Intelligent optimization method for neutron beam shaping assembly

Abstract

The invention discloses an intelligent optimization method for a neutron beam shaping assembly. The intelligent optimization method comprises the following steps: establishing a beam shaping assembly basic model; randomly generating and combining a plurality of beam shaping component models within the value range of each optimization parameter; using a Monte Carlo method to simulate beam outlet parameters of the plurality of beam shaping assembly models, and extracting related calculation results; taking the correlation calculation result as input, training a neural network, and directly predicting the result according to optimization parameters; and carrying out multi-objective optimization by using a BSA multi-objective optimization method in which a multi-objective optimization algorithm and a neural network are coupled, and evaluating Pareto optimal solutions under various optimization methods through a hyper-volumetric method to obtain a BSA optimal combination scheme. According to the method, a large number of combination models can be efficiently generated within the value range of the optimization parameters, the optimal combination parameters are screened out, the optimization effect and optimization efficiency of the beam shaping assembly are improved, and the workload of a designer is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of boron neutron capture therapy, and particularly to an intelligent optimization method for a neutron beam shaping assembly. Background Art

[0002] When protons are accelerated to a certain energy by an accelerator and bombarded on a target, neutrons can be generated. The neutrons generated by proton bombardment of the target have relatively high energy, strong anisotropy, and are accompanied by a certain amount of γ rays, and cannot be directly used for boron neutron capture therapy (BNCT). The beam shaping assembly (BSA) is one of the key components of a boron neutron capture therapy device. It is located behind the target and its function is to shape and slow down the high-energy neutron beam generated by proton bombardment of the target into a thermal neutron beam (0.5 eV < E < 10 keV) that meets the treatment standard. The BSA mainly consists of a moderator, a reflector, a filter, a collimator, etc.

[0003] However, it is difficult to achieve a thermal neutron beam that meets the IAEA recommended values by traditional step-by-step optimization methods. Since various reaction cross-sections such as scattering and absorption of materials vary with energy, when neutrons pass through materials, scattering and absorption occur simultaneously, and the effects of materials of different sizes are also different. This results in the mutual influence of various parameters when optimizing the BSA. Therefore, the materials recommended by traditional methods in different regions can only be qualitative conclusions, and these conclusions are often only valid for neutrons in a certain energy range, and have little effect on the optimization design. Since there are many types of materials and the adjustable range of sizes is huge, the number of solutions obtained by simultaneously adjusting sizes and materials is almost infinite, and this process will take a lot of time. Moreover, the essence of the traditional step-by-step optimization method is to iteratively find the optimal solution from a small number of individuals, which is prone to falling into a local optimal solution and difficult to optimize multiple parameters simultaneously.

[0004] At the same time, although the neutron beam parameters at the collimator outlet of the BSA scheme obtained by the traditional step-by-step optimization method are relatively good, there are still many harmful beams, mainly neutrons, above and below the collimator outlet. During irradiation treatment, these neutrons will be absorbed by normal human tissues, and the dose generated will cause certain damage to normal tissues. Therefore, after designing the structure of the BSA, it is necessary to consider the shielding above and below the collimator outlet, and consider the optimization of light weight and radiation protection, which will also take a lot of time. Summary of the Invention

[0005] In order to overcome the deficiencies of the prior art, the present invention proposes an intelligent optimization method for a neutron beam shaping assembly. The technical solution of the present invention is realized as follows: An intelligent optimization method for a neutron beam shaping assembly includes the steps of:

[0006] S1: Establish a basic model of the beam shaping assembly, and the model includes a moderator, a reflector, a filter, and a collimator of the beam assembly;

[0007] S2: determining the value range of each optimization parameter, and randomly generating and combining multiple beam shaping component models within the value range of each optimization parameter;

[0008] S3: using a Monte Carlo method to simulate the beam outlet parameters of the multiple beam shaping component models, and extracting relevant calculation results;

[0009] S4: using the relevant calculation results as input, training a neural network, and directly predicting the results according to the optimized parameters;

[0010] S5: Use the BSA multi-objective optimization method coupled with a multi-objective optimization algorithm and a neural network to perform multi-objective optimization. Use the hypervolume method to evaluate the Pareto optimal solutions under various optimization methods, analyze the optimization degree of each optimization objective in the optimal solution compared with the original solution, and obtain the BSA optimal combination solution.

[0011] Furthermore, the S2 comprises the steps of: determining the secondary neutron source term generated by the proton beam target according to the beam shaping component model; and the S3 comprises the steps of:

[0012] S31: In combination with the secondary neutron source term, the material type and geometric parameters of each part in the beam shaping component model are changed, and the Monte Carlo method is used to perform simulation calculations, and the calculation results are compared with the IAEA epithermal neutron beam recommended values ​​to determine the material selection and optimized size range of the moderator, reflector, filter, and collimator in the device model;

[0013] S32: randomly generating a specified number of Monte Carlo calculation samples of the beam shaping component model within the optimization interval according to the set optimization variables and the corresponding optimization interval;

[0014] S33: inputting the generated Monte Carlo samples of the specified number of beam shaping component models into the Monte Carlo program to simulate the neutron and photon transport process of the beam shaping component model during operation, and extracting and counting the simulation calculation results of the neutron parameters at the beam outlet of the beam shaping component model.

[0015] Furthermore, the neural network in S4 is an error back propagation neural network, and S4 comprises the steps of:

[0016] S41: The free path number of particles in different parts replaces the thickness and material type characteristics, selects the energy values ​​that may exist when the particles are transported in them, these energy values ​​are mainly distributed between 0.5eV and 1MeV, and calculates the mean free path number under these energies as the input layer parameters of the neural network;

[0017] S42: using the calculation results of the beam outlet of the beam shaping component model obtained by statistics as the output layer parameters of the neural network;

[0018] S43: setting appropriate training set ratio, validation set ratio, test set ratio, learning rate and number of iterations parameters to perform neural network training;

[0019] S44: When the error of the validation set drops to a specific value, the iteration ends and a binary file is generated, which saves the neuron parameter information of the neural network.

[0020] Furthermore, the S5 comprises the steps of:

[0021] S51: The optimization target is set as the physical quantities of the neutron beam at the BSA beam outlet, which are the epithermal neutron flux density of the neutron beam, the fast neutron dose corresponding to the epithermal neutron per unit flux density, the gamma dose corresponding to the epithermal neutron per unit flux density, the ratio of the neutron flux and flux density, and the thermal neutron flux density corresponding to the epithermal neutron per unit flux density. The appropriate population size, number of iterations, optimization variables, and optimization target are set to start the optimization.

[0022] S52: Generate a new population based on the existing population through the operation operator in the genetic algorithm, and the new population is the regenerated optimization model;

[0023] S53: If the number of iterations meets the termination condition, multiple optimization models under the last number of iterations are outputted as the optimization result in descending order of fitness, that is, the best combination parameters of each part of the beam shaping assembly device are obtained; otherwise, the iteration is continued until the number of iterations meets the termination condition;

[0024] S54: Evaluate and analyze the parameters of the Pareto optimal solution finally obtained, and determine the optimal combination scheme of BSA according to the recommended values ​​given by the International Atomic Energy Agency for the boron neutron capture therapy device.

[0025] Furthermore, the genetic algorithm in S52 includes but is not limited to selection operation, crossover operation and mutation operation.

[0026] Furthermore, the selection operation operator in S52 includes but is not limited to: roulette, tournament, and sorting.

[0027] An intelligent optimization method for a neutron beam shaping component further includes the steps of: S6: generating an optimal shielding scheme above and below the collimator outlet, including the steps of:

[0028] S61: Based on the BSA optimal combination solution, the particle trajectory information passing through the upper and lower parts of the collimator outlet is recorded as the initial source term of the shielding calculation, the corresponding shielding layer material is selected, the value range of the optimization parameter is determined, and multiple calculation samples are randomly generated within the optimization interval, and Monte Carlo calculation is performed;

[0029] S62: extract the calculation results, calculate the mean free path number to replace the material and thickness of the shielding layer as the input layer data of the neural network, use the photon flux data as the output layer data of the neural network, and use the small batch stochastic gradient descent algorithm with momentum to perform neural network learning;

[0030] S63: The S-metric is used to select the evolutionary multi-objective optimization algorithm for optimization. Finally, the Pareto optimal solutions under various optimization methods are evaluated by the hypervolume method, and the optimal shielding solution above and below the collimator exit is obtained under the BSA optimal combination model.

[0031] The beneficial effect of the present invention is that, compared with the prior art, the intelligent optimization method of the neutron beam shaping component of the present invention can efficiently generate a large number of combination models within the value range of the optimization parameters, and screen out the optimal combination parameters therefrom, thereby improving the optimization effect and optimization efficiency of the beam shaping component and reducing the workload of the designer. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a flow chart of the intelligent optimization method of the neutron beam shaping component of the present invention;

[0033] Figure 2 is a schematic diagram of a device model of a beam shaping assembly in one embodiment of the present invention;

[0034] Figure 3 is a schematic diagram of a neural network structure in one embodiment of the present invention;

[0035] Figure 4 It is a flow chart of an optimal shielding method for generating the upper and lower parts of the collimator outlet in one embodiment of the present invention. DETAILED DESCRIPTION

[0036] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not 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.

[0037] See also Figure 1 The present invention provides an intelligent optimization method for a neutron beam shaping component, comprising the steps of:

[0038] S1: Establish the basic model of beam shaping components, including the moderator, reflector, filter and collimator of the beam component;

[0039] S2: determining the value range of each optimization parameter, and randomly generating and combining multiple beam shaping component models within the value range of each optimization parameter;

[0040] S3: using a Monte Carlo method to simulate the beam outlet parameters of the multiple beam shaping component models, and extracting relevant calculation results;

[0041] S4: using the relevant calculation results as input, training a neural network, and directly predicting the results according to the optimized parameters;

[0042] S5: Use the BSA multi-objective optimization method coupled with a multi-objective optimization algorithm and a neural network to perform multi-objective optimization. Use the hypervolume method to evaluate the Pareto optimal solutions under various optimization methods, analyze the optimization degree of each optimization objective in the optimal solution compared with the original solution, and obtain the BSA optimal combination solution.

[0043] See also Figure 2 The model of the beam shaping assembly established in the present invention includes multiple parts such as a moderator, a reflector, a filter, and a collimator. According to these parts, the secondary neutron source term generated by the proton beam target can be determined. In combination with this neutron source term, the material type and geometric parameters of each part in the device model are changed, and the Monte Carlo method is used for simulation calculation. The calculation results are compared with the IAEA epithermal neutron beam recommended values ​​to determine the material selection and optimized size range of the moderator, reflector, filter, and collimator in the device model. According to the set optimization variables and the corresponding optimization interval, the Monte Carlo calculation samples of the specified number of device models are randomly generated within the optimization interval. The Monte Carlo samples of the specified number of device models are input into the Monte Carlo program to simulate the neutron and photon transport process when they are working, and the simulation calculation results of the neutron parameters at the beam outlet of the device model are extracted and counted.

[0044] See also Figure 3 , the neural network described in S4 is an error back propagation neural network. Error back propagation neural network (BP-Net) is one of the most widely used and successful neural network algorithms. BP-Net consists of two main parts: weight forward propagation and result calculation, error back propagation and weight update.

[0045] Weight forward propagation and calculation is also called node output model. The nodes of the neural network are divided into input layer and hidden layer. Each node retains certain weight information. The output of the hidden layer node is:

[0046]

[0047] The number of output layer nodes is:

[0048]

[0049] In the above two formulas, It is an activation function (or transfer function), which is generally divided into linear activation function and nonlinear activation function. Nonlinear activation function is often used in practical problems due to its advantages of nonlinear expression. is the bias (or threshold) of the neural unit; and They represent the weights of each node in the hidden layer and the output layer respectively.

[0050] Error back propagation and weight update mainly include error calculation and weight adjustment.

[0051] Due to the characteristics of the BP algorithm, it has the advantages of adaptive learning, strong nonlinear mapping ability and strong generalization ability. The trained neural network has strong generalization ability and can make reasonable predictions for data that has not been encountered in the training process. It can be used to process approximate solutions to some complex problems and has broad application prospects in optimization problems. BP-Net consists of an input layer, a hidden layer and an output layer.

[0052] Since the material type, thickness / diameter of each part of the device model are related to the number of free paths of the particles in it, the number of free paths of the particles in different parts is used to replace the thickness and material type characteristics, and some energy values ​​that may exist when the particles are transported in it are selected. These energy values ​​are mainly distributed between 0.5eV and 1MeV, and the average free path number under these energies is calculated as the input layer parameter of the neural network. At the same time, the calculation results of the beam outlet of the device model are statistically obtained as the output layer parameter of the neural network. Set appropriate training set ratio, validation set ratio, test set ratio, learning rate, number of iterations and other parameters of the neural network, and perform neural network training. When the error of the validation set drops to a specific value, the iteration ends and a binary file is generated, which saves the neuron parameter information of the neural network.

[0053] The hypervolume index is also called the S metric. The basic idea of ​​the algorithm is to process only one-dimensional targets at a time, so as to accurately calculate the hypervolume value. The evaluation indicators of multi-objective optimization are divided into convergence index, uniformity index and extensiveness index. The hypervolume index can simultaneously evaluate the convergence, uniformity and extensiveness of the optimization solution, and give a comprehensive evaluation result of the optimization solution. The larger the hypervolume value, the better the comprehensive performance of the corresponding optimization solution.

[0054] The S-metric selection evolutionary multi-objective optimization algorithm (SMS-EMOA) draws on the non-dominated sorting genetic algorithm with elite strategy (NSGA-Ⅱ), uses non-dominated sorting as the sorting criterion, and uses hypervolume index as the selection criterion. It is a typical hypervolume-based multi-objective evolutionary algorithm. In each non-dominated block in the population iteration process, the slicing method is used to calculate the hypervolume value of all individuals. The hypervolume value of the individual's independent contribution is obtained by subtracting the hypervolume value of the individual after removing the hypervolume value of the individual from the hypervolume value of all individuals in the entire dominated block. In the SMS-EMOA algorithm, the independent contribution hypervolume value of each individual is used as the fitness of the individual, and the fitness determines whether the individual is retained. The fitness function value is used to evaluate the quality of the problem solution, and the global search of the problem solution space is achieved by using selection, crossover, and mutation operators.

[0055] The step S5 of the present invention includes the steps: S51: the optimization target is the physical quantity of the neutron beam at the BSA beam outlet, which are the epithermal neutron flux density of the neutron beam, the fast neutron dose corresponding to the epithermal neutron per unit flux density, the gamma dose corresponding to the epithermal neutron per unit flux density, the ratio of the neutron flow rate to the flux density, and the thermal neutron flux density corresponding to the epithermal neutron per unit flux density. Set appropriate parameters such as population size, number of iterations, optimization variables, and optimization target to start optimization.

[0056] S52: Generate a new population based on the existing population through the operation operators in the genetic algorithm such as selection operation, crossover operation, and mutation operation. The new population is the regenerated optimization model. The selection operation embodies the idea of ​​survival of the fittest in nature. It is an operation of eliminating the fittest for individuals. Individuals with higher fitness are inherited to the next generation population, while individuals with lower fitness are eliminated. Commonly used selection operators include: roulette, tournament, sorting, etc., and those skilled in the art can choose according to actual conditions, without limitation. Crossover operation refers to the operation of replacing and recombining the partial structures of two parent individuals to generate new individuals, that is, replacing and recombining the values ​​of the parameters to be optimized in the two existing optimization models to generate a new optimization model. Mutation operation refers to the change of the offspring gene caused by a small probability disturbance. The mutation operation changes the gene values ​​on certain loci of individuals in the population, that is, changes certain selected values ​​of the parameters to be optimized in the optimization model. There are two purposes for introducing mutation in the genetic algorithm: one is to enable the genetic algorithm to have local random search capabilities. When the genetic algorithm has approached the optimal solution neighborhood through the crossover operator, the local random search capability of the mutation operator can be used to accelerate the convergence to the optimal solution. The second is to enable the genetic algorithm to maintain population diversity to prevent premature convergence.

[0057] S53: Determine whether the optimization meets the termination condition. If the number of iterations meets the termination condition, multiple optimization models under the last number of iterations are output as the optimization result in descending order of fitness, that is, the best combination parameters of each part of the beam shaping component device are obtained. Otherwise, continue to iterate, repeat the selection operation, crossover operation, mutation operation and other operations until the number of iterations meets the termination condition.

[0058] S54: Evaluate and analyze the parameters of the Pareto optimal solution finally obtained, and determine the optimal combination scheme of BSA according to the recommended values ​​given by the International Atomic Energy Agency (IAEA) for the boron neutron capture therapy device.

[0059] See also Figure 4 An important function of the present invention is to provide an optimal shielding scheme for the upper and lower parts of the collimator outlet. The shielding scheme includes the steps of:

[0060] S61: Based on the above BSA optimal combination model, the particle trajectory information passing through the upper and lower parts of the collimator outlet is recorded as the initial source term of the shielding calculation, a more suitable shielding layer material is selected, the value range of the optimization parameter is determined, and multiple calculation samples are randomly generated within the optimization interval, and Monte Carlo calculation is performed;

[0061] S62: extract the calculation results, calculate the mean free path number to replace the material and thickness of the shielding layer as the input layer data of the neural network, use the photon flux data as the output layer data of the neural network, and use the small batch stochastic gradient descent algorithm with momentum to perform neural network learning;

[0062] S63: Use the S-metric to select the evolutionary multi-objective optimization algorithm to find the best solution. Finally, the Pareto optimal solution under various optimization methods is evaluated by the hypervolume method, and the optimal shielding solution above and below the collimator outlet is obtained under the BSA optimal combination model. The S metric is a hypervolume indicator, which is used to evaluate the convergence, uniformity and extensiveness of the optimization solution, and gives a comprehensive evaluation result of the optimization solution. The larger the hypervolume value, the better the overall performance of the corresponding optimization solution. The SMS-EMOA algorithm uses non-dominated sorting as the sorting criterion and the hypervolume indicator as the selection criterion. It is a typical multi-objective evolutionary algorithm based on the hypervolume indicator.

[0063] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. An intelligent optimization method for a neutron beam shaping component, characterized in that: Includes steps: S1: Establish the basic model of beam shaping components, including the moderator, reflector, filter and collimator of the beam component; S2: determining the value range of each optimization parameter, and randomly generating and combining multiple beam shaping component models within the value range of each optimization parameter; S3: using a Monte Carlo method to simulate the beam outlet parameters of the multiple beam shaping component models, and extracting relevant calculation results; S4: using the relevant calculation results as input, training a neural network, and directly predicting the results according to the optimized parameters; S5: Use the BSA multi-objective optimization method coupled with a multi-objective optimization algorithm and a neural network to perform multi-objective optimization. Use the hypervolume method to evaluate the Pareto optimal solutions under various optimization methods, analyze the optimization degree of each optimization objective in the optimal solution compared with the original solution, and obtain the BSA optimal combination solution.

2. The intelligent optimization method for a neutron beam shaping assembly according to claim 1, characterized in that: The S2 comprises the steps of: determining the secondary neutron source term generated by proton beam target shooting according to the beam shaping component model; and the S3 comprises the steps of: S31: In combination with the secondary neutron source term, the material type and geometric parameters of each part in the beam shaping component model are changed, and the Monte Carlo method is used to perform simulation calculations, and the calculation results are compared with the IAEA epithermal neutron beam recommended values ​​to determine the material selection and optimized size range of the moderator, reflector, filter, and collimator in the device model; S32: randomly generating a specified number of Monte Carlo calculation samples of the beam shaping component model within the optimization interval according to the set optimization variables and the corresponding optimization interval; S33: inputting the generated Monte Carlo samples of the specified number of beam shaping component models into the Monte Carlo program to simulate the neutron and photon transport process of the beam shaping component model during operation, and extracting and counting the simulation calculation results of the neutron parameters at the beam outlet of the beam shaping component model.

3. The intelligent optimization method for a neutron beam shaping assembly according to claim 1, characterized in that: The neural network in S4 is an error back propagation neural network, and S4 comprises the steps of: S41: The free path number of particles in different parts replaces the thickness and material type characteristics, selects the energy values ​​that may exist when the particles are transported in them, these energy values ​​are mainly distributed between 0.5eV and 1MeV, and calculates the mean free path number under these energies as the input layer parameters of the neural network; S42: using the calculation results of the beam outlet of the beam shaping component model obtained by statistics as the output layer parameters of the neural network; S43: setting appropriate training set ratio, validation set ratio, test set ratio, learning rate and number of iterations parameters to perform neural network training; S44: When the error of the validation set drops to a specific value, the iteration ends and a binary file is generated, which saves the neuron parameter information of the neural network.

4. The intelligent optimization method for a neutron beam shaping assembly according to claim 1, characterized in that: The S5 comprises the steps of: S51: The optimization target is set as the physical quantities of the neutron beam at the BSA beam outlet, which are the epithermal neutron flux density of the neutron beam, the fast neutron dose corresponding to the epithermal neutron per unit flux density, the gamma dose corresponding to the epithermal neutron per unit flux density, the ratio of the neutron flux and flux density, and the thermal neutron flux density corresponding to the epithermal neutron per unit flux density. The appropriate population size, number of iterations, optimization variables, and optimization target are set to start the optimization. S52: Generate a new population based on the existing population through the operation operator in the genetic algorithm, and the new population is the regenerated optimization model; S53: If the number of iterations meets the termination condition, multiple optimization models under the last number of iterations are outputted as the optimization result in descending order of fitness, that is, the best combination parameters of each part of the beam shaping assembly device are obtained; otherwise, the iteration is continued until the number of iterations meets the termination condition; S54: Evaluate and analyze the parameters of the Pareto optimal solution finally obtained, and determine the optimal combination scheme of BSA according to the recommended values ​​given by the International Atomic Energy Agency for the boron neutron capture therapy device.

5. The intelligent optimization method for a neutron beam shaping assembly according to claim 4, characterized in that: The genetic algorithm in S52 includes but is not limited to selection operation, crossover operation and mutation operation.

6. The intelligent optimization method for a neutron beam shaping assembly according to claim 1, characterized in that: The selection operation operator in S52 includes but is not limited to: roulette, tournament, and sorting.

7. The intelligent optimization method for a neutron beam shaping assembly according to claim 1, characterized in that: The step S6 is also included: generating an optimal shielding scheme above and below the collimator outlet, including the steps of: S61: Based on the BSA optimal combination solution, the particle trajectory information passing through the upper and lower parts of the collimator outlet is recorded as the initial source term of the shielding calculation, the corresponding shielding layer material is selected, the value range of the optimization parameter is determined, and multiple calculation samples are randomly generated within the optimization interval, and Monte Carlo calculation is performed; S62: extract the calculation results, calculate the mean free path number to replace the material and thickness of the shielding layer as the input layer data of the neural network, use the photon flux data as the output layer data of the neural network, and use the small batch stochastic gradient descent algorithm with momentum to perform neural network learning; S63: The S-metric is used to select the evolutionary multi-objective optimization algorithm for optimization. Finally, the Pareto optimal solutions under various optimization methods are evaluated by the hypervolume method, and the optimal shielding solution above and below the collimator exit is obtained under the BSA optimal combination model.

Images