Rapid inversion method and system for beam parameter distribution of ultrasonic molecular beam

Abstract

The invention provides a beam parameter distribution rapid inversion method and system of an ultrasonic molecular beam, relates to the technical field of feeding control of a magnetic confinement nuclear fusion device, and solves the problem of various limitations of beam parameter distribution acquisition in the existing SMBI technology. The method comprises the following steps: based on historical data, acquiring at least two different types of beam parameter distributions with corresponding relation, and forming an inversion data set; a U-net neural network is used as a basic framework, an attention mechanism module is additionally arranged, and an inversion calculation model is constructed; model training is carried out, so that the inversion calculation model can capture a mapping relation between at least two types of beam parameter distribution; and when a subsequent new experiment is carried out and multiple types of beam parameter distribution need to be measured, the model outputs corresponding inversion results of other required types of beam parameter distribution only by inputting an actual measurement result of one type of beam parameter distribution into the inversion calculation model. According to the invention, rapid inversion prediction of any beam parameter distribution of the ultrasonic molecular beam can be realized.

Description

Technical Field

[0001] This invention relates to the field of feeding control technology for magnetic confinement nuclear fusion devices, specifically to a method and system for rapid inversion of beam parameter distribution of ultrasonic molecular beams. Background Technology

[0002] Supersonic Molecular Beam Injection (SMBI), a key technique for achieving efficient fuel loading in fusion devices, plays an irreplaceable role in improving plasma confinement performance and extending discharge time due to its unique advantages: the ease of operation of conventional gas delivery systems combined with the deep penetration of projectile injection systems. As fusion energy research progresses towards higher parameters and more stable operation, the need for optimization of this technology becomes increasingly urgent. Among these, parameters such as the profile characteristics and internal structure of the fuel gas beam have become core factors determining loading efficiency and plasma response. These parameters directly affect the interaction process between the beam and plasma, thus influencing the overall performance of the fusion device.

[0003] In experimental research and engineering applications, beam density distribution and velocity distribution are recognized as two major physical parameters affecting the effectiveness of SMBI technology. Density distribution determines the spatial aggregation of fuel particles, while velocity distribution is related to the beam's penetration depth and energy transfer efficiency. However, because fusion devices must maintain a high vacuum environment, and the beam itself possesses inherent characteristics of high-speed motion and low particle density, directly obtaining the complete spatial distribution of these parameters presents a significant challenge. Vacuum conditions not only limit the feasibility of traditional contact measurements but also exacerbate signal attenuation and background noise interference, making the accurate acquisition of fully distributed parameters exceptionally difficult.

[0004] Currently, the field primarily relies on schlieren technology to indirectly observe beam density distribution. This method reflects the density gradient through changes in light refraction, but its effective measurement range is limited by the resolution and dynamic response capability of the optical system, making it difficult to cover the complete evolution of the beam from the nozzle exit to the far field. For obtaining velocity distribution, particle tracking technology or discharge-based diagnostic methods are commonly used. The former tracks tracer particles through high-speed imaging, while the latter uses ionization signals generated by the interaction between the beam and residual gas for estimation. However, these methods generally have significant limitations: the spatial coverage of the measurement parameters is narrow, often only sampling local areas of the beam; data processing is time-consuming, failing to meet real-time control requirements; and when simultaneously measuring multiple parameters such as density and velocity, hardware compatibility and data synchronization issues between different diagnostic systems are prominent, leading to highly complex system structures and high maintenance costs.

[0005] The high-speed characteristics of the beam and the coupling effect with the vacuum environment further amplify the measurement challenges. High-speed motion results in extremely short particle dwell times, requiring diagnostic equipment with nanosecond-level response capabilities, while the low-density characteristics lead to weak signal strength, easily submerged by environmental noise. This not only increases the economic costs of equipment development and operation but also makes the parameter inversion process prone to introducing large errors. Furthermore, existing technologies struggle to balance temporal and spatial resolution, often sacrificing one for the other, failing to fully reflect the physical essence of the beam's dynamic behavior. These factors collectively restrict a deeper understanding of the beam formation mechanism and hinder the precise control and performance breakthroughs of SMBI technology in fusion devices.

[0006] In summary, existing diagnostic methods face multiple bottlenecks in acquiring beam distribution parameters, including low efficiency, system redundancy, and insufficient physical interpretability. Achieving rapid and accurate acquisition of key parameters without significantly increasing system complexity has become an urgent need driving the development of ultrasound molecular beam injection technology. This highlights the necessity of developing novel diagnostic strategies to effectively improve the controllability and reliability of the fusion fueling process while ensuring physical rationality. Summary of the Invention

[0007] The purpose of this invention is to address several limitations in the acquisition of beam parameter distribution in existing SMBI (Supersonic Molecular Beam Induction) technology, such as high cost, poor measurement accuracy, and difficulties in multi-device coordination. Therefore, a rapid inversion method and system for beam parameter distribution in ultrasonic molecular beams is proposed. This invention utilizes artificial intelligence technology to simulate the ultrasonic molecular beam nozzle structure and obtain different beam parameter distributions. The algorithm is then trained to capture the mapping relationships between various parameters, achieving full beam parameter distribution calculation, meeting integrated analysis requirements, and providing reliable technical support for building a feeding database.

[0008] The present invention employs the following technical solutions to achieve its objective: A method for rapid inversion of beam parameter distribution of an ultrasonic molecular beam, the method comprising the following steps: S1. Based on historical data of ultrasonic molecular beams, obtain at least two different types of beam parameter distributions that are related to each other, and form an inversion dataset for model training. S2. Using the U-net neural network as the basic framework, an attention mechanism module is added to construct an inversion calculation model; S3. Use the inversion dataset to train the inversion calculation model so that the trained inversion calculation model can capture the mapping relationship between at least two types of beam parameter distributions; S4. When conducting new experiments with ultrasonic molecular beams and needing to measure multiple types of beam parameter distributions, only the measured results of one type of beam parameter distribution are input into the inversion calculation model, and the inversion calculation model will output the corresponding inversion results of the other required types of beam parameter distributions.

[0009] Specifically, the beam parameter distribution of ultrasonic molecular beams includes density parameter distribution, velocity parameter distribution, pressure parameter distribution and temperature parameter distribution. There are corresponding relationships between different types of beam parameter distributions that can form a mapping relationship.

[0010] Optionally, in step S1, the density parameter distribution of the ultrasonic molecular beam is obtained based on historical data. and velocity parameter distribution Density parameter distribution and velocity parameter distribution There is a relationship between them. Corresponding relationship, function This is the mapping function between the two.

[0011] Preferably, in step S1, the historical data of the ultrasonic molecular beam comes from the simulation or experimental measurement of the ultrasonic molecular beam injector. The ultrasonic molecular beam injector has different historical data under different gas source pressure, background pressure, throat diameter and divergence angle. The types of historical data include simulation data and experimental measurement data, which are respectively formed into simulation inversion dataset and experimental measurement inversion dataset.

[0012] Specifically, in step S1, the distribution of multiple beam parameters in the inversion dataset is stored in the form of a visual image.

[0013] Specifically, in step S2, the basic framework of the U-net neural network includes downsampling units, upsampling units, and skip connection units; where: The downsampling unit is used to extract information from the visualized image through the first convolutional layer and the max pooling layer; The upsampling unit is used to restore information in the visualized image through the second convolutional layer and the upsampling layer; Skip connection units are used to transfer information extracted from each layer in the downsampling unit to the corresponding layer in the upsampling unit to prevent the loss of image spatial information.

[0014] Preferably, based on the basic framework of the U-net neural network, three types of attention mechanism modules are added: CA attention mechanism module, SE attention mechanism module and CBAM attention mechanism module. These three types of attention mechanism modules form three corresponding image processing sub-models on the basic framework of the U-net neural network, and the inversion calculation model is constructed based on these three image processing sub-models.

[0015] Specifically, the CA attention mechanism module is used to embed a coordinate-aware feature extraction unit in the encoder path of U-net. It generates a one-dimensional coordinate attention map by performing global average pooling along the height and width directions, calculates the cross-correlation weights of the horizontal and vertical position coordinates in the visualization image of the beam parameter distribution, strengthens the spatial structure representation of the ultrasound molecular beam hotspot region, and outputs the coordinate-weighted feature map to the decoder path of U-net. The SE attention mechanism module is used to insert channel recalibration components in the skip connections of U-net. It compresses the spatial dimension through the Squeeze operation to generate channel descriptors, learns the sensitivity coefficients of the beam parameter distribution through the fully connected layer, dynamically amplifies the feature channel weights related to their intensity gradients, suppresses low-sensitivity channels, and improves the inversion robustness to beam parameter distribution non-uniformity. The CBAM attention mechanism module has additional channel attention submodules and spatial attention submodules. The CBAM attention mechanism module is used to cascade a channel-spatial dual attention mechanism at the end of the decoder path of U-net. First, the channel attention submodule fuses the channel weights of global average pooling and max pooling to filter the key physical features of the beam parameter distribution. Then, the spatial attention submodule uses a pre-sized convolution kernel to generate a binary spatial mask and focus on the core region of the beam parameter distribution, thus collaboratively suppressing background noise and outputting a high-fidelity beam parameter distribution inversion result.

[0016] Preferably, in step S4, the trained inversion calculation model is encapsulated in a GUI visualization so that the inversion calculation model outputs the corresponding inversion results of the beam parameter distribution through the GUI visualization interface; the GUI visualization interface is provided with display boxes for at least two types of beam parameter distributions, a calculation accuracy display box, control icons for selecting input data, and control icons for selecting different sub-models in the inversion calculation model.

[0017] The present invention also provides a computer system, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the aforementioned rapid inversion method for the beam parameter distribution of ultrasonic molecular beams.

[0018] In summary, due to the adoption of this technical solution, the beneficial effects of this invention are as follows: This invention significantly enhances the diagnostic and control capabilities of ultrasonic molecular beam injection systems, enabling the full distribution calculation of beam velocity parameters. It provides a clear visual representation of the beam's spatial contour and internal structural features, offering comprehensive and accurate parameter support for the fuel feeding process in fusion devices. Compared to traditional methods that can only acquire local or discrete data, this invention effectively overcomes the challenge of obtaining various beam parameter distributions in a vacuum environment. This results in more complete and reliable fuel feeding control parameters, optimizing the interaction efficiency between the beam and plasma, and laying the foundation for improving the confinement performance and operational stability of fusion devices. Its rapid response characteristics support real-time operation requirements, enabling the completion of evolutionary mapping analysis of various parameters within milliseconds and timely feedback to the control system. This significantly enhances the dynamic control capability of the fusion experiment process, avoids performance fluctuations caused by data delays, and greatly improves the continuity and reliability of device operation.

[0019] This invention also introduces a user-friendly visual operating environment, allowing users to easily complete parameter settings, model switching, and result interpretation through an intuitive interface. This greatly simplifies complex diagnostic processes, lowers the operational threshold, and improves work efficiency. The generated data profile possesses complete spatial coverage characteristics and can be seamlessly embedded into existing hardware control systems or numerical simulation code frameworks, demonstrating high compatibility and practicality. Supported by mapping technology based on artificial intelligence, the integration capabilities of this invention not only reduce reliance on additional diagnostic equipment but also promote data collaboration between multiple systems. This provides flexible and efficient technical support for the engineering application of fusion energy research, effectively promoting the refined and intelligent development of ultrasonic molecular beam injection technology in fusion devices. Attached Figure Description

[0020] The present invention is further described in detail with reference to the following figures, which include four figures as follows: Figure 1 This is a schematic diagram illustrating the overall process of the rapid inversion method for beam parameter distribution of the present invention; Figure 2 This is a schematic diagram illustrating the logical principle of the fast inversion method for beam parameter distribution of the present invention; Figure 3 This is a layout example diagram of the GUI visual interface in the method of the present invention; Figure 4 This is a schematic diagram illustrating the effect of the GUI visualization interface in presenting the inversion results in the method of the present invention. Detailed Implementation

[0021] 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 clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0022] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0023] A fast inversion method for beam parameter distribution of ultrasonic molecular beams. Figure 1 The overall process of this method is briefly described below, which can be viewed concurrently. The key steps of this method can be summarized as follows: S1. Based on historical data of ultrasonic molecular beams, obtain at least two different types of beam parameter distributions that are related to each other, and form an inversion dataset for model training. S2. Using the U-net neural network as the basic framework, an attention mechanism module is added to construct an inversion calculation model; S3. Use the inversion dataset to train the inversion calculation model so that the trained inversion calculation model can capture the mapping relationship between at least two types of beam parameter distributions; S4. When conducting new experiments with ultrasonic molecular beams and needing to measure multiple types of beam parameter distributions, only the measured results of one type of beam parameter distribution are input into the inversion calculation model, and the inversion calculation model will output the corresponding inversion results of the other required types of beam parameter distributions.

[0024] Figure 2 The logical principles underlying this method in practical applications can be understood concurrently with the detailed description of this embodiment. The corresponding beam parameter distributions can refer to the density and velocity distributions of the same ultrasonic molecular beam, which are mutually inverse, from density distribution to velocity distribution or vice versa. Therefore, in this embodiment, the types of beam parameter distributions of the ultrasonic molecular beam include density parameter distribution, velocity parameter distribution, pressure parameter distribution, and temperature parameter distribution, and different types of beam parameter distributions have corresponding mapping relationships between each other.

[0025] In ultrasonic molecular beams, the beam parameter distribution describes the spatial distribution characteristics of physical quantities within the molecular beam. These different types of parameter distributions have an inherent correspondence for the same ultrasonic molecular beam, forming a mutual mapping relationship, thus enabling the derivation of other distributions from one distribution; where: The density parameter distribution describes the spatial variation of molecular number density in the ultrasonic molecular beam, reflecting the local concentration characteristics of molecular aggregation. The velocity parameter distribution describes the distribution of the magnitude and direction of molecular velocity in the beam, reflecting the dynamic characteristics of directional molecular flow. The pressure parameter distribution characterizes the gradient variation of gas pressure along the beam path, demonstrating the transition behavior from a high-pressure source to a low-pressure region. The temperature parameter distribution reveals the spatial distribution pattern of the intensity of molecular thermal motion, and is related to the local uniformity of molecular kinetic energy.

[0026] These beam parameter distributions each describe the internal state of the ultrasonic molecular beam from different physical dimensions, and for the same ultrasonic molecular beam, there is an inherent correlation between them.

[0027] For example, in step S1, the density parameter distribution is obtained based on historical data of the ultrasonic molecular beam. and velocity parameter distribution Then the density parameter distribution and velocity parameter distribution There is a relationship between them. Corresponding relationship, function This is the mapping function between the two. The key to the inversion calculation model lies in this function. The mapping training enables the model to have the ability to quickly predict images from one image to another. In this case, the model parameters are relatively small, which is suitable for the needs of this implementation method in the field of fast prediction.

[0028] In a preferred embodiment, in step S1, the historical data of the ultrasonic molecular beam originates from simulations or experimental measurements of the ultrasonic molecular beam injector. The ultrasonic molecular beam injector generates different historical data under different gas source pressures, background pressures, throat diameters, and divergence angles. The types of historical data include simulation data and experimental measurement data, corresponding to simulation inversion datasets and experimental measurement inversion datasets. Furthermore, the distribution of various beam parameters in the inversion datasets is stored in the form of visualized images.

[0029] These visualizations can clearly present the continuous variation characteristics of beam parameters in space through intuitive two-dimensional or three-dimensional forms. For example, density parameter distribution images show the density difference of molecular aggregation through light and dark levels, velocity distribution images reflect the direction and speed of motion through color gradients, pressure distribution images show the gradual transition of pressure gradients, and temperature distribution images reveal the regional distribution law of thermal motion intensity.

[0030] During data processing, images undergo denoising, enhancement, and standardization to eliminate interference from measurement noise and simulation errors, ensuring compatibility and consistency of data from different sources. Simulation data generates a large number of high-precision samples by systematically adjusting key variables such as gas source pressure, background pressure, throat diameter, and divergence angle. For example, simulation data of an ultrasonic molecular beam injector with a throat diameter of 1 mm, a divergence angle of 30 degrees, and hydrogen as the gas source is obtained. This large number of samples covers a wide range of operating conditions. Experimental measurement data can also be obtained under similar conditions, relying on actual equipment to collect real beam responses. This allows simulation and measurement to complement each other, forming a highly complementary and comprehensive inversion data foundation. This data construction strategy in this implementation not only improves the representativeness and robustness of the inversion dataset but also provides high-quality input for subsequent neural network training, effectively supporting the model's accurate learning and generalization capabilities regarding complex mapping relationships between multiple parameter distributions.

[0031] In step S2, this implementation uses Python to build the U-net neural network, specifically the PyTorch neural network framework. The basic framework of the U-net neural network includes downsampling units, upsampling units, and skip connection units; wherein: The downsampling unit is used to extract information from the visualized image through the first convolutional layer and the max pooling layer; The upsampling unit is used to restore information in the visualized image through the second convolutional layer and the upsampling layer; Skip connection units are used to transfer information extracted from each layer in the downsampling unit to the corresponding layer in the upsampling unit to prevent the loss of image spatial information.

[0032] In this embodiment, when constructing the U-net neural network, the downsampling unit mines local feature patterns in the visualized image through multi-level convolutional operations, while gradually reducing the spatial resolution with the help of max pooling layers to enhance the scale invariance of features and ensure that key information is preserved during compression. The upsampling unit uses upsampling technology combined with convolutional layers to gradually restore the spatial details of the image and finely reconstruct the continuous changes in the parameter distribution, avoiding structural blurring caused by resolution reduction. The skip connection unit directly bridges the high-resolution features extracted from each level of the downsampling stage to the corresponding upsampling level, effectively bridging the information gaps in the deep network and ensuring the complete transmission and accurate reconstruction of the spatial details of the beam parameter distribution.

[0033] To enhance the physical feature recognition capability of the final inversion calculation model, this implementation incorporates an attention mechanism module into the U-net framework. This module intelligently focuses on core regions in the molecular beam where parameters such as density and velocity change significantly by dynamically calculating the weight distribution of different regions in the feature map. Simultaneously, it suppresses background noise and irrelevant interference, enabling the model to more accurately capture the nonlinear mapping patterns between multiple parameter distributions. The entire U-net-based network is efficiently implemented within the PyTorch framework, fully utilizing its flexible dynamic computation graph mechanism and rich neural network components to support rapid iterative optimization and stable training of the inversion calculation model, thus providing reliable technical support for high-precision inversion of beam parameter distributions.

[0034] As a preferred embodiment, based on the U-net neural network framework, three types of attention mechanism modules are added through a file-based construction method. These three attention mechanisms can be embedded into the skip connections of U-net. The three types of attention mechanism modules are: CA attention mechanism module, SE attention mechanism module, and CBAM attention mechanism module. These three attention mechanism modules form three corresponding image processing sub-models on the U-net neural network framework, and the inversion calculation model is constructed based on these three image processing sub-models. The following is a detailed description of these three types of image processing sub-models.

[0035] The CA attention mechanism module is used to embed a coordinate-aware feature extraction unit in the encoder path of U-net. It generates a one-dimensional coordinate attention map by performing global average pooling along the height and width directions, calculates the cross-correlation weights of the horizontal and vertical position coordinates in the visualization image of the beam parameter distribution, enhances the spatial structure representation of the ultrasound molecular beam hotspot region, and outputs the coordinate-weighted feature map to the decoder path of U-net.

[0036] The SE attention mechanism module is used to insert channel recalibration components in the skip connections of U-net. It generates channel descriptors by compressing the spatial dimension through the Squeeze operation, learns the sensitivity coefficients of the beam parameter distribution through the fully connected layer, dynamically amplifies the feature channel weights related to their intensity gradients, suppresses low-sensitivity channels, and improves the inversion robustness to beam parameter distribution non-uniformity.

[0037] The CBAM attention mechanism module has additional channel attention submodules and spatial attention submodules. The CBAM attention mechanism module is used to cascade a channel-spatial dual attention mechanism at the end of the decoder path of U-net. First, the channel attention submodule fuses the channel weights of global average pooling and max pooling to filter the key physical features of the beam parameter distribution. Then, the spatial attention submodule uses a pre-sized convolution kernel to generate a binary spatial mask and focus on the core region of the beam parameter distribution, thus collaboratively suppressing background noise and outputting a high-fidelity beam parameter distribution inversion result.

[0038] In this embodiment, a coordinate attention mechanism module, namely the aforementioned CA module, is embedded in the encoder path of the U-net neural network. This module achieves a fine characterization of the spatial structure of the beam parameter distribution through a coordinate-aware feature extraction unit. This unit performs global average pooling operations along the height and width directions of the image to generate a one-dimensional coordinate attention map, thereby quantifying the cross-correlation weights of the lateral and vertical position coordinates. This design enables the resulting sub-model to dynamically identify hotspot regions with significant density or velocity gradients in the ultrasound molecular beam, such as the core beam region with dense molecular aggregation or the transition zone at the diffusion edge, and adaptively weight the feature map according to the coordinate weights. The principle is to use coordinate information to strengthen the physical correlation of spatial positions, avoid the neglect of position-sensitive features by traditional convolution, and ensure that the structural details of the molecular beam hotspot region are fully preserved during the feature extraction process. In implementation, this module is directly integrated after the convolutional layer of the encoder, generating the coordinate attention map through lightweight computation, which not only does not significantly increase the model complexity but also significantly improves the ability to model the spatial continuity of the beam parameter distribution.

[0039] A channel attention mechanism module, the aforementioned SE module, is inserted into the skip connection path of the U-net neural network. The core of this module is a channel recalibration component designed to optimize the response characteristics of feature channels to beam parameter inhomogeneities. This component first compresses the spatial dimension of the feature map through a Squeeze operation, generating descriptors reflecting the global statistical characteristics of each channel. Then, it uses its fully connected layers to learn the sensitivity coefficients of the beam parameter distribution. This process dynamically evaluates the correlation between different feature channels and parameter intensity gradients. For example, in velocity distribution inversion, channels reflecting high-speed flow regions are given higher weights, while low-sensitivity channels associated with background noise or low-varying regions are suppressed. Its key technical feature is that, through a channel-level recalibration mechanism, it adaptively highlights key physical features in the parameter distribution, such as steep gradients in density distribution or local peaks in temperature distribution, thereby enhancing the model's robustness to common non-uniform disturbances in experimental measurements. In practical implementation, this module is embedded into skip connections in a plug-and-play manner, requiring only a few parameters to achieve real-time adjustment of channel weights, effectively mitigating the distortion of beam parameter distribution caused by measurement noise.

[0040] At the end of the decoder path of the U-net neural network, a channel-space dual attention mechanism module, namely the aforementioned CBAM module, is cascaded. This module outputs high-fidelity inversion results through the synergistic effect of the channel attention submodule and the spatial attention submodule. The channel attention submodule first fuses the channel statistical information of global average pooling and max pooling to generate comprehensive weights to filter key features highly correlated with beam physics properties. For example, in pressure distribution inversion, it prioritizes retaining the dominant channel reflecting the pressure gradient. Subsequently, the spatial attention submodule uses convolutional kernels of a preset size to analyze the spatial structure of the feature map and generate a binary spatial mask to accurately focus on the core region of the molecular beam, such as the main channel of directional flow in velocity distribution or high-concentration regions in density distribution. This dual-stage attention mechanism, through joint optimization of channel and spatial dimensions, can both suppress the interference of background noise and enhance the core physical features of parameter distribution. In terms of implementation details, this module is deployed in a cascaded structure at the end of the decoder to ensure fine calibration of features in the final stage of image reconstruction. Its output is directly used as the inversion result, significantly improving the spatial continuity and boundary clarity of parameters such as density and velocity distribution.

[0041] This implementation uses the three attention mechanism modules described above to form complementary image processing sub-models on the U-net framework. These sub-models can also be used independently and selected by the user for single or multiple comparisons, thus collectively constructing a robust inversion computation model. The coordinate attention mechanism module focuses on spatial structure modeling during the encoding stage, the channel attention mechanism module strengthens feature channel optimization in skip connections, and the channel-space dual attention mechanism module focuses on detail restoration at the decoding end. These three modules cover the entire process of feature extraction, transfer, and reconstruction, and can be used in combination, taking into account prediction time and resource consumption costs. This design fully utilizes the physical characteristics of beam parameter distributions, such as the spatial locality of density distribution, the directional dependence of velocity distribution, and the gradient sensitivity of pressure and temperature distribution, enabling the model to adaptively capture the nonlinear mapping patterns between different parameters. Within the PyTorch framework, these modules are efficiently integrated through a dynamic computation graph, supporting the inversion computation model to automatically learn physical correlations in historical data during training. Ultimately, this achieves the ability to derive the inversion of other parameter distributions with high accuracy from measured images of only one type of parameter distribution.

[0042] The model training process in step S3 can employ common and mature model training methods in this field. This implementation does not impose excessive restrictions on them, but a brief example can be introduced: During the model training phase, the inversion dataset is divided into training samples and validation samples. The model learns the physical mapping between the input beam parameter distribution image and the target distribution image through repeated iterations. During training, the mean squared error loss function is used to quantify the deviation between the predicted results and the true distribution, and an adaptive optimization algorithm is used to dynamically adjust the network weights to minimize the error. At the same time, the learning rate is adjusted in a timely manner based on the performance of the validation set to prevent the model from overfitting. After multiple rounds of training iterations, the inversion calculation model gradually masters the nonlinear correlation characteristics between the distributions of parameters such as density, velocity, pressure, and temperature, ultimately achieving a high-precision cross-parameter extrapolation capability for the physical state of the beam.

[0043] Finally, in step S4, this embodiment performs the following steps on the trained inversion calculation model: Figure 3 The GUI visualization encapsulation shown enables the inversion calculation model to output the corresponding inversion results of the beam parameter distribution through the GUI visualization interface, as shown in the image. Figure 4 As shown. In this embodiment, the GUI visualization interface is provided with display boxes for at least two types of beam parameter distributions, a calculation accuracy display box, control icons for selecting input data, and control icons for selecting different sub-models in the inversion calculation model.

[0044] The GUI visual design and encapsulation can be implemented using PyQt5, a Python binding library. PyQt5 is used to create graphical user interfaces and is based on the cross-platform GUI framework Qt5. It has cross-platform characteristics, and graphical user interfaces developed using PyQt5 can usually run on different operating systems with little or no code modification.

[0045] When a user uses the inversion calculation model that has completed model training and visualization, they can select the model and data in sequence through the GUI visualization interface and click the start prediction icon. The model will then display the distribution map of the input data and the distribution map of the inversion results in the GUI visualization interface. At the same time, auxiliary information such as the model calculation accuracy will be displayed through the corresponding calculation indicators. In this way, the user can realize the inversion process between various beam parameter distributions and meet the corresponding experimental requirements.

[0046] The method steps described in this embodiment can be effectively applied to a computer system, which includes a memory, a processor, and a computer program stored in the memory. When the processor executes the computer program, the steps of the rapid inversion method for the beam parameter distribution of ultrasonic molecular beams of this embodiment can be implemented. These computer programs or instructions can be stored in a computer-readable storage medium that can guide a computer or other programmable data processing device to operate in a specific manner, causing the instructions stored in the computer-readable storage medium to produce an article of manufacture including instruction means that implement the function specified in one or more steps of the method.

Claims

1. A method for rapid inversion of beam parameter distribution of an ultrasonic molecular beam, characterized in that, The method includes the following steps: S1. Based on historical data of ultrasonic molecular beams, obtain at least two different types of beam parameter distributions that are related to each other, and form an inversion dataset for model training. S2. Using the U-net neural network as the basic framework, an attention mechanism module is added to construct an inversion calculation model; S3. Use the inversion dataset to train the inversion calculation model so that the trained inversion calculation model can capture the mapping relationship between at least two types of beam parameter distributions; S4. When conducting new experiments with ultrasonic molecular beams and needing to measure multiple types of beam parameter distributions, only the measured results of one type of beam parameter distribution are input into the inversion calculation model, and the inversion calculation model will output the corresponding inversion results of the other required types of beam parameter distributions.

2. The rapid inversion method for beam parameter distribution according to claim 1, characterized in that: The types of beam parameter distributions of ultrasonic molecular beams include density parameter distribution, velocity parameter distribution, pressure parameter distribution and temperature parameter distribution. There are corresponding relationships between different types of beam parameter distributions that can form a mapping relationship.

3. The rapid inversion method for beam parameter distribution according to claim 1, characterized in that: In step S1, the density parameter distribution is obtained based on historical data of the ultrasonic molecular beam. and velocity parameter distribution Density parameter distribution and velocity parameter distribution There is a relationship between them. Corresponding relationship, function This is the mapping function between the two.

4. The fast inversion method for beam parameter distribution according to claim 1, characterized in that: In step S1, the historical data of the ultrasonic molecular beam comes from the simulation or experimental measurement of the ultrasonic molecular beam injector. The ultrasonic molecular beam injector has different historical data under different gas source pressure, background pressure, throat diameter and divergence angle. The types of historical data include simulation data and experimental measurement data, which are respectively formed into simulation inversion dataset and experimental measurement inversion dataset.

5. The rapid inversion method for beam parameter distribution according to claim 1, characterized in that: In step S1, the distribution of multiple beam parameters in the inversion dataset is stored in the form of a visual image.

6. The rapid inversion method for beam parameter distribution according to claim 5, characterized in that: In step S2, the basic framework of the U-net neural network includes downsampling units, upsampling units, and skip connection units; wherein: The downsampling unit is used to extract information from the visualized image through the first convolutional layer and the max pooling layer; The upsampling unit is used to restore information in the visualized image through the second convolutional layer and the upsampling layer; Skip connection units are used to transfer information extracted from each layer in the downsampling unit to the corresponding layer in the upsampling unit to prevent the loss of image spatial information.

7. The rapid inversion method for beam parameter distribution according to claim 6, characterized in that: Based on the U-net neural network framework, three types of attention mechanism modules are added: CA attention mechanism module, SE attention mechanism module and CBAM attention mechanism module; These three types of attention mechanism modules form three corresponding image processing sub-models on the U-net neural network framework, and the inversion calculation model is constructed based on these three types of image processing sub-models.

8. The rapid inversion method for beam parameter distribution according to claim 7, characterized in that: The CA attention mechanism module is used to embed a coordinate-aware feature extraction unit in the encoder path of U-net. It generates a one-dimensional coordinate attention map by performing global average pooling along the height and width directions, calculates the cross-correlation weights of the horizontal and vertical position coordinates in the visualization image of the beam parameter distribution, strengthens the spatial structure representation of the ultrasound molecular beam hotspot region, and outputs the coordinate-weighted feature map to the decoder path of U-net. The SE attention mechanism module is used to insert channel recalibration components in the skip connections of U-net. It compresses the spatial dimension through the Squeeze operation to generate channel descriptors, learns the sensitivity coefficients of the beam parameter distribution through the fully connected layer, dynamically amplifies the feature channel weights related to their intensity gradients, and suppresses low-sensitivity channels. The CBAM attention mechanism module has additional channel attention submodules and spatial attention submodules; The CBAM attention mechanism module is used to cascade a channel-space dual attention mechanism at the end of the decoder path of U-net. First, the channel attention submodule fuses the channel weights of global average pooling and max pooling to filter the key physical features of the beam parameter distribution. Then, the spatial attention submodule uses a pre-sized convolution kernel to generate a binary spatial mask and focus on the core region of the beam parameter distribution, which together suppresses background noise and outputs the inversion result of the beam parameter distribution.

9. The rapid inversion method for beam parameter distribution according to claim 1, characterized in that: In step S4, the trained inversion calculation model is encapsulated in a GUI visualization so that the inversion calculation model can output the corresponding inversion results of the beam parameter distribution through the GUI visualization interface. The GUI visualization interface has display boxes for at least two types of beam parameter distributions, a calculation accuracy display box, control icons for selecting input data, and control icons for selecting different sub-models in the inversion calculation model.

10. A computer system comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the rapid inversion method for beam parameter distribution of ultrasonic molecular beams according to any one of claims 1-9.

Images