Radiation field acceleration calculation method and system based on denoising neural network, processing equipment and storage medium

Abstract

The invention relates to a radiation field acceleration calculation method and system based on a denoising neural network, processing equipment and a storage medium, and the method comprises the steps: obtaining a low particle number MC simulation result in a to-be-detected scene, and carrying out the preprocessing; the preprocessed low-particle-number MC simulation result is input into a pre-trained and optimized deep radiation field denoising neural network model, a high-particle-number MC simulation result in the scene to be measured is obtained, radiation field acceleration calculation is completed, the deep radiation field denoising neural network model is obtained by adopting an improved UNet convolutional neural network, and the depth radiation field denoising neural network model is obtained by adopting an improved UNet convolutional neural network. The method has breakthrough improvement in calculation efficiency, the prediction result completely meets the precision requirement of engineering application, and the method has excellent generalization performance and can be widely applied to the technical field of nuclear application.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of nuclear applications, in particular to a radiation field accelerated calculation method and system based on a denoising neural network, a processing device and a storage medium. BACKGROUND

[0002] In the field of radiation field calculation, traditional methods are mainly divided into two categories: deterministic methods and non-deterministic methods. Among them, deterministic methods such as discrete ordinate method, characteristic line method, point kernel integration method, etc. obtain the distribution of neutron and photon radiation field by directly solving the radiation transport equation. The discrete ordinate method and the characteristic line method can accurately solve the radiation field, but when dealing with complex geometric structures, the grid division is difficult, and the calculation amount increases sharply with the complexity of the geometry. The point kernel integration method has fast calculation speed, but it is only suitable for homogeneous medium or simple geometry, and the accuracy decreases significantly in complex shielding (such as multi-layer reactor structure). Monte Carlo (MC) method as a representative of non-deterministic method, through random particle transport simulation to calculate the radiation field, can accurately handle any complex geometry, and is regarded as the "gold standard" of radiation field calculation.

[0003] However, the MC method faces the problem of deep penetration when calculating thick shielding (such as reactor pressure vessel, concrete biological shielding layer), that is, the statistical fluctuation is large and the calculation efficiency is extremely low when high-energy particles penetrate thick medium. Although a variety of variance reduction methods (such as importance sampling, splitting and roulette, exponential transformation, etc.) have been theoretically proposed, a large amount of computing resources is still needed to obtain convergent results. At the engineering application level, although the use of GPU parallel computing can greatly improve the efficiency of Monte Carlo calculation, it needs to be fully transformed from serial code, which is a huge amount of work, and there is still a lack of mature GPU accelerated Monte Carlo calculation complete scheme.

[0004] Therefore, it is urgent to develop an efficient and high-precision radiation field accelerated calculation method to significantly reduce the calculation cost while ensuring the calculation accuracy, in order to meet the real-time simulation needs in the fields of reactor radiation shielding design and medical physics. SUMMARY

[0005] In order to solve the above problems, the purpose of the present application is to provide a radiation field accelerated calculation method, system, processing device and storage medium based on a denoising neural network, which can significantly reduce the calculation cost while ensuring the calculation accuracy.

[0006] To achieve the above purpose, the present application adopts the following technical scheme: in the first aspect, a radiation field accelerated calculation method based on a denoising neural network is provided, comprising: obtaining the low-particle-number MC simulation result under the to-be-measured scene, and preprocessing; The preprocessed low-particle-number MC simulation results are input into a pre-trained and optimized deep radiation field denoising neural network model to obtain high-particle-number MC simulation results for the test scene, thus completing the accelerated calculation of the radiation field. The deep radiation field denoising neural network model is obtained by using an improved UNet convolutional neural network.

[0007] Furthermore, the training and optimization process of the deep radiation field denoising neural network model is as follows: The key factors affecting the distribution of the radiation field are modeled, and the Monte Carlo simulation method is used to generate a high-quality dataset and perform preprocessing. An improved UNet convolutional neural network is used to construct a deep radiation field denoising neural network model; Based on the preprocessed high-quality dataset, the constructed deep radiation field denoising neural network model is trained and optimized.

[0008] Furthermore, the modeling of key factors affecting the radiation field distribution, and the use of Monte Carlo simulation to generate a high-quality dataset and perform preprocessing, includes: Model the source term information and the shielding arrangement; Based on the source term information and shielding arrangement from the model, Monte Carlo simulation was performed with a particle number of 10. 6 The calculations were performed on a scale of magnitude to obtain low-particle-number MC simulation results; Based on the source term information and shielding arrangement from the model, Monte Carlo simulation was performed with a particle number of 10. 9 Calculations on the order of magnitude or higher yielded high-particle-number MC simulation results; The generated low-particle-number MC simulation results and high-particle-number MC simulation results are preprocessed in a unified manner.

[0009] Furthermore, the source information includes point sources, surface sources, and volume sources; The shielding arrangement includes several shielding layouts and material combinations with variations in materials, geometry, and thickness.

[0010] Furthermore, the unified preprocessing of the generated low-particle-number MC simulation results and high-particle-number MC simulation results includes: The particle dose rate or flux distribution data obtained from the low particle number MC simulation results and the high particle number MC simulation results are discretized into a predefined regular three-dimensional voxel grid to form three-dimensional tensor data; The radiation field values ​​within the three-dimensional voxel mesh are normalized to obtain preprocessed low-particle-number MC simulation results and high-particle-number MC simulation results.

[0011] Furthermore, the deep radiation field denoising neural network model includes: The input layer is used to input data; The spatial attention module, located in the input layer, is used to assign learnable weights to different spatial locations by learning the spatial features of the input data. The encoder is used to progressively extract features from the input data and reduce the spatial dimension while increasing the number of feature channels through successive convolutional layers and average pooling layers. The decoder is used to gradually restore the spatial dimension through successive upsampling operations, while reducing the number of feature channels through convolution; Skip connection layers are used to directly connect feature maps in the encoder to feature maps of the corresponding scale in the decoder; The output layer uses a linear activation function to output data that restores the spatial dimension as the prediction result.

[0012] Furthermore, the training and optimization of the constructed deep radiation field denoising neural network model based on the preprocessed high-quality dataset includes: Based on the preprocessed high-quality dataset, a weighted combined loss function is used to train the constructed deep radiation field denoising neural network model. The Adam optimizer is used to optimize the trained deep radiation field denoising neural network model.

[0013] Secondly, a radiation field acceleration calculation system based on a denoising neural network is provided, comprising: The data acquisition module is used to acquire low-particle-number MC simulation results for the test scenario and perform preprocessing. The calculation module is used to input the preprocessed low-particle-number MC simulation results into the pre-trained and optimized deep radiation field denoising neural network model to obtain the high-particle-number MC simulation results of the scene under test. The deep radiation field denoising neural network model is obtained by using an improved UNet convolutional neural network.

[0014] Thirdly, a processing device is provided, including computer program instructions, wherein when the computer program instructions are executed by the processing device, they are used to implement the steps corresponding to the above-described accelerated calculation method for radiation fields based on a denoising neural network.

[0015] Fourthly, a computer-readable storage medium is provided, wherein computer program instructions are stored on the computer-readable storage medium, wherein the computer program instructions, when executed by a processor, are used to implement the steps corresponding to the above-described accelerated calculation method for radiation fields based on a denoising neural network.

[0016] The present invention has the following advantages due to the adoption of the above technical solutions: 1. Breakthrough in computational efficiency: This invention innovatively combines low-particle-number Monte Carlo simulation with deep learning networks, significantly improving computational efficiency while maintaining computational accuracy. Compared to traditional Monte Carlo simulation methods that require substantial computational resources to achieve convergence, this invention only requires pre-computation of a small number of particles combined with rapid neural network inference to obtain results comparable to high-precision simulations, achieving an order-of-magnitude improvement in computational efficiency.

[0017] 2. Accelerated Calculation of Deep Penetration Problems: Addressing the technical challenge of requiring a large number of particles to ensure convergence and resulting in long computation times in calculations involving thick shielding due to the deep penetration effect, this invention introduces an adaptive attention mechanism. It focuses on key locations such as the area behind the shielding where personnel are active, and employs a special hierarchical feature fusion network structure design, significantly improving the prediction accuracy of low-particle-number simulations in deep penetration regions. This invention effectively solves the problem of large statistical fluctuations in the far-end region of the shielding in traditional Monte Carlo simulations. While ensuring the continuity of the spatial distribution of the radiation field, it ensures that the prediction results fully meet the accuracy requirements of engineering applications, providing an efficient and reliable solution for calculating the radiation field of thick shielding.

[0018] 3. Excellent generalization performance: Thanks to the systematic data generation strategy and network architecture design, this invention exhibits strong generalization ability. It maintains stable prediction performance in tests with different shielding materials, radiation source types, and geometric configurations. Especially for engineering scenarios where no training was conducted, this invention can still provide reliable prediction results, fully demonstrating the practical value of the method.

[0019] 4. Efficient utilization of computing resources: This invention significantly reduces the demand for computing resources through optimized data representation and network design. Compared to traditional methods that require a large number of computing nodes and storage space, this invention can run efficiently on conventional computing devices, significantly reducing computing energy consumption and hardware investment, and providing an economical and feasible solution for engineering applications.

[0020] 5. Ease of Engineering Application: This invention has excellent engineering applicability and can be easily integrated into existing Monte Carlo simulation workflows. It achieves improved computational efficiency without requiring large-scale modifications to existing calculation code, providing engineers in radiation shielding design, medical physics, and other fields with an efficient and convenient computational tool, thus contributing to technological advancements in these areas.

[0021] In summary, this invention can be widely applied in the field of nuclear application technology. Attached Figure Description

[0022] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. In the drawings: Figure 1 This is a schematic diagram of the model structure provided in an embodiment of the present invention. Detailed Implementation

[0023] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the invention and to fully convey the scope of the invention to those skilled in the art.

[0024] It should be understood that the terminology used herein is for the purpose of describing particular exemplary embodiments only and is not intended to be limiting. Unless the context clearly indicates otherwise, the singular forms “a,” “an,” and “described” as used herein may also include the plural forms. The terms “comprising,” “including,” “containing,” and “having” are inclusive and therefore indicate the presence of the stated features, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, elements, components, and / or combinations thereof. The method steps, processes, and operations described herein are not construed as requiring them to be performed in a particular order described or illustrated unless the order of performance is explicitly indicated. It should also be understood that additional or alternative steps may be used.

[0025] Although terms such as first, second, third, etc., may be used in this document to describe multiple elements, components, regions, layers, and / or segments, these elements, components, regions, layers, and / or segments should not be limited by these terms. These terms may be used only to distinguish one element, component, region, layer, or segment from another. Unless the context clearly indicates otherwise, terms such as "first," "second," and other numerical terms used herein do not imply order or sequence. Therefore, the first element, component, region, layer, or segment discussed below may be referred to as the second element, component, region, layer, or segment without departing from the teachings of the exemplary embodiments.

[0026] Currently, the Monte Carlo (MC) method faces the problem of deep penetration when calculating thick shields, i.e., high-energy particles experience large statistical fluctuations and extremely low computational efficiency when penetrating thick media. Although various variance reduction methods have been theoretically proposed, a large amount of computational resources are still required to obtain convergence results. In engineering applications, although GPU parallel computing can significantly improve the efficiency of Monte Carlo calculations, it requires a complete overhaul of the serial code, which is a huge workload. Currently, there is a lack of mature and complete solutions for GPU-accelerated Monte Carlo calculations. This invention provides a radiation field acceleration calculation method based on a denoising neural network, including: obtaining low-particle-number MC simulation results of the test scene and performing preprocessing; inputting the preprocessed low-particle-number MC simulation results into a pre-trained and optimized deep radiation field denoising neural network model to obtain high-particle-number MC simulation results of the test scene, thus completing the accelerated radiation field calculation. The deep radiation field denoising neural network model is obtained using an improved UNet convolutional neural network. This invention treats low-particle-number Monte Carlo simulation results with high statistical fluctuations as noise samples and high-particle-number Monte Carlo simulation results with low statistical fluctuations as clean target samples. It utilizes a denoising neural network based on the UNet framework to learn the mapping relationship between the two, achieving a breakthrough improvement in computational efficiency. This invention trains the pre-calculated results using a specially designed deep radiation field denoising neural network model, achieving rapid conversion from low-precision simulations to high-precision results, effectively solving the problem of low computational efficiency in traditional methods while ensuring the accuracy requirements of the calculated results. Through the above implementation methods, this invention can effectively utilize the speed of low-particle-number Monte Carlo simulation results, combined with the powerful denoising and feature extraction capabilities of deep learning, to achieve rapid prediction of high-precision radiation field distributions, significantly accelerating the fine radiation field calculation process. Figure 1 The examples "Scenario 1" and "Scenario 2" in the text intuitively demonstrate the effect of this method in transforming the low particle number simulation results with obvious statistical noise on the left into the smooth, clear, and near-realistic high particle number simulation results on the right after processing by a neural network.

[0027] Example 1 This embodiment provides a method for accelerating radiation field calculation based on a denoising neural network, including the following steps: 1) Key factors influencing the radiation field distribution are modeled, and Monte Carlo simulation is used to generate a high-quality dataset, which is then preprocessed. The high-quality dataset includes both low-particle-number (MC) and high-particle-number (MC) simulation results, specifically: 1.1) To enhance data diversity, source item information and shielding arrangement are modeled.

[0028] Specifically, the source term information includes different spatial distribution forms such as point sources, surface sources, and volume sources, and the energy spectrum coverage is extensive, including but not limited to thermal neutrons, fast neutrons, and gamma rays of different energies, to meet the simulation needs of different types of radiation. Among them, point sources are used to simulate the emission point of a single radioactive particle, surface sources are used to simulate planar radioactive contamination or material surface activation, and volume sources are used to simulate radioactive materials or activated regions with a certain volume.

[0029] Specifically, the shielding arrangement includes several common materials, geometric shapes, and thickness variations in shielding layouts and material combinations to simulate complex shielding structures in real-world engineering scenarios. More specifically, it involves the combined use of various commonly used shielding materials such as concrete, lead, water, and steel; the geometric shapes consider combinations of basic shapes such as walls, plates, and cylinders, as well as scenarios with complex structures including holes, gaps, and bends; the thickness varies within a range of 1 to 5 mean free paths to simulate different shielding performance requirements, with particular attention to the deep penetration problem of thick shielding.

[0030] For example: Figure 1 The terms "Scenario 1" and "Scenario 2" in this invention refer to two typical modeling scenarios in the training data used in this invention, which are scenarios with different source term parameters (energy) under a specific shielding structure.

[0031] 1.2) Based on the source term information and shielding arrangement from the model, mainstream Monte Carlo simulation software (such as MCNP, Geant4, FLUKA, JMCT, etc.) is used to simulate a particle number of 10. 6 A rough calculation of the order of magnitude yields MC simulation results for low particle number.

[0032] Specifically, these low-particle-number MC simulation results exhibit significant statistical fluctuations, manifested as follows: Figure 1 The noise and speckled distribution shown in the "Low Particle Number MC Simulation" are relatively fast, and these results will be used as input data for the neural network described below.

[0033] 1.3) For the same modeling scenario as the low-precision sample, based on the source term information and shielding arrangement, the same mainstream Monte Carlo simulation software is used to perform a particle number of 10. 9 Fine-grained calculations on the order of magnitude or higher yielded high-particle-number MC simulation results.

[0034] Specifically, to improve computational efficiency and ensure true value accuracy, various variance reduction techniques (such as weighted windows, splitting or Russian roulette, importance sampling, etc.) can be used in combination to reduce computational error, with the statistical error needing to be less than 5%. The computational goal is to reduce the statistical error of the region of interest (e.g., the area accessible to people) to below 5%, ensuring that this data is reliable as the baseline true value for neural network training. These results are manifested as follows: Figure 1The smoother, more detailed distribution shown in the "high particle number MC simulation" will serve as the target data for the output of the neural network.

[0035] 1.4) Perform unified preprocessing on the generated low-particle-number MC simulation results and high-particle-number MC simulation results: 1.4.1) Discretize the particle dose rate or flux distribution data obtained from the low particle number MC simulation results and the high particle number MC simulation results into a predefined regular three-dimensional voxel grid to form three-dimensional tensor data.

[0036] Specifically, Figure 1 The "scene mesh size" shown (e.g., 220x52x70) represents the voxel mesh dimensions of a scene. Although Figure 1 What is displayed is a 2D slice or projection, but what is actually processed is 3D voxel data.

[0037] 1.4.2) Normalize the radiation field values ​​in the three-dimensional voxel grid, for example, by scaling them to the range of [0, 1] or [-1, 1], in order to improve the stability and efficiency of neural network training and obtain preprocessed low-particle-number MC simulation results and high-particle-number MC simulation results.

[0038] Specifically, normalization methods can include maximum-minimum normalization, Z-score normalization, or logarithmic normalization, and the normalization parameters are recorded for inverse normalization of the prediction results.

[0039] 2) such as Figure 1 As shown, an improved UNet convolutional neural network is used to construct a deep radiation field denoising neural network model, specifically: 2.1) Construct the basic UNet convolutional neural network.

[0040] Specifically, the basic UNet convolutional neural network adopts a standard encoder-decoder structure, which is symmetrically U-shaped and includes an input layer, encoder, decoder, skip connection layer and output layer.

[0041] Specifically, the input layer is used to input data.

[0042] encoder ( Figure 1 The left side of the image is used to pass through consecutive convolutional layers and downsampling operations. Figure 1 The red arrow in the middle indicates "average pooling") which progressively extracts features from the input data and reduces the spatial dimension, while increasing the number of feature channels. Figure 1The numbers above or below the blue boxes indicate the number of channels, ranging from 1 to 64, then to 128, and finally to 1024. For example, an input size of 220x52x70 (1 channel) becomes 64 channels after passing through a convolutional layer, while the size remains the same; after downsampling through an average pooling layer, the size becomes 110x26x35 (128 channels), and so on, until the bottom layer with 13x3x4 (1024 channels). Figure 1 (Indicated by the blue arrow) It typically includes a 3x3 convolutional kernel, a batch normalization layer, and a ReLU activation function.

[0043] decoder ( Figure 1 The right side of the image is used for continuous upsampling operations. Figure 1 The green arrow in the middle indicates "upsampling", which can be transposed convolution or interpolated convolution to gradually restore the spatial dimension, while reducing the number of feature channels through convolution (e.g., from 1024 to 512, then to 256, etc.).

[0044] Skip Connections layers are used to directly connect high-resolution, detailed feature maps from the encoder to corresponding scale feature maps in the decoder (typically after upsampling in the decoder). For example, a 27x6x8 feature map with 512 channels in the encoder is spliced ​​with a 27x6x8 feature map with 512 channels upsampled in the decoder via a skip connection layer, forming a 1024-channel feature map. These skip connections help the decoder retain the fine details captured by the encoder when recovering spatial information, avoiding the loss of crucial information during downsampling, which is essential for accurately predicting the spatial distribution of the radiation field.

[0045] The output layer is used to output data that restores the spatial dimension as the prediction result.

[0046] 2.2) Improve the basic UNet convolutional neural network by setting a spatial attention module in the input layer and using a linear activation function in the output layer to obtain the improved UNet convolutional neural network.

[0047] Specifically, addressing the challenges of large statistical fluctuations and computational difficulties in calculating deep penetration regions using Monte Carlo simulation methods for thick shielding structures, this invention incorporates a spatial attention module in the input layer of the basic UNet convolutional neural network. This module learns the spatial features of the input data and assigns learnable weights to different spatial locations. The network automatically assigns higher attention weights to areas with low signal-to-noise ratios and significant impact on results in traditional Monte Carlo simulations, such as the area behind the shielding structure, corners, or gaps. This guides the network to focus more on denoising and detail recovery in these critical areas, significantly improving the prediction accuracy for deep penetration regions and areas accessible to personnel. The spatial attention module can be implemented based on existing spatial attention mechanisms, such as learning a spatial weight map through convolutional layers and multiplying it with the input feature map.

[0048] Specifically, this invention addresses the issue in the last output layer of the basic UNet convolutional neural network (…). Figure 1 The 1x1 convolutional layer (reducing the number of channels from 64 to 1) uses a linear activation function (Identity Function) instead of the more common ReLU or Sigmoid. This is because radiation field values ​​(such as dose rate) can be any non-negative real number without a fixed upper limit. Using a linear activation function ensures that the network's prediction output is not limited by the numerical range of the activation function, and can completely and accurately represent the entire radiation field range from near zero to high values, preserving the physical properties and dynamic range of the radiation field to the greatest extent.

[0049] 2.3) An improved UNet convolutional neural network is used to construct a deep radiation field denoising neural network model.

[0050] Specifically, the deep radiation field denoising neural network model includes an input layer, a spatial attention module, an encoder, a decoder, a skip connection layer, and an output layer. The input layer is used to input the data. The spatial attention module, located in the input layer, learns the spatial features of the input data and assigns learnable weights to different spatial locations. The encoder progressively extracts features from the input data and reduces the spatial dimension while increasing the number of feature channels through successive convolutional and average pooling layers. The decoder progressively restores the spatial dimension through successive upsampling operations while reducing the number of feature channels through convolution. The skip connection layer directly connects the feature maps with high-resolution details in the encoder to the corresponding scale feature maps in the decoder. The output layer uses a linear activation function to output the data with restored spatial dimensions as the prediction result.

[0051] 3) Based on the preprocessed high-quality dataset, the constructed deep radiation field denoising neural network model is trained and optimized, specifically as follows: 3.1) Based on the preprocessed high-quality dataset, a weighted combined loss function is used to train the constructed deep radiation field denoising neural network model.

[0052] Specifically, to balance the numerical accuracy of the prediction results with the similarity of the spatial structure, this invention employs a weighted combined loss function: total loss = 70% * MAE + 30% * SSIM. Here, MAE (Mean Absolute Error) is the average of the absolute differences between the predicted voxel values ​​and the true voxel values. MAE directly measures the numerical accuracy of the prediction results, enabling the network to learn to predict values ​​close to the baseline. SSIM (Structural Similarity Index Measure) evaluates the similarity between the predicted and true radiation field distributions in terms of brightness, contrast, and structure. SSIM is particularly suitable for preserving the texture and structural features of the three-dimensional spatial distribution. Including SSIM as part of the loss (usually in the form of 1-SSIM) encourages the network to generate a spatially smoother radiation field distribution with details closer to the true result. This is crucial for accurately describing the physical diffusion and attenuation processes of the radiation field, and is especially suitable for evaluating... Figure 1 The radiation field distribution shown has a complex spatial structure. The weighting ratio (70% MAE, 30% SSIM) can be fine-tuned according to the characteristics of the actual dataset and the training effect.

[0053] 3.2) The Adam optimizer is used to optimize the trained deep radiation field denoising neural network model.

[0054] Specifically, the Adam optimizer combines the advantages of AdaGrad and RMSprop, performing well in handling sparse gradients and non-stationary objectives. The initial learning rate of the Adam optimizer is set to 1e-4.

[0055] Specifically, a cosine annealing learning rate scheduling strategy is employed to manage the learning rate. During training, the learning rate decreases according to a cosine function curve, decreasing slowly at the beginning, accelerating in the middle, and slowing down again at the end. This strategy helps the model escape local optima and achieve better convergence.

[0056] 4) Obtain the low particle number MC simulation results in the scene to be tested, and perform the preprocessing described in step 1.4) above.

[0057] 5) Input the preprocessed low-particle-number MC simulation results into the trained and optimized deep radiation field denoising neural network model to obtain the high-particle-number MC simulation results of the test scene, and complete the accelerated calculation of the radiation field.

[0058] The trained and optimized deep radiation field denoising neural network model obtained through the data preparation and training process described above in this invention can be widely applied in fields requiring rapid and accurate radiation field assessment. Nuclear power plant shielding design: Quickly assess the radiation field distribution under different shielding schemes during the design phase, optimize the shielding layout, and reduce design iteration time.

[0059] Medical radiation therapy: Predicting dose distribution in treatment plans helps doctors develop more precise radiation therapy plans and reduce damage to healthy tissues.

[0060] Radiation protection optimization for industrial testing equipment: Rapid simulation and optimization of shielding design for industrial X-ray testing equipment to ensure operator safety.

[0061] Radiation accident emergency response: Based on low-precision data obtained from rapid on-site measurements and combined with modeling information, quickly predict the distribution of radiation fields over a large area to assist in decision-making.

[0062] Example 2 This embodiment provides a radiation field acceleration calculation system based on a denoising neural network, including: The data acquisition module is used to acquire low-particle-number MC simulation results in the test scenario and perform preprocessing.

[0063] The calculation module is used to input the preprocessed low-particle-number MC simulation results into the pre-trained and optimized deep radiation field denoising neural network model to obtain the high-particle-number MC simulation results of the scene under test. The deep radiation field denoising neural network model is obtained by using an improved UNet convolutional neural network.

[0064] The system provided in this embodiment is used to execute the above-described method embodiments. For specific processes and details, please refer to the above embodiments, which will not be repeated here.

[0065] Example 3 This embodiment provides a processing device corresponding to the radiation field acceleration calculation method based on denoising neural network provided in Embodiment 1. The processing device can be applied to client processing devices, such as mobile phones, laptops, tablets, desktop computers, etc., to execute the method of Embodiment 1.

[0066] The processing device includes a processor, a memory, a communication interface, and a bus. The processor, memory, and communication interface are connected via the bus to enable communication between them. The memory stores computer programs that can run on the processing device. When the processing device runs the computer program, it executes the radiation field acceleration calculation method based on a denoising neural network provided in Embodiment 1.

[0067] In some implementations, the memory may be high-speed random access memory (RAM), and may also include non-volatile memory, such as at least one disk storage device.

[0068] In other implementations, the processor can be any type of general-purpose processor, such as a central processing unit (CPU) or a digital signal processor (DSP), and there is no limitation here.

[0069] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0070] Those skilled in the art will understand that the structure of the above-described computing device is only a partial structure related to the present invention and does not constitute a limitation on the computing device to which the present invention is applied. A specific computing device may include more or fewer components, or combine certain components, or have different component arrangements.

[0071] Example 4 This embodiment provides a computer program product corresponding to the radiation field acceleration calculation method based on a denoising neural network provided in Embodiment 1. The computer program product may include a computer-readable storage medium on which computer-readable program instructions for executing the radiation field acceleration calculation method based on a denoising neural network described in Embodiment 1 are loaded.

[0072] A computer-readable storage medium can be a tangible device that holds and stores instructions for use by an instruction execution device. A computer-readable storage medium can be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any combination thereof.

[0073] The computer-readable storage medium provided in the above embodiments has a similar implementation principle and technical effect to the above method embodiments, and will not be described again here.

[0074] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0075] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0076] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. ​ One or more processes and / or boxes ​ The steps of the function specified in one or more boxes.

[0077] The above embodiments are only used to illustrate the present invention. The structure, connection method and manufacturing process of each component can be varied. All equivalent transformations and improvements made on the basis of the technical solution of the present invention should not be excluded from the protection scope of the present invention.

Claims

1. A method for accelerating the calculation of radiation fields based on a denoising neural network, characterized in that, include: Obtain the low-particle-number MC simulation results for the scene under test and perform preprocessing; The preprocessed low-particle-number MC simulation results are input into a pre-trained and optimized deep radiation field denoising neural network model to obtain high-particle-number MC simulation results for the test scene, thus completing the accelerated calculation of the radiation field. The deep radiation field denoising neural network model is obtained by using an improved UNet convolutional neural network.

2. The accelerated calculation method for radiation fields based on a denoising neural network as described in claim 1, characterized in that, The training and optimization process of the deep radiation field denoising neural network model is as follows: The key factors affecting the distribution of the radiation field are modeled, and the Monte Carlo simulation method is used to generate a high-quality dataset and perform preprocessing. An improved UNet convolutional neural network is used to construct a deep radiation field denoising neural network model; Based on the preprocessed high-quality dataset, the constructed deep radiation field denoising neural network model is trained and optimized.

3. The accelerated calculation method for radiation fields based on a denoising neural network as described in claim 2, characterized in that, The process involves modeling the key factors influencing the radiation field distribution, employing Monte Carlo simulation to generate a high-quality dataset, and preprocessing it, including: Model the source term information and the shielding arrangement; Based on the source term information and shielding arrangement from the model, Monte Carlo simulation was performed with a particle number of 10. 6 The calculations were performed on a scale of magnitude to obtain low-particle-number MC simulation results; Based on the source term information and shielding arrangement from the model, Monte Carlo simulation was performed with a particle number of 10. 9 Calculations on the order of magnitude or higher yielded high-particle-number MC simulation results; The generated low-particle-number MC simulation results and high-particle-number MC simulation results are preprocessed in a unified manner.

4. The accelerated calculation method for radiation fields based on a denoising neural network as described in claim 3, characterized in that, The source information includes point sources, surface sources, and volume sources; The shielding arrangement includes several shielding layouts and material combinations with variations in materials, geometry, and thickness.

5. The accelerated calculation method for radiation fields based on a denoising neural network as described in claim 3, characterized in that, The unified preprocessing of the generated low-particle-number MC simulation results and high-particle-number MC simulation results includes: The particle dose rate or flux distribution data obtained from the low particle number MC simulation results and the high particle number MC simulation results are discretized into a predefined regular three-dimensional voxel grid to form three-dimensional tensor data; The radiation field values ​​within the three-dimensional voxel mesh are normalized to obtain preprocessed low-particle-number MC simulation results and high-particle-number MC simulation results.

6. The accelerated calculation method for radiation fields based on a denoising neural network as described in claim 2, characterized in that, The deep radiation field denoising neural network model includes: The input layer is used to input data; The spatial attention module, located in the input layer, is used to assign learnable weights to different spatial locations by learning the spatial features of the input data. The encoder is used to progressively extract features from the input data and reduce the spatial dimension while increasing the number of feature channels through successive convolutional layers and average pooling layers. The decoder is used to gradually restore the spatial dimension through successive upsampling operations, while reducing the number of feature channels through convolution; Skip connection layers are used to directly connect feature maps in the encoder to feature maps of the corresponding scale in the decoder; The output layer uses a linear activation function to output data that restores the spatial dimension as the prediction result.

7. The accelerated calculation method for radiation fields based on a denoising neural network as described in claim 2, characterized in that, The deep radiation field denoising neural network model, based on the preprocessed high-quality dataset, is trained and optimized, including: Based on the preprocessed high-quality dataset, a weighted combined loss function is used to train the constructed deep radiation field denoising neural network model. The Adam optimizer is used to optimize the trained deep radiation field denoising neural network model.

8. A radiation field acceleration calculation system based on a denoising neural network, characterized in that, include: The data acquisition module is used to acquire low-particle-number MC simulation results for the test scenario and perform preprocessing. The calculation module is used to input the preprocessed low-particle-number MC simulation results into the pre-trained and optimized deep radiation field denoising neural network model to obtain the high-particle-number MC simulation results of the scene under test. The deep radiation field denoising neural network model is obtained by using an improved UNet convolutional neural network.

9. A processing device, characterized in that, It includes computer program instructions, wherein when executed by a processing device, the computer program instructions are used to implement the steps corresponding to the radiation field acceleration calculation method based on a denoising neural network as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions, wherein when the computer program instructions are executed by a processor, they are used to implement the steps corresponding to the radiation field acceleration calculation method based on a denoising neural network as described in any one of claims 1-7.

Images