Multi-channel feature inversion method and device for non-Gaussian ultrashort pulse

Abstract

The invention provides a non-Gaussian ultrashort pulse-oriented multi-channel feature inversion method and device, and the method comprises the steps: obtaining a first FROG trace graph, carrying out the multi-channel feature extraction of the first FROG trace graph, and generating a three-channel input feature graph; inputting the three-channel input feature map into an intensity prediction network, outputting a predicted pulse intensity envelope, inputting the first FROG trace map into a phase prediction network, outputting predicted pulse phase distribution, and filtering the predicted pulse intensity envelope and pulse phase distribution to smooth noise; and reconstructing a complex electric field based on the filtered intensity and phase, regenerating to obtain a second FROG trace graph, and calculating an error index between the first FROG trace graph and the second FROG trace graph, the error index comprising an FROG error and an intensity correlation coefficient. According to the invention, the requirement of an advanced ultrafast laser system for high-precision online diagnosis of non-ideal pulses is met, and key technical support is provided for online diagnosis and closed-loop control of the advanced ultrafast laser system.

Description

Technical Field

[0001] This specification relates to one or more embodiments in the field of ultrafast optics and artificial intelligence, and in particular to a multi-channel feature inversion method and apparatus for non-Gaussian ultrashort pulses. Background Technology

[0002] Ultrashort laser pulses are key tools in fields such as precision machining, attosecond science, bioimaging, and strong-field physics, and their accurate characterization is crucial for ensuring the reliability of experiments and systems. As a current mainstream ultrafast pulse measurement technique, FROG (Flatform-Oriented Radiation Spectroscopy) can theoretically reconstruct the intensity and phase information of pulses completely by acquiring two-dimensional time-frequency coupled traces generated by nonlinear processes.

[0003] However, in practical applications, the pulses under test are often not ideally Gaussian, but rather exhibit complex intensity envelopes such as triangular, parabolic, bimodal, or perturbed modulation patterns. These non-Gaussian pulses are commonly found in fiber amplifier outputs, pulse shaping systems, or laser links subject to environmental disturbances. Their steep edges, asymmetric attenuation, and other geometric characteristics pose a significant challenge to accurate diagnosis.

[0004] Traditional FROG inversion algorithms (such as Principal Component Generalized Projection Algorithm PCGPA) are prone to getting trapped in local minima when dealing with such complex envelopes, resulting in slow convergence and a heavy reliance on initial guesses. Even under high signal-to-noise ratio conditions, they struggle to accurately reconstruct inflection points, peaks, or truncation boundaries in intensity curves, leading to waveform distortion. For example, triangular pulses are often smoothed into Gaussian-like shapes, losing crucial information about the leading and trailing edges; parabolic pulses, on the other hand, fail to maintain a strict quadratic profile, affecting the accuracy of subsequent dispersion compensation or nonlinear simulations.

[0005] In recent years, deep learning techniques have been introduced into FROG inversion tasks, significantly improving computational efficiency. However, most existing convolutional neural network-based methods are based on the Gaussian impulse assumption, and the training data only contains variants with slight chirps or small perturbations. This results in insufficient generalization ability when faced with diverse non-Gaussian intensity distributions in real-world scenes, failing to effectively capture the geometric essence of the intensity envelope. Furthermore, most networks treat the FROG trace map merely as a regular image, ignoring the prior information of the global intensity distribution (such as edge gradient distribution), thus lacking the ability to discriminate between different envelope types and adaptive adjustment mechanisms. Summary of the Invention

[0006] This invention describes a multi-channel feature inversion method and apparatus for non-Gaussian ultrashort pulses, which can solve the above-mentioned technical problems.

[0007] According to the first aspect, a multi-channel feature inversion method for non-Gaussian ultrashort pulses is provided, including: Obtain a first FROG trace, wherein the first FROG trace contains multiple non-Gaussian intensity envelope types, including at least two of triangular envelope, parabolic envelope, rectangular envelope, and perturbation envelope; Multi-channel feature extraction is performed on the first FROG trace map to generate a three-channel input feature map; wherein, the three-channel input feature map includes the first FROG trace map, an edge map extracted based on the first FROG trace map, and a gradient magnitude map; The three-channel input feature map is input into the intensity prediction network, which outputs the predicted pulse intensity envelope. The intensity prediction network is a convolutional neural network. The first FROG trace is input into the phase prediction network, which outputs the predicted pulse phase distribution, wherein the phase prediction network is a convolutional neural network. The predicted pulse intensity envelope and pulse phase distribution are filtered to smooth out noise. Based on the filtered intensity and phase, the complex electric field is reconstructed and regenerated to obtain the second FROG trace. The error index between the first FROG trace and the second FROG trace is calculated, and the error index includes FROG error and intensity correlation coefficient.

[0008] Based on the above embodiments, the edge map is obtained by processing the first FROG trace map using the Canny edge detection algorithm; The gradient magnitude map is obtained by calculating the first FROG trace map based on the Sobel operator.

[0009] Based on the above embodiments, the intensity prediction network specifically includes: The input layer is used to receive three-channel input feature maps; The feature extraction module, which includes three levels of convolutional layers and pooling layers, is used to extract local features; The self-attention module is used to guide attention weights by utilizing prior information from the edge map and the gradient magnitude map, focusing on key regions related to the envelope shape; A global average pooling layer is used to aggregate global context information; The regression output layer is used to output a 256-dimensional pulse intensity vector as the predicted pulse phase distribution.

[0010] Based on the above embodiments, the input of the phase prediction network is the first FROG trace map, and its network structure includes a Dropout layer to suppress overfitting, and outputs a 256-dimensional pulse phase vector as the predicted pulse phase distribution through a fully connected regression head.

[0011] Based on the above embodiments, a network training step is also included: During the training phase, the real impulse intensity envelope is used as a label, and the real impulse intensity envelope is adjusted to a scale suitable for network learning by normalization. The Adam optimization algorithm is used to drive the intensity prediction network to learn the mapping relationship from the original FROG trace map to the label in an end-to-end and parameter-updated manner.

[0012] Based on the above embodiments, a network training step is also included: During the training phase, the real pulse phase distribution is used as a label, and the real pulse phase distribution is adjusted to a scale suitable for network learning by normalization. The Adam optimization algorithm is used to drive the phase prediction network to learn the mapping relationship from the original FROG trace map to the label in an end-to-end and parameter-updated manner.

[0013] Based on the above embodiments, it also includes: The joint training of the intensity prediction network and the phase prediction network employs a composite loss function that includes intensity reconstruction error, phase reconstruction error, and FROG trace reconstruction error.

[0014] According to the second aspect, a multi-channel feature inversion device for non-Gaussian ultrashort pulses is provided, comprising: A first processing module is used to acquire a first FROG trace map, wherein the first FROG trace map contains multiple non-Gaussian intensity envelope types, including at least two of triangular envelope, parabolic envelope, rectangular envelope and perturbation envelope. The second processing module is used to perform multi-channel feature extraction on the first FROG trace map to generate a three-channel input feature map; wherein, the three-channel input feature map includes the first FROG trace map, an edge map extracted based on the first FROG trace map, and a gradient magnitude map; The third processing module is used to input the three-channel input feature map into the intensity prediction network and output the predicted pulse intensity envelope, wherein the intensity prediction network is a convolutional neural network; The fourth processing module is used to input the first FROG trace image into the phase prediction network and output the predicted pulse phase distribution, wherein the phase prediction network is a convolutional neural network. The fifth processing module is used to apply filtering to the predicted pulse intensity envelope and pulse phase distribution to smooth noise; The sixth processing module is used to reconstruct the complex electric field based on the filtered intensity and phase, and regenerate it to obtain the second FROG trace. It also calculates the error index between the first FROG trace and the second FROG trace, which includes the FROG error and the intensity correlation coefficient.

[0015] Based on the above embodiments, the edge map is obtained by processing the first FROG trace map using the Canny edge detection algorithm; The gradient magnitude map is obtained by calculating the first FROG trace map based on the Sobel operator.

[0016] Based on the above embodiments, the intensity prediction network specifically includes: The input layer is used to receive three-channel input feature maps; The feature extraction module, which includes three levels of convolutional layers and pooling layers, is used to extract local features; The self-attention module is used to guide attention weights by utilizing prior information from the edge map and the gradient magnitude map, focusing on key regions related to the envelope shape; A global average pooling layer is used to aggregate global context information; The regression output layer is used to output a 256-dimensional pulse intensity vector as the predicted pulse phase distribution.

[0017] Based on the above embodiments, the input of the phase prediction network is the first FROG trace map, and its network structure includes a Dropout layer to suppress overfitting, and outputs a 256-dimensional pulse phase vector as the predicted pulse phase distribution through a fully connected regression head.

[0018] Based on the above embodiments, a seventh processing module is also included for network training: During the training phase, the real impulse intensity envelope is used as a label, and the real impulse intensity envelope is adjusted to a scale suitable for network learning by normalization. The Adam optimization algorithm is used to drive the intensity prediction network to learn the mapping relationship from the original FROG trace map to the label in an end-to-end and parameter-updated manner.

[0019] Based on the above embodiments, a seventh processing module is also included for network training: During the training phase, the real pulse phase distribution is used as a label, and the real pulse phase distribution is adjusted to a scale suitable for network learning by normalization. The Adam optimization algorithm is used to drive the phase prediction network to learn the mapping relationship from the original FROG trace map to the label in an end-to-end and parameter-updated manner.

[0020] Based on the above embodiments, the sixth processing module is further configured to use a composite loss function that includes intensity reconstruction error, phase reconstruction error and FROG trace reconstruction error for the joint training of the intensity prediction network and the phase prediction network.

[0021] According to a third aspect, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the program, it implements a multi-channel feature inversion method for non-Gaussian ultrashort pulses as described in the above technical solutions.

[0022] According to a fourth aspect, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the program, it implements a multi-channel feature inversion method for non-Gaussian ultrashort pulses as described in the above technical solutions.

[0023] In the methods described in the embodiments of this specification, robust, efficient and high-precision reconstruction of diverse intensity profiles is achieved through collaborative design feature preprocessing, attention guidance and global context aggregation mechanisms. This meets the requirements of advanced ultrafast laser systems for high-precision online diagnosis of non-ideal pulses and provides key technical support for online diagnosis and closed-loop control of advanced ultrafast laser systems. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart illustrating a multi-channel feature inversion method for non-Gaussian ultrashort pulses provided by the present invention. Figure 2 This invention provides triangular, parabolic, and sinusoidal perturbation-type FROG trace maps and their reconstruction results using a CNN model. Figure 3 This is a schematic diagram of a module for a multi-channel feature inversion device for non-Gaussian ultrashort pulses provided by the present invention. Detailed Implementation

[0026] The solution provided in this specification will now be described with reference to the accompanying drawings.

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be described below with reference to the accompanying drawings.

[0028] In the description of the embodiments of the present invention, the words "exemplary," "for example," or "for instance" are used to indicate that they are examples, illustrations, or descriptions. Any embodiment or design that is described as "exemplary," "for example," or "for instance" in the embodiments of the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Rather, the use of the words "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a specific manner.

[0029] In the description of the embodiments of the present invention, the term "and / or" is merely a description of the relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, B existing alone, and A and B existing simultaneously. Furthermore, unless otherwise stated, the term "multiple" means two or more.

[0030] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and their variations all mean "including but not limited to," unless otherwise specifically emphasized.

[0031] Currently, traditional FROG inversion algorithms (such as PCGPA) are prone to getting trapped in local minima, have slow convergence, and are sensitive to initial guesses when dealing with complex intensity envelopes. Even under high signal-to-noise ratio conditions, they often struggle to accurately recover inflection points, shoulders, or truncation boundaries in the intensity curve, leading to reconstruction distortion. For example, triangular pulses are easily smoothed into Gaussian-like shapes, losing leading / trailing edge information; parabolic pulses, on the other hand, struggle to maintain accurate quadratic contours, affecting the accuracy of subsequent dispersion compensation or nonlinear process simulations.

[0032] In recent years, deep learning techniques have been introduced into FROG inversion tasks, significantly improving processing efficiency. However, most existing methods based on convolutional neural networks (CNNs) are built on the Gaussian impulse assumption, and the training data only contains variations with slight perturbations or chirps. This results in insufficient generalization ability when faced with diverse non-Gaussian intensity distributions in real-world scenes, making it difficult to effectively capture the geometric features of the intensity envelope. Furthermore, most methods treat the FROG trace map as a regular image, ignoring the prior information about the global intensity distribution (such as edge gradient distribution), leaving the model lacking the ability to recognize and adapt to different envelope shapes.

[0033] In summary, existing FROG trace inversion methods face the following main problems when processing ultrashort laser pulses with complex intensity envelopes, such as triangular, parabolic, bimodal, or multi-perturbation modulations: Traditional iterative algorithms, such as generalized projection methods, easily get trapped in local minima when inverting such non-Gaussian pulses, making it difficult to accurately reconstruct key geometric features such as steep edges, asymmetrical contours, or multi-peak structures, leading to distortion of the intensity envelope. Existing deep learning methods based on convolutional neural networks (CNNs) are mostly trained on Gaussian pulse assumptions, making it difficult for the models to adapt to the complex and varied intensity contour shapes in real-world scenes. Existing deep learning models typically treat FROG traces as ordinary images, failing to effectively integrate prior physical information about pulse intensity distribution, such as envelope shape and gradient features, thus limiting their ability to discriminate and reconstruct complex envelopes.

[0034] Therefore, how to develop an end-to-end inversion method that can explicitly sense and adapt to various complex intensity envelope morphologies, and achieve high-precision, high-intensity fidelity, and accurate phase reconstruction of non-ideal pulses, thereby meeting the urgent need of advanced ultrafast laser systems for pulse online diagnosis, is a technical problem that urgently needs to be solved.

[0035] Therefore, this invention proposes a method for inverting complex intensity envelope ultrashort pulse FROG trace maps based on convolutional neural networks. Geometric structural information of the intensity envelope is explicitly extracted through physics-inspired feature engineering, and an attention mechanism guides the network to focus on time-frequency regions sensitive to envelope morphology. Specifically, the original FROG trace map, its Canny edge map, and gradient magnitude map are fused into a three-channel input, enabling the network to perceive the topological features of the envelope, such as abrupt edge transitions, plateau regions, and symmetry breaks, from the initial stage. A lightweight CNN-attention-global average pooling (GAP) architecture is employed. This design combines the local feature extraction capability of convolutional neural networks, the dynamic focusing capability of self-attention mechanisms on key regions, and the global context perception capability of the GAP layer, thereby effectively integrating global information while preserving local details. High-fidelity reconstruction can be achieved without iteration, accurately recovering complex intensity contours such as triangles, parabolic shapes, and sinusoidal perturbations that are difficult to handle by traditional methods, while simultaneously achieving accurate reconstruction of phase information. This method simultaneously supports the recovery of phase information, and the overall process includes four stages: data construction, feature enhancement, dual-task network training, and pulse reconstruction. Through three mechanisms—cooperatively designed feature preprocessing, attention guidance, and global context aggregation—this method aims to achieve robust, efficient, and high-precision reconstruction of diverse non-Gaussian intensity profiles, such as triangular, parabolic, bimodal, and modulated profiles. This method can provide key technical support for the online diagnosis and closed-loop control of advanced ultrafast laser systems.

[0036] Figure 1This is a flowchart illustrating a multi-channel feature inversion method for non-Gaussian ultrashort pulses proposed in this invention, as shown below. Figure 1 As shown, it specifically includes: Step 110: Obtain the FROG trace map.

[0037] The FROG trace data includes various non-Gaussian intensity envelope types, such as triangular, parabolic, bimodal, and modulated FROG trace data, which are used to construct datasets for training and testing.

[0038] Step 120: Perform multi-channel feature extraction on the FROG trace map to generate a three-channel input feature map.

[0039] Specifically, the three-channel input feature map includes the original FROG trace map, the edge map, and the gradient magnitude map. The original FROG trace map serves as the basic time-frequency distribution information; the edge map is used to extract contour features using edge detection algorithms such as Canny, highlighting the abrupt changes in the intensity envelope; and the gradient magnitude map reflects the intensity of local changes in the time-frequency distribution, enhancing sensitivity to envelope details.

[0040] Specifically, the original FROG trace image is processed by Canny edge detection to extract stripe boundary information and obtain an edge map; gradient magnitude calculation is performed on the original FROG trace image, for example, based on the Sobel operator to capture the rate of intensity change.

[0041] The original FROG trace map, edge map, and gradient map are concatenated into a three-channel input tensor, which is used as the input to the intensity prediction network, explicitly encoding the envelope geometric prior.

[0042] Step 130: Input the three-channel feature map into the intensity prediction network and output the predicted intensity through the CNN-self-attention-GAP structure.

[0043] Specifically, the three-channel feature maps are used to extract local features using CNN, and a self-attention mechanism is introduced to dynamically focus on the key time-frequency regions of the envelope shape. Finally, global average pooling (GAP) is used to aggregate global context information and output the predicted pulse intensity envelope.

[0044] Specifically, the intensity prediction network predicts the intensity envelope shape of an ultrashort laser pulse in the time domain.

[0045] It is a convolutional neural network specifically designed to handle complex non-Gaussian intensity contours (such as triangles, parabolas, bimodals, etc.). The network architecture includes: Feature Extraction (CNN): First, through multiple layers of convolution and pooling operations, local features from shallow to deep layers are extracted from the input image. Key Region Focusing (Self-Attention Mechanism): This is its innovative feature. The network utilizes cues learned from edge maps and gradient maps, dynamically judging and focusing on the regions in the FROG trace map that are most critical to determining the shape of the pulse envelope, such as time-frequency locations corresponding to sharp increases or decreases in intensity, through a self-attention mechanism. Information Aggregation, Global Average Pooling (GAP): The attention-weighted feature maps are globally average pooled, summarizing feature information from all locations to form a compact feature vector that represents the entire pulse envelope. Finally, a fully connected layer maps this feature vector into a 256-dimensional intensity vector. This vector corresponds to the intensity values ​​of the pulse at 256 time points, i.e., the complete temporal intensity envelope curve.

[0046] Specifically, the original FROG trace map is a whole, containing a wealth of information. Conventional convolutional neural networks (CNNs) process all regions equally or can only implicitly learn important features, resulting in low efficiency and poor interpretability. For pulses with complex shapes, key features (such as abrupt edges) may occupy only a small pixel area in the FROG map, easily becoming obscured. Physically guided self-attention mechanisms consist of two parts: physical cues provide guiding information, and the self-attention mechanism performs dynamic focusing. Physical cues are provided by edge maps and gradient magnitude maps. Edge maps delineate the contour boundaries of FROG stripes. Edge responses are strongest where stripes abruptly change, break, or intersect. Gradient magnitude maps quantify the intensity change at each point in the FROG map. The more dramatic the change, the larger the gradient value. For example, the steep leading / trailing edge of a triangular pulse corresponds to areas in the FROG map where stripes rapidly compress or expand; these areas have strong edge responses and large gradient values. The trough of a bimodal pulse may correspond to areas where stripes are blurred or weak, and its gradient changes may also exhibit specific patterns. The intensity prediction network receives both the original FROG map and two cue maps simultaneously at the input stage. Through initial learning via convolutional layers, the network extracts high-level, envelope-geometric-related feature representations from these cue maps. Next, a self-attention mechanism dynamically focuses the network, comparing each location in the feature maps with all other locations in the same map to calculate a relevance score, thus determining which other locations should receive more attention when integrating information. After several convolutional layers, the network learns "what kind of stripes are these" from the original map and "where the boundaries are and where the changes are rapid" from the edge / gradient maps. These features are fused together in the middle layers of the network. Then, attention weights are calculated; the self-attention mechanism computes an attention weight matrix. This matrix generation utilizes the edge and gradient change information contained in the fused features from the previous step. Specifically, higher weights are assigned to locations with strong edge / gradient features and those highly correlated with these strong feature regions. For example, a pixel located in a region with smooth pulses might be found to be highly correlated with a pixel at a steep edge. Therefore, when processing pixels in this flat region, the intensity prediction network assigns high weights to the information of the edge pixels. Using the calculated attention weights, all features are weighted, summed, and reconstructed. As a result, features identified by physical cues as key regions, such as edges, have their influence amplified, while the influence of background or secondary region features is relatively weakened.

[0047] For different input FROG maps, such as triangles vs. parabolic shapes, the intensity prediction network calculates different attention weight maps in real time based on their specific edge and gradient patterns. For triangular pulses, attention is strongly focused on the two narrow regions corresponding to the leading and trailing edges; for parabolic pulses, attention may exhibit a smoother gradient distribution. The intensity prediction network learns specific edge-gradient pattern combinations from the data through training, corresponding to which key geometric features in the pulse envelope, such as inflection points and shoulders, and automatically adjusts its focus.

[0048] Unlike traditional iterative algorithms, the intensity prediction model is a forward inference model. It takes the FROG map as input and outputs the intensity directly without repeated iterative optimization. Therefore, it is extremely fast and completely avoids the problems of getting trapped in local minima and relying on initial guesses.

[0049] Unlike simple deep learning models, intensity prediction networks explicitly guide the network to learn and pay attention to the geometric structure of pulses through multi-channel feature input and self-attention mechanism, thereby having stronger reconstruction ability and higher accuracy for non-Gaussian pulses with complex shapes such as steep edges and plateaus.

[0050] Specifically, first, the intensity prediction network receives a 256×256×3 feature map as input. Then, the feature map undergoes three convolutional-pooling modules for local feature extraction. To enhance the model's ability to focus on key regions, a self-attention mechanism is introduced after feature extraction. This mechanism utilizes edge and gradient cues extracted in the previous feature engineering to guide the attention weights, enabling the network to focus more on regions highly correlated with the envelope shape. Next, a global average pooling layer aggregates the global contextual information of the entire feature map. Finally, a lightweight fully connected regression head outputs a 256-dimensional intensity vector, completing the prediction of the pulse intensity envelope.

[0051] Step 140: Input the original FROG trace image into the phase prediction network and output the predicted phase through a lightweight CNN structure.

[0052] Specifically, the original FROG trace image is input into the phase prediction network, which includes a lightweight CNN that efficiently extracts phase-related features and outputs the predicted pulse phase distribution.

[0053] Joint training or phased training is adopted, the loss function takes into account both the reconstruction errors of intensity and phase, and physical consistency constraints (such as FROG error minimization) are introduced.

[0054] Specifically, the phase prediction network is another core deep learning module alongside the intensity prediction network. It predicts the phase distribution of ultrashort laser pulses in the time domain from the original FROG trace.

[0055] Unlike intensity networks, phase prediction networks take a single-channel raw FROG trace as input.

[0056] The core architecture is a relatively standard lightweight convolutional neural network. The phase prediction network extracts and abstracts features from the FROG map step by step through multiple convolutional layers. Dropout layers are introduced at a deeper level. This is because phase prediction tasks are prone to overfitting to specific noise or patterns in the training data. Dropout forces the network to learn more robust and general features by randomly shutting down some neurons during training, thus significantly improving the model's ability to generalize to unknown data. The network terminates at a fully connected layer, outputting a 256-dimensional phase vector. This vector defines the phase value of the pulse at 256 time points.

[0057] During the training phase, the phase network and the intensity network are typically trained independently end-to-end. Their parameter updates are separate, which gives the model greater flexibility.

[0058] The distribution of phase data (typically around 0 to 2π) is quite different from that of intensity data. Therefore, the phase labels are independently normalized before training to accommodate the network's optimization process.

[0059] During the inference (prediction) phase, the trained intensity network and phase network run in parallel. Their respective output intensity vectors and phase vectors are directly combined to form a complete complex electric field description, which is used for pulse reconstruction and verification.

[0060] The phase prediction network is a dedicated module focused on decoding the hidden phase information in the FROG trace. Its design emphasizes extracting robust features from the global structure and preventing overfitting through regularization techniques such as Dropout. It works in close collaboration with the intensity prediction network, which focuses on geometry, forming an intelligent dual-branch system that enables end-to-end, high-precision, and fast full-waveform reconstruction of ultrashort pulses.

[0061] Specifically, the phase prediction network takes a single-channel, 256×256 original FROG trace as input. Its network structure employs a standard convolutional neural network for feature extraction. Dropout layers are introduced at deeper layers to enhance generalization and suppress overfitting. The network terminates with a fully connected regression head, outputting a 256-dimensional phase vector.

[0062] Before training, the intensity and phase labels of the intensity and phase prediction networks are independently normalized according to their respective data distributions. The Adam optimizer can also be used to train the intensity and phase prediction branches independently end-to-end, meaning the parameter updates of the two branches are separated from each other.

[0063] Specifically, labels need to be prepared for each training sample before training begins.

[0064] For each FROG trace image generated through simulation, the corresponding real temporal intensity envelope used to generate that trace image is known in advance. This real envelope is a vector of length 256, which is extracted as the label that the intensity prediction network needs to learn.

[0065] The phase label is the real temporal phase distribution used to generate the FROG trace map, which is extracted as the learning target of the phase prediction network.

[0066] Intensity labels are typically positive, and their range, for example, from 0 to a certain peak, can vary considerably depending on the pulse energy and shape. Phase labels are usually expressed in radians and are mainly distributed within the cyclic intervals of [-π, π] or [0, 2π].

[0067] The intensity labels of all training samples are aggregated and normalized as a whole, for example, scaled to the [0,1] interval. Simultaneously, the phase labels of all training samples are aggregated and normalized independently as a separate whole, for example, scaled to the [-1,1] interval.

[0068] End-to-end independent training and the Adam optimizer mean that the entire mapping relationship from the original input, FROG trace map / three-channel feature map, to the final output, the predicted intensity / phase vector, is learned directly by the neural network without intermediate manual intervention or step-by-step iteration.

[0069] The training process includes forward propagation, which uses a batch of training data to obtain predicted intensity and phase by passing the intensity prediction branch and the phase prediction branch respectively.

[0070] Loss calculations are performed separately for intensity prediction loss (such as the mean squared error (MSE) between predicted intensity and true intensity labels) and phase prediction loss (such as the MSE between predicted phase and true phase labels or periodic cosine similarity loss). The final total loss is a weighted sum of these two losses.

[0071] Backpropagation and parameter updates are used during backpropagation of the total loss. The Adam optimizer independently updates the trainable parameters of the intensity prediction branch and the phase prediction branch based on the calculated gradients. Although the loss is joint, the gradient flow and parameter updates are performed independently for each branch.

[0072] Adam is an adaptive learning rate optimization algorithm that combines the advantages of momentum and adaptive learning rate, automatically adjusting the learning step size for each parameter. In the training of complex two-branch networks, Adam can effectively handle the different learning dynamics that different layers and branches may require, promoting smooth and fast model convergence.

[0073] Step 150: Apply Savitzky-Golay filtering to the predicted intensity and phase results to smooth out noise.

[0074] Specifically, intensity and phase are predicted on the test set, and Savitzky-Golay filtering is applied to the prediction results to smooth noise. Then, the complex electric field is reconstructed, and finally, the FROG trace is recalculated based on the reconstructed complex electric field and compared with the original trace to verify the consistency and accuracy of the reconstruction results.

[0075] Specifically, Savitzky-Golay filtering is applied to the intensity and phase of the network output to suppress noise interference while preserving the geometric features of the packet profile.

[0076] Step 160: Based on the filtered intensity and phase, reconstruct the complex electric field and regenerate the FROG trace map to verify the consistency of the reconstruction.

[0077] Specifically, based on the filtered intensity and phase, the complex electric field is reconstructed to obtain the time-domain pulse waveform; the reconstructed pulse is then used to regenerate the FROG trace map, which is compared with the original input to calculate the reconstruction error (such as FROG error, intensity correlation coefficient, etc.) and verify the reliability and accuracy of the method.

[0078] This invention achieves accurate reconstruction of the intensity and phase of ultrashort pulses through deep learning. First, FROG trace data containing various complex intensity envelopes (such as triangular, parabolic, bimodal, and modulated types) is generated as the foundation dataset for training and validating the deep learning model. Data preprocessing and feature engineering techniques are used to enhance the original FROG traces, including multi-channel feature fusion (such as edge maps and gradient magnitude maps), to improve the image's information representation and physical interpretability, providing richer input to the network. A two-branch neural network is designed and trained, used to predict pulse intensity and phase separately. Both branches are based on convolutional neural network structures and employ an end-to-end independent training strategy to achieve high-precision reconstruction of the original pulse waveform. The reconstructed results output by the network are compared with actual measurement or simulation data. Quantitative indicators (such as FROG error and intensity correlation coefficient) are used to systematically evaluate the reliability and accuracy of the method under different complexity and diversity conditions.

[0079] Finally, the trained model is deployed in an ultrafast laser system to achieve real-time monitoring and reconstruction of pulse waveforms. Based on the reconstruction results, the dispersion, modulation and other parameters of the laser are adjusted to achieve active optimization and control of the pulse shape.

[0080] Compared to traditional methods, this invention addresses the problem of traditional iterative inversion algorithms easily getting trapped in local minima and suffering from low reconstruction accuracy when dealing with non-Gaussian intensity envelopes such as triangles, parabolic shapes, and sinusoidal perturbations. This invention employs an end-to-end deep learning framework, achieving high-precision recovery of various complex envelope shapes, avoiding the dependence on artificial initial values ​​and the instability of convergence inherent in traditional methods. Unlike existing deep learning methods based on the Gaussian impulse assumption, which suffer from insufficient generalization, this invention constructs a diverse training set containing various typical non-ideal intensity contours and introduces physically inspired multi-channel feature engineering (original trace map, edge map, and gradient map) to explicitly guide the self-attention mechanism to focus on key regions related to envelope geometry, thereby significantly enhancing the model's ability to perceive and reconstruct complex structures. Through a dual-branch independent prediction architecture for intensity and phase, combined with label-independent normalization and the Adam optimization strategy, the detailed features of the intensity contour are effectively restored while ensuring phase reconstruction accuracy. The entire reconstruction process requires no iterative calculations, has a fast inference speed, and is suitable for online diagnosis and real-time feedback control of ultrafast laser systems. It provides efficient and reliable technical support for application scenarios that rely on the accurate characterization of ultrashort pulses, such as monitoring of new pollutants and precision machining.

[0081] Figure 2 This is a schematic diagram illustrating the triangular, parabolic, and sinusoidal perturbation-type FROG traces proposed in this invention and their reconstruction results using a CNN model, as shown below. Figure 2 As shown, it specifically includes: First, construct a diverse training dataset.

[0082] Set the time parameters: number of sampling points Nt=256, time step Δt=2fs, generate time axis t=(-Nt / 2:Nt / 2-1)Δt.

[0083] Based on this, pulse samples with four types of complex intensity envelopes are randomly generated: triangular, parabolic, rectangular, and perturbation-type.

[0084] Based on the FROG trace maps corresponding to the aforementioned pulses, a total of 1000 sets of samples were constructed to form a structured training and testing dataset.

[0085] Next, multi-channel feature extraction and enhancement are performed.

[0086] For each FROG trace image sample, the following processing is performed: the edge map is extracted using the Canny operator to highlight the stripe boundaries; the gradient magnitude map is calculated using the Sobel operator to capture local intensity changes; The original FROG trace map, edge map, and gradient map are concatenated along the channel dimension to form a 256×256×3 three-channel tensor, which is used as the input to the intensity prediction network.

[0087] Subsequently, the dual-branch neural network was trained. The input to the intensity prediction network is a 256×256×3 feature tensor; Specifically, the structure of the intensity prediction network includes a three-level convolutional-pooling module, a self-attention mechanism, global average pooling, and a fully connected layer; The output is a 256-dimensional intensity vector.

[0088] The input to the phase prediction network is a 256×256 original FROG trace map; The phase prediction network has a standard CNN structure with Dropout layers, which can prevent overfitting. The output is a 256-dimensional phase vector.

[0089] Specifically, the Adam optimizer can be used to train the two branches independently end-to-end.

[0090] Finally, the intensity and phase predicted by the network are obtained on the test set. Savitzky-Golay filtering is applied to the prediction results to smooth the noise. The pulse is reconstructed based on the filtered intensity and phase, and the FROG trace is recalculated. The reconstructed trace is visualized and quantitatively compared with the original trace (e.g., SSIM, MAE, etc.) to verify the consistency of reconstruction.

[0091] like Figure 2 As shown, Figure 2 (a), (d), (g), (i) FROG traces of triangles, parabolas, rectangles and sinusoidal perturbations; (b), (e), (h), (k) reconstructed traces after input into the CNN network; (c), (f), (i), (l) predicted curves of the intensity and phase of the real impulse.

[0092] Figure 2 FROG trace maps with triangular, parabolic, and sinusoidal perturbations are presented, along with their reconstruction results using a CNN model. Experimental results show that this invention can achieve high-precision reconstruction under complex envelopes such as triangles, parabolas, rectangles, and sinusoidal perturbations, significantly outperforming the PCGPA algorithm and traditional deep learning methods.

[0093] In summary, based on the above embodiments, robust, efficient, and high-precision reconstruction of diverse intensity profiles is achieved through collaborative design of feature preprocessing, attention guidance, and global context aggregation mechanisms. This meets the requirements of advanced ultrafast laser systems for high-precision online diagnosis of non-ideal pulses and provides key technical support for the online diagnosis and closed-loop control of advanced ultrafast laser systems.

[0094] Figure 3 This is a schematic diagram of a module for a multi-channel feature inversion device for non-Gaussian ultrashort pulses proposed in this invention, as shown below. Figure 3 As shown, it includes: A first processing module is used to acquire a first FROG trace map, wherein the first FROG trace map contains multiple non-Gaussian intensity envelope types, including at least two of triangular envelope, parabolic envelope, rectangular envelope and perturbation envelope. The second processing module is used to perform multi-channel feature extraction on the first FROG trace map to generate a three-channel input feature map; wherein, the three-channel input feature map includes the first FROG trace map, an edge map extracted based on the first FROG trace map, and a gradient magnitude map; The third processing module is used to input the three-channel input feature map into the intensity prediction network and output the predicted pulse intensity envelope, wherein the intensity prediction network is a convolutional neural network; The fourth processing module is used to input the first FROG trace image into the phase prediction network and output the predicted pulse phase distribution, wherein the phase prediction network is a convolutional neural network. The fifth processing module is used to apply filtering to the predicted pulse intensity envelope and pulse phase distribution to smooth noise; The sixth processing module is used to reconstruct the complex electric field based on the filtered intensity and phase, and regenerate it to obtain the second FROG trace. It also calculates the error index between the first FROG trace and the second FROG trace, which includes the FROG error and the intensity correlation coefficient.

[0095] Based on the above embodiments, the edge map is obtained by processing the first FROG trace map using the Canny edge detection algorithm; The gradient magnitude map is obtained by calculating the first FROG trace map based on the Sobel operator.

[0096] Based on the above embodiments, the intensity prediction network specifically includes: The input layer is used to receive three-channel input feature maps; The feature extraction module, which includes three levels of convolutional layers and pooling layers, is used to extract local features; The self-attention module is used to guide attention weights by utilizing prior information from the edge map and the gradient magnitude map, focusing on key regions related to the envelope shape; A global average pooling layer is used to aggregate global context information; The regression output layer is used to output a 256-dimensional pulse intensity vector as the predicted pulse phase distribution.

[0097] Based on the above embodiments, the input of the phase prediction network is the first FROG trace map, and its network structure includes a Dropout layer to suppress overfitting, and outputs a 256-dimensional pulse phase vector as the predicted pulse phase distribution through a fully connected regression head.

[0098] Based on the above embodiments, a seventh processing module is also included for network training: During the training phase, the real impulse intensity envelope is used as a label, and the real impulse intensity envelope is adjusted to a scale suitable for network learning by normalization. The Adam optimization algorithm is used to drive the intensity prediction network to learn the mapping relationship from the original FROG trace map to the label in an end-to-end and parameter-updated manner.

[0099] Based on the above embodiments, a seventh processing module is also included for network training: During the training phase, the real pulse phase distribution is used as a label, and the real pulse phase distribution is adjusted to a scale suitable for network learning by normalization. The Adam optimization algorithm is used to drive the phase prediction network to learn the mapping relationship from the original FROG trace map to the label in an end-to-end and parameter-updated manner.

[0100] Based on the above embodiments, the sixth processing module is further configured to use a composite loss function that includes intensity reconstruction error, phase reconstruction error and FROG trace reconstruction error for the joint training of the intensity prediction network and the phase prediction network.

[0101] According to another embodiment, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the program, it implements a multi-channel feature inversion method for non-Gaussian ultrashort pulses as described in the above technical solution.

[0102] According to another embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed in a computer, causes the computer to perform a multi-channel feature inversion method for non-Gaussian ultrashort pulses.

[0103] Those skilled in the art will recognize that, in one or more of the examples above, the functions described in this invention can be implemented using hardware, software, firmware, or any combination thereof. When implemented in software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium.

[0104] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-channel feature inversion method for non-Gaussian ultrashort pulses, characterized in that, include: Obtain a first FROG trace, wherein the first FROG trace contains multiple non-Gaussian intensity envelope types, including at least two of triangular envelope, parabolic envelope, rectangular envelope, and perturbation envelope; Multi-channel feature extraction is performed on the first FROG trace map to generate a three-channel input feature map; wherein, the three-channel input feature map includes the first FROG trace map, an edge map extracted based on the first FROG trace map, and a gradient magnitude map; The three-channel input feature map is input into the intensity prediction network, which outputs the predicted pulse intensity envelope. The intensity prediction network is a convolutional neural network. The first FROG trace is input into the phase prediction network, which outputs the predicted pulse phase distribution, wherein the phase prediction network is a convolutional neural network. The predicted pulse intensity envelope and pulse phase distribution are filtered to smooth out noise. Based on the filtered intensity and phase, the complex electric field is reconstructed and regenerated to obtain the second FROG trace. The error index between the first FROG trace and the second FROG trace is calculated, and the error index includes FROG error and intensity correlation coefficient.

2. The method according to claim 1, characterized in that, The edge map is obtained by processing the first FROG trace map using the Canny edge detection algorithm; The gradient magnitude map is obtained by calculating the first FROG trace map based on the Sobel operator.

3. The method according to claim 1, characterized in that, The intensity prediction network specifically includes: The input layer is used to receive three-channel input feature maps; The feature extraction module, which includes three levels of convolutional layers and pooling layers, is used to extract local features; The self-attention module is used to guide attention weights by utilizing prior information from the edge map and the gradient magnitude map, focusing on key regions related to the envelope shape; A global average pooling layer is used to aggregate global context information; The regression output layer is used to output a 256-dimensional pulse intensity vector as the predicted pulse phase distribution.

4. The method according to claim 1, characterized in that, The input to the phase prediction network is the first FROG trace map. Its network structure includes a Dropout layer to suppress overfitting, and outputs a 256-dimensional pulse phase vector as the predicted pulse phase distribution through a fully connected regression head.

5. The method according to claim 1, characterized in that, It also includes network training steps: During the training phase, the real impulse intensity envelope is used as a label, and the real impulse intensity envelope is adjusted to a scale suitable for network learning by normalization. The Adam optimization algorithm is used to drive the intensity prediction network to learn the mapping relationship from the original FROG trace map to the label in an end-to-end and parameter-updated manner.

6. The method according to claim 1, characterized in that, It also includes network training steps: During the training phase, the real pulse phase distribution is used as a label, and the real pulse phase distribution is adjusted to a scale suitable for network learning by normalization. The Adam optimization algorithm is used to drive the phase prediction network to learn the mapping relationship from the original FROG trace map to the label in an end-to-end and parameter-updated manner.

7. The method according to claim 1, characterized in that, Also includes: The joint training of the intensity prediction network and the phase prediction network employs a composite loss function that includes intensity reconstruction error, phase reconstruction error, and FROG trace reconstruction error.

8. A multi-channel feature inversion device for non-Gaussian ultrashort pulses, characterized in that, include: A first processing module is used to acquire a first FROG trace map, wherein the first FROG trace map contains multiple non-Gaussian intensity envelope types, including at least two of triangular envelope, parabolic envelope, rectangular envelope and perturbation envelope. The second processing module is used to perform multi-channel feature extraction on the first FROG trace map to generate a three-channel input feature map; wherein, the three-channel input feature map includes the first FROG trace map, an edge map extracted based on the first FROG trace map, and a gradient magnitude map; The third processing module is used to input the three-channel input feature map into the intensity prediction network and output the predicted pulse intensity envelope, wherein the intensity prediction network is a convolutional neural network; The fourth processing module is used to input the first FROG trace image into the phase prediction network and output the predicted pulse phase distribution, wherein the phase prediction network is a convolutional neural network. The fifth processing module is used to apply filtering to the predicted pulse intensity envelope and pulse phase distribution to smooth noise; The sixth processing module is used to reconstruct the complex electric field based on the filtered intensity and phase, and regenerate it to obtain the second FROG trace. It also calculates the error index between the first FROG trace and the second FROG trace, which includes the FROG error and the intensity correlation coefficient.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the multi-channel feature inversion method for non-Gaussian ultrashort pulses as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, causes the processor to perform the multi-channel feature inversion method for non-Gaussian ultrashort pulses as described in any one of claims 1 to 7.

Images