Laser coherent combination method based on graph neural network iterative optimization

Abstract

The invention relates to the technical field of intelligent phase control in the field of laser coherent synthesis, and discloses a laser coherent synthesis method based on graph neural network iterative optimization, which utilizes intelligent reasoning of GNN based on physical prior to replace random probing of SPGD based on physical disturbance, so that each step of iteration is more accurate and faster. Through the scheme, a brand-new iterative optimization mode is realized, the dependence of SPGD on physical disturbance is eliminated, and meanwhile, the model has higher expansibility due to internal physical priori.

Description

Technical Field

[0001] This invention relates to the field of intelligent phase control technology in the field of laser coherent combining, and specifically to a laser coherent combining method based on graph neural network iterative optimization. Background Technology

[0002] Existing technologies typically model the phase control problem of N-beam lasers as an optimization problem, aiming to maximize the far-field coherent combined light intensity. In the SPGD scheme, the system uses hardware platforms such as FPGAs to apply positive and negative random voltage perturbations (U0) to each channel in each iteration. + and U - ), and respectively measure the corresponding far-field light intensity J + and J - Using the difference δJ = J + J - The gradient direction is estimated, and the control voltage is then updated. This process relies on two additional physical measurements, forming a closed-loop optimization. In contrast, CNN-based methods attempt to bypass physical perturbations: they train an end-to-end model using a large amount of data (far-field patterns, phase labels) in the offline phase; in the online phase, only the current single-frame far-field image is input, and the CNN can directly output the phase correction values ​​for each channel, achieving perturbation-free control.

[0003] Currently, (1) the SPGD algorithm is inefficient and introduces noise: each iteration must perform two physical perturbations and measurements, which severely limits the control bandwidth; at the same time, continuous perturbations are equivalent to injecting noise into the system, resulting in a decrease in phase-locked loop accuracy and a jitter in the synthesized energy.

[0004] (2) The CNN method has a "topological blind spot": CNN treats the far-field pattern as an ordinary image and ignores the specific topological structure of the N beams in physical space (such as hexagonal or rectangular arrangement). This neglect of geometric prior makes it difficult for the model to efficiently learn the spatial correlation between beams, resulting in weak generalization ability and poor scalability - when the array size increases, the control accuracy decreases significantly.

[0005] The aforementioned shortcomings are particularly prominent in high-channel-count and high-dynamic environments, and there is an urgent need for a new optimization method that can both avoid physical disturbances and explicitly integrate physical topological priors. Summary of the Invention

[0006] In view of this, the present invention provides a laser coherent synthesis method based on graph neural network iterative optimization to solve the above problems.

[0007] To address the above technical problems, this invention provides a laser coherent synthesis method based on iterative optimization using graph neural networks, comprising: Construct a physical topology diagram of the laser array. The physical topology diagram uses the laser beam voltage control channel as a node, establishes edges based on the beam spatial spacing and configures complex weights. During the iteration process, the edge weights are dynamically updated based on the beam phase difference. A graph neural network gradient prediction model incorporating physical priors was constructed and trained. The model includes a convolutional neural network encoder for extracting global features of far-field images, a GNN incorporating optical laws, and an output layer that outputs gradient prediction vectors. The training adopts a multi-objective hybrid loss function. Initialize the phase modulator driving voltage, acquire the far-field spot image and the current voltage, input them into the trained model to obtain the gradient prediction vector, update the voltage according to the preset step size and send it down, and if the coherent synthesis performance index meets the standard and is stable, the synthesis is completed; otherwise, repeat the acquisition and update steps.

[0008] When establishing an edge, the edge is only constructed when the spatial distance between the two laser beams is less than a preset threshold, which is twice the spot radius.

[0009] The complex weights of the edges include a Gaussian function and a phase propagation factor. The Gaussian function is used to characterize the attenuation of the intensity of the mutual influence between beams as the spatial distance increases; the phase propagation factor is used to reflect the optical path difference phase relationship between beams caused by the difference in propagation distance.

[0010] Dynamically updating edge weights includes: Increase the weight of edges between nodes with large phase differences, so that the graph neural network will prioritize regions with prominent phase errors when passing messages.

[0011] The working process of a convolutional neural network encoder includes: The input single-frame far-field light spot intensity image is subjected to multi-layer convolution and pooling. After extracting high-level visual semantic features, a global feature vector that condenses the overall structural information of the far-field interference pattern is output.

[0012] It also includes initializing the node features of the GNN: The current control voltage of the laser channel corresponding to each node is concatenated with the global feature vector output by the convolutional neural network encoder to obtain the initial features of the node, so that the node contains both local voltage state and global interference context information.

[0013] The integration of optical interference laws is achieved through a phase interaction operator, which is a sine or cosine function of the phase difference between beams at corresponding nodes. At the same time, attention weights that depend on node features and edge weights are combined to perform weighted aggregation of information transmitted by neighboring nodes.

[0014] GNN is a two-level structure with a pre-defined number of layers for learning the local coupling relationship between adjacent beams; its output layer is a shared multilayer perceptron, which is used to map the updated features of each node into scalars, and all scalars are combined to form a gradient prediction vector.

[0015] The multi-objective hybrid loss function includes orientation alignment loss, magnitude regularization term, and physical performance-driven loss; The orientation alignment loss ensures that the predicted gradient is in the same direction as the true gradient, the magnitude regularization term constrains the magnitudes of the two to be close, and the physical performance driven loss ensures that the gradient update can effectively improve the coherent synthesis effect.

[0016] The orientation alignment loss is obtained by calculating the cosine similarity between the predicted gradient and the true gradient; The amplitude regularization term is obtained by calculating the square of the difference between the two L2 norms; The physical performance-driven loss is obtained by calculating the complement of the ratio of the power in the bucket before and after the gradient update.

[0017] As an optional approach, the preset step size is either a fixed value or a value that decreases adaptively with the number of iterations; the coherent synthesis performance index is the power in the bucket, and the standard for achieving the target threshold is that the power in the current bucket reaches the target threshold and remains stable for three or more consecutive iterations.

[0018] The beneficial effects of this invention are as follows: This invention utilizes GNN's intelligent reasoning based on physical priors to replace SPGD's random probing based on physical perturbations, making each iteration more accurate and faster. This approach achieves a novel iterative optimization mode, freeing the model from SPGD's dependence on physical perturbations, while also making it more scalable due to its inherent physical priors. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the laser coherent synthesis method based on graph neural network iterative optimization according to the present invention. Detailed Implementation

[0020] 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.

[0021] Please see Figure 1This embodiment discloses a laser coherent synthesis method based on iterative optimization using graph neural networks. It aims to address the technical problems of low efficiency and noise introduction in existing stochastic parallel gradient descent (SPGD) algorithms, as well as topological blind spots in convolutional neural network (CNN) methods. By explicitly fusing the physical topology prior of the laser array with the intelligent reasoning capability of the graph neural network (GNN), it achieves efficient and perturbation-free laser coherent synthesis control. The specific implementation process is as follows: First, a physical topology graph of the laser array is constructed. This topology graph forms the foundational framework for the subsequent learning of physical laws by the GNN model. During construction, the voltage control channel corresponding to each laser beam is used as a node in the graph, ensuring that each node is directly associated with the control parameters of the specific beam. In this embodiment, the node is defined as follows: Abstract the N laser beams as N nodes in a graph. , , ,…, Node set V: Each node corresponds to the voltage control channel of one beam.

[0022] Edges are constructed based on the spatial spacing between beams. An edge is only constructed when the spatial spacing between two laser beams is less than a preset threshold, which is set to twice the spot radius. This condition filters out beam pairs with significant spatial interaction, avoiding redundant correlations between unrelated beams that could affect model efficiency. Simultaneously, a complex weight is assigned to each edge, consisting of two parts: a Gaussian function characterizing the attenuation of beam interaction intensity as spatial spacing increases (i.e., the larger the spacing, the weaker the beam interaction); and a phase propagation factor reflecting the optical path difference phase relationship caused by the difference in propagation distance between beams, ensuring that the edge weight accurately reflects the physical correlation characteristics between beams.

[0023] As an optional approach, edge features: for each edge Define complex weights

[0024] in It is a Gaussian function representing the intensity of the interaction between beams as a function of distance. It increases while decreasing. k is the wavenumber, used to characterize the phase relationship introduced by the optical path difference. This is the phase factor for free-space spherical wave propagation. σ is a parameter characterizing the spatial correlation attenuation scale.

[0025] In subsequent iterative optimization, the edge weights need to be dynamically updated based on the beam phase difference. Specifically, when the beam phase difference between two nodes is large, the weight of the edge between these two nodes will be increased, so that the graph neural network will automatically prioritize the region with prominent phase error when passing messages, thereby achieving adaptive correction of system error and improving the pertinence of phase control.

[0026] Next, a graph neural network gradient prediction model incorporating physical priors is constructed and trained. In this embodiment, the model includes a convolutional neural network (CNN) encoder, which performs multi-layer convolution and pooling operations on the input single-frame far-field light spot intensity image. Convolution extracts local features from the image, pooling compresses feature dimensions while retaining key information, and finally extracts high-level visual semantic features to form a global feature vector that can condense the overall structural information of the far-field interference pattern, providing a global interference scene context for subsequent node features.

[0027] In this embodiment, the CNN encoder uses a single frame of far-field light spot intensity image. As input, high-level visual semantic features are extracted through multi-layer convolution and pooling operations, and finally a global feature vector is output. This vector encapsulates the overall structural information of the current far-field interferometry pattern, providing context awareness for subsequent node features.

[0028] And GNNs that incorporate optical principles (embedding optical principles such as laser interference and propagation into the structure and computational logic) require initialization of node features before operation. Specifically, this involves concatenating the current control voltage of the laser channel corresponding to each node with the global feature vector output by the CNN encoder. This ensures that the initial node features contain both local voltage state information and global interference context information, laying the foundation for accurate feature updates. During message passing, this layer incorporates optical interference principles through phase interaction operators. These phase interaction operators are sine or cosine functions of the phase difference between the beams represented by the corresponding nodes, directly simulating the physical laws of light field superposition. Simultaneously, attention weights are used to weighted aggregate the information transmitted by neighboring nodes. The calculation of these attention weights depends on the node's own features and the complex weights of the edges, further enhancing the transmission of effective information. In essence, the core of GNNs is their message passing (node ​​update) mechanism. Features in layer l+1 Due to its own characteristics in layer l and all its neighboring nodes ( The information transmitted is aggregated. This process explicitly incorporates physical laws.

[0029] Among them Attention weights depend on node features and edge weights. , It is a phase interaction operator based on physical interference, such as or The phase difference Voltages contained in node characteristics and The calculation yielded this result. This operator is used to explicitly model the superposition law of light fields.

[0030] Furthermore, the GNN in this embodiment is set as a two-level structure, with a network of a specific number of layers preset to specifically learn the local coupling relationship between adjacent beams, ensuring that the model can capture the close-range correlation between beams.

[0031] In this embodiment, the output layer uses a shared multilayer perceptron (MLP). Its function is to map the features of each node after being updated through multilayer message passing into individual scalars. All scalars are combined in the order of the corresponding beam channels to form the final gradient prediction vector, which provides a directional basis for voltage updates.

[0032] As an optional approach, after L-layer message passing, each node obtains a final representation rich in global context and local topological relationships. Then it is mapped to a scalar through a shared MLP. The scalars from all channels are combined to form the final gradient prediction vector. .

[0033] During the model training phase, a multi-objective hybrid loss function is adopted.

[0034] The loss function comprises three parts: orientation alignment loss, magnitude regularization term, and physics-driven loss. The orientation alignment loss is obtained by calculating the cosine similarity between the predicted and true gradients, ensuring their orientations are consistent and preventing the gradient direction from deviating from the optimal optimization path. Its formula is as follows:

[0035] The amplitude regularization term is obtained by calculating the square of the difference between the L2 norm of the predicted gradient and the true gradient. It is used to constrain the amplitudes of the two to be close, preventing excessively large gradients from causing iterative oscillations or excessively small gradients from causing slow convergence. Its formula is:

[0036] The physics-driven loss is obtained by calculating the complement of the ratio of the power in the bucket (PIB) before and after the gradient update. This ensures that the gradient update effectively improves the coherent synthesis effect, preventing the model from merely mathematically fitting the gradient and deviating from actual physical performance requirements. Training with this multi-objective loss function allows the model's predicted gradients to simultaneously meet the requirements of accurate direction, reasonable magnitude, and effective performance. Its formula is:

[0037] Finally, through perturbation-free iterative optimization of closed-loop control, the actual control of laser coherent synthesis is achieved: First, the drive voltage vectors of the N-channel phase modulators are initialized. = Typically, a zero-point bias or midpoint voltage is selected to ensure that the coherent beam synthesis output is within the measurable range in the initial state, and the iteration counter is initialized to t=0.

[0038] Subsequently, a far-field spot image of the current moment is acquired using a CCD camera. At the same time, read the current drive voltage vector from the controller. ,Will( As input to the trained graph neural network gradient prediction model, the CNN... Extracting the global context vector GNNs, based on the topological graph structure and edge weights, incorporate node features... As input, the multi-layer physics-enhanced GNN updates the node features layer by layer according to the node update formula described above, and outputs the gradient components of each node. , forming the predicted gradient vector .

[0039] The controller updates the driving voltage according to the gradient prediction vector in preset steps. The preset step size α can be set to a fixed value or a value that decreases adaptively with the number of iterations. A fixed step size is easy to operate, while an adaptively decreasing step size can quickly approach the optimal value in the early stage of iteration and achieve stable convergence in the later stage. The updated driving voltage is sent to the N-channel phase modulators to adjust the phase state of each beam. Then, the current coherent combining performance index is calculated. This index is set as the power in the barrel, which can intuitively reflect the energy concentration of far-field coherent combining.

[0040] Determine whether the power in the current bucket has reached the preset target threshold and remains stable in three or more consecutive iterations. If it does, it means that the system has achieved stable phase-locked loop and coherent synthesis is complete. If it does not, return to the acquisition step, reacquire the far-field spot image and the current voltage, and repeat the above gradient prediction and voltage update process until the system reaches a stable coherent synthesis state.

[0041] Through the complete implementation process described above, this embodiment does not require applying physical perturbations and performing two light intensity measurements in each iteration as the SPGD algorithm does, thus completely avoiding noise introduced by perturbations and improving control bandwidth. Furthermore, by explicitly modeling the physical topology and incorporating optical laws, it solves the topological blind spot problem of the CNN method, maintaining high control accuracy and good scalability even in high-channel-count and high-dynamic environments, ultimately achieving efficient, accurate, and stable laser coherent synthesis.

[0042] 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 laser coherent synthesis method based on iterative optimization using graph neural networks, characterized in that, include: A physical topology diagram of the laser array is constructed, wherein the physical topology diagram uses the laser beam voltage control channel as a node, establishes edges based on the beam spatial spacing and configures complex weights, and dynamically updates the edge weights according to the beam phase difference during the iteration process. A graph neural network gradient prediction model that incorporates physical priors is constructed and trained. The model includes a convolutional neural network encoder for extracting global features of far-field images, a GNN that incorporates optical laws, and an output layer that outputs gradient prediction vectors. The training adopts a multi-objective hybrid loss function. Initialize the phase modulator driving voltage, acquire the far-field spot image and the current voltage, input them into the trained model to obtain the gradient prediction vector, update the voltage according to the preset step size and send it down, and if the coherent synthesis performance index meets the standard and is stable, the synthesis is completed; otherwise, repeat the acquisition and update steps.

2. The laser coherent combining method based on graph neural network iterative optimization according to claim 1, characterized in that, When establishing an edge, the edge is only constructed when the spatial distance between the two laser beams is less than a preset threshold, which is twice the spot radius.

3. The laser coherent combining method based on graph neural network iterative optimization according to claim 1, characterized in that, The complex weights of the edges include a Gaussian function and a phase propagation factor. The Gaussian function is used to characterize the attenuation of the intensity of the mutual influence between beams as the spatial distance increases; the phase propagation factor is used to reflect the optical path difference phase relationship between beams caused by the difference in propagation distance.

4. The laser coherent combining method based on graph neural network iterative optimization according to claim 1, characterized in that, The dynamically updated edge weights include: Increase the weight of edges between nodes with large phase differences, so that the graph neural network will prioritize regions with prominent phase errors when passing messages.

5. The laser coherent combining method based on graph neural network iterative optimization according to claim 1, characterized in that, The working process of the convolutional neural network encoder includes: The input single-frame far-field light spot intensity image is subjected to multi-layer convolution and pooling. After extracting high-level visual semantic features, a global feature vector that condenses the overall structural information of the far-field interference pattern is output.

6. The laser coherent combining method based on graph neural network iterative optimization according to claim 1, characterized in that, It also includes initializing the node features of the GNN: The current control voltage of the laser channel corresponding to each node is concatenated with the global feature vector output by the convolutional neural network encoder to obtain the initial features of the node, so that the node contains both local voltage state and global interference context information.

7. The laser coherent combining method based on graph neural network iterative optimization according to claim 1, characterized in that, The integration of optical interference laws is achieved through a phase interaction operator, which is a sine or cosine function of the phase difference between beams at corresponding nodes. At the same time, attention weights that depend on node features and edge weights are combined to perform weighted aggregation of information transmitted by neighboring nodes.

8. The laser coherent combining method based on graph neural network iterative optimization according to claim 1, characterized in that, The GNN is a two-level structure with a preset number of layers for learning the local coupling relationship between adjacent beams; its output layer is a shared multilayer perceptron, which is used to map the updated features of each node into scalars, and all scalars are combined to form a gradient prediction vector.

9. The laser coherent combining method based on graph neural network iterative optimization according to claim 1, characterized in that, The multi-objective hybrid loss function includes orientation alignment loss, magnitude regularization term and physical performance driving loss; The orientation alignment loss ensures that the predicted gradient is in the same direction as the true gradient, the magnitude regularization term constrains the magnitudes of the two to be close, and the physical performance driven loss ensures that the gradient update can effectively improve the coherent synthesis effect.

10. A laser coherent combining method based on iterative optimization of a graph neural network according to claim 9, characterized in that, The orientation alignment loss is obtained by calculating the cosine similarity between the predicted gradient and the true gradient; The amplitude regularization term is obtained by calculating the square of the difference between the two L2 norms; The physical performance-driven loss is obtained by calculating the complement of the ratio of the power in the bucket after the gradient update to that before the update.

11. The laser coherent combining method based on iterative optimization of graph neural networks according to claim 1, characterized in that, The preset step size is a fixed value or a value that decreases adaptively with the number of iterations; the coherent synthesis performance index is the power in the bucket, and the standard for achieving the target is that the power in the current bucket reaches the target threshold and remains stable for three or more consecutive iterations.

Images