A method for quickly constructing neutron spectrum in simulated work sites
Through the technology combined with deep learning and Monte Carlo method, the problems of large calculation errors and long time-consuming in the construction of neutron spectrum on the simulation work site are solved, and rapid construction of neutron spectrum and accurate slow configuration scheme design are realized.
Patent Information
- Application Number
- CN202210527270.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-16
- Publication Date
- 2025-04-18
- Estimated Expiration
- 2042-05-16
AI Technical Summary
When simulating neutron spectrum on the work site, the existing technology has large calculation errors, long time-consuming and difficult to quickly design satisfactory slow configuration schemes, resulting in inaccurate construction of neutron spectrum.
Deep learning technology combined with Monte Carlo method is adopted to achieve rapid construction of neutron spectrum by making sample data sets, training neural network models, deploying models, and using multi-objective optimization algorithms to find global optimal slow configuration solutions.
The rapid construction of neutron spectrum is achieved, the calculation accuracy is close to that of the Monte Carlo method, and the calculation speed is greatly improved, and the slow configuration solution can be optimized according to economics, material quality and volume requirements.
Smart Images
Figure CN115166811B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the fields of radiation safety assessment, neutron detection instrument calibration field construction and deep learning, and relates to a method for quickly constructing a neutron spectrum of a simulated work site. Background Art
[0002] Accurately measuring the neutron dose level at the work site is particularly important for radiation safety assessment and radiation protection of workers. Usually, the neutron energy spectrum at the work site is wide, with an energy range of 10 -9 ~20MeV; the neutron ambient dose equivalent rate H*(10) is closely related to the neutron energy density. Measurement studies conducted by the European Union have shown that the on-site measurement error of neutron dosimeters based on the GB / T14055 series (international standard ISO 8259 series) standard reference radiation field scale is large. At different sites such as reactors, spent fuel transport containers, and spent fuel reprocessing, the measurement results of most commonly used neutron dosimeters differ from the reference values by about 2 times. If the energy spectrum of the calibration source is different from the energy spectrum of the work site where the dosimeter will be used, it is considered that the dosimeter reading is incorrectly determined.
[0003] The simulated work site neutron spectrum technology uses a laboratory neutron source to configure a moderator, scatterer and shield to establish a calibration field that is more representative of the work field energy spectrum. It is an effective solution to the large error between the reference value and the measured value of the dose meter.
[0004] The construction of a neutron spectrum calibration field for a simulated work site requires the design of neutron moderation. The physical process of neutron moderation is relatively complex, and the moderation calculation error based on empirical formulas or attenuation coefficients is large. Usually, a large moderation margin is required, resulting in inaccurate neutron spectrum at the simulated work site. The Monte Carlo method (MC) can accurately calculate neutron moderation to obtain the neutron spectrum of the simulated site, but it is extremely computer-intensive and time-consuming. In addition, the neutron energy spectrum characteristics of different workplaces are different, so the moderation materials and structural configuration schemes are different. It is cumbersome to design moderation configuration schemes for specific sites one by one, which requires a calculation method to achieve a high calculation accuracy while ensuring that computer resources are less occupied and time-consuming. Summary of the invention
[0005] In view of the defects existing in the prior art, the purpose of the present invention is to provide a method for quickly constructing a neutron spectrum of a simulated work site. By combining deep learning technology with the Monte Carlo method, a moderation configuration scheme can be quickly designed according to the neutron source energy spectrum and the work site neutron energy spectrum. At the same time, it can also ensure that the obtained simulated on-site neutron spectrum and the actual workplace neutron spectrum meet the set accuracy requirements.
[0006] To achieve this object, the present invention provides a method for quickly constructing a neutron spectrum of a simulated work site, the method comprising the following steps:
[0007] S1, sample data set for making neural network model;
[0008] S2, the neural network model after training and learning;
[0009] S3, deploying the learned neural network model to hardware;
[0010] S4. Use a multi-objective optimization algorithm to find the global optimal slowing configuration solution.
[0011] Furthermore, the sample data set preparation process includes the following two steps:
[0012] S11. Based on the Monte Carlo method, simulate the neutron energy spectrum after neutrons are slowed down, scattered and shielded to obtain sample data and sample labels;
[0013] The sample data includes energy spectrum sample data and moderation configuration scheme, the energy spectrum sample data includes neutron source energy spectrum and simulated on-site neutron energy spectrum data, and the moderation configuration scheme is a sample label;
[0014] S12, preprocessing the multiple groups of energy spectrum sample data and sample labels to generate the sample data set;
[0015] The sample data set includes two parts: a training set and a test set, wherein the training set is used to train the neural network model, and the test set is used to verify the accuracy of the neural network model in predicting the slowing configuration scheme and the sample label.
[0016] Furthermore, in step S11, a Monte Carlo simulation particle transport tool is used to model the moderation structure of the isotope neutron source or the accelerator neutron source, and a series of simulation calculations are carried out for a variety of moderation configuration schemes to obtain simulated on-site neutron energy spectrum data under different moderation configuration schemes;
[0017] The moderation configuration scheme includes moderation configuration features using various moderation materials, various size structures and shapes, and various moderation material placement sequences, and the moderation configuration scheme forms a sample label.
[0018] Furthermore, in step 12, the method for preprocessing the multiple sets of energy spectrum sample data and sample labels includes: normalizing the energy spectrum sample data, and numerically representing the sample labels to produce a sample data set for neural network model training and testing.
[0019] Furthermore, in step 12, the method for preprocessing the energy spectrum sample data and sample labels also includes: before inputting into the neural network model, a random shuffling and batch processing process of the sample data sorting in the sample data set.
[0020] Further, the step S2 includes the following four steps:
[0021] S21. Construct a multi-layer neural network model, wherein the neural network model includes but is not limited to a fully connected neural network model;
[0022] S22, construct a loss function, process input data through a multi-layer network structure, and calculate the loss value;
[0023] S23, model training, iteratively calculating the input training set, finding the minimum value of the variance of the loss value through logistic regression gradient descent, back-propagating the gradient to the parameters of the neural network model, and updating the weight values in the network according to the update rule; tuning the hyperparameters of the neural network model to obtain the learned neural network model;
[0024] S24. Verifying the accuracy of the learned neural network model through the test set.
[0025] Further, in step S21, the neural network model is a fully connected neural network model, and the fully connected neural network model adopts a 5-layer structure: 1 input layer, 3 hidden layers and 1 output layer. The function of the input layer is to obtain the energy spectrum sample data, the hidden layer is to further abstract the energy spectrum sample data, and the output layer outputs a predicted slowing configuration plan based on the energy spectrum sample data.
[0026] Further, in step S22, the loss function is used to calculate the difference between the predicted slowing configuration scheme output by the neural network model and the sample label, and the difference is the loss value;
[0027] If the number of parameters of the predicted slowing configuration scheme output by the neural network model is 2, the loss function adopts a logarithmic loss function;
[0028] If the number of parameters of the predicted slowing configuration scheme output by the neural network model is greater than 2, the loss function adopts the cross entropy loss function.
[0029] Further, in step S23, the model training is to find the minimum value of the variance of the loss value in the training set, that is, the minimum point of the cost function, by iteratively updating the neural network weight value, and obtain a neural network model with generalization ability based on this;
[0030] In the iterative calculation, the gradient of the cost function is calculated, and the gradient is reversely transferred to the neural network model, and the network weight value is updated in combination with the learning rate hyperparameter;
[0031] The cost function is a variance function of the loss value, and a regularization method is used to prevent the neural network model from overfitting.
[0032] Further, in step S24, the test set is used to verify the accuracy of the learned neural network model. During the verification process, there is no need to update the network weight value. After the verification is passed, the learned neural network model can be deployed.
[0033] Further, in step S3, after the learned neural network model is deployed to the hardware, the on-site neutron spectrum and neutron source energy spectrum data required for simulation are input, and the learned neural network model will output a predicted moderation configuration plan that meets the set accuracy requirements.
[0034] Further, in step S4, the multi-objective optimization algorithm seeks a global optimal shielding configuration solution based on actual engineering implementation factors, wherein the actual engineering implementation factors include but are not limited to economic cost, quality of moderator material and volume of moderator.
[0035] The beneficial effect of the present invention is that the method for quickly constructing the neutron spectrum of the simulated work site provided by the present invention adopts the Monte Carlo method to simulate neutron moderation, scattering and shielding, obtain the neutron spectrum after moderation, and combine the neutron source energy spectrum, moderation configuration scheme and the neutron energy spectrum result after moderation used in the simulation to make a data set, and then use the data set in combination with the deep learning method to train the neural network model, and predict the moderation configuration scheme according to the required work site neutron spectrum and neutron source energy spectrum combined with the network model, and finally use the multi-objective optimization algorithm to find the global optimal configuration scheme. The method for quickly constructing the neutron spectrum of the simulated work site provided by the present invention can quickly calculate the neutron spectrum moderation configuration scheme of the simulated work site for any laboratory standard neutron source, and the configuration scheme obtained has a calculation accuracy similar to that of the Monte Carlo method, and can greatly improve the calculation speed. The method provided by the present invention can perform multi-objective optimization on the moderation configuration scheme according to the economic efficiency, the quality of the moderation material and the volume requirements, and find the global optimal solution. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 A workflow diagram of a method for rapidly constructing a neutron spectrum of a simulated work site provided in a specific embodiment of the present invention.
[0037] Figure 2 A structural diagram of a fully connected neural network model provided for a specific implementation of the present invention.
[0038] Figure 3 A calculation flow chart of a multi-objective optimization algorithm provided for a specific implementation of the present invention. DETAILED DESCRIPTION
[0039] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0040] like Figures 1 to 3 As shown, an embodiment of the present invention provides a method for quickly constructing a simulated work site neutron spectrum. The method simulates neutron moderation, scattering and shielding based on the Monte Carlo method, obtains the neutron energy spectrum after moderation, and combines the neutron source energy spectrum, the moderation configuration scheme and the neutron spectrum after moderation (i.e., the simulated on-site neutron spectrum) to produce a data set; then, the relationship between the moderation configuration scheme and the neutron energy spectrum is deeply learned by combining the neural network model; the neural network model after learning is deployed to the hardware, and the neural network model can be used to quickly construct a moderation configuration scheme according to the neutron energy spectrum and neutron source energy spectrum of the work site. When there are multiple moderation configuration schemes that meet the accuracy requirements for the same on-site neutron spectrum, the multi-objective optimization algorithm can be further combined to comprehensively consider factors such as economy, quality and volume of moderation materials to find the global optimal moderation configuration scheme.
[0041] The method mainly comprises the following steps:
[0042] S1, sample data set for making neural network model;
[0043] The process of making the input sample data set includes the following two steps:
[0044] S11. Based on the Monte Carlo method, simulate the neutron energy spectrum after neutrons are slowed down, scattered and shielded to obtain sample data and sample labels;
[0045] In this embodiment, the neutron spectrum data after neutrons are moderated, scattered and shielded based on the Monte Carlo method in step S11 needs to be modeled by a Monte Carlo simulation particle transport tool for a moderation structure of an isotope neutron source or an accelerator neutron source, and a series of simulation calculations are carried out for a variety of moderation configurations to obtain simulated on-site neutron energy spectrum data under different moderation configurations. The moderation configuration includes the use of various moderation materials, various size structures and shapes, and the order in which various moderation materials are placed.
[0046] The sample data includes energy spectrum sample data and moderation configuration scheme, and the moderation configuration scheme is a sample label. The energy spectrum sample data includes neutron source energy spectrum and simulated field neutron energy spectrum data.
[0047] In this article, neutron source spectrum and simulated on-site neutron spectrum refer to the neutron source energy spectrum and simulated on-site neutron energy spectrum respectively.
[0048] S12, preprocessing the multiple groups of energy spectrum sample data and sample labels to generate the sample data set;
[0049] The sample data set includes two parts: a training set and a test set, wherein the training set is used to train the neural network model, and the test set is used to verify the accuracy of the neural network model in predicting the slowing configuration and sample labels.
[0050] In this embodiment, the method of preprocessing the energy spectrum sample data and sample labels in step S12 includes: normalizing the energy spectrum sample data, and numerically representing the sample labels, and preparing a sample data set for neural network model training and testing. That is, the neutron source spectrum and simulated field neutron spectrum data are normalized, and the moderation configuration features such as the type, size structure, and placement order of the moderator material are numerically represented to prepare a sample data set and a sample label set for neural network model training.
[0051] Furthermore, the method for preprocessing the energy spectrum sample data and sample labels also includes: before inputting into the neural network model, a random shuffling and batch processing process of the sample data sorting in the sample data set.
[0052] S2, the neural network model after training and learning;
[0053] The step S2 includes the following two steps:
[0054] S21. Construct a multi-layer neural network model, wherein the neural network model includes but is not limited to a fully connected neural network model;
[0055] In this implementation, step S21 constructs a multi-layer fully connected neural network model, which adopts a 5-layer structure: 1 input layer, 3 hidden layers and 1 output layer. The function of the input layer is to obtain the energy spectrum sample data, the hidden layer is to further abstract the energy spectrum sample data, and the output layer outputs a predicted slowing configuration plan based on the energy spectrum sample data.
[0056] S22, construct a loss function, process the input data through a multi-layer network structure, and calculate the loss value;
[0057] In step S22 of this embodiment, the loss function is used to calculate the difference between the predicted moderation configuration scheme (output value) output by the fully connected neural network model and the sample label (true value), and this difference is the loss value. If the number of parameters of the predicted moderation configuration scheme output by the fully connected neural network model is 2, the logarithmic loss function is adopted for the loss function; if the number of parameters of the predicted moderation configuration scheme output by the fully connected neural network model is greater than 2, the cross-entropy loss function is adopted for the loss function. For example, when there are two types of moderation materials, the logarithmic loss function is adopted for the loss function; when the number of types of moderation materials is greater than 2, the cross-entropy loss function is adopted for the loss function. The specific form of the cross-entropy loss function is as follows:
[0058]
[0059] Among them, L represents the cross-entropy loss function; n represents the dimension of the output predicted moderation configuration data; x represents the x-th dimension, where 0 < x < n; y represents the numerically converted sample label value of the x-th dimension; a represents the predicted value of the data of the x-th dimension.
[0060] S23. Model training: Perform iterative calculations on the input training set, find the minimum value of the variance of the loss value through logistic regression gradient descent, backpropagate the gradient into the parameters of the fully connected neural network model, and update the weight values in the network according to the update rule; optimize the hyperparameters of the fully connected neural network model to obtain the learned fully connected neural network model;
[0061] In this embodiment, in step S23 of model training, the minimum value of the variance of the loss value in the training set, that is, the minimum value point of the cost function, is found by iteratively updating the neural network weight values, and based on this, a neural network model with generalization ability is obtained. In the iterative calculation, the gradient of the cost function is calculated and the gradient is backpropagated into the neural network, and the neural network weight values are updated in combination with the learning rate hyperparameter. The cost function is the variance function of the loss value, and overfitting of the neural network model is prevented through regularization methods.
[0062] S24. Verify the accuracy of the learned neural network model through the test set.
[0063] In step S24 of this embodiment, the test set is used to verify the accuracy of the learned fully connected neural network model. During the verification process, the network weight values do not need to be updated. After passing the verification, the learned fully connected neural network model can be deployed.
[0064] S3. Deploy the learned neural network model to the hardware;
[0065] The learned fully connected neural network model is deployed to the hardware, and the on-site neutron spectrum and neutron source energy spectrum data required for simulation are input. The learned fully connected neural network model will output a predicted moderation configuration plan that meets the set accuracy requirements.
[0066] S4. Use a multi-objective optimization algorithm to find the global optimal slowing configuration solution.
[0067] In this embodiment, the purpose of using the multi-objective optimization algorithm in step S4 is to perform a global optimal search for the predicted moderation configuration scheme output by the learned fully connected neural network model based on factors such as economic cost, moderator material quality and moderator volume.
[0068] In a specific embodiment, the workflow diagram of the method for quickly constructing a neutron spectrum of a simulated work site is as follows: Figure 1 As shown, the method for quickly constructing a simulated on-site neutron spectrum in this embodiment combines the advantages of high calculation accuracy of the Monte Carlo method and fast calculation speed of the deep learning method. The Monte Carlo method is used to simulate neutron spectrum data under different slowing configuration conditions, and the neutron source energy spectrum, slowing configuration scheme and neutron energy spectrum results after slowing down used in the simulation are combined to produce sample data sets and sample label sets for deep learning neural network model training and learning; the learned fully connected neural network model is deployed to the hardware, and the required simulated on-site neutron spectrum and neutron source energy spectrum data are input. The learned fully connected neural network model will output a predicted slowing configuration scheme that meets certain confidence requirements. Since the fully connected neural network model is based on probability and statistical theory, there will be a variety of predicted slowing configuration schemes that meet the requirements, rather than a unique solution. The optimal slowing configuration scheme needs to meet the quality, volume and economic requirements at the same time, which is a multi-objective optimization problem.
[0069] The quality of the sample data set determines the accuracy of the predicted moderation configuration scheme output by the network model. The sample data set production process is relatively complicated. First, it is necessary to model the neutron source and moderation configuration structure based on the simulated on-site neutron spectrum and conduct Monte Carlo simulation of particle transport. Then, the neutron spectrum output after moderation is normalized, and the characteristic numerical information such as the type, thickness and placement order of the moderation material is combined with the normalized neutron energy spectrum into a sample data set and a numerical sample label set. Before the sample data set is input into the fully connected neural network model, the sample data set and the digitized sample label set need to be randomly shuffled and batched. For example, 10 sample data sets are defined as a batch. After a batch of sample data sets are trained with the network model, the weight values of the neurons are updated.
[0070] The fully connected neural network model structure provided in this embodiment is as follows Figure 2As shown, the fully connected network model adopts a 5-layer structure: 1 input layer, 3 hidden layers and 1 output layer. The input layer receives the sample data set, and the number of neurons in the input layer is the number of frequencies in the energy spectrum histogram; the hidden layer abstracts the sample data set, and the number of hidden layers and neurons is variable, but the number of neurons in each hidden layer is required to be greater than the number of neurons in the input layer; the output layer outputs the predicted slowing configuration scheme, and the number of neurons in the output layer is equal to the number of parameters after the characteristics such as the number of types of slowing materials, size structure and shape, and the order of slowing material placement are digitized. The activation function of the fully connected neural network model runs on the neurons of the fully connected neural network model, and is responsible for mapping the input of the neuron to the output end. The activation function of the fully connected neural network model adopts the ReLU function, and the convergence speed is faster than the Sigmod and tanh functions. The training of the fully connected neural network model adopts the stochastic gradient descent algorithm with small batch processing to find the minimum point of the error between the true value and the predicted value; the adaptive learning rate parameter adjustment method is adopted to achieve the purpose of rapid convergence of the training model. Since the number of parameters of the predicted slowing configuration scheme output by the fully connected neural network model is greater than 2, the cross entropy loss function is used to evaluate the transmission and accumulation of errors during the training process of the fully connected neural network model.
[0071] The calculation flow chart of the multi-objective optimization algorithm provided in this embodiment is as follows: Figure 3 As shown in the figure, a simulated work site neutron spectrum is obtained. The moderation configuration scheme predicted for the neutron source is not the only solution, so the results of multiple calculations of the fully connected neural network model after learning are different; considering the actual engineering implementation needs, such as the volume, mass and economic cost of the moderator, it is necessary to find the optimal moderation configuration scheme that meets these conditions at the same time. To obtain the optimal moderation configuration scheme, it is necessary to weigh multiple aspects such as volume, mass and economic cost, because when the volume is optimal, the mass may be large and the cost may be high. The simplest multi-objective optimization method is the linear weighted method, but this method requires careful selection of weights and the results obtained are not necessarily the global optimal solution. The implementation method of the present invention adopts the non-dominated sorting genetic algorithm NSGA-II. First, the volume, mass and economic cost are calculated according to the candidate slowing configuration scheme to generate the initial population, the population size is set to 100, and the number of iterations is 1000. Then, non-dominated sorting is performed. Referring to the theory of biological evolution, the crowding degree of the non-dominated sorted population is compared, selected, crossed and mutated. The offspring population is generated according to the elite strategy. It is judged whether the termination condition is met. If so, the Pareto solution set with the global optimal shielding solution is output. Otherwise, it jumps to the non-dominated sorting step and continues genetic iteration until the termination condition is met. The algorithm gradually eliminates inferior solutions and retains high-quality solutions through genetic crossover mixing. The offspring of high-quality solutions are still high-quality, and the population formed by them is the Pareto solution set.
[0072] In this embodiment, the inventor uses machine learning technology combined with the Monte Carlo simulation particle transport method to effectively solve the problem of the conflict between the construction accuracy and speed of the neutron spectrum slowing configuration scheme for the simulated work site. The Monte Carlo method is used to accurately design the simulated site neutron spectrum slowing configuration scheme, combined with the neutron source energy spectrum, slowing configuration scheme and the neutron energy spectrum result data set after slowing down used in the simulation, and then the data set is combined with deep learning to obtain a neural network model, so as to use the neural network model to quickly predict any site neutron spectrum slowing configuration design scheme. The method provided in this embodiment combines the rapid design characteristics of machine learning with the characteristics of accurate calculation results of the Monte Carlo method.
[0073] The above embodiments are merely illustrative of the present invention, and those skilled in the art may make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.
Claims
1. A method for quickly constructing a neutron spectrum at a simulated work site, characterized in that: The method comprises the following steps: S1, sample data set for making neural network model; The process of making the sample data set includes the following two steps: S11. Based on the Monte Carlo method, simulate the neutron energy spectrum after neutrons are slowed down, scattered and shielded to obtain sample data and sample labels; The sample data includes energy spectrum sample data and moderation configuration scheme, the energy spectrum sample data includes neutron source energy spectrum and simulated on-site neutron energy spectrum data, and the moderation configuration scheme is a sample label; S12, preprocessing the multiple groups of energy spectrum sample data and sample labels to generate the sample data set; The sample data set includes two parts: a training set and a test set, wherein the training set is used to train the neural network model, and the test set is used to verify the accuracy of the neural network model in predicting the slowing configuration scheme and the sample label; S2, the neural network model after training and learning; S3, deploying the learned neural network model to hardware; S4. Use a multi-objective optimization algorithm to find the global optimal slowing configuration solution.
2. The method for quickly constructing a neutron spectrum of a simulated work site according to claim 1, characterized in that: In the step S11, a Monte Carlo simulation particle transport tool is used to model the moderation structure of the isotope neutron source or the accelerator neutron source, and a series of simulation calculations are carried out for a variety of moderation configuration schemes to obtain simulated on-site neutron energy spectrum data under different moderation configuration schemes; The moderation configuration scheme includes moderation configuration features using various moderation materials, various size structures and shapes, and various moderation material placement sequences, and the moderation configuration scheme forms a sample label.
3. The method for quickly constructing a neutron spectrum of a simulated work site according to claim 1, characterized in that: In step 12, the method for preprocessing the multiple sets of energy spectrum sample data and sample labels includes: normalizing the energy spectrum sample data, and numerically representing the sample labels to produce a sample data set for neural network model training and testing.
4. The method for quickly constructing a neutron spectrum at a simulated work site according to claim 1, characterized in that: In step 12, the method for preprocessing the energy spectrum sample data and sample labels also includes: before inputting into the neural network model, randomly shuffling and batch processing the sample data in the sample data set.
5. The method for quickly constructing a neutron spectrum of a simulated work site according to claim 1, characterized in that: The step S2 includes the following four steps: S21. Construct a multi-layer neural network model, wherein the neural network model includes but is not limited to a fully connected neural network model; S22, construct a loss function, process the input data through a multi-layer network structure, and calculate the loss value; S23, model training, iteratively calculating the input training set, finding the minimum value of the variance of the loss value through logistic regression gradient descent, back-propagating the gradient to the parameters of the neural network model, and updating the weight values in the network according to the update rule; tuning the hyperparameters of the neural network model to obtain the learned neural network model; S24. Verifying the accuracy of the learned neural network model through the test set.
6. A method for rapidly constructing a neutron spectrum at a simulated work site according to claim 5, characterized in that: In step S21, the neural network model is a fully connected neural network model, which adopts a 5-layer structure: 1 input layer, 3 hidden layers and 1 output layer. The function of the input layer is to obtain the energy spectrum sample data, the hidden layer is to further abstract the energy spectrum sample data, and the output layer outputs a predicted slowing configuration plan based on the energy spectrum sample data.
7. The method for quickly constructing a neutron spectrum of a simulated work site according to claim 5, characterized in that: In step S22, the loss function is used to calculate the difference between the predicted slowing configuration scheme output by the neural network model and the sample label, and the difference is the loss value; If the number of parameters of the predicted slowing configuration scheme output by the neural network model is 2, the loss function adopts a logarithmic loss function; If the number of parameters of the predicted slowing down configuration scheme output by the neural network model is greater than 2, the loss function adopts the cross entropy loss function.
8. The method for quickly constructing a neutron spectrum of a simulated work site according to claim 5, characterized in that: In step S23, the model training is to find the minimum value of the variance of the loss value in the training set, that is, the minimum point of the cost function, by iteratively updating the neural network weight value, and obtain a neural network model with generalization ability based on this; In the iterative calculation, the gradient of the cost function is calculated, and the gradient is reversely transferred to the neural network model, and the network weight value is updated in combination with the learning rate hyperparameter; The cost function is a variance function of the loss value, and a regularization method is used to prevent the neural network model from overfitting.
9. The method for quickly constructing a neutron spectrum of a simulated work site according to claim 5, characterized in that: In step S24, the test set is used to verify the accuracy of the learned neural network model. During the verification process, there is no need to update the network weight value. After the verification is passed, the learned neural network model can be deployed.
10. A method for quickly constructing a neutron spectrum of a simulated work site according to any one of claims 1 to 9, characterized in that: In step S3, after the learned neural network model is deployed to the hardware, the on-site neutron spectrum and neutron source energy spectrum data required for simulation are input, and the learned neural network model will output a predicted moderation configuration plan that meets the set accuracy requirements.
11. A method for quickly constructing a neutron spectrum at a simulated work site according to claim 10, characterized in that: In step S4, the multi-objective optimization algorithm searches for a global optimal shielding configuration solution based on actual engineering implementation factors, wherein the actual engineering implementation factors include but are not limited to economic cost, moderator material quality, and moderator volume.