Neural network based radiation dose prediction method for reactor shielding optimization

Abstract

The application discloses a kind of neural network-based radiation dose prediction methods for reactor shielding optimization, by simultaneously sampling the total number of shielding layers in the reactor shielding area, the thickness of each shielding layer, the material composition of the shielding layer, the combination mode and the arrangement order and other multi-parameters, generate data samples to train the neural network of the predictable radiation dose correction factor. For any shielding scheme in the shielding optimization design, the equivalent one-dimensional geometric radiation dose at the radiation dose evaluation target point is quickly obtained by one-dimensional discrete longitudinal scale method program for one-dimensional transport equivalent calculation, and then the corresponding radiation dose correction factor is predicted based on the trained neural network and the actual total radiation dose size at the radiation dose evaluation target point is obtained by correction calculation. Compared with the traditional method, the method can realize efficient and high-precision prediction of the radiation dose of the reactor shielding target area, and can be applied to important scenarios such as small nuclear reactor shielding optimization design in large parameter space.

Description

Technical Field

[0001] This invention belongs to the field of nuclear reactor engineering technology, specifically relating to a neural network-based radiation dose prediction method for reactor shielding optimization. Background Technology

[0002] Shielding optimization design is a crucial step in the design of small or micro nuclear reactors. Its goal is to simultaneously optimize multiple parameters within the reactor's shielding area, including the number of shielding layers, the thickness of each layer, and the composition, type, combination, and relative arrangement of shielding materials. This optimization aims to ensure that the radiation dose generated by ionizing radiation, primarily composed of neutrons and photons, emanating from the reactor core reaches safe levels in the target area, while minimizing the total thickness and mass of the shielding layers. This optimization operates within a large parameter space, involving extensive three-dimensional neutron-photon coupling transport calculations. Currently, the main computational methods are the Monte Carlo method and the discrete ordinate method. However, regardless of the method used, obtaining a high-accuracy solution typically requires tens to hundreds of nuclei hours to complete a single shielding scheme calculation, resulting in excessive computational costs that make it difficult to directly apply to reactor shielding optimization design tasks within a large parameter space. Summary of the Invention

[0003] To address the aforementioned problems, this invention provides a neural network-based radiation dose prediction method for reactor shielding optimization. This method first generates a shielding dataset by simultaneously performing random and uniform sampling of multiple parameters in a large parameter space, training a fully connected neural network model to predict the radiation dose correction factor. Next, for any given shielding scheme, a one-dimensional discrete ordinate method is used to perform one-dimensional transport equivalence calculations to quickly obtain the equivalent one-dimensional geometric radiation dose at the target point for radiation dose assessment. Finally, the trained fully connected neural network model is used to predict and output the corresponding radiation dose correction factor, and the actual total radiation dose at the target point for radiation dose assessment is calculated. This method can quickly predict the radiation dose in the target area for radiation dose assessment, and the results are similar to the radiation dose values ​​obtained from three-dimensional neutron-photon coupled transport calculations, achieving efficient and high-precision prediction of the radiation dose in the target area.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] A neural network-based radiation dose prediction method for reactor shielding optimization includes the following steps:

[0006] Step 1: Read the source term distribution information, geometric information, material information, and spatial coordinates of the target point for radiation dose assessment from the three-dimensional nuclear reactor shielding model to be simulated:

[0007] (a) Source term distribution information: spatial distribution parameters and energy distribution parameters of the source terms;

[0008] (b) Geometric information: Describes the geometric dimensions and spatial coordinates of the unshielded area and the components of the shielded area of ​​the nuclear reactor;

[0009] (c) Material information: the types, composition and spatial distribution of materials arranged in the unshielded area of ​​the nuclear reactor; a library of shielding materials that can be used in the shielded area of ​​the nuclear reactor, including the density and nuclide composition parameters of single materials and composite materials, wherein composite materials also include the composition range, types, nuclide composition and density parameters of the matrix material and reinforcing material that make up the composite material;

[0010] Step 2: Based on the relevant information obtained in Step 1 and the accessed shielding material library, several shielding schemes are generated in the nuclear reactor shielding area by first randomly and uniformly sampling the total number of shielding layers, the relative arrangement order and geometric dimensions of each shielding layer, and the types and components of the shielding materials used in each shielding layer. Then, a Monte Carlo program or a three-dimensional discrete ordinate method is used to perform three-dimensional neutron-photon coupling transport calculations on these shielding schemes to obtain the actual total radiation dose at the target point corresponding to each shielding scheme. Finally, a one-dimensional discrete ordinate method is used to perform one-dimensional transport equivalence calculations on these shielding schemes. The equivalent one-dimensional geometric flux and equivalent one-dimensional geometric radiation dose at the radiation dose assessment target point for each of these shielding schemes are obtained. The actual total radiation dose at the radiation dose assessment target point for each of these shielding schemes is divided by the equivalent one-dimensional geometric radiation dose at the same target point to obtain the radiation dose correction factor for each shielding scheme. Then, the total absorbed mean free path number and the total mean free path number for each energy group for each shielding scheme are calculated. Thus, a shielding dataset for neural network training, validation, and testing is established.

[0011] Step 3: Based on the shielding dataset obtained in Step 2, establish and train a fully connected neural network model with the total absorbed mean free path number of each energy group corresponding to each shielding scheme, the total mean free path number of each energy group, and the equivalent one-dimensional geometric flux at the radiation dose assessment target point as inputs, and the radiation dose correction factor at the radiation dose assessment target point as output.

[0012] Step 4: For any shielding scheme to be evaluated in the reactor shielding optimization design, based on the source term distribution information, geometric information and material information of the unshielded area of ​​the nuclear reactor obtained in Step 1, use the one-dimensional discrete ordinate method to perform one-dimensional transport equivalent calculation on the shielding scheme, and quickly obtain the equivalent one-dimensional geometric flux FLUX of the shielding scheme to be evaluated at the radiation dose assessment target point. evalEquivalent one-dimensional geometric radiation dose (Dose) of the shielding scheme to be evaluated 1D,eval The total mean free path number (NAMFP) of the energy groups of the shielding scheme to be evaluated was calculated. eval The total mean free path number (NTMFP) of each energy group of the shielding scheme to be evaluated. eval The radiation dose is then fed into the fully connected neural network model trained in step 3, and the predicted output yields the radiation dose correction factor CF at the target point for evaluating the shielding scheme. eval The predicted value of the actual total radiation dose at the target point of the radiation dose assessment for this shielding scheme to be evaluated is calculated according to equation (1). eval :

[0013] Dose eval =Dose 1D,eval ·CF eval (1).

[0014] The shielding area of ​​the three-dimensional nuclear reactor shielding model in step 1 is cylindrical shell in shape. In step 2, inside the nuclear reactor shielding area, there are several shielding layers extending outward from the reactor core. Each shielding layer is cylindrical shell in shape and adjacent shielding layers are in close contact with each other. The types and compositions of shielding materials used in each shielding layer are uniform.

[0015] In step 3, the input to the fully connected neural network model is specifically a 3N-length input. G The vector x, where N G The sum of the energy group numbers of neutrons and photons; this vector x is defined as:

[0016] x=(NAMFP,NTMFP,FLUX) (2)

[0017] NAMFP—The total mean free path number of the energy groups corresponding to the shielding scheme;

[0018] NTMFP—Total mean free path number of each energy group corresponding to the shielding scheme;

[0019] FLUX—Equivalent one-dimensional geometric flux at the radiation dose assessment target point corresponding to the shielding scheme;

[0020] The specific formal definitions of NAMFP, NTMFP, and FLUX are as follows:

[0021]

[0022]

[0023]

[0024] NSL —The total number of shielding layers inside the nuclear reactor shielding area corresponding to the shielding scheme;

[0025] Σ ai,g —The shielding material of the i-th shielding layer inside the nuclear reactor shielding area corresponds to the macroscopic absorption cross section of the g-th energy group;

[0026] Σ ti,g —The shielding material of the i-th shielding layer inside the nuclear reactor shielding area, corresponding to the shielding scheme, is located at the macroscopic total cross section of the g-th energy group;

[0027] d i —The thickness of the i-th shielding layer inside the nuclear reactor shielding area corresponding to the shielding scheme;

[0028] φ g —The equivalent one-dimensional geometric flux at the radiation dose assessment target point corresponding to the shielding scheme is the flux of the g-th energy group.

[0029] In step 2, when performing one-dimensional transport equivalent calculations using the one-dimensional discrete ordinate method, one-dimensional cylindrical geometry is used for radial calculations, and one-dimensional planar geometry is used for axial calculations.

[0030] Compared with the prior art, the present invention has the following outstanding advantages:

[0031] 1. For any given shielding scheme, when predicting the actual total radiation dose at the target point of radiation dose assessment, this invention can quickly obtain the equivalent one-dimensional geometric radiation dose at the target point of radiation dose assessment and correct it based on one-dimensional transport equivalent calculation and a fully connected neural network model. Compared with traditional methods, it does not require three-dimensional neutron-photon coupling transport calculation, which can save a lot of computing resources and effectively reduce computing costs.

[0032] 2. This invention adopts a "two-stage" calculation and prediction mode, which first uses one-dimensional transport equivalent calculation and then uses a fully connected neural network model for radiation dose correction. This mode can achieve a calculation speed similar to that of one-dimensional transport equivalent calculation and obtain an accuracy similar to that of the actual total radiation dose obtained by three-dimensional neutron-photon coupled transport calculation. The whole process takes 1-2 seconds, realizing efficient and high-precision prediction of the actual total radiation dose at the target point of radiation dose assessment.

[0033] 3. This invention uses a fully connected neural network model to predict the output radiation dose correction factor, rather than directly predicting the actual total radiation dose at the radiation dose assessment target point. This avoids the negative impact of multi-scale features, where the radiation dose values ​​in the shielding dataset span multiple orders of magnitude, on the prediction accuracy of the neural network under a large parameter space. Furthermore, the neural network uses the total absorbed mean free path number of each energy group corresponding to the shielding scheme, the total mean free path number of each energy group, and the equivalent one-dimensional geometric flux at the radiation dose assessment target point as inputs. These inputs can effectively reflect the physical characteristics of a shielding scheme and have good universality for shielding schemes using different shielding materials and with different material arrangement orders. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the overall process of the method of the present invention.

[0035] Figure 2 This is a schematic diagram of the shielding area of ​​a nuclear reactor and its internal shielding layer, showing the axial and radial directions. Detailed Implementation

[0036] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0037] This invention discloses a neural network-based radiation dose prediction method for reactor shielding optimization, the overall process of which is as follows: Figure 1 As shown. To achieve rapid prediction and calculation of the actual total radiation dose at the target point for radiation dose assessment in a large parameter space, this invention adopts a two-stage calculation and prediction mode when evaluating any given shielding scheme. This mode first uses one-dimensional transport equivalent calculation and then uses a neural network for radiation dose correction. The specific implementation steps are as follows:

[0038] Step 1: Read the source term distribution information, geometric information, material information, and spatial coordinates of the target point for radiation dose assessment from the three-dimensional nuclear reactor shielding model to be simulated:

[0039] (a) Source term distribution information: spatial distribution parameters and energy distribution parameters of the source terms;

[0040] (b) Geometric information: Describes the geometric dimensions and spatial coordinates of the unshielded area and the components of the shielded area of ​​the nuclear reactor;

[0041] (c) Material information: the types, composition and spatial distribution of materials arranged in the unshielded area of ​​the nuclear reactor; a library of shielding materials that can be used in the shielded area of ​​the nuclear reactor, including the density and nuclide composition parameters of single materials and composite materials, wherein composite materials also include the composition range, types, nuclide composition and density parameters of the matrix material and reinforcing material that make up the composite material;

[0042] Step 2: Based on the relevant information obtained in Step 1 and the accessed shielding material library, several shielding schemes are generated in the nuclear reactor shielding area by first randomly and uniformly sampling the total number of shielding layers, the relative arrangement order and geometric dimensions of each shielding layer, and the types and components of the shielding materials used in each shielding layer. Then, a Monte Carlo program or a three-dimensional discrete ordinate method is used to perform three-dimensional neutron-photon coupling transport calculations on these shielding schemes to obtain the actual total radiation dose at the target point corresponding to each shielding scheme. Finally, a one-dimensional discrete ordinate method is used to perform one-dimensional transport equivalence calculations on these shielding schemes. The equivalent one-dimensional geometric flux and equivalent one-dimensional geometric radiation dose at the radiation dose assessment target point for each of these shielding schemes are obtained. The actual total radiation dose at the radiation dose assessment target point for each of these shielding schemes is divided by the equivalent one-dimensional geometric radiation dose at the same target point to obtain the radiation dose correction factor for each shielding scheme. Then, the total absorbed mean free path number and the total mean free path number for each energy group for each shielding scheme are calculated. Thus, a shielding dataset for neural network training, validation, and testing is established.

[0043] Step 3: Based on the shielding dataset obtained in Step 2, establish and train a fully connected neural network model with the total absorbed mean free path number of each energy group corresponding to each shielding scheme, the total mean free path number of each energy group, and the equivalent one-dimensional geometric flux at the radiation dose assessment target point as inputs, and the radiation dose correction factor at the radiation dose assessment target point as output.

[0044] Step 4: For any shielding scheme to be evaluated in the reactor shielding optimization design, based on the source term distribution information, geometric information, and material information of the unshielded area of ​​the nuclear reactor obtained in Step 1, similar to Step 2, the one-dimensional discrete ordinate method is used to perform one-dimensional transport equivalent calculation on the shielding scheme to quickly obtain the equivalent one-dimensional geometric flux FLUX of the shielding scheme to be evaluated at the radiation dose assessment target point. eval Equivalent one-dimensional geometric radiation dose (Dose) of the shielding scheme to be evaluated 1D,eval The total mean free path number (NAMFP) of the energy groups of the shielding scheme to be evaluated was calculated. eval The total mean free path number (NTMFP) of each energy group of the shielding scheme to be evaluated. eval The radiation dose is then fed into the fully connected neural network model trained in step 3, and the predicted output yields the radiation dose correction factor CF at the target point for evaluating the shielding scheme. eval The predicted value of the actual total radiation dose at the target point of the radiation dose assessment for this shielding scheme to be evaluated is calculated according to equation (1). eval :

[0045] Dose eval =Dose 1D,eval ·CF eval (1).

[0046] In practical engineering, since the core of a nuclear reactor is typically cylindrical, the shielding area of ​​the three-dimensional nuclear reactor shielding model in step 1 is a cylindrical shell, enclosing the reactor core, to shield radiation from the core as much as possible. To achieve better shielding against neutrons and photons from the core, a multi-layered shielding structure is usually required to adequately moderate and absorb high-energy particles. To reduce the difficulty of manufacturing the shielding layers, several shielding layers exist inside the nuclear reactor shielding area in step 2, extending outwards from the reactor core. Each shielding layer is cylindrical, and adjacent layers are tightly bonded. The type and composition of the shielding material used in each layer are uniform. Figure 2 As shown.

[0047] Therefore, based on the overall shape of the nuclear reactor shielding area, in step 2, in order to quickly obtain the equivalent one-dimensional geometric flux and equivalent one-dimensional geometric radiation dose at the radiation dose assessment target point, when performing one-dimensional transport equivalent calculations using the one-dimensional discrete ordinate method program, one-dimensional cylindrical geometry is used for calculations in the radial direction, and one-dimensional planar geometry is used for calculations in the axial direction.

[0048] To better reflect the physical characteristics of a shielding scheme and to ensure good universality for shielding schemes using different shielding materials and with different material arrangement orders, the input to the neural network in step 3 is designed as a 3N-length input. G The vector x, where N G It is the sum of the energy group numbers of neutrons and photons. And this vector x is defined as:

[0049] x=(NAMFP,NTMFP,FLUX) (2)

[0050] NAMFP—The total mean free path number of the energy groups corresponding to the shielding scheme;

[0051] NTMFP—Total mean free path number of each energy group corresponding to the shielding scheme;

[0052] FLUX—Equivalent one-dimensional geometric flux at the radiation dose assessment target point corresponding to the shielding scheme;

[0053] The specific formal definitions of NAMFP, NTMFP, and FLUX are as follows:

[0054]

[0055]

[0056]

[0057] N SL —The total number of shielding layers inside the nuclear reactor shielding area corresponding to the shielding scheme;

[0058] Σ ai,g —The shielding material of the i-th shielding layer inside the nuclear reactor shielding area corresponds to the macroscopic absorption cross section of the g-th energy group;

[0059] Σ ti,g —The shielding material of the i-th shielding layer inside the nuclear reactor shielding area, corresponding to the shielding scheme, is located at the macroscopic total cross section of the g-th energy group;

[0060] d i —The thickness of the i-th shielding layer inside the nuclear reactor shielding area corresponding to the shielding scheme;

[0061] φ g —The equivalent one-dimensional geometric flux at the radiation dose assessment target point corresponding to the shielding scheme is the flux of the g-th energy group.

[0062] In practical reactor shielding optimization design applied to large parameter spaces, taking a reactor shield with a rectangular grid size of 158×158×128 in the X×Y×Z direction as the object, obtaining the actual total radiation dose at the target point corresponding to any shielding scheme to be evaluated using the three-dimensional neutron-photon coupled transport calculation method requires approximately 5.5 minutes to complete on 512 CPU cores, i.e., consuming 46.93 CPU cores. However, using the prediction method of this application, completing the first-stage one-dimensional transport equivalent calculation on a single CPU core takes only 1 second, and completing the second-stage radiation dose correction based on a fully connected neural network model on a single GPU takes only 3.33 × 10⁻⁶ seconds. -6 The prediction method described in this application can obtain the actual total radiation dose at the target point of any shielding scheme to be evaluated in just 1-2 seconds on a single CPU core and GPU, significantly reducing the computational cost.

[0063] Regarding method accuracy, the actual total radiation dose at the target point for radiation dose assessment corresponding to each shielding scheme under a large parameter space is obtained based on a three-dimensional neutron-photon coupled transport calculation method, and this is used as the benchmark result. In reactor shielding optimization design under a large parameter space, the currently widely used method based on a fully connected neural network model to directly predict the actual total radiation dose at the target point for radiation dose assessment has an average absolute percentage error between the prediction results and the benchmark results of 20%-40%. In contrast, the prediction results using the method of this application have an average absolute percentage error of less than 10%, which means that the prediction method of this application ensures both high efficiency and high accuracy in radiation dose prediction.

[0064] Finally, it should be noted that although specific embodiments of the present invention have been described and illustrated, those skilled in the art will understand that various modifications can be made to these embodiments without departing from the purpose and spirit of the present invention, such as changing the total number of transport computing energy groups, etc. The scope of the present invention is defined by the claims and their equivalents.

Claims

1. A neural network-based radiation dose prediction method for reactor shielding optimization, characterized in that: Includes the following steps: Step 1: Read the source term distribution information, geometric information, material information, and spatial coordinates of the target point for radiation dose assessment from the three-dimensional nuclear reactor shielding model to be simulated: (a) Source term distribution information: spatial distribution parameters and energy distribution parameters of the source terms; (b) Geometric information: Describes the geometric dimensions and spatial coordinates of the unshielded area and the components of the shielded area of ​​the nuclear reactor; (c) Material information: the types, composition and spatial distribution of materials arranged in the unshielded area of ​​the nuclear reactor; a library of shielding materials that can be used in the shielded area of ​​the nuclear reactor, including the density and nuclide composition parameters of single materials and composite materials, wherein composite materials also include the composition range, types, nuclide composition and density parameters of the matrix material and reinforcing material that make up the composite material; Step 2: Based on the relevant information read in Step 1 and the read shielding material library, in the nuclear reactor shielding area, firstly, several shielding schemes are generated by randomly and uniformly sampling the total number of shielding layers, the relative arrangement order and geometric dimensions of each shielding layer, and the types and components of shielding materials used in each shielding layer; secondly, three-dimensional neutron-photon coupling transport calculations are performed on these shielding schemes using a Monte Carlo program or a three-dimensional discrete ordinate method program to obtain the actual total radiation dose at the target point corresponding to each of these shielding schemes. Then, the one-dimensional discrete ordinate method is used to perform one-dimensional transport equivalent calculations on these shielding schemes to obtain the equivalent one-dimensional geometric flux and equivalent one-dimensional geometric radiation dose at the radiation dose assessment target point for each shielding scheme. The actual total radiation dose at the radiation dose assessment target point for each shielding scheme is divided by the equivalent one-dimensional geometric radiation dose at the radiation dose assessment target point for each shielding scheme to obtain the radiation dose correction factor at the radiation dose assessment target point for each shielding scheme. Next, the total absorbed mean free path number and the total mean free path number for each energy group for each shielding scheme are calculated. Thus, a shielding dataset for neural network training, verification, and testing is established. Step 3: Based on the shielding dataset obtained in Step 2, establish and train a fully connected neural network model with the total absorbed mean free path number of each energy group corresponding to each shielding scheme, the total mean free path number of each energy group, and the equivalent one-dimensional geometric flux at the radiation dose assessment target point as inputs, and the radiation dose correction factor at the radiation dose assessment target point as output. Step 4: For any shielding scheme to be evaluated in the reactor shielding optimization design, based on the source term distribution information, geometric information and material information of the unshielded area of ​​the nuclear reactor obtained in Step 1, use the one-dimensional discrete ordinate method to perform one-dimensional transport equivalent calculation on the shielding scheme, and quickly obtain the equivalent one-dimensional geometric flux FLUX of the shielding scheme to be evaluated at the radiation dose assessment target point. eval Equivalent one-dimensional geometric radiation dose (Dose) of the shielding scheme to be evaluated 1D,eval The total mean free path number (NAMFP) of the energy groups of the shielding scheme to be evaluated was calculated. eval The total mean free path number (NTMFP) of each energy group of the shielding scheme to be evaluated. eval ; The input is fed into the fully connected neural network model trained in step 3, and the predicted output yields the radiation dose correction factor CF at the target point for radiation dose assessment of the shielding scheme to be evaluated. eval The predicted value of the actual total radiation dose at the target point of the radiation dose assessment for this shielding scheme to be evaluated is calculated according to equation (1). eval : Dose eval *Dose 1D,eval ·CF eval (1) 2. The neural network-based radiation dose prediction method for reactor shielding optimization according to claim 1, characterized in that: The shielding area of ​​the three-dimensional nuclear reactor shielding model in step 1 is cylindrical shell in shape. In step 2, inside the nuclear reactor shielding area, there are several shielding layers extending outward from the reactor core. Each shielding layer is cylindrical shell in shape and adjacent shielding layers are in close contact with each other. The types and compositions of shielding materials used in each shielding layer are uniform.

3. The neural network-based radiation dose prediction method for reactor shielding optimization according to claim 1, characterized in that: In step 3, the input to the fully connected neural network model is specifically a 3N-length input. G The vector x, where N G It is the sum of the energy groups of neutrons and photons; This vector x is defined as: x=(NAMFP,NTMFP,FLUX) (2) NAMFP—The total mean free path number of the energy groups corresponding to the shielding scheme; NTMFP—Total mean free path number of each energy group corresponding to the shielding scheme; FLUX—Equivalent one-dimensional geometric flux at the radiation dose assessment target point corresponding to the shielding scheme; The specific formal definitions of NAMFP, NTMFP, and FLUX are as follows: N SL —The total number of shielding layers inside the nuclear reactor shielding area corresponding to the shielding scheme; Σ ai,g —The shielding material of the i-th shielding layer inside the nuclear reactor shielding area corresponds to the macroscopic absorption cross section of the g-th energy group; Σ ti,g —The shielding material of the i-th shielding layer inside the nuclear reactor shielding area, corresponding to the shielding scheme, is located at the macroscopic total cross section of the g-th energy group; d i —The thickness of the i-th shielding layer within the nuclear reactor shielding area corresponding to the shielding scheme; φ g —The equivalent one-dimensional geometric flux at the radiation dose assessment target point corresponding to the shielding scheme is the flux of the g-th energy group.

4. The neural network-based radiation dose prediction method for reactor shielding optimization according to claim 1, characterized in that: In step 2, when performing one-dimensional transport equivalent calculations using the one-dimensional discrete ordinate method, one-dimensional cylindrical geometry is used for radial calculations, and one-dimensional planar geometry is used for axial calculations.

Images