to enable the reader to better understand the present invention. However, the technical solutions claimed in the present invention can be implemented even without these technical details and various changes and modifications based on the following embodiments.

[0027] The main purpose of laser coherent combining technology is to improve both the power and beam quality of the output laser by controlling the phase of multiple laser beams to maintain consistency. In practical applications, laser propagation faces various uncertainties, such as different types of noise interference, variations in the intensity of atmospheric turbulence, and adjustments to the number of laser channels during equipment upgrades. These factors all affect the effectiveness of phase control, so the phase control method must be able to cope with these changes.

[0028] However, the two commonly used active phase control techniques both have significant shortcomings in adapting to different environments. A representative traditional optimization algorithm is the stochastic parallel gradient descent algorithm, which works by applying physical random perturbations to the system and then determining the optimization direction based on changes in the evaluation function, thus making gradual adjustments. This algorithm has a relatively simple structure and does not require precise phase measurement. However, as the number of laser channels increases, the algorithm requires more steps to converge, and the growth rate is rapid. This slows down the control response of large-scale laser arrays, failing to meet the requirements of real-time control. Furthermore, the physical perturbations themselves introduce additional noise into the system, causing the system to remain in a fluctuating state when approaching ideal control performance. Even if the environment changes slowly, the control may fail to keep up with the changes.

[0029] Deep reinforcement learning-based intelligent algorithms typically introduce decision-making logic through the construction of relevant frameworks. While they perform well in single environments, a significant problem arises in practical applications: they can only adapt to the specific environment during training. This is because the training process of most existing deep reinforcement learning models relies on a single and fixed simulation environment. After training, the network parameters are no longer adjusted, essentially learning only control methods specific to that environment. However, real-world environments are often unpredictable: when a system is moved from indoors to outdoors, the intensity of atmospheric turbulence may suddenly change; after prolonged operation, equipment may drift due to heat, affecting its original working state; and the number of laser channels may increase or decrease during system maintenance or upgrades. In these new situations, the original control methods become ineffective, and the control effect drops significantly. This necessitates re-collecting data for complete training, a process that is both time-consuming and resource-intensive, greatly increasing the cost of practical applications.

[0030] To address the shortcomings of existing technologies in environmental adaptability and reduce costs during scene switching, this embodiment is based on meta-reinforcement learning, selecting a meta-agent as the subject performing phase control tasks (the adaptive policy network is the functional carrier for the meta-agent to achieve cross-environmental adaptability). All actions of the meta-agent, including environment identification, historical interaction memory, and phase control decision-making, are implemented through the modular design and operation process of the adaptive policy network. The design concept of this embodiment is not to teach the meta-agent only how to cope with a specific environment, but to enable the meta-agent to quickly adapt to new environments through targeted training of the adaptive policy network. By combining meta-training and online deployment, the goal of quickly adapting to different environments without retraining is achieved. The specific implementation process is as follows: This embodiment first needs to construct an environment distribution containing multiple parameter combinations. Used to generate various simulation environments during meta-training. The results of this distribution should ideally cover various scenarios that laser coherent combining might encounter. In some scenarios, the noise needs to include different types such as Gaussian white noise and pink noise, while allowing the noise intensity and frequency of variation to vary randomly within a reasonable range; from the perspective of the propagation environment, it is necessary to simulate atmospheric turbulence of different intensities by randomly selecting the atmospheric refractive index structure constant. To recreate real propagation conditions; in terms of hardware configuration, it should cover different numbers of laser channels such as 8, 10, and 16, and can also set the situation of randomly shutting down some channels, so as to simulate the working state after equipment failure or upgrade.

[0031] Once the environment distribution is constructed, a periodic training process begins. The purpose of this training cycle is to allow the meta-agent to accumulate cross-environment adaptation experience; the specific settings are not limited in this embodiment. This embodiment defines a training cycle as the meta-agent learning from the environment distribution... A specific environment randomly selected from In this process, the interaction is completed over K time steps to maximize the expected cumulative reward.

[0032] Where r t The agent receives a reward at time step t (or control cycle), using either the bucket power or the peak intensity of the far-field main lobe. θ represents all the parameters (weights and biases) of the agent (i.e., the neural network). These parameters determine the agent's "policy," i.e., what phase correction command it will output when it sees a certain far-field pattern. This represents the specific environment in which the agent is currently located. For example, environment A might be "low turbulence, 8 beams," while environment B might be "high noise, 16 beams with 1 beam failure." Even if the agent performs the same action in different environments, the reward may differ. Let be the discount factor at time step t-1, with a value range of [0,1]. This factor is used to adjust the weight of future rewards in the cumulative reward, and its value increases with time step t. The size will decrease, and the intelligent agent will place more emphasis on current, near-term rewards, while paying less attention to rewards far in the future.

[0033] At the start of each training cycle, a specific environment is randomly selected from the environment distribution. This ensures that each training scenario is unique, allowing the meta-agent to continuously encounter diverse situations. Next, the hidden states of the LSTM in the adaptive policy network are... Reset to the initial value because each cycle corresponds to an independent environment, and resetting the memory can prevent the interaction traces of the previous environment from interfering with the current training; Subsequently, meta-agents in Execution action sequence The corresponding far-field diffraction pattern sequence observed simultaneously and reward sequence The interaction data at each step is fully recorded to form the trajectory. The beam quality index can be measured using the far-field main lobe peak intensity or the power in the bucket. The main training objective is to maximize the cumulative beam quality index within each cycle. Then, based on the collected trajectory data, reinforcement learning algorithms such as PPO are used to update the parameters θ of the Actor-Critic network, allowing the policy to better adapt to various scenarios in the current environment. This process is repeated until the network parameters θ converge and stabilize. At this point, the adaptive policy network possesses a general adaptability that is independent of a specific environment and can flexibly adapt to different scenarios.

[0034] This embodiment considers that the meta-agent, through systematic training in diverse environmental distributions, learns "how to quickly adapt to new environments," rather than specific action combinations for a fixed environment. Therefore, after meta-training, all parameters θ (including network weights and biases) of the adaptive policy network are fixed and no further adjustments are made. Parameter θ determines the phase correction decision logic of the meta-agent when observing a specific far-field diffraction pattern. Its optimization objective already covers various scenarios in the environmental distribution, thus eliminating the need for retraining to adapt to new environments.

[0035] As an optional approach, the architecture of the adaptive policy network in this embodiment is configured as follows: The overall structure adopts an Actor-Critic architecture, which consists of a policy network (Actor) and a value network (Critic). The two share the feature extraction part of the front end, and are equipped with a recurrent structure to give the network memory ability. This allows the meta-agent to evaluate the value of the current state while outputting control actions, ensuring that policy optimization has both a clear direction and effective feedback.

[0036] The network's front end is configured with a convolutional neural network (CNN) encoder, whose input is the far-field diffraction pattern I_t for each time step t. The far-field diffraction pattern reflects the phase synchronization state of multiple lasers, but the original image data contains a lot of redundant information, making it difficult to use directly for policy decision-making. Therefore, this embodiment uses a CNN to perform multi-layer convolution to process it, which can extract high-order spatial features related to the spot distribution and phase error.

[0037] The network architecture uses recurrent neural networks such as Long Short-Term Memory (LSTM) or Gated Recurrent Units (GRUs). Specifically:

[0038] in, It is the hidden state from the previous moment. a is the far-field diffraction pattern feature extracted by the front-end encoder. t-1 and r t-1 These are the phase correction action and the corresponding reward from the previous time step. Through this input combination, the hidden state at the current time step is determined. It can continuously accumulate historical interaction information and current environmental characteristics to form an internal "cognition" of the environment, just like building a dynamically updated memory bank for the meta-agent, enabling it to adjust current decisions based on past experience.

[0039] Hidden state of LSTM output It will connect to two parallel fully connected (MLP) branches: the Actor branch and the Critic branch. The Actor branch is used to generate specific phase control actions and receive... After several fully connected layers of computation, the phase correction values ​​for the N laser beams are output. Considering the physical limitations of the laser phase range, to prevent the correction from exceeding the effective range and causing control failure, a tanh activation function is added at the end of the network to normalize the phase correction to the range of [-π, π], so that the phase adjustment of each laser path conforms to physical laws.

[0040] The Critic branch works in parallel with the Actor branch, and its function is to evaluate the value of the current state and output a scalar value. This scalar value is an expected estimate of the cumulative reward the meta-agent might obtain in the current hidden state, used to reflect the quality of the current phase control strategy. A higher value indicates that the current strategy is moving towards improving the bundle effect; a lower value suggests that the action direction needs to be adjusted. This evaluation result is used to guide the policy update of the Actor branch, reducing meaningless exploration by the agent and improving training efficiency and policy stability.

[0041] During training, the reward r t Defined as the beam quality index for coherent laser combining, this embodiment optionally uses either the far-field main lobe peak intensity or the barrel power. These two indices reflect the phase synchronization effect and beam combining quality of multiple lasers. A higher far-field main lobe peak intensity indicates more concentrated laser energy; a higher barrel power indicates higher effective energy after beam combining. By using these two indices as positive rewards, the meta-agent will naturally learn and correct itself during training in the direction of maximizing beam quality (maximizing cumulative reward), gradually forming a control strategy capable of achieving precise phase synchronization.

[0042] Through periodic training under diverse environmental distributions, the parameters of the adaptive policy network have converged and solidified. The meta-agent learns general adaptability across environments through this process, no longer limited to fixed control logic in a single scenario. Based on this training result, when deploying the meta-agent to a real laser coherent synthesis system, no additional adjustments to the network parameters are required; environmental characteristic identification and policy adaptation can be completed solely through minimal initial interactions with the real environment. Therefore, only the trained meta-agent LSTM hidden state needs to be reset. This allows for direct deployment to real systems: Once in a real-world work environment, the meta-agent achieves phase control through the following three steps.

[0043] The meta-agent first undergoes rapid adaptation, during which all parameters θ of the adaptive policy network remain fixed, and decisions are output solely through forward inference, without any gradient descent or parameter update operations. The agent then outputs an initial phase correction command (phase correction amount) based on the current policy. (As mentioned above, this instruction has been normalized to the [-π, π] interval using the tanh activation function), and then applied to the laser coherent combining system. After executing the instruction, the system will immediately provide feedback on the new far-field diffraction pattern. and the corresponding beam quality index r t , where r t The peak intensity of the far-field main lobe or the power in the barrel is used as a metric to directly reflect the actual effect of the current phase synchronization.

[0044] The agent extracts features from the far-field diffraction pattern It generated in this interaction using a CNN, and then compares it with the phase correction command a from the previous time step. t-1 The corresponding reward r t-1 Input them together into the LSTM recurrent unit, according to the formula Update internal hidden state This step allows the LSTM to continuously integrate historical interaction information with current environmental characteristics, gradually building an internal understanding of the real environment. In typical scenarios, after repeating this interaction process for 5 to 10 control steps, the LSTM's hidden state is sufficient to capture the key characteristics of the current environment, including the type and intensity of noise, the actual state of atmospheric turbulence, and the channel configuration of the laser array. The corresponding agent's output phase correction command will also gradually approach the ideal control effect of the current environment as the hidden state is optimized.

[0045] After rapid adaptation, the hidden state of the LSTM The network has converged to a stable state that matches the current real-world environment. This state fully encodes key information such as the noise characteristics, atmospheric turbulence intensity, and laser array configuration of the current environment, forming a stable internal understanding of the current environment. At this point, the Actor branch of the adaptive policy network is based on this stable hidden state. Continuously outputs precise phase correction values. Since the network parameter θ remains constant, the peak intensity of the far-field main lobe or the power in the barrel remains at a high level, and the laser coherent combining system thus enters a highly efficient and stable phase-locked state.

[0046] However, in real-world applications, the environment is not static. Atmospheric turbulence intensity may change slowly over time, and equipment may overheat during prolonged operation, all of which can affect the synchronization state of the laser phase. This embodiment, however, requires no additional intervention. When new environmental changes occur, the laser coherent synthesis system provides a new far-field diffraction pattern and a corresponding immediate reward. The meta-intelligence then feeds this new interaction data, along with the phase correction command from the previous moment, into the LSTM, continuously updating the hidden state. No gradient descent or network weight adjustments are needed at this stage; the new interaction information is integrated through forward propagation, allowing the hidden state to smoothly adapt to environmental changes, thereby driving the Actor branches to dynamically adjust the phase correction command. No offline retraining or manual debugging is required throughout the process.

[0047] By employing the above approach, this embodiment addresses the overfitting problem inherent in traditional reinforcement learning for laser coherent synthesis. It eliminates the need for retraining in new environments, enabling rapid adaptation with only minimal initial interactions, significantly reducing system deployment and scene switching debugging costs. It also significantly enhances system robustness and scalability, adapting to varying noise levels, turbulence intensities, and laser array sizes, consistently maintaining stable and efficient phase-locked control.

[0048] The embodiments of the present invention have been described in detail above. The various embodiments in the specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section. It should be noted that those skilled in the art can make several improvements and modifications to the present invention without departing from the principles of the invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A cross-environment adaptive laser coherent combining method, characterized in that, include: Construct an environment distribution containing multiple parameter combinations, and randomly sample an environment from the environment distribution at the beginning of each training cycle; An adaptive policy network with memory capabilities is trained based on sampled environments; the adaptive policy network is used to learn a general laser phase control strategy across environments. The trained adaptive policy network is deployed to the laser coherent synthesis system, and the network parameters of the adaptive policy network are fixed. The adaptive policy network updates its internal memory state through initial interaction with the current environment and outputs phase correction commands to achieve phase control of laser coherent synthesis.

2. The cross-environment adaptive laser coherent combining method according to claim 1, characterized in that, The construction of the environment distribution, which includes multiple parameter combinations, involves setting variable ranges for noise parameters, atmospheric turbulence intensity parameters, and laser array configuration parameters, and generating multiple simulation environments according to the variable ranges to form the environment distribution.

3. The cross-environment adaptive laser coherent combining method according to claim 1, characterized in that, The adaptive policy network with memory capabilities includes: At the beginning of each training cycle, the internal memory state of the adaptive policy network is reset, enabling the adaptive policy network to perform a preset number of interactions in the sampled environment, recording the observation data, output actions and corresponding rewards for each step; The policy parameters of the adaptive policy network are updated based on the observed data, output actions, and rewards. Repeat the above training cycles until the policy parameters of the adaptive policy network converge and / or the training objective of each training cycle is to maximize the cumulative reward within that cycle.

4. The cross-environment adaptive laser coherent combining method according to claim 3, characterized in that, The interaction includes: Receive the far-field diffraction pattern of laser coherent synthesis under the current environment as observation data, and output a phase correction command based on the observation data; Obtain the immediate reward after the phase correction command is applied to the environment, and store the observation data, phase correction command, and immediate reward together.

5. The cross-environment adaptive laser coherent combining method according to claim 4, characterized in that, Also includes: After collecting interactive data for multiple training cycles, the policy parameters of the adaptive policy network are updated based on the observation data, output actions, and rewards of all training cycles. This step is repeated until the policy parameters converge.

6. The cross-environment adaptive laser coherent combining method according to claim 1, characterized in that, Deploying the trained adaptive policy network to the laser coherent synthesis system includes fixing the network parameter information of the adaptive policy network and resetting the internal memory state of the adaptive policy network.

7. The cross-environment adaptive laser coherent combining method according to claim 1, characterized in that, The initial interaction between the adaptive policy network and the current environment to update its internal memory state includes: Output an initial phase correction command, receive observation data and real-time rewards from the laser coherent combining system, update the internal memory state based on the initial phase correction command, observation data and real-time rewards, and repeat the above operation until the initial interaction is completed a preset number of times.

8. The cross-environment adaptive laser coherent combining method according to claim 7, characterized in that, Also includes: After the initial interaction is completed, the adaptive policy network continuously outputs phase correction instructions based on the updated internal memory state; if the environment changes, it updates the internal memory state and adjusts the phase correction instructions through new observation data, output actions and rewards.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the cross-environment adaptive laser coherent synthesis method as described in any one of claims 1-8.

10. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the cross-environment adaptive laser coherent synthesis method as described in any one of claims 1-8.

Images