Intelligent design method of radiation shielding material based on cross-scale error feedback lock step

Abstract

The invention relates to an intelligent design method of a radiation shielding material based on cross-scale error feedback lock step, which comprises the following steps: determining a material component set and acquiring an atomic-scale graph structure of the material component set, and extracting intrinsic performance parameters such as heat conductivity and attenuation coefficient through an atomic-scale prediction module; and then, constructing a design variable vector in combination with a multi-physical field condition, calculating equivalent performance parameters of the composite material, establishing a cross-scale lock step mechanism, and triggering a molecular dynamics and Monte Carlo recalculation module to carry out correction when the deviation between a predicted value and a mesoscopic recalculation result exceeds a limit. Based on the correction parameters, a candidate material scheme is generated by using a multi-objective optimization algorithm; and performing gamma-ray and neutron shielding performance simulation on the candidate schemes through a Monte Carlo radiation transfer model accelerated by a deep neural network, and comparing with an experimental result to realize self-adaptive material design driven by cross-scale errors. According to the method, the design precision and efficiency of the radiation shielding material are remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radiation shielding material design, and particularly relates to an intelligent design method for a radiation shielding material based on cross-scale error feedback lockstep. BACKGROUND

[0002] Advanced nuclear energy materials are long-term served under high temperature, strong irradiation and complex chemical environment, and the performance thereof is directly related to the safety and reliability of nuclear energy systems. In order to meet the increasingly strict performance requirements of nuclear energy devices, researchers generally adopt two main means: one is atomic scale calculation simulation based on quantum mechanics, molecular dynamics or Monte Carlo method, which is used to predict the microstructure evolution and property change of materials under specific service conditions; the other is to measure the mechanical, thermal and irradiation performance parameters of materials on the macroscopic scale through experimental means. However, the traditional atomic scale calculation simulation has a large amount of calculation, and it is difficult to realize large-scale and rapid prediction in a multi-dimensional design space; and the experimental test has a long period and high cost, and it is difficult to obtain experimental data under extreme working conditions, resulting in a long material development period.

[0003] In recent years, artificial intelligence and machine learning technology has been introduced into the field of material science, which can improve the performance prediction efficiency to a certain extent by establishing the mapping relationship between structure and performance. However, such methods usually rely on a single data source or single scale sample, lack of cross-scale data fusion and effective physical constraints, resulting in insufficient accuracy and generalization ability of the prediction results under complex multi-physical field conditions.

[0004] In the prior art, Chinese patent CN118095084A discloses a multi-objective optimization design and preparation method of a neutron shielding composite material, comprising the following steps: step 1, establishing a target function, a fitness function and a constraint; step 2, setting up an input file of a radiation shielding calculation program; step 3, calling the radiation shielding calculation program in the optimization process; step 4, calculating the target function value of the population by using the obtained neutron transmission rate and other results, and judging whether the constraint condition is met, if not, calculating the penalty function value; step 5, generating a child population through selection, mutation and crossover operation of the population; step 6, completing the set iteration number to obtain the design scheme of the shielding material; step 7, designing the proportion of the powder according to the selected scheme and mixing uniformly; step 8, and observing the microstructure and testing the mechanical properties of the prepared neutron shielding composite material.

[0005] However, the method has the following limitations: 1. The method mainly relies on Monte Carlo calculation and genetic algorithm iterative optimization, and the calculation and experiment are only connected in the final verification stage, and there is no dynamic interaction feedback mechanism between calculation and experimental data, which leads to the drift of the multi-scale model in the iteration process, and it is difficult to guarantee the convergence and stability of the prediction result. 2. In the existing multi-scale optimization design, when there is a deviation between the calculation result and the experimental result, there is no effective real-time correction strategy, and when the error exceeds the acceptable range, the low-scale fine calculation or reiteration cannot be automatically triggered, which makes it difficult to guarantee the prediction accuracy of the model. 3. Although this technology combines calculation optimization and experimental verification, it does not establish a multi-source data through mechanism from microstructure characterization to molecular / Monte Carlo simulation to macroscopic performance measurement, and cannot realize online model correction based on experimental feedback, and the design-verification chain is broken.

[0006] In summary, the existing technology still has problems such as cross-scale model drift, insufficient prediction accuracy, and insufficient integration of experimental data into design iteration in multi-scale material design and radiation shielding composite material optimization, and a new design method is needed to realize dynamic coupling between multi-scale models, build an efficient closed loop between experiments-simulation-prediction, and reduce prediction deviation, in order to improve the accuracy and efficiency of material design. SUMMARY

[0007] The purpose of the present application is to overcome the defects of the prior art and provide an intelligent design method for radiation shielding materials based on cross-scale error feedback lockstep.

[0008] The purpose of the present application can be achieved by the following technical solutions:

[0009] The present application provides an intelligent design method for radiation shielding materials based on cross-scale error feedback lockstep, comprising the following steps:

[0010] A multi-agent collaborative optimization framework for cross-scale error feedback lockstep is constructed, which includes an atomic level prediction module, a mesoscopic level recalculation module, an optimization agent module and a verification agent module;

[0011] A set of components for constituting a radiation shielding material is determined, the atomic level graph structure of each component is obtained, and the intrinsic performance parameters are extracted by the atomic level prediction module;

[0012] Multi-physical field condition data is obtained, and a design variable vector is constructed based on the intrinsic performance parameters and the multi-physical field conditions. The equivalent performance parameters of the composite material are calculated based on the design variables, and a cross-scale lockstep condition is established. If the deviation of the equivalent performance parameters and the calculation results of the mesoscopic level recalculation module exceeds the limit, the mesoscopic level recalculation module is triggered to recalculate until the set error threshold is met;

[0013] Based on the corrected equivalent performance parameters, the design variables are optimized by the optimization agent module using a multi-objective optimization algorithm to generate a candidate material scheme;

[0014] The verification agent module constructs a solid geometry and component model based on the candidate material scheme, performs radiation shielding performance simulation using a deep neural network accelerated Monte Carlo radiation transport model, compares the simulation results with the experimental test results, and outputs the radiation shielding material design scheme when the deviation meets the set threshold. If the deviation is out of limit, the experimental feedback results are used to correct the parameters of the mesoscale re-calculation module, and the optimization step is continued to iterate.

[0015] Further, the components include metal-based components, non-metal-based components, and composite additive components for constituting the radiation shielding material. The metal-based components include high atomic number metal elements or their alloys, and the non-metal-based components include light elements or high hydrogen content compounds.

[0016] The intrinsic performance parameters include gamma-ray attenuation coefficient, neutron absorption cross section, thermal conductivity, electrical conductivity, linear thermal expansion coefficient, and elastic modulus.

[0017] Further, the atomic-level prediction module is constructed based on a graph neural network and trained by the following steps:

[0018] The atomic-level graph structure of a plurality of different components is used as a training sample. The nodes in the atomic-level graph structure are atoms in the components, the node attributes include element type, valence electron number and isotope information, the edges are chemical bonds in the components, and the edge attributes include bond length, bond order and bond energy.

[0019] The intrinsic performance parameters corresponding to the plurality of different components are used as the supervision labels of each training sample.

[0020] The training sample is input into the multi-layer graph convolutional neural network of the atomic-level prediction module for feature extraction. The training sample is input into the multi-layer graph convolutional neural network of the atomic-level prediction module for feature extraction. The multi-layer graph convolutional neural network includes a plurality of graph convolutional layers, a global average pooling layer and a fully connected output layer in sequence. The output layer outputs the predicted value of the intrinsic performance parameter. According to the predicted value of the intrinsic performance parameter and the supervision label, a loss function is constructed based on the mean absolute error and the physical constraint regularization term, and the network parameters of the multi-layer graph convolutional neural network are updated by the back propagation algorithm. The atomic-level prediction module reflecting the mapping relationship between the material components and the microstructure and the intrinsic performance parameters is established.

[0021] The loss function is:

[0022]

[0023] wherein, L is a loss function, N is the total number of training samples, K is the number of intrinsic performance parameters, y i,j represents the jth intrinsic performance parameter in the supervised label of the ith training sample; is the jth intrinsic performance parameter prediction value of the ith training sample output by the material performance prediction model; λ1, λ2, λ3, λ4, λ5 are preset weight coefficients; L nonneg is a non-negative constraint penalty term; L WF is a heat conduction-electricity conduction constraint penalty term based on the Wiedemann-Franz law; L SA is a cross-section conservation constraint penalty term; L therm is a thermodynamic stability constraint penalty term, L phy is a physical residual term;

[0024] The non-negative constraint penalty term has a formula as follows:

[0025]

[0026] wherein, is a non-negative intrinsic performance parameter set, including thermal conductivity gamma-ray attenuation coefficient neutron absorption cross section ReLU is a rectified linear unit function, which outputs the absolute value when the input is negative, and outputs zero when the input is non-negative;

[0027] The heat conduction-electricity conduction constraint penalty term based on the Wiedemann-Franz law has a formula as follows:

[0028]

[0029] wherein, is the predicted thermal conductivity of the ith training sample, is the predicted electrical conductivity of the ith sample, L0 is the Lorentz constant, T i is the temperature of the ith sample;

[0030] The cross-section conservation constraint penalty term has a formula as follows:

[0031]

[0032] wherein, is the predicted gamma-ray attenuation coefficient of the ith sample, is the predicted neutron absorption cross section of the ith sample, C total is a preset total cross-section conservation constant;

[0033] The thermodynamic stability constraint penalty term has a formula as follows:

[0034]

[0035] wherein, are the predicted linear thermal expansion coefficient and the predicted elastic modulus of the i-th sample, respectively, and max , E max are preset threshold values of the linear thermal expansion coefficient and the elastic modulus, respectively;

[0036] The physical residual term L phy is given by:

[0037] L phy = w γ L γ + w n L n

[0038] L γ =∑ i |E in,i –(E abs,i +E scat,i +E trans,i )|2

[0039] L n =∑ i |Φ in,i –(Φ abs,i +Φ scat,i +Φ leak,i )|2

[0040] wherein, w γ , w n are the weight coefficients of the gamma energy conservation residual and the neutron particle number conservation residual, respectively, L γ , L n are the gamma energy conservation residual and the neutron particle number conservation residual, respectively, ∑ i denotes the summation over all training samples i, and denotes the cumulative residual over the entire dataset, E in,i , E abs,i , E scat,i , E trans,i denote the total incident gamma energy, the absorbed gamma energy, the scattered gamma energy, and the transmitted gamma energy in the i-th training sample, respectively, and Φ in,i , Φ abs,i , Φ scat,i , Φ leak,i denote the incident neutron flux, the absorbed neutron flux, the scattered neutron flux, and the leaked neutron flux in the i-th training sample, respectively.

[0041] Further, the multi-physical field condition data comprises the encoded neutron energy spectrum Φ n (E), gamma spectrum Φγ (E), temperature gradient Mechanical stress σ mech .

[0042] Further, the construction design variable vector:

[0043] x = {V i , t j , g j , S k , M k}

[0044] wherein x is a material design variable vector, each material design variable vector x represents a candidate material scheme, V i is the mass percentage of the i-th component, t j represents the layer thickness of the j-th layer in the material, g j is the interlayer gradient parameter of the j-th layer in the material, S k is the bonding strength of the k-th interface in the material, M k is the thermal expansion coefficient matching degree of the k-th interface in the material;

[0045] The equivalent performance parameters include:

[0046] Effective thermal conductivity λ eff (x), the formula is:

[0047]

[0048] wherein w i is the mass fraction of the i-th component, λ i is the thermal conductivity of the i-th component, ζ j is a continuous correction factor calibrated by experiment, 0.8 ≤ ζ j ≤ 5.0, is the interface thermal resistance calculated by molecular dynamics, d j is the thickness of the j-th layer of material;

[0049] Macroscopic neutron total cross section ∑ t (x), the formula is:

[0050]

[0051] wherein, is the neutron self-shielding factor calculated by the Monte Carlo method, σ n,i is the neutron macroscopic cross section of the i-th component;

[0052] Macroscopic γ-ray attenuation coefficient μ t,eff (x), the formula is:

[0053] μ t,eff(x) = ∑ i w i μ γ,i

[0054] wherein μ γ,i is the gamma-ray linear attenuation coefficient of the i-th component.

[0055] Further, the training set of the continuously modified experimental calibration factor ζ j contains no less than 100 groups of three-element pairing data composed of scanning electron microscope (SEM) images, interface thermal resistance results calculated by molecular dynamics (MD), and corresponding measured thermal resistance values, and the value range of ζ j is determined by statistical analysis of the MD calculated values and measured values in the three-element data, and based on the upper and lower limits of the 95% confidence interval.

[0056] Further, the optimization agent module optimizes the design variables by using a multi-objective optimization algorithm to generate a candidate material scheme, specifically including:

[0057] Based on the equivalent performance parameters of the constructed composite material, a material optimization multi-objective function is defined:

[0058] min F(x) = [f1(x), f2(x), f3(x), f4(x), f5(x)]

[0059]

[0060] f2(x) = -[1-exp(-μ t,eff (x)·t total )]

[0061] f3(x) = -[1-exp(-∑ t (x)·t total )]

[0062] f44(x) = σ thermal (x)

[0063] σ thermal = |Δα| × ΔT × E eff

[0064] f5(x) = Φ(-β)

[0065] wherein f1(x) is the first optimization objective, representing the total density of the material, ρ i is the density of the i-th component; f2(x) is the second optimization objective, representing the value of the gamma-ray shielding efficiency after taking the negative sign, μ t,eff (x) represents the macroscopic gamma-ray attenuation coefficient corresponding to the design variable x, t totalf1(x) is the first optimization objective, representing the total thickness of the material; f3(x) is the third optimization objective, representing the neutron shielding efficiency, Σ t (x) is the neutron macroscopic total cross section of the material; f4(x) is the fourth optimization objective, representing the interlayer thermal stress, Δα is the difference between the linear thermal expansion coefficients of the adjacent two layers of material, ΔT is the temperature gradient, E eff is the equivalent elastic modulus of the composite material; f5(x) is the fifth optimization objective, representing the service reliability failure probability, Φ is the standard normal cumulative distribution function, and β is the reliability index obtained by linearization of the second moment;

[0066] The constraint conditions are set as:

[0067]

[0068] wherein, is the minimum / maximum thickness of the jth layer of the material, is the gradient parameter value range of the jth layer of the material, is the value range of the kth interface bonding strength of the material, is the value range of the kth interface thermal expansion coefficient matching degree of the material;

[0069] The material optimization multi-objective function is iteratively solved by using a non-dominated sorting genetic algorithm, and a candidate material scheme is generated based on the Pareto optimal solution set of the design variable vector x under the premise of meeting the constraint conditions.

[0070] Further, the mesoscale recalculation module comprises:

[0071] A molecular dynamics recalculation unit is configured to, when the cross-scale lockstep condition is triggered, obtain interface structure information of each layer from the atomic-level prediction module, and automatically execute molecular dynamics simulation to recalculate the interface thermal resistance The thermal resistance value is obtained according to the interatomic potential energy function and the heat flow statistical method under a constant temperature and pressure system;

[0072] A Monte Carlo recalculation unit is configured to, when the cross-scale lockstep condition is triggered, obtain neutron cross section information of each component from the atomic-level prediction module, and automatically execute Monte Carlo neutron transport simulation to recalculate the neutron self-shielding factor , which simulates the input of a multi-energy section neutron source, and statistically obtains the neutron absorption, scattering and leakage flux of each layer of the material;

[0073] A recalculation result covering and feedback unit is configured to cover and the corresponding predicted values output by the atomic-level prediction module, and preferentially use the recalculation results in subsequent equivalent performance parameter calculation and multi-objective optimization, to drive the cross-scale error feedback lockstep closed loop until the prediction error meets the set threshold;

[0074] The lockstep condition is:

[0075]

[0076] wherein, are the effective thermal conductivity and the macroscopic neutron total cross section of the composite material output by the atomistic prediction module respectively, is the effective thermal conductivity of the composite material calculated by the molecular dynamics reevaluation unit, is the macroscopic neutron total cross section of the composite material calculated by the Monte Carlo reevaluation unit.

[0077] Further, the mesoscopic reevaluation module is only started when the cross-scale lockstep condition is met, and the reevaluation structure covers the corresponding predicted values output by the atomistic prediction module, to serve as the priority input for subsequent composite material equivalent performance parameter calculation, forming an error-driven adaptive closed loop.

[0078] Further, the verification agent module establishes a solid geometry and component model based on the candidate material scheme, and performs radiation shielding performance simulation using a deep neural network accelerated Monte Carlo radiation transport model, specifically including:

[0079] Based on the design variable vector x of the candidate material scheme, a solid geometry and component model is established, and on the basis of the solid geometry and component model, a deep neural network accelerated Monte Carlo radiation transport model is used to simulate the transport process of γ-rays and neutrons in the material. The deep neural network is used to predict the energy deposition distribution and scattering path probability of particles in each layer of material;

[0080] According to the particle transmittance and dose distribution output by the simulation and simulation, the shielding performance index of the material is calculated:

[0081]

[0082] wherein, η γ , η n are the γ-ray and neutron shielding efficiency of the material, T γ , T n are the γ-ray and neutron transmittance obtained by simulation, D r is the dose decay rate, D in , D out are the incident surface dose and transmitted surface dose obtained by simulation;

[0083] The shielding efficiency and dose decay rate are compared with the preset target threshold value, if and then output the radiation shielding material scheme that meets the requirements, if not, return to the optimization agent module for continued iterative optimization;

[0084] The stereogeometry and component model adopts octree voxelization representation, voxel size ≤ 50 μm, and is encrypted to ≤ 10 μm at the interlayer interface.

[0085] Compared with the prior art, the present application has the following advantages:

[0086] (1) The present application adds a cross-scale error feedback lockstep condition in the multi-agent collaborative optimization framework. When the effective thermal conductivity of the composite material or the total cross section of the neutron output by the graph neural network (atomic level prediction module) deviates more than 10% from the calculation results of the mesoscopic level molecular dynamics re-calculation unit (MD) and the Monte Carlo re-calculation unit (MC), the graph neural network weight is automatically frozen and the MD / MC re-calculation is forcibly triggered until the error is less than 5%. This solves the problem of lack of dynamic control mechanism based on error threshold in the prior art, realizes real-time closed-loop correction of atomic level prediction and mesoscopic level high-precision simulation data, seamlessly connects the cross-scale model, ensures the continuous accuracy of equivalent performance parameter prediction, and significantly improves the model convergence speed.

[0087] (2) The present application establishes a three-way paired training set containing no less than 100 groups of SEM images-MD thermal resistance calculation results-actual measured thermal resistance in the experimental calibration module, and determines the continuous correction factor ζ j using the 95% confidence interval. The measured interface thermal resistance and the simulation results are used to correct the graph neural network prediction to form a complete data chain of SEM images-MD / MC re-calculation-experimental measurement-GNN online update, which solves the problem of simulation results and experimental data being disconnected in the prior art, and the model cannot dynamically absorb measured information. It realizes the full-process closed-loop fusion of multi-source data, greatly improves the generalization ability and precision of material performance prediction.

[0088] (3) The present application combines the above lockstep mechanism and three-way closed-loop calibration scheme with a multi-objective optimization algorithm, and solves the design variables (component mass fraction, layer thickness, interlayer gradient, interface strength and matching degree) in the optimization agent module using a multi-objective genetic algorithm. Comprehensive balance of the five indicators solves the problem of model drift and inability to consider multiple objectives when balancing micro-macro performance in traditional multi-scale design methods, realizes the collaborative optimization of thermal conductivity, radiation shielding efficiency, thermal stress and service reliability and other multi-dimensional performance under multi-physical field conditions, and significantly improves the precision and iteration efficiency of material design.

[0089] (4) The atomic level prediction module of the present application is constructed based on a graph neural network, and through the introduction of multi-layer graph convolution feature extraction of node and edge attributes, and the training of a composite loss function combining the mean absolute error and five physical constraint regularization terms (non-negativity constraint, Wiedemann-Franz heat conduction-conduction constraint, cross-section conservation constraint, thermodynamic stability constraint and physical residual constraint), the problem of the model relying only on a single supervised label, lacking physical consistency and multiple constraints in the prior art is solved, high-precision learning of the mapping relationship between the material component microstructure and intrinsic performance parameters is realized, the accuracy and physical reasonableness of the predicted thermal conductivity, attenuation coefficient and absorption cross section and other key parameters are significantly improved, and the generalization ability and stability of the model under complex multi-physical field conditions are enhanced.

[0090] (5) The verification agent module of the present application realizes fast and high-precision simulation of the gamma ray and neutron transport process by mapping the design variable vector of the candidate material scheme into an octree voxelized three-dimensional geometry and component model, and introducing a three-dimensional U-Net deep neural network accelerated Monte Carlo radiation transport simulation based on the fine model; the voxel size is controlled to be ≤50 μm, and the interface is encrypted to ≤10 μm, which not only ensures the accurate restoration of the micro-interface characteristics, but also effectively reduces the calculation and storage overhead, breaking the bottleneck of the traditional whole gridding method that resolution and efficiency are difficult to balance; the particle transmission rate and dose distribution output by the simulation are calculated online for shielding efficiency and dose attenuation rate, and are directly compared with the preset threshold in real time, so that the scheme can be output in real time when the performance requirements are met, and the optimization module is returned to continue iteration when the requirements are not met, thereby forming a closed loop process of simulation-judgment-optimization, greatly improving the speed and reliability of material design verification. BRIEF DESCRIPTION OF DRAWINGS

[0091] Figure 1 The radiation shielding material intelligent design method flowchart of the embodiment of the present application is shown in the figure.

[0092] Figure 2 The equivalent performance parameter calculation flowchart of the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0093] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the scope of protection of the present application.

[0094] Embodiment 1:

[0095] The embodiment provides a radiation shielding material intelligent design method based on cross-scale error feedback lockstep, as shown in the formula (I), comprising the following steps: Figure 1

[0096] Step S1: constructing a multi-agent collaborative optimization framework of cross-scale error feedback lockstep, the multi-agent collaborative optimization framework comprising an atomic level prediction module, a mesoscopic level recalculation module, an optimization agent module and a verification agent module;

[0097] Step S2: determining a set of components for constituting the radiation shielding material, obtaining the atomic level graph structure of each component, and extracting intrinsic performance parameters through the atomic level prediction module;

[0098] Wherein, the components include metal-based components, non-metal-based components and composite additive components for constituting the radiation shielding material, the metal-based components include high atomic number metal elements or alloys thereof, and the non-metal-based components include light elements or high hydrogen content compounds; the intrinsic performance parameters include gamma ray attenuation coefficient, neutron absorption cross section, thermal conductivity, electrical conductivity, linear thermal expansion coefficient and elastic modulus.

[0099] In this step, the atomic level prediction module is constructed based on a graph neural network, and is trained through the following steps:

[0100] The atomic level graph structure of a plurality of different components is taken as a training sample, the nodes in the atomic level graph structure are atoms in the components, the node attributes include element type, valence electron number and isotope information, the edges are chemical bonds in the components, and the edge attributes include bond length, bond order and bond energy;

[0101] The intrinsic performance parameters corresponding to the plurality of different components are taken as the supervision labels of each training sample;

[0102] The training sample is input into the multi-layer graph convolutional neural network of the atomic level prediction module for feature extraction; the training sample is input into the multi-layer graph convolutional neural network of the atomic level prediction module for feature extraction, the multi-layer graph convolutional neural network comprises a plurality of graph convolutional layers, a global average pooling layer and a fully connected output layer in sequence, the output layer outputs a predicted value of the intrinsic performance parameter, a loss function is constructed based on the mean absolute error and the physical constraint regularization term according to the predicted value of the intrinsic performance parameter and the supervision label, and the network parameters of the multi-layer graph convolutional neural network are iteratively updated through a back propagation algorithm, so as to establish the atomic level prediction module reflecting the mapping relationship between the material components and the microstructure and the intrinsic performance parameters; wherein, the loss function of the atomic level prediction module is:

[0103]

[0104] Wherein, L is the loss function, N is the total number of training samples, K is the number of intrinsic performance parameters, y i,j ​the jth intrinsic property parameter in the supervision label of the ith training sample; the jth intrinsic property parameter prediction value of the ith training sample output by the material property prediction model; λ1, λ2, λ3, λ4, λ5 are preset weight coefficients; L nonneg is a non-negative constraint penalty term; L WF is a heat conduction-electricity conduction constraint penalty term based on the Wiedemann-Franz law; L SA is a cross-section conservation constraint penalty term; L therm is a thermodynamic stability constraint penalty term, L phy is a physical residual term;

[0105] The non-negative constraint penalty term has a formula as follows:

[0106]

[0107] wherein, The non-negative intrinsic property parameter set includes thermal conductivity Gamma-ray attenuation coefficient Neutron absorption cross section ReLU is a rectified linear unit function, which outputs the absolute value when the input is negative, and outputs zero when the input is non-negative;

[0108] The heat conduction-electricity conduction constraint penalty term based on the Wiedemann-Franz law has a formula as follows:

[0109]

[0110] wherein, is the predicted thermal conductivity of the ith training sample, is the predicted electrical conductivity of the ith sample, L0 is the Lorentz constant, T i is the temperature of the ith sample;

[0111] The cross-section conservation constraint penalty term has a formula as follows:

[0112]

[0113] wherein, is the predicted gamma-ray attenuation coefficient of the ith sample, is the predicted neutron absorption cross section of the ith sample, C total is a preset total cross-section conservation constant;

[0114] The thermodynamic stability constraint penalty term has a formula as follows:

[0115]

[0116] wherein, respectively the predicted linear thermal expansion coefficient and the predicted elastic modulus of the i-th sample, α max , E max respectively the preset linear thermal expansion coefficient threshold and the elastic modulus threshold;

[0117] the physical residual term L phy , the formula is:

[0118] L phy = w γ L γ + w n L n

[0119] L γ = Σ i |E in,i -(E abs,i +E scat,i +E trans,i )| 2

[0120] L n = Σ i |Φ in,i- (Φ abs,i +Φ scat,i +Φ leak,i )| 2

[0121] wherein w γ , w n are the weight coefficients of the gamma-ray energy conservation residual and the neutron particle number conservation residual respectively, L γ , L n are the gamma-ray energy conservation residual and the neutron particle number conservation residual respectively, ∑ i denotes the accumulation of all training samples i, denotes the cumulative residual of the entire data set, E in,i , E abs,i , E scat,i , E trans,i denote the total energy of the incident gamma rays, the absorbed gamma ray energy, the scattered gamma ray energy and the transmitted gamma ray energy in the i-th training sample, Φ in,i , Φ abs,i , Φ scat,i , Φ leak,i denote the incident neutron flux, the absorbed neutron flux, the scattered neutron flux and the leaked neutron flux in the i-th training sample.

[0122] Step S3: Obtain multi-physical field condition data, construct a design variable vector based on intrinsic performance parameters and multi-physical field conditions, calculate the equivalent performance parameters of the composite material based on the design variables, and establish a cross-scale lock-in condition. If the equivalent performance parameters deviate from the calculation results of the mesoscopic level recalculation module beyond the limit, trigger the mesoscopic level recalculation module to recalculate until the set error threshold is met;

[0123] Wherein, the multi-physical field condition data includes the coded neutron energy spectrum Φ n (E), the gamma spectrum Φ Υ (E), the temperature gradient The mechanical stress σ mech .

[0124] The design variable vector is:

[0125] x = {V i , t j , g j , S k , M k}

[0126] Wherein, x is the material design variable vector, each material design variable vector x represents a candidate material scheme, V i is the mass percentage of the i-th component, t j represents the layer thickness of the j-th layer in the material, g j is the interlayer gradient parameter of the j-th layer in the material, S k is the bonding strength of the k-th interface in the material, and M k is the thermal expansion coefficient matching degree of the k-th interface in the material;

[0127] The equivalent performance parameters include: effective thermal conductivity λ eff (x), macroscopic neutron total cross section ∑ t (x) and macroscopic gamma ray attenuation coefficient μ t,eff (x);

[0128] The effective thermal conductivity λ eff (x) is calculated by the formula:

[0129]

[0130] Wherein, w i is the mass fraction of the i-th component, λ i is the thermal conductivity of the i-th component, ζ j is the continuous correction factor calibrated by experiment, 0.8 ≤ ζ j ≤ 5.0, is the interface thermal resistance calculated by molecular dynamics, d j is the thickness of the j-th layer of material;

[0131] Experimentally calibrated continuous correction factor ζ j The training set contains no less than 100 groups of ternary paired data composed of scanning electron microscope (SEM) images, interface thermal resistance results calculated by molecular dynamics (MD), and corresponding measured thermal resistance values, ζ j The value range of ζ is determined by statistical analysis of MD calculated values and measured values in the ternary data, and based on the upper and lower limits of the 95% confidence interval.

[0132] Macroscopic neutron total cross section ∑ t (x), the formula is:

[0133]

[0134] Wherein, is the neutron self-shielding factor calculated by the Monte Carlo method, σ n,i is the neutron macroscopic cross section of the i-th component;

[0135] Macroscopic gamma-ray attenuation coefficient μ t,eff (x), the formula is:

[0136] μ t,eff (x) = ∑ i w i μ γ,i

[0137] Wherein, μ γ,i is the gamma-ray linear attenuation coefficient of the i-th component.

[0138] The mesoscale recalculation module comprises:

[0139] A molecular dynamics recalculation unit, configured to obtain interface structure information of each layer from the atomic level prediction module when the cross-scale lockstep condition is triggered, and automatically execute molecular dynamics simulation to recalculate the interface thermal resistance Simulate the thermal resistance value according to the interatomic potential energy function and the heat flow statistical method in the constant temperature and pressure system;

[0140] A Monte Carlo recalculation unit, configured to obtain neutron cross section information of each component from the atomic level prediction module when the cross-scale lockstep condition is triggered, and automatically execute Monte Carlo neutron transport simulation to recalculate the neutron self-shielding factor , simulate the neutron absorption, scattering and leakage flux of each layer of the material with a multi-energy section neutron source as input;

[0141] A recalculation result covering and feedback unit, configured to cover and the corresponding predicted values output by the atomic level prediction module, and preferentially use the recalculation results in subsequent equivalent performance parameter calculation and multi-objective optimization to drive the cross-scale error feedback lockstep closed loop until the prediction error meets the set threshold.

[0142] The lockstep condition is:

[0143]

[0144] wherein, are the effective thermal conductivity and the macroscopic neutron total cross section of the composite material output by the atomic-level prediction module, is the effective thermal conductivity of the composite material calculated by the molecular dynamics re-calculation unit, is the macroscopic neutron total cross section of the composite material calculated by the Monte Carlo re-calculation unit.

[0145] The meso-scale re-calculation module is only started when the cross-scale lockstep condition is met, and the re-calculation structure covers the corresponding predicted values output by the atomic-level prediction module, serving as the preferred input for subsequent calculation of the equivalent performance parameters of the composite material, forming an error-driven adaptive closed loop.

[0146] As shown in Figure 2 , the specific process of step S3 is as follows:

[0147] Step S301: Obtain multi-physical field condition data, including neutron energy spectrum, gamma ray energy spectrum, temperature gradient and mechanical stress in the material, and encode and store them into a field condition tensor according to a predetermined format;

[0148] Step S302: According to the intrinsic performance parameters of each component output by the atomic-level prediction module and the multi-physical field conditions, a design variable vector is constructed, which is composed of mass percentage of each component, layered thickness, interlayer gradient parameter, interfacial bonding strength of each layer and interfacial thermal expansion matching degree, to completely characterize the microstructure and macrostructure characteristics of the candidate material scheme;

[0149] Step S303: Using the design variables, the equivalent thermal conductivity, the equivalent neutron total cross section and the equivalent gamma ray attenuation ability of the composite material are calculated respectively, and the obtained equivalent performance parameters are the material macroscopic performance indicators input into the cross-scale optimization algorithm;

[0150] Step S304: According to the preset lockstep threshold, the above equivalent performance parameters are compared with the corresponding performance derived from the high-precision interfacial thermal resistance and neutron self-shielding factor regenerated by the meso-scale re-calculation module through molecular dynamics and Monte Carlo simulation, to determine whether the error limit is exceeded;

[0151] Step S305: When the deviation of the judgment index exceeds the threshold, the meso-scale re-calculation module is automatically triggered: the molecular dynamics re-calculation unit re-calculates the interfacial thermal resistance, the Monte Carlo re-calculation unit re-generates the neutron self-shielding factor, and the two cover the original predicted values;

[0152] Step S306: feedback the recalculation result to the equivalent performance calculation process, and recalculate the equivalent thermal conductivity, neutron cross section and γ-ray attenuation ability;

[0153] Step S307: repeat steps S304-S306 until the deviation of the equivalent performance parameters after recalculation and the atomic-level prediction result is controlled within the preset error threshold, and the cross-scale lockstep closed loop is completed;

[0154] Step S308: after the closed loop correction is completed, output the final equivalent performance parameters to the optimization agent module, and start the subsequent multi-objective optimization iteration.

[0155] Step S4: based on the corrected equivalent performance parameters, the optimization agent module is used to optimize the design variables by using a multi-objective optimization algorithm to generate a candidate material scheme;

[0156] The optimization agent module is used to optimize the design variables by using a multi-objective optimization algorithm to generate a candidate material scheme, which specifically includes:

[0157] Based on the equivalent performance parameters of the composite material, a material optimization multi-objective function is defined:

[0158] minF(x)=[f1(x),f2(x),f3(x),f4(x),f5(x)]

[0159]

[0160] f2(x)=-[1-exp(-μ t,eff (x)·t total )]

[0161] f3(x)=-[1-exp(-Σ t (x)·t total )]

[0162] f4(x)=σ thermal (x)

[0163] σ thermal =|Δα|×ΔT×E eff

[0164] f5(x)=Φ(-β)

[0165] Wherein, f1(x) is the first optimization target, representing the total density of the material, ρ i is the density of the i-th component; f2(x) is the second optimization target, representing the value of the γ-ray shielding efficiency after taking the negative sign, μ t,eff (x) represents the macroscopic γ-ray attenuation coefficient corresponding to the design variable x, t totalf1(x) is the first optimization objective, representing the total thickness of the material; f3(x) is the third optimization objective, representing the neutron shielding efficiency, Σ t (x) is the neutron macroscopic total cross section of the material; f4(x) is the fourth optimization objective, representing the interlayer thermal stress, Δα is the difference between the linear thermal expansion coefficients of the adjacent two layers of material, ΔT is the temperature gradient, E eff is the equivalent elastic modulus of the composite material; f5(x) is the fifth optimization objective, representing the service reliability failure probability, Φ is the standard normal cumulative distribution function, and β is the reliability index obtained by linearization of the second moment;

[0166] The constraint conditions are set as:

[0167]

[0168] wherein, is the minimum / maximum thickness of the jth layer of the material, is the gradient parameter value range of the jth layer of the material, is the value range of the kth interface bonding strength of the material, is the value range of the kth interface thermal expansion coefficient matching degree of the material;

[0169] The non-dominated sorting genetic algorithm is used to iteratively solve the material optimization multi-objective function, and the candidate material scheme is generated based on the Pareto optimal solution set of the design variable vector x under the premise of meeting the constraint conditions.

[0170] Step S5: The verification agent module establishes a solid geometry and component model based on the candidate material scheme, uses a deep neural network accelerated Monte Carlo radiation transport model to simulate the radiation shielding performance, compares the simulation results with the experimental test results, and outputs the radiation shielding material design scheme when the deviation meets the set threshold; if the deviation is out of limit, the experimental feedback results are used to correct the parameters of the mesoscopic level re-calculation module, and the optimization step is returned to continue iteration.

[0171] The verification agent module establishes a solid geometry and component model based on the candidate material scheme, uses a deep neural network accelerated Monte Carlo radiation transport model to simulate the radiation shielding performance, which specifically includes:

[0172] Based on the design variable vector x of the candidate material scheme, a solid geometry and component model is established, and based on the solid geometry and component model, a deep neural network accelerated Monte Carlo radiation transport model is used to simulate the transport process of γ-rays and neutrons in the material. The deep neural network is used to predict the energy deposition distribution and scattering path probability of particles in each layer of material.

[0173] According to the particle transmittance and dose distribution output by the simulation and simulation, the shielding performance index of the material is calculated:

[0174]

[0175] wherein η γ , η n are the γ-ray and neutron shielding efficiency of the material, T γ , T n are the simulated γ-ray and neutron transmissivity, D r is the dose attenuation rate, D in , D out are the simulated incident and transmitted surface dose;

[0176] The shielding efficiency and the dose attenuation rate are compared with a preset target threshold, if and a radiation shielding material scheme meeting the requirements is output, if not, the optimization agent module continues to iterate and optimize;

[0177] The stereoscopic geometry and component model is represented by octree voxelization, the voxel size is ≤50μm, and the voxel size is encrypted to ≤10μm at the interlayer interface.

[0178] Embodiment 2:

[0179] The embodiment realizes an intelligent design system for radiation shielding materials based on cross-scale error feedback lockstep, which comprises:

[0180] The processor cluster adopts a CPU-GPU heterogeneous computing architecture, wherein the GPU is dedicated to accelerating mesoscopic molecular dynamics (MD) and Monte Carlo (MC) re-calculation tasks to improve re-calculation efficiency and throughput;

[0181] The non-transient storage medium is used to persistently store the lockstep trigger script, the multi-physical field experiment database, and the graph neural network (GNN) model weight, realizing online iterative updating of the model and the data;

[0182] The multi-physical field data acquisition module is used to acquire and encode neutron energy spectrum, γ-ray energy spectrum, temperature gradient, and mechanical stress information, and generate a standardized field condition tensor;

[0183] The atomic level prediction module is based on graph neural network to extract and infer the atomic level graph structure of each component, and outputs intrinsic performance parameters such as thermal conductivity and attenuation coefficient;

[0184] The mesoscopic level re-calculation module includes a molecular dynamics re-calculation unit and a Monte Carlo re-calculation unit, which calls the GPU accelerated re-calculation interface thermal resistance and neutron self-screening factor when the cross-scale lockstep condition is triggered, and covers the original prediction with the results;

[0185] An optimization agent module iteratively optimizes the design variable vector based on the equivalent performance parameters and multi-objective functions defined in the claims using a non-dominated sorting genetic algorithm to generate candidate material solutions;

[0186] A verification agent module builds a body geometry and composition model based on the candidate solution, uses a three-dimensional U-Net accelerated Monte Carlo simulation to evaluate the shielding performance, and drives the next round of iteration through online comparison with experimental thresholds;

[0187] A communication bus and scheduler for efficient data exchange and job scheduling coordination between the CPU, GPU, functional modules, and storage media;

[0188] A user interface for real-time display of calculation progress, simulation results, and candidate solutions, and for receiving experimental feedback data to complete error-driven adaptive closed-loop optimization.

[0189] The above functions, if implemented in the form of software function units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the present application or the part of the prior art that essentially contributes to the present application or part of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various program code storage media.

[0190] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present application, and these modifications or replacements should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for intelligent design of radiation shielding materials based on cross-scale error feedback lock-in, characterized by, The method comprises the following steps: A multi-agent collaborative optimization framework of cross-scale error feedback lockstep is constructed, which comprises an atomic level prediction module, a mesoscopic level recalculation module, an optimization agent module and a verification agent module; A set of components for constituting a radiation shielding material is determined, atomic level graph structures of the components are obtained, and intrinsic performance parameters are extracted by the atomic level prediction module; Multi-physics field condition data is obtained, a design variable vector is constructed based on the intrinsic performance parameters and the multi-physics field conditions, the equivalent performance parameters of the composite material are calculated based on the design variables, and a cross-scale lockstep condition is established. If the deviation of the equivalent performance parameters from the calculation results of the mesoscopic level recalculation module exceeds the limit, the mesoscopic level recalculation module is triggered to recalculate until the set error threshold is met; Based on the corrected equivalent performance parameters, the optimization agent module uses a multi-objective optimization algorithm to optimize the design variables, and generates a candidate material scheme; The verification agent module constructs a solid geometry and component model based on the candidate material scheme, performs radiation shielding performance simulation using a deep neural network accelerated Monte Carlo radiation transport model, compares the simulation results with experimental test results, and outputs a radiation shielding material design scheme when the deviation meets the set threshold; If the deviation exceeds the limit, the experimental feedback results are used to correct the parameters of the mesoscopic level recalculation module, and the optimization step is continued to iterate.

2. The method of claim 1, wherein, The components include metal-based components, non-metal-based components and composite additive components for constituting a radiation shielding material, the metal-based components include high atomic number metal elements or their alloys, and the non-metal-based components include light elements or high hydrogen content compounds. The intrinsic performance parameters include gamma ray attenuation coefficient, neutron absorption cross section, thermal conductivity, electrical conductivity, linear thermal expansion coefficient and elastic modulus.

3. The method of claim 1, wherein, The atomic level prediction module is constructed based on a graph neural network and trained by the following steps: The atomic level graph structures of a plurality of different components are used as training samples, the nodes in the atomic level graph structures are atoms in the components, the node attributes include element type, valence electron number and isotope information, the edges are chemical bonds in the components, and the edge attributes include bond length, bond order and bond energy; The intrinsic performance parameters corresponding to the plurality of different components are used as the supervision labels of each training sample; The training samples are input into a multi-layer graph convolutional neural network of the atomic level prediction module for feature extraction; the training samples are input into the multi-layer graph convolutional neural network of the atomic level prediction module for feature extraction, the multi-layer graph convolutional neural network comprises a plurality of graph convolutional layers, a global average pooling layer and a fully connected output layer in sequence, the output layer outputs a predicted value of the intrinsic performance parameter, a loss function is constructed based on the mean absolute error and a physical constraint regularization term according to the predicted value of the intrinsic performance parameter and the supervision label, and the network parameters of the multi-layer graph convolutional neural network are updated iteratively by a back propagation algorithm, thereby establishing the atomic level prediction module reflecting the mapping relationship between the material components and microstructure and the intrinsic performance parameters; The loss function is: wherein L is a loss function, N is the total number of training samples, K is the number of intrinsic performance parameters, y i,j represents the jth intrinsic performance parameter in the supervised label of the ith training sample; is the jth intrinsic performance parameter prediction value of the ith training sample output by the material performance prediction model; λ1, λ2, λ3, λ4, λ5 are preset weight coefficients; L nonneg is a non-negative constraint penalty term; L WF is a heat conduction-electricity conduction constraint penalty term based on the Wiedemann-Franz law; L SA is a cross-section conservation constraint penalty term; L therm is a thermodynamic stability constraint penalty term, L phy is a physical residual term; The non-negative constraint penalty term is: wherein, is a set of non-negative intrinsic property parameters including thermal conductivity gamma-ray attenuation coefficient neutron absorption cross-section ReLU is a rectified linear unit function that outputs the absolute value of the input when it is negative and zero when the input is non-negative; The heat conduction-electricity conduction constraint penalty term based on the Wiedemann-Franz law has a formula as follows: wherein, is the predicted thermal conductivity for the i-th training sample, is the predicted electrical conductivity for the i-th sample, L0is the Lorenz constant, T i is the temperature of the i-th sample; The cross-section conservation constraint penalty term has a formula as follows: wherein, is a predicted gamma-ray attenuation coefficient for the i-th sample, is a predicted neutron absorption cross section for the i-th sample, C total is a predetermined total cross section conservation constant; The thermodynamic stability constraint penalty term has a formula as follows: wherein, respectively, the predicted linear thermal expansion coefficient and the predicted elastic modulus of the i-th sample, a max , E max respectively, a preset linear thermal expansion coefficient threshold value and an elastic modulus threshold value; The physical residual term L phy The formula is: L phy = w γ L γ + w n L n L γ =∑ i |E in,i -(E abs,i +E scat,i +E trans,i )| 2 L n =∑ i |Φ in,i -(Φ abs,i +Φ scat,i +Φ leak,i )| 2 where w γ , w n are the weight coefficients of the gamma-ray energy conservation residual and the neutron particle number conservation residual, respectively, L γ , L n are the gamma-ray energy conservation residual and the neutron particle number conservation residual, respectively, ∑ i denotes the accumulation for all training samples i, denotes the cumulative residual for the entire data set, E in,i , E abs,o , E scat,i , E trans,i denote the total energy of the incident gamma rays, the energy of the gamma rays absorbed by the material, the energy of the gamma rays scattered by the material, and the energy of the transmitted gamma rays in the i-th training sample, respectively, Φ in,i , Φ abs,i , Φ scat,i , Φ leak,o denote the incident neutron flux, the neutron flux absorbed by the material, the neutron flux scattered by the material, and the leaked neutron flux in the i-th training sample, respectively.

4. The method of claim 1, wherein, said multi-physical field condition data comprises encoded neutron energy spectrum Φ n (E), gamma energy spectrum Φ γ (E), temperature gradient mechanical stress σ mech .

5. The method of claim 1, wherein, The design variable vector is constructed as follows: x = {V i , t j , g j , S k , M k} wherein x is a material design variable vector, each material design variable vector x represents a candidate material scheme, V i is the mass percentage of the i-th component, t j represents the layer thickness of the j-th layer in the material, g j is the interlayer gradient parameter of the j-th layer in the material, S k is the bonding strength of the k-th interface in the material, M k is the thermal expansion coefficient matching degree of the k-th interface in the material; The equivalent performance parameters include: Effective thermal conductivity λ eff (x), where the formula is: where w i is the mass fraction of the i-th component, λ i is the thermal conductivity of the i-th component, ζ j is the experimentally calibrated continuous correction factor, 0.8 ≤ ζ j ≤ 5.0, is the interfacial thermal resistance calculated by molecular dynamics, d j is the thickness of the j-th layer material; Macroscopic total cross section ∑ t (x), given by the formula: wherein, is the neutron self-shielding factor calculated by the Monte Carlo method, σ n,i is the neutron macroscopic cross section of the i-th component; macroscopic gamma-ray attenuation coefficient μ t,eff (x), where the formula is: μ t,eff (x) =∑ i w i μ γ,i where μ γ,o is the gamma-ray linear attenuation coefficient of the i-th component.

6. The method of claim 5, wherein, The experimentally calibrated continuous correction factor ζ j The training set contains no less than 100 groups of ternary paired data composed of scanning electron microscope (SEM) images, interface thermal resistance results calculated by molecular dynamics (MD), and corresponding measured thermal resistance values. The ζ j The value range of ζ is determined by statistical analysis of the MD calculated values and measured values in the ternary data, and based on the upper and lower limits of the 95% confidence interval.

7. The method of claim 1, wherein, The optimization agent module adopts a multi-objective optimization algorithm to optimize the design variables to generate a candidate material scheme, and specifically includes the following steps: Based on the equivalent performance parameters of the constructed composite material, a material optimization multi-objective function is defined as follows: minF(x)=[f1(x),f2(x),f3(x),f4(x),f5(x)] f2(x) = -[1 - exp(-μ t,eff (x) · t total )] f3(x) = - [1 - exp(-∑ t (x) · t total )] f4(x) = σ thermal (x) σ thermal = |Δα| x ΔT x E eff f5(x)=Φ(-β) wherein f1(x) is a first optimization objective, representing the total density of the material, p i is the density of the i-th component; f2(x) is a second optimization objective, representing the value of the gamma ray shielding efficiency after taking the negative sign, p t,eff (x) represents the macroscopic gamma ray attenuation coefficient corresponding to the design variable x, t total is the total thickness of the material; f3(x) is a third optimization objective, representing the neutron shielding efficiency, p t (x) is the macroscopic total cross section of the material; f4(x) is a fourth optimization objective, representing the interlaminar thermal stress, D is the difference between the linear thermal expansion coefficients of the adjacent two layers of material, D is the temperature gradient, E eff is the equivalent elastic modulus of the composite material; f5(x) is a fifth optimization objective, representing the service reliability failure probability, F is the standard normal cumulative distribution function, and b is the reliability index obtained by linearization of the second moment. The constraint conditions are set as follows: wherein, is the minimum / maximum thickness of the jth layer of material, is the range of values for the gradient parameter of the jth layer of material, is the range of values for the kth interfacial bonding strength of the material, is the range of values for the kth interfacial coefficient of thermal expansion matching of the material; The material optimization multi-objective function is iteratively solved by using a non-dominated sorting genetic algorithm, and a candidate material scheme is generated based on a Pareto optimal solution set of the design variable vector x under the premise of meeting the constraint conditions.

8. The method of claim 1, wherein, The mesoscale recalculation module includes: A molecular dynamics re-calculation unit is configured to obtain interface structure information of each layer from the atomic level prediction module when the cross-scale lockstep condition is triggered, and automatically perform a molecular dynamics simulation to re-calculate the interface thermal resistance The thermal resistance value is obtained according to an interatomic interaction potential energy function and a heat flow statistical method under a constant temperature and constant pressure system. a Monte Carlo recalculation unit for automatically performing a Monte Carlo neutron transport simulation to recalculate the neutron self-shielding factor from the atomic-level prediction module when the cross-scale lockstep condition is triggered , the simulation takes a multi-energy section neutron source as input, and counts the neutron absorption, scattering and leakage flux of each layer of the material; The re-computation result is covered with a feedback unit for covering and corresponding prediction values of the atomic level prediction module output, and the re-computation result is preferentially used in subsequent equivalent performance parameter calculation and multi-objective optimization to drive the cross-scale error feedback lockstep closed loop until the prediction error meets the set threshold. The lockstep condition is as follows: wherein, are the effective thermal conductivity of the composite material and the macroscopic total cross section of neutrons, respectively, output by the atomistic prediction module, is the effective thermal conductivity of the composite material calculated by the molecular dynamics re-evaluation unit, is the macroscopic total cross section of neutrons of the composite material calculated by the Monte Carlo re-evaluation unit.

9. The method of claim 8, wherein, The mesoscale recalculation module is started only when the cross-scale lockstep condition is met, and the recalculation structure covers the corresponding predicted values output by the atomic-level prediction module to serve as a priority input for subsequent calculation of the equivalent performance parameters of the composite material, forming an error-driven adaptive closed loop.

10. The method of claim 1, wherein, The verification agent module constructs a solid geometry and component model based on the candidate material scheme, and performs radiation shielding performance simulation by using a deep neural network accelerated Monte Carlo radiation transport model, and specifically includes the following steps: A solid geometry and component model is constructed based on the design variable vector x of the candidate material scheme, and a deep neural network is used to predict the energy deposition distribution and scattering path probability of particles in each layer of material based on the solid geometry and component model. According to the particle transmittance and dose distribution output by the simulation, the shielding performance indicators of the material are calculated as follows: wherein η γ , η n are the γ-ray and neutron shielding efficiency of the material, respectively, T γ , T n are the simulated γ-ray and neutron transmission, respectively, D r is the dose attenuation rate, D in , D out are the simulated incident and transmitted surface dose, respectively; The shielding efficiency and dose attenuation rate are compared with preset target threshold values, and if and a radiation shielding material scheme meeting requirements is output, and if not, the optimization agent module is returned to continue iterative optimization. The solid geometry and component model is represented by octree voxelization, and the voxel size is ≤50 μm, and is encrypted to ≤10 μm at the layer interface.

Images