Data-driven small modular prismatic high temperature gas cooled reactor core optimization design method and system

Abstract

The application discloses a kind of based on data-driven small modular prismatic high temperature gas cooled reactor core optimization design method and system, comprising: constructing training data set;Training a data-driven proxy model is used to replace the execution high-cost simulation calculation;With a multi-objective evolutionary algorithm, and with the proxy model as its fitness function evaluator, to explore design space and generate a set of Pareto optimal candidate solution.The key features of the application are an iterative validation and model updating closed loop: the candidate solution is verified by high-fidelity simulation, and if the prediction accuracy does not meet the convergence criteria, the new data points after verification are expanded to the training data set, and the proxy model is retrained.The application can significantly reduce the calculation cost, efficiently explore the complex design space, and obtain a physically reliable and more optimal reactor design scheme.

Description

Technical Field

[0001] This invention relates to a data-driven method and system for optimizing the core design of a small modular prismatic high-temperature gas-cooled reactor, belonging to the field of nuclear reactor design and multiphysics coupling analysis technology. Background Technology

[0002] High-temperature gas-cooled reactors (HTGRs), as a fourth-generation advanced nuclear energy system, have attracted much attention due to their inherent safety characteristics and high outlet temperatures of 750-950°C. Among them, small modular HTGRs using prismatic fuel assemblies show great promise in distributed clean energy supply due to their compact design and flexible deployment. In recent years, breakthroughs in new high-temperature resistant and radiation-resistant materials such as SiC / SiC composite cladding and oxide dispersion reinforced steel (ODS) have provided the physical basis for achieving higher operating parameters and longer refueling cycles in prismatic HTGRs. However, this also places higher demands on core thermal design, especially under small modular conditions. The core power density is high and the physical processes are complex, requiring refined and coordinated optimization of parameters such as fuel assembly spacing, reflector thickness, fuel enrichment distribution, and coolant flow distribution to balance operating efficiency and safety margins.

[0003] The design of a prism reactor core is a typical multiphysics problem with strong coupling, and its performance is highly dependent on the close interaction between the neutron physics field and the thermal fluid field. To ensure the accuracy and safety of the design, especially to ensure that the peak fuel temperature does not exceed the safety limit (e.g., 1600°C) under accident conditions, high-fidelity multiphysics coupling simulation tools are required to conduct three-dimensional full-scale simulation analysis. Although these high-fidelity simulation tools are accurate, they have computational bottlenecks, namely, extremely high computational costs, with a single full-scale simulation potentially taking hours or even days.

[0004] Traditional optimization processes, such as directly coupling genetic algorithms with high-fidelity simulations, require thousands of performance evaluations, which is impractical in engineering practice. Therefore, designers are forced to rely on simplified models, engineering experience, or limited parameter scans. This often leads to conservative design solutions, hindering a systematic exploration of the entire design space and making it difficult to quantify the complex trade-offs between multiple conflicting design objectives (such as power density, temperature safety, and fuel economy). Consequently, this severely restricts the full realization of the performance potential of advanced reactors.

[0005] In recent years, data-driven technologies, especially AI-based surrogate models, have provided new approaches to solving the aforementioned challenges. Research shows that neural network-based surrogate models can reduce the time for a single performance evaluation from hours to milliseconds while maintaining prediction accuracy, achieving an efficiency improvement of several orders of magnitude. Combined with multi-objective genetic algorithms (such as NSGA-II), rapid global search of the design space can be achieved, obtaining a series of Pareto optimal solutions, and revealing the trade-offs between objectives.

[0006] In recent years, data-driven technologies have offered new approaches to solving these challenges. Data-driven surrogate models (such as neural network-based models) can reduce the time for a single performance evaluation by several orders of magnitude while maintaining acceptable predictive accuracy. Meanwhile, multi-objective evolutionary algorithms (such as NSGA-II) are mature tools for solving multi-objective optimization problems. However, although these technological components are known in their respective fields, simply combining them linearly is insufficient to address the unique design challenges faced by small modular prismatic stacks. These challenges include the strong coupling between neutronics and thermal-hydraulics, complex geometries, and conflicting design objectives under extreme conditions.

[0007] Therefore, the unresolved issue in current technology is the lack of a fully integrated, automated, and adaptively corrective intelligent design framework capable of efficiently and reliably exploring the complex multiphysics design space of prismatic high-temperature gas-cooled reactors. Existing methods are often one-off, open-loop, or focused on a single physics field, lacking a closed-loop mechanism that can continuously improve its prediction accuracy and optimization performance through iterative learning. Summary of the Invention

[0008] Objective of the Invention: To overcome the shortcomings of existing technologies, this invention provides a data-driven method and system for optimizing the core design of a small, modular, prismatic high-temperature gas-cooled reactor. The invention aims to achieve efficient, global, and multi-objective optimization of core design parameters by constructing an automated, adaptive closed-loop process of "modeling-simulation-learning-optimization-verification," thereby maximizing the reactor's performance potential while ensuring safety.

[0009] Technical Solution: To solve the above-mentioned technical problems, the present invention provides a data-driven optimization design method for a small modular prismatic high-temperature gas-cooled reactor core, which includes the following steps:

[0010] (1) A three-dimensional multiphysics coupled numerical model of a small modular prismatic high-temperature gas-cooled reactor core is provided, wherein the model contains at least one neutron physics module and one thermal-hydraulic module.

[0011] (2) Define a set of design variables that constitute a multidimensional design space; and define a set of performance indicators that are calculated by the numerical model, wherein the performance indicators include at least two conflicting optimization objectives and at least one design constraint.

[0012] (3) Apply experimental design methods such as Latin hypercube sampling (LHS) to generate a set of sample points in the design space; for each sample point in the set of sample points, perform numerical simulation of the multiphysics coupling model described in step (1) to generate a set of corresponding performance index values ​​for each sample point; and compile the sample points and their corresponding performance index values ​​into a training dataset.

[0013] (4) A neural network is trained using the training dataset, wherein the trained neural network constitutes a surrogate model, which is configured to establish a mapping relationship from the design variables to the performance metrics;

[0014] (5) Provide the surrogate model to a multi-objective evolutionary algorithm as the fitness function evaluator of the algorithm; and execute the multi-objective evolutionary algorithm, wherein the optimization objective is used as the fitness function and the design constraints are used as the limiting conditions, thereby generating a Pareto optimal solution set;

[0015] (6) Select at least one candidate design scheme from the Pareto optimal solution set; verify the candidate design scheme by performing numerical simulation of the multiphysics coupling model described in step (1) to obtain a set of verification performance index values; compare the verification performance index values ​​with the performance index values ​​predicted by the surrogate model to determine whether a preset convergence criterion is met; and if the convergence criterion is not met, expand the training dataset using the candidate design scheme and the corresponding set of verification performance index values, and retrain the data-driven surrogate model using the expanded training dataset by repeating steps (4) to (6).

[0016] Furthermore, in step (1), the multiphysics coupled numerical model is coupled with a Monte Carlo neutron transport program for determining the core power distribution and a computational fluid dynamics (CFD) program for determining the three-dimensional temperature field; the neutron physics module and the thermal hydraulic module exchange data through an external coupling interface and calculate until convergence according to a preset iterative strategy.

[0017] Furthermore, the aforementioned coupling interface is an interface program or script configured to: parse the output of the neutron physics module to obtain the power density of the volume element or mesh element, and write it into the user-defined function (UDF) or equivalent source term input of the thermal-hydraulic module; extract material temperature and density data from the output of the thermal-hydraulic module, and update the input file or cross-section library associated parameters of the neutron physics module accordingly.

[0018] Furthermore, the design parameters in step (2) include geometric parameters and operating parameters; the geometric parameters are selected from at least one of the following: fuel assembly geometry and size, number and diameter of coolant channels and their arrangement, fuel enrichment, TRISO coated particle filling ratio, and reflective layer thickness; the operating parameters include at least one of the following: coolant mass flow rate, coolant inlet temperature, and outlet pressure.

[0019] Furthermore, the optimization objectives in step (2) include at least two of the following: maximizing core power density, minimizing peak fuel temperature, minimizing fuel enrichment, and minimizing coolant pressure drop; the design constraints include at least one of the following: peak fuel temperature limit, non-positive moderator temperature coefficient requirement, effective multiplication factor operating range, and structural material stress limit.

[0020] Furthermore, in step (3), the sample size of the Latin hypercube sampling satisfies the requirement of stratified uniform coverage of the value range of each design parameter. The sample size is ≥ 5 times the dimension of the design variable to ensure stratified uniform coverage of the value range of each design parameter. The sample dataset is divided according to the ratio of training set, validation set and test set, and normalization or standardization preprocessing is performed on the input and output.

[0021] Furthermore, the neural network proxy model in step (4) is a deep neural network (DNN) or a multilayer perceptron (MLP), including at least one input layer, at least one hidden layer with a non-linear activation function and at least one output layer, and meets a preset accuracy threshold on the test set.

[0022] Furthermore, the training of the aforementioned neural network agent model employs either the Adam optimization algorithm or the L-BFGS optimization algorithm, and uses learning rate scheduling.

[0023] Furthermore, the multi-objective evolution algorithm in step (5) is the Non-Dominated Sorting Genetic Algorithm II (NSGA-II).

[0024] Furthermore, the method of step (6) includes: selecting a final design scheme from the Pareto optimal solution set by using one or a combination of compromise point selection, weighted and scalar quantification or weighted multi-index ranking, and performing a safety constraint verification process on the final scheme.

[0025] Furthermore, the verification in step (6) includes performing mesh independence verification and boundary condition sensitivity analysis on the candidate design scheme, and using the relative error between the high-fidelity model and the surrogate model in the predicted values ​​of key indicators as the convergence criterion, and using the relative error between the high-fidelity model and the surrogate model in the predicted values ​​of key indicators ≤5% as the convergence criterion.

[0026] Furthermore, the present invention also provides a system for implementing the above method, the system comprising:

[0027] The parametric modeling module is configured to automatically generate a three-dimensional geometric model of the reactor core based on the input design parameters and to generate input files for multiphysics simulation software.

[0028] The multiphysics coupled solver module is deployed on a high-performance computing cluster and configured to perform high-fidelity coupled simulations of neutron physics and computational fluid dynamics on sample points and output corresponding performance index data.

[0029] The data-driven agent model module is configured to build and train neural network agent models based on sample datasets and provide fast prediction services.

[0030] The multi-objective optimization module is configured to call the proxy model to perform NSGA-II multi-objective optimization operations and generate a Pareto optimal solution set;

[0031] The results verification and visualization module is configured to analyze the Pareto optimal solution set output by the multi-objective optimization module, select a final design scheme, perform high-fidelity verification of the final design scheme by the multi-physics coupled solver module, and visualize the optimization results to output the final optimized design.

[0032] The following is a detailed explanation of the working process of the system's five modules, along with the flowchart:

[0033] 1. Parametric Modeling Module

[0034] Input: Receive design parameters (such as core geometry, material properties, coolant parameters, etc.) provided by the main control script.

[0035] Processing: Automatically generate a three-dimensional geometric model of the reactor core based on the input parameters, and generate standardized simulation input files (such as mesh generation, physical field boundary condition definition, etc.) for multiphysics simulation software.

[0036] Output: Passes the simulation input file to the multiphysics coupled solver module for subsequent high-fidelity simulation calculations.

[0037] 2. Multiphysics Coupled Solver Module

[0038] Input: Receives the simulation input file generated by the parametric modeling module.

[0039] Processing: Deployed on a high-performance computing cluster, it performs high-fidelity coupled simulations of neutron physics and computational fluid dynamics on sample points (such as simulating the coupled processes of neutron transport, heat transfer, and coolant flow within the reactor core).

[0040] Output: Output the corresponding performance index data (such as core power distribution, fuel temperature, coolant flow rate, etc.) as "high-fidelity training data" and pass it to the data-driven proxy model module;

[0041] In the "Result Verification Phase", the candidate solutions selected by the Result Verification and Visualization Module are received, high-fidelity verification is performed, and the verification results are returned.

[0042] 3. Data-driven agent model module

[0043] Input: Receives a high-fidelity sample dataset output from the multiphysics coupling solver module.

[0044] Processing: Based on the sample dataset, construct and train a neural network proxy model (such as using a deep learning model to fit the mapping relationship between "design parameters and performance indicators") to enable the model to quickly predict the core performance.

[0045] Output: After training is complete, it receives the scheduling training instructions from the main control script to continuously optimize the model accuracy; during the optimization phase, it receives the call request from the multi-objective optimization module to quickly predict the performance of candidate design schemes and provide efficient support for optimization operations.

[0046] 4. Multi-objective optimization module

[0047] Input: Receives the startup optimization command from the main control script; calls the trained proxy model from the data-driven proxy model module.

[0048] Processing: Perform NSGA-II multi-objective optimization operations based on the surrogate model (simultaneously optimizing multiple objectives such as core thermal efficiency, safety, and economy), and search for the optimal solution in the design space.

[0049] Output: Generate a Pareto optimal solution set (i.e., "a set of design schemes that achieve optimal balance among multiple objectives") and pass it to the result verification and visualization module.

[0050] 5. Result Validation and Visualization Module

[0051] Input: Receive the Pareto optimal solution set output by the multi-objective optimization module.

[0052] Processing: Perform multi-dimensional analysis on the Pareto optimal solution set (such as technical feasibility, economic cost, operation and maintenance difficulty, etc.) and select a final design scheme;

[0053] The multiphysics coupling solver module was called to perform high-fidelity verification of the scheme to confirm the consistency between the scheme performance and the surrogate model prediction; the optimization results were visualized (such as comparison of 3D core models, performance index curve analysis, etc.).

[0054] Output: Output the final optimized design and feed the verification results back to the closed loop of the process to complete the entire core optimization design process.

[0055] Through the closed-loop collaboration of the above modules in the "input-processing-output" process, the system realizes fully automated core optimization design from "parametric modeling → high-fidelity simulation → proxy model acceleration → multi-objective optimization → result verification".

[0056] Furthermore, the data-driven agent model module is built based on the TensorFlow or PyTorch deep learning framework.

[0057] Furthermore, the parametric modeling module, multiphysics coupled solver module, data-driven proxy model module, and multi-objective optimization module are scheduled and controlled through a unified master control script.

[0058] Furthermore, the result verification and visualization module is configured to generate Pareto front plots and comparative cloud plots of key physical fields (such as temperature field and power field) of the core before and after optimization.

[0059] Furthermore, the system includes one or more processors and instructions stored on a non-transitory computer-readable medium, which, when executed by the one or more processors, cause the system module to be executed.

[0060] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0061] High efficiency: By replacing most of the high-cost simulations with proxy models, design efficiency is greatly improved, making it possible to complete global multi-objective optimization within an acceptable engineering timeframe.

[0062] Adaptability and accuracy: The unique iterative verification and model update closed loop of this invention is essentially an active learning framework. It can intelligently and accurately add new high-fidelity data points to the design space region where the surrogate model's prediction is the least accurate, but which is also identified by the optimizer as the most promising region. This continuously corrects and improves the global prediction accuracy of the surrogate model, ensuring the physical reliability of the final optimization result.

[0063] Automation: Through a modular system architecture, parametric modeling, coupled simulation, surrogate model training, optimization calculation and verification are integrated into a fully automated intelligent design platform. Attached Figure Description

[0064] Figure 1 A schematic diagram of the data-driven optimization design process for small modular prismatic high-temperature gas-cooled reactor cores.

[0065] Figure 2 A block diagram of a data-driven, small, modular, prismatic high-temperature gas-cooled reactor core optimization design system. Detailed Implementation

[0066] The invention will now be further described with reference to the accompanying drawings.

[0067] Example 1

[0068] This embodiment details a data-driven optimization design method for small modular prismatic high-temperature gas-cooled reactor cores, the process of which can be found in the following reference. Figure 1 It mainly includes the following steps:

[0069] Step (1): Establish a parameterized three-dimensional multiphysics coupled numerical model that can automatically update geometry and settings based on a set of input design variables.

[0070] To achieve automated optimization, a parameterized, small-scale, modular, prismatic 3D model of the high-temperature reactor core must first be constructed. This model precisely describes the key geometric entities within the core, including fuel assemblies (containing fuel channels, coolant passages, and a graphite matrix), control rods, combustible poison rods, and side reflectors. Through parameterization techniques, geometric design variables (such as coolant passage diameter, fuel assembly spacing, and reflector thickness) are directly correlated with the model's geometric dimensions, enabling the model to automatically update and generate based on a given set of input parameters.

[0071] Specifically, the multiphysics coupling numerical model in this embodiment is constructed as follows:

[0072] Neutron Physics Module: Implemented using the Monte Carlo neutron transport program OpenMC. The OpenMC model defines material composition, geometry, and computational settings by reading XML-formatted input files. The model needs to accurately describe the microstructure and macroscopic arrangement of fuel elements (such as TRISO-coated particles). A key aspect is handling temperature feedback effects: through data exchange with the thermal-hydraulic module, the temperature and density distribution of materials in different regions of the reactor core are obtained. OpenMC uses this data to perform Doppler broadening calculations in real time, updating the neutron cross section, thereby accurately simulating the effect of temperature on reactivity.

[0073] Thermal-hydraulic module: Commercial or open-source computational fluid dynamics (CFD) software is used to generate high-quality meshes for both the fluid domain (coolant channels) and the solid domain (fuel and structural materials). To ensure computational accuracy, structured meshes are preferred, and rigorous mesh independence verification is essential to ensure that the calculation results do not significantly change with mesh density. The Reynolds-averaged Navier-Stokes equations, including mass, momentum, and energy conservation equations, are solved. An appropriate turbulence model is selected based on the flow regime. The heat transfer model must consider heat conduction within the solids, convective heat transfer at the fluid-solid interface, and thermal radiation at high temperatures. Boundary conditions are set according to reactor design parameters, such as given mass flow rate and temperature at the coolant inlet and given pressure at the outlet.

[0074] Multiphysics Coupling Strategy: This invention employs a flexible and practically applicable external coupling (or loose coupling) scheme, using a master script (such as a Python script) to schedule the alternating operation of two independent solvers: neutron physics and thermal hydraulics. The specific process includes:

[0075] (1) The three-dimensional power distribution of the reactor core was calculated by the neutron physics module.

[0076] (2) The main control script parses its output file, extracts the power density value of each computational grid or volume element, and writes it into an input file that the CFD module can read, such as loading it into the energy equation as a source term of a user-defined function (UDF).

[0077] (3) The CFD module calculates the converged three-dimensional temperature field and density field of the core.

[0078] (4) The main control script extracts the average temperature and density of each material region from the CFD output results and updates the OpenMC input file for the next neutron cross section calculation.

[0079] (5) The classic Picard iteration method is used to repeat the above process until the relative change of key coupling parameters (such as effective multiplication factor, fuel peak temperature, power peak factor, etc.) between two adjacent iterations is less than the preset convergence criterion (e.g., 0.1%). Then the multiphysics coupling calculation is considered to have converged.

[0080] Step (2): Define the design space and performance space

[0081] First, identify the key design parameters that affect the thermal fluid performance of the reactor core and determine their value range to form a multi-dimensional design space.

[0082] Secondly, performance indicators used to evaluate the merits of design schemes are determined, thus forming a performance space.

[0083] Design variables include geometric parameters (such as fuel assembly size, coolant channel diameter, reflector thickness, etc.) and operating parameters (such as coolant mass flow rate, inlet temperature, etc.).

[0084] Performance metrics include conflicting optimization objectives (such as "maximizing thermal power density" to improve economy and "minimizing fuel peak temperature" to improve safety) and design constraints that must be met (such as a fuel peak temperature limit of <1600℃ and a life-end kJ / mJ limit). eff >1.0 etc.), the conflicting optimization objectives selected in this embodiment are: "maximizing thermal power density" to improve economy and "minimizing fuel peak temperature" to improve safety; the design constraint that must be met is: upper limit of fuel peak temperature <1600℃.

[0085] Step (3): Construct the dataset using experimental design methods such as Latin hypercube sampling (LHS).

[0086] Within the design space defined in step (2), a set of sample points with good space-filling and uniformity is generated using the Latin hypercube sampling (LHS) method. Compared to simple random sampling, LHS can capture global features within the design space more effectively with fewer sample points, laying the foundation for building a high-quality surrogate model.

[0087] For each sample point generated by LHS (i.e., a specific combination of design parameters), a complete simulation calculation is performed using the high-fidelity multiphysics coupling model in step (1) until convergence. The design variable combination and its corresponding performance index (values ​​of optimization objectives and constraints) for each sample point are recorded to form an "input-output" data pair. All data pairs together constitute the dataset used for training and validating the surrogate model.

[0088] Step (4): Data-driven agent model construction and training

[0089] This step uses the dataset constructed in step (3) to train a data-driven proxy model. This proxy model is one of the core components of this invention.

[0090] Data-driven surrogate models encompass all mathematical or computational models capable of learning input-output mappings from data. In a preferred embodiment of the invention, the model may be a deep neural network (DNN) or a multilayer perceptron (MLP). However, those skilled in the art will understand that other types of models are equally applicable, such as Gaussian process (GP) models, radial basis function (RBF) networks, or support vector regression (SVR), etc.

[0091] The specific implementation method of this embodiment is as follows:

[0092] Model Architecture: Employs a Deep Neural Network (DNN), also known as a Multilayer Perceptron (MLP). The network structure includes: Input Layer: The number of neurons equals the dimension of the design variables. Hidden Layers: Contains one or more hidden layers, each consisting of several neurons. Neurons are connected by weights and biases, and non-linear activation functions (such as ReLU) are used to capture the complex non-linear relationship between the design variables and the performance metric. Output Layer: The number of neurons equals the dimension of the performance metric to be predicted.

[0093] Data preprocessing: Before feeding the data into the neural network for training, the input data (design variables) and output data (performance metrics) are normalized or standardized. This step can eliminate the influence of different parameter units, accelerate the convergence speed of model training, and improve the stability of the model.

[0094] Model training and validation: Divide the generated dataset into a certain proportion (e.g., 70% training set, 15% validation set, 15% test set).

[0095] (a) The network is trained using a training set, and the network weights are continuously adjusted using a backpropagation algorithm and an optimizer (such as the Adam algorithm) to minimize the loss function (such as mean squared error, MSE). A validation set is used to monitor model performance during training and prevent overfitting.

[0096] (b) After training, evaluate the final generalization ability of the surrogate model using an independent test set. Evaluation metrics include root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). The model is considered successfully trained and can be used for subsequent optimization steps only if the surrogate model's value on the test set is higher than a preset accuracy threshold (e.g., 0.99).

[0097] Step (5): Apply multi-objective evolutionary algorithm for multi-objective optimization

[0098] "Multi-objective evolutionary algorithm": In the preferred embodiment of the present invention, the Non-dominated sorting genetic algorithm II (NSGA-II) is used due to its robustness and wide application. However, other algorithms, such as the Strength Pareto Evolutionary Algorithm 2 (SPEA2) and the Decomposition-based Multi-objective Evolutionary Algorithm (MOEA / D), can also be applied to the method of the present invention.

[0099] First, the core design problem is formalized into a standard multi-objective optimization problem. This involves clearly defining the design variables and their ranges, conflicting optimization objectives, and the design constraints that must be satisfied.

[0100] The trained surrogate model is used as the fitness function evaluator for the NSGA-II algorithm. In each generation of the algorithm, for each individual in the population (representing a set of design variables), time-consuming high-fidelity simulations are no longer required; instead, the surrogate model is directly invoked to quickly calculate its corresponding performance metrics. The NSGA-II algorithm stratifies and selects the population based on non-dominated sorting and crowding calculations, and generates new offspring by simulating genetic operators such as binary crossover and polynomial mutation, gradually evolving towards the Pareto optimal front.

[0101] After sufficient generations of evolution, the NSGA-II algorithm outputs a Pareto-optimal solution set. This solution set contains a series of "non-dominated" design schemes, each representing an optimal trade-off between different optimization objectives. Designers can visualize and analyze the Pareto front plot to make decisions based on engineering preferences.

[0102] Step (6): High-fidelity verification and decision-making

[0103] To ensure the reliability of the final design, the results obtained from the optimization of the proxy model must be rigorously verified with high fidelity.

[0104] Candidate selection and verification: Several representative candidate design schemes are selected from the final Pareto front, such as extreme schemes at both ends of the front and compromise schemes in the "inflection point" region. For each candidate scheme, a complete numerical simulation is performed using the high-fidelity multiphysics coupling model described in Part II to obtain its "real" performance index values.

[0105] Convergence Criteria and Model Iterative Updates: Compare the calculation results of the high-fidelity model with the prediction results of the surrogate model for the same scheme, and calculate key performance indicators (such as peak fuel temperature, kJ / m³). eff The relative error (etc.) is considered. If the error exceeds the preset convergence criterion (e.g., 5%), it indicates that the surrogate model's prediction accuracy in that region of the design space is insufficient. In this case, the validated "candidate solution-real performance" data pair is added to the original training dataset, and the process returns to step 4. The neural network surrogate model is retrained using the expanded dataset, and then multi-objective optimization is re-executed. This iterative process continuously "corrects" the prediction bias of the surrogate model until all candidate solutions on the Pareto front can be accurately validated by the high-fidelity model.

[0106] Final Solution Decision: After the verification process converges, designers can select a final design solution from the verified Pareto optimal solution set based on specific engineering requirements and preferences (e.g., different emphases on safety or economy). Methods such as trade-off point selection can be used for decision-making, and a final, comprehensive safety constraint check is performed on the solution to complete the entire optimization design process and output the final optimization design report. To illustrate the implementation of this invention more specifically, a hypothetical and simplified optimization design process is provided below.

[0107] Problem definition: The objective is to optimize the fuel enrichment and coolant channel diameter of a small prismatic stack to achieve the dual goals of maximizing power density and minimizing peak fuel temperature, while meeting the safety constraint that the fuel temperature does not exceed 1600°C.

[0108] Initial Sampling: Generate 30 sample points in the design space using LHS. Run a high-fidelity simulation (e.g., a program coupled with OpenMC and Fluent) on these 30 points to construct the initial training dataset.

[0109] Proxy Model Training (Iteration 1): Train a deep neural network (DNN) surrogate model using these 30 data points. Multi-Objective Optimization (Iteration 1): Use this DNN model as the fitness function and run the NSGA-II algorithm (e.g., population size 100, evolution 200 generations). The algorithm outputs a Pareto front containing multiple trade-off solutions.

[0110] Validation and Update (Iteration 1): Select 5 representative points on the Pareto front (e.g., solutions at the two endpoints and three inflection points). Run high-fidelity simulations on these 5 points. It is found that the surrogate model predicts a temperature of 1550℃ for one solution, while the high-fidelity simulation result is 1610℃, indicating a large error and a violation of the constraints. Add these 5 new, validated data points to the training dataset, now containing 35 points. Iterative Loop: Return to step 3 and retrain the DNN model using the 35 data points. Repeat steps 4 and 5. After several iterations, the surrogate model's prediction error for all candidate solutions on the Pareto front is less than 3%, satisfying the convergence criterion. Final Decision: At this point, the validated Pareto front is considered reliable. The designer can select the final design solution from it based on engineering preferences.

[0111] To enable those skilled in the art to implement the invention without excessive experimentation, Table 1 provides a set of illustrative parameters for guiding implementation.

[0112] Table 1 Examples of Design Variables, Optimization Objectives, and Constraints

[0113]

[0114] Example 2

[0115] This embodiment will refer to Figure 2 This paper describes an optimization design system used to implement the above methods. This system automates and modularizes the entire optimization process, forming a closed loop. For example... Figure 2 As shown, the system architecture consists of five core functional modules that work together:

[0116] Parametric Modeling Module: This module is the starting point for the automated design process. Based on a set of input design variables (covering geometric and operational parameters), it automatically generates a 3D geometric model of the reactor core and produces the necessary input files for downstream multiphysics simulation software. For example, it generates XML-formatted geometry and material description files for neutron physics calculation programs (such as OpenMC) and mesh files and setup scripts for computational fluid dynamics (CFD) software.

[0117] Multiphysics Coupled Solver Module: This module is the core for acquiring high-fidelity physics data and is typically deployed on a high-performance computing (HPC) cluster. It is responsible for performing accurate but time-consuming neutron physics-thermal-hydraulic coupled simulations. By performing high-fidelity simulations of sample points in the design space, this module provides a high-quality training dataset for building data-driven surrogate models.

[0118] Data-Driven Surrogate Model Module: The core of this module is a neural network model. Built on deep learning frameworks such as TensorFlow or PyTorch, this module uses a "design variable-performance metric" dataset generated by the multiphysics coupling solver module for training, learning and constructing a mathematical model (surrogate model) capable of rapidly predicting reactor performance. This model can replace high-fidelity simulations that take hours or even days with millisecond-level speeds, making it crucial for achieving efficient optimization.

[0119] Multi-objective optimization module: This module integrates advanced optimization algorithms, such as NSGA-II. It calls upon a data-driven surrogate model as the fitness function evaluator to perform a fast, large-scale iterative search in a multi-dimensional design space to identify a series of Pareto optimal solutions that achieve the best trade-off between different optimization objectives.

[0120] The Results Verification and Visualization module is responsible for post-processing and decision support of the optimization results. It selects representative candidate design schemes from the Pareto optimal solution set and calls the multiphysics coupled solver module to perform high-fidelity verification to ensure the accuracy of the surrogate model's predictions. Simultaneously, this module visualizes the optimization results (such as Pareto front plots and key physics comparison contour maps), providing designers with intuitive decision-making support and outputting the final optimized design scheme.

[0121] Preferably, all the above modules are scheduled and controlled through a unified master script (e.g., a Python script), achieving full automation from parametric modeling to final verification. Through the closed-loop collaboration of the above modules' "input-processing-output," the system realizes fully automated core optimization design from "parametric modeling → high-fidelity simulation → proxy model acceleration → multi-objective optimization → result verification."

[0122] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A data-driven optimization design method for a small, modular, prismatic high-temperature gas-cooled reactor core, characterized in that, Includes the following steps: (1) A three-dimensional multiphysics coupled numerical model of a small modular prismatic high-temperature gas-cooled reactor core is given, wherein the coupled numerical model contains at least one neutron physics module and one thermal-hydraulic module. (2) Define a set of design variables, which constitute a multidimensional design space; And define a set of performance metrics, which are calculated by the coupled numerical model, wherein the performance metrics include at least two conflicting optimization objectives and at least one design constraint. (3) Apply experimental design method to generate a set of sample points in the design space; for each sample point in the set of sample points, perform numerical simulation of the multiphysics coupling model described in step (1) to generate a set of corresponding performance index values ​​for each sample point. And compile the sample points and their corresponding performance index values ​​into a training dataset; (4) Use the training dataset to train a data-driven surrogate model, wherein the surrogate model is configured to establish a mapping relationship from the design variables to the performance metrics; (5) Provide the surrogate model to a multi-objective evolutionary algorithm as the fitness function evaluator of the algorithm; and execute the multi-objective evolutionary algorithm, wherein the optimization objective is used as the fitness function and the design constraints are used as the limiting conditions, thereby generating a Pareto optimal solution set; (6) Select at least one candidate design scheme from the Pareto optimal solution set; The candidate design scheme is verified by performing numerical simulation of the multiphysics coupling model described in step (1) to obtain a set of verification performance index values; the verification performance index values ​​are compared with the performance index values ​​predicted by the surrogate model to determine whether a preset convergence criterion is met; and if the convergence criterion is not met, the training dataset is expanded using the candidate design scheme and the corresponding set of verification performance index values, and the data-driven surrogate model is retrained using the expanded training dataset by repeating steps (4) to (6) until the final optimized design scheme is output. In step (1), the multiphysics coupled numerical model couples a Monte Carlo neutron transport program for determining the core power distribution and a computational fluid dynamics (CFD) program for determining the three-dimensional temperature field. The neutron physics module and the thermal hydraulic module exchange data through an external coupling interface and calculate to convergence according to a preset iteration strategy. The coupling interface is an interface program or script configured to: parse the output of the neutron physics module to obtain the power density of the volume element or mesh element, and write it into the user-defined function UDF or equivalent source term input of the thermal hydraulic module; extract material temperature and density data from the output of the thermal hydraulic module, and update the input file or cross-section library associated parameters of the neutron physics module accordingly.

2. The data-driven optimization design method for small modular prismatic high-temperature gas-cooled reactor cores according to claim 1, characterized in that: The design variables in step (2) include geometric parameters and operating parameters; the geometric parameters are selected from at least one of the following: fuel assembly geometry and size, number and diameter of coolant channels and their arrangement, fuel enrichment, TRISO coated particle filling ratio, and reflective layer thickness; the operating parameters include at least one of the following: coolant mass flow rate, coolant inlet temperature, and outlet pressure.

3. The data-driven optimization design method for small modular prismatic high-temperature gas-cooled reactor cores according to claim 1, characterized in that: The optimization objectives in step (2) include at least two of the following: maximizing core power density, minimizing peak fuel temperature, minimizing fuel enrichment, and minimizing coolant pressure drop; the design constraints include at least one of the following: peak fuel temperature limit, non-positive moderator temperature coefficient requirement, effective multiplication factor operating range, and structural material stress limit.

4. The data-driven optimization design method for small modular prismatic high-temperature gas-cooled reactor cores according to claim 1, characterized in that: The sample size of the training dataset in step (3) satisfies the requirement of hierarchical uniform coverage of the value range of each design parameter. The sample dataset is divided into training, validation, and test sets according to their proportions, and normalization or standardization preprocessing is performed on the inputs and outputs.

5. The data-driven optimization design method for small modular prismatic high-temperature gas-cooled reactor cores according to claim 1, characterized in that, The surrogate model in step (4) is a deep neural network (DNN) or a multilayer perceptron (MLP), including at least one input layer, at least one hidden layer with a nonlinear activation function, and at least one output layer, and meets a preset accuracy threshold on the test set.

6. The data-driven optimization design method for small modular prismatic high-temperature gas-cooled reactor cores according to claim 1, characterized in that, Step (6) further includes: selecting a final design scheme from the Pareto optimal solution set by using one or a combination of compromise point selection, weighted and scalar quantification or weighted multi-index ranking, and performing a safety constraint verification process on the final scheme.

7. The data-driven optimization design method for small modular prismatic high-temperature gas-cooled reactor cores according to claim 1, characterized in that, The verification in step (6) includes performing mesh independence verification and boundary condition sensitivity analysis on the candidate design scheme, and using the relative error between the high-fidelity model and the surrogate model in the predicted values ​​of key indicators as the convergence criterion.

8. A system for optimizing the core design of a small modular prismatic high-temperature gas-cooled reactor, comprising executing the data-driven optimization design method for a small modular prismatic high-temperature gas-cooled reactor core as described in any one of claims 1 to 7, characterized in that, include: The parametric modeling module is configured to automatically generate a three-dimensional geometric model of the reactor core based on the input design parameters and to generate input files for multiphysics simulation software. The multiphysics coupled solver module is deployed on a high-performance computing cluster and configured to perform high-fidelity coupled simulations of neutron physics and computational fluid dynamics on sample points and output corresponding performance index data. The data-driven proxy model module is configured to construct and train a neural network proxy model based on the sample dataset output by the multiphysics coupling solver module, and provide fast prediction services. The multi-objective optimization module is configured to call the proxy model to perform NSGA-II multi-objective optimization operations and generate a Pareto optimal solution set; The result verification and visualization module is configured to analyze the Pareto optimal solution set output by the multi-objective optimization module to select the final design scheme, perform high-fidelity verification of the final design scheme by the multi-physics coupled solver module, and visualize the optimization results to output the final optimized design.

Images