A method for locating foreign objects in a nuclear reactor in-core component based on a physical AI engine

Abstract

The present application relates to a kind of nuclear reactor in-vessel component foreign matter positioning method based on physical AI engine, including constructing CFD model and CAD real model;Carry out steady-state CFD model calculation to form snapshot data set, POD is reduced to snapshot data, and ANN training is carried out to establish the mapping relationship of operating condition parameter to modal coefficient, complete surface triangulation, normal extraction, local path expression and space query organization to CAD real model, and encapsulation under unified coordinate and unified call interface, construct physical AI engine;Call physical AI engine to carry out flow field reconstruction, form frozen background flow field set;Carry out monte carlo lagrange particle tracking;Call CAD real model to carry out collision detection, speed response and stagnation determination;The stagnation three-dimensional coordinates under each operating condition are counted and the high-probability stagnation region coordinate is output. Realize the unified call of millisecond level online flow field reconstruction and real geometry collision information, and give consideration to the calculation efficiency of multiple operating conditions and the geometric reality of local collision determination.

Description

Technical Field

[0001] This invention belongs to the field of nuclear reactor operation monitoring and predictive maintenance technology, and particularly relates to a method for locating foreign objects in nuclear reactor internals based on a physics AI engine. Background Technology

[0002] The reactor internals are crucial components of a pressurized water reactor nuclear power plant. They comprise various solid components, including control rod guide tubes, core support plates and their drainage holes, instrumentation lines, and the core shroud, with densely packed fuel assemblies occupying most of the space. Under normal operating conditions, the high-temperature, high-pressure coolant (approximately 280–295°C and 15.5 MPa) flows through the internals at a high velocity. If loose metallic foreign objects (such as bolts, nuts, or metal fragments) are present in the primary loop, they will enter the internals region with the coolant and may become stuck in the drainage holes of the control rod guide tubes or between fuel assemblies. This can lead to flow-induced vibration and fretting wear, posing a risk to the safe insertion of control rods and the overall safe operation of the nuclear power plant.

[0003] Currently, most existing methods for dealing with loose foreign objects in reactor internals adopt a "find first, then remove" approach. For example, referring to Chinese Patent Publication No. CN119763871A, published on April 4, 2025, a method for handling foreign objects in the lower reactor internals is disclosed. Specifically, this includes: using a video inspection mechanism to perform a full-area video inspection of the lower reactor internals to determine the presence of foreign objects; determining the location, size, and type of foreign objects based on the full-area images of the lower reactor internals obtained by the video inspection mechanism; using a data processing module of a position navigation mechanism to determine the corresponding drainage hole area number of the target foreign object, the type of the robotic arm body of the robotic arm mechanism, the type of gripper of the grasping mechanism, and the initial motion posture of the robotic arm mechanism; and using a remote control mechanism to operate the robotic arm mechanism to move and operate the grasping mechanism to accurately grasp the target foreign object.

[0004] The specific measures of this technical solution are to use a video inspection agency to conduct a full-area video inspection of the internal components of the lower part of the reactor to obtain full-area images, determine the location, size and type of foreign objects based on the full-area images, and then determine the installation position and grasping posture of the robotic arm based on the three-dimensional digital model, and finally the robotic arm completes the grasping of foreign objects.

[0005] Acoustic signal detection technology based on the Loose Parts Monitoring System (LPMS) is used to detect acoustic signals generated by foreign object impacts in real time by installing an acceleration sensor on the outer wall of the reactor pressure vessel, thereby determining whether there are loose foreign objects in the primary loop during operation.

[0006] In the field of numerical prediction of foreign object trajectories, the academic community has explored combining CFD flow field calculations with Monte Carlo particle tracking. Previous studies have proposed a method based on a combination of full-order CFD flow field analysis and Monte Carlo particle tracking to predict the collision location of loose foreign objects at the tube sheet of a pressurized water reactor steam generator.

[0007] However, the aforementioned prior art has the following drawbacks: 1. Existing foreign object handling solutions lack the ability to predict the location of foreign objects in advance, and can only rely on "post-event search" of full-area video flaw detection, resulting in low maintenance efficiency and high radiation dose during nuclear power plant overhauls. 2. Existing numerical prediction methods use full-order CFD calculations, which take hundreds of core hours and are difficult to support rapid flow field prediction under different operating conditions. 3. If only a simplified flow field model is used, it is difficult to retain the real structural geometry information required for collision, and therefore it cannot be reliably applied to the control rod guide tube region with a complex three-dimensional spatial structure.

[0008] Given the aforementioned deficiencies in the existing technology, it is necessary to conduct research and resolve these deficiencies. Summary of the Invention

[0009] To address the aforementioned issues, this invention proposes a method for locating foreign objects in nuclear reactor internals based on a physics AI engine. This method aims to overcome the limitations of existing technologies, which lack the ability to predict the location of stuck foreign objects in advance and struggle to support rapid flow field prediction under different operating conditions. A porous media CFD model is used to generate multi-condition flow field snapshots and predict background flow, while a CAD-based realistic geometric model is used to characterize local wall positions, minimum diameters, and collision normals. This achieves a balance between computational efficiency across multiple operating conditions and the geometric realism required for local collision determination within the same computational system.

[0010] To achieve the above objectives, the technical solution adopted by this invention is as follows: A method for locating foreign objects in nuclear reactor internals based on a physics AI engine includes: an offline construction phase and an online tracking phase; the offline construction phase includes, S1. Construct a CFD model based on porous media to characterize the flow resistance of the reactor core, perform steady-state CFD model calculations to form a velocity field snapshot dataset, perform POD order reduction on the snapshot data and perform ANN training to establish the mapping relationship between operating parameters and modal coefficients. S2. Simultaneously construct a real CAD model, perform surface triangulation, normal extraction, local path expression, and spatial query organization on the real CAD model, and encapsulate it under a unified coordinate and unified calling interface; build a physical AI engine based on the CFD model and the real CAD model. The online tracking phase includes, S3. Call the physical AI engine, select representative working conditions, perform online flow field reconstruction on the representative working conditions, obtain a set of frozen background flow fields consistent with the CFD model, and perform Monte Carlo Lagrange particle tracking on each frozen background flow field in the set of frozen background flow fields. S4. Call the CAD real model to perform collision detection, speed response and jamming determination; count the jamming three-dimensional coordinates under each working condition and output the coordinates of the high probability jamming area.

[0011] In a preferred embodiment of the present invention, step S1 includes: Determine the parameters of the porous medium CFD model, which consist of porosity, viscous drag coefficient, and inertial drag coefficient, and characterize the porous medium domain; calculate the porosity and calibrate the viscous drag coefficient and inertial drag coefficient; set the solver for the CFD model.

[0012] In a preferred embodiment of the present invention, step S1 further includes: Select operating condition variables and sampling space, perform Latin hypercube sampling, set boundary conditions for the CFD model for the samples of the operating condition, run the steady-state CFD model to converge and extract the velocity field; calculate fluid properties.

[0013] In a preferred embodiment of the present invention, step S1 further includes: The snapshot data were subjected to snapshot intrinsic orthogonal decomposition using the Sirovich snapshot method to extract the spatial modal basis of the flow field.

[0014] As a preferred embodiment of the present invention, the step of performing snapshot intrinsic orthogonal decomposition on the snapshot data using the Sirovich snapshot method includes: constructing the covariance matrix C of the snapshot data; performing eigenvalue decomposition on the covariance matrix C to calculate the spatial mode basis vectors; truncating the spatial modes according to the energy contribution rate; and calculating the mode coefficients of the training samples.

[0015] In a preferred embodiment of the present invention, step S2 includes: constructing a multilayer perceptron neural network, the multilayer perceptron neural network consisting of an input layer, a hidden layer and an output layer; preparing training data and randomly dividing the training data; setting a loss function and an optimizer; and organizing and processing the geometric structure information of the CAD real model to form a unified data organization and calling interface.

[0016] As a preferred embodiment of the present invention, step S3, which involves calling the physical AI engine to select representative working conditions and reconstructing the flow field online for those conditions, includes: normalizing the input working condition parameters for each working condition, calling ANN forward inference to obtain modal coefficients, performing linear reconstruction, obtaining the reconstructed velocity field of the CFD model and organizing it into a nodal velocity vector field; recording the reconstructed velocity field as a frozen background flow field set; using spatial interpolation to obtain the fluid velocity at the particle's location for flow field spatial interpolation; and then performing Monte Carlo Lagrange particle tracking.

[0017] As a preferred embodiment of the present invention, the Monte Carlo Lagrange particle tracking includes: constructing a geometric model of a solid component of a nuclear reactor; setting the basic parameters of the foreign object; performing Monte Carlo random sampling of the initial position and initial velocity of the foreign object; solving the Lagrange equations of motion using RK4 integration in the frozen background flow field; and performing collision detection and velocity response using the restitution coefficient method.

[0018] In a preferred embodiment of the present invention, the jamming determination in step S4 includes, Rule 1: When the equivalent diameter of the foreign object is greater than the local channel diameter at the current location, it is determined to be a jam. Criterion 2: Calculate the oscillation standard deviation of the particle's centroid position over N consecutive time steps. When the oscillation standard deviation of the particle's centroid position is less than the convergence threshold r, it is determined that the foreign object is stuck in the region.

[0019] As a preferred embodiment of the present invention, the step S4 of statistically analyzing the three-dimensional coordinates of the jamming under each working condition further includes, when the jamming determination is met, recording the three-dimensional coordinates of the jamming, and recording the jamming type and the corresponding working condition; calculating the jamming probability, obtaining a set of high-probability jamming regions, and outputting the coordinates; and taking the union of the sets of high-probability jamming regions for each working condition.

[0020] The beneficial effects of this invention are: 1. Achieve unified access to millisecond-level online flow field reconstruction and real geometric collision information through a physics AI engine.

[0021] 2. By coordinating the work of CFD models and CAD models, the computational feasibility of batch snapshot generation and the geometric accuracy of collision determination are both taken into account.

[0022] 3. The reconstructed flow field and collision detection are completed in the same coordinate system, which can be applied to complex three-dimensional regions such as control rod guide tubes.

[0023] 4. It can output multiple stuck areas under various operating conditions for maintenance and troubleshooting. Attached Figure Description

[0024] Figure 1This is an overall flowchart of an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram illustrating the construction and coupling relationship of the physical AI engine in an embodiment of the present invention.

[0026] Figure 3 This is a schematic diagram of the three-dimensional geometric structure of the solid component in the lower chamber of a three-loop megawatt-class pressurized water reactor according to an embodiment of the present invention.

[0027] Figure 4 This is a schematic diagram of the probability distribution of foreign object entrapment in the lower chamber of a three-loop megawatt-class pressurized water reactor according to an embodiment of the present invention.

[0028] In the figure: 1-1, control rod guide tube; 1-2, core support plate; 1-3, upper vortex suppression plate; 1-4, lower vortex suppression plate; 1-5, secondary support structure. Detailed Implementation

[0029] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0030] like Figures 1-4 As shown, a method for locating foreign objects in nuclear reactor internals based on a physics AI engine includes: an offline construction phase and an online tracking phase; the offline construction phase includes, S1. Construct a CFD model based on porous media to characterize the flow resistance of the reactor core, perform steady-state CFD model calculations to form a velocity field snapshot dataset, perform POD order reduction on the snapshot data and perform ANN training to establish the mapping relationship between operating parameters and modal coefficients. S2. Simultaneously construct a real CAD model, perform surface triangulation, normal extraction, local path expression, and spatial query organization on the real CAD model, and encapsulate it under a unified coordinate and unified calling interface; build a physical AI engine based on the CFD model and the real CAD model. The online tracking phase includes, S3. Call the physical AI engine, select representative working conditions, perform online flow field reconstruction on the representative working conditions, obtain a set of frozen background flow fields consistent with the CFD model, and perform Monte Carlo Lagrange particle tracking on each frozen background flow field in the set of frozen background flow fields. S4. Call the CAD real model to perform collision detection, speed response and jamming determination; count the jamming three-dimensional coordinates under each working condition and output the coordinates of the high probability jamming area.

[0031] In Example 1, firstly, for the reactor internals of the target nuclear power plant, a porous media CFD model for flow field calculation and a CAD realistic model for collision detection are established. These two types of models are required to be built under the same equipment coordinate reference and aligned within the common spatial range involved in foreign object tracking. Offline CFD snapshots are generated based on the porous media CFD model. The physics AI engine reconstructs the background flow field on the velocity degrees of freedom of the CFD model, providing the background flow information required for velocity interpolation, drag, and added mass terms in the particle motion equations. Collision detection calls the geometric information organization module obtained from the CAD realistic geometric model, using the organized triangular facet geometry, wall normals, and local path information to process the collision response and jamming determination between foreign objects and walls. The physics AI engine forms a unified calculation system for background flow field reconstruction and real structural geometry calling under the same equipment coordinate reference. The porous media CFD model and the CAD realistic model for collision detection work collaboratively according to their respective roles in the foreign object localization task.

[0032] Among them, the porous medium CFD model treats the entire fuel assembly region of the reactor core as a porous medium continuum. Its drag parameters can be set separately in the axial and radial directions as needed, avoiding fine meshing of millions of fuel rods. This model covers the main flow path from the inlet section of the cold tube section to the outlet section of the core, and is responsible for the overall pressure drop, flow redistribution and main transport background characterization. It is used to generate multi-condition flow field snapshots and support subsequent flow field reconstruction. The CAD real model is also a CAD real geometric model. It is a real three-dimensional geometric model of the solid components in the lower chamber that have a significant impact on the analysis of foreign object movement trajectory (control rod guide tubes, core support plates and flow holes, vortex suppression plates, instrumentation columns, etc.) according to the reactor design CAD drawings. This model is responsible for the expression of collision boundaries, rebound normals and local minimum diameters, and is used for subsequent collision detection and local diameter determination.

[0033] Step S1 includes: determining the CFD model parameters for the porous medium. The CFD model parameters for the porous medium are fully characterized by three parameters: porosity ε (dimensionless), viscous resistance coefficient Dvisc=1 / K (unit: m). -2 ) and inertial drag coefficient C iner (Unit: m) -1 Porosity is defined as the volume fraction through which fluid can pass and characterizes the porous medium domain; the porosity is calculated, and the viscous drag coefficient and inertial drag coefficient are calibrated; the solver for the CFD model is set.

[0034] In calibrating the viscous drag coefficient Dvisc and the inertial drag coefficient C inerDuring the process, the axial (flow direction) flow resistance is described by the Darcy-Forchheimer equation. Furthermore, the CFD model is solved using one or more of ANSYS Fluent, STAR-CCM+, and OpenFOAM. The preferred governing equation is the steady-state incompressible Reynolds-averaged Navier-Stokes (RANS) equation, and the appropriate turbulence model is selected based on the specific operating conditions and mesh conditions.

[0035] When selecting operating variables and sampling space, a total loop volumetric flow rate Q and a cold pipe inlet temperature T are chosen. in As the operating condition variable, the system pressure is fixed at the rated value P=15.5Mpa. Using Latin hypercube sampling, M uniformly distributed sample points are generated in the two-dimensional operating condition space above. Latin hypercube sampling can achieve more uniform spatial coverage with fewer samples.

[0036] Boundary conditions of the CFD model are set for the sample of the working condition, the steady-state CFD model is run to convergence, the three-dimensional velocity vectors of all n grid nodes in the computational domain are extracted, and the fluid properties are calculated.

[0037] The snapshot data is subjected to snapshot intrinsic orthogonal decomposition using the Sirovich snapshot method to extract the spatial modal basis of the flow field. Specifically, this includes: constructing the covariance matrix C of the snapshot data; performing eigenvalue decomposition on the covariance matrix C to calculate the spatial modal basis vectors; truncating the spatial modes according to the energy contribution rate; and calculating the modal coefficients of the training samples.

[0038] A multilayer perceptron neural network is constructed, consisting of an input layer, hidden layers, and an output layer. In the input layer, the working condition parameter x has a dimension of 2. Before being fed into the network, it undergoes normalization to compress the input value to the [0,1] interval, with the normalization boundary conditions consistent with those described above. In the hidden layer, a multilayer fully connected structure is adopted, and the number of layers, the number of neurons per layer, and the activation function can be adjusted according to the training error and inference efficiency requirements. In the output layer, there are N POD modal coefficients with a dimension of N, and linear activation.

[0039] Training data is prepared and randomly divided into training data at an 8:2 ratio. The mean squared error (MSE) is set as the loss function, and the Adam optimizer is used. Network parameters θ are updated iteratively using mini-batch gradient descent. After each training iteration, the relative error of the velocity field reconstruction is calculated on the validation basis, and training convergence is achieved. The geometric information of the CAD real model is organized and processed to form a unified data organization and calling interface. After ANN training, the structural information of the CAD real geometric model is organized and processed. Specifically, the CAD model surface is triangulated, and the spatial position and normal information of each facet are extracted. Local path representation and spatial query structures are established for collision-sensitive locations such as the gap between control rod guide tubes, the drainage holes in the core support plate, and the adjacent area of ​​the vortex suppression plate. Subsequently, the processed CAD real geometric information is encapsulated into a geometric information organization module Mgeo, and integrated with the porous media CFD model, offline snapshot library U, POD spatial modal basis Ψ, and ANN mapping network f. θ The association is completed under the same device coordinate system, forming a unified data organization and calling interface.

[0040] When selecting representative operating conditions, J representative steady-state operating conditions are selected from the operating condition space Q. Online flow field reconstruction is performed on the representative operating conditions, including: normalizing the input operating condition parameters according to each operating condition point j, calling ANN forward inference to obtain modal coefficients, performing linear reconstruction, obtaining the reconstructed velocity field of the CFD model and organizing it into a nodal velocity vector field; the J reconstructed velocity fields are denoted as the frozen background flow field set, and the freezing is defined as: the flow field time derivative in the subsequent foreign object tracking stage. The spatial distribution of the flow field remains unchanged (steady-state assumption). This assumption applies when the characteristic size d of the foreign object remains constant. p Much smaller than the characteristic scale of the computational domain, under which the disturbance of the flow field by foreign objects is negligible. This set serves as the background flow field source for the subsequent particle tracking stage; collision boundary, wall normal, and local path information are provided by a geometric information organization module obtained from the processing of the CAD real geometric model.

[0041] During the tracking process, the position x of the foreign particle p Typically, the fluid velocity at the particle's location does not coincide with the CFD mesh nodes. Spatial interpolation is used to obtain the fluid velocity at the particle's location for flow field spatial interpolation; then Monte Carlo Lagrange particle tracking is performed. The specific interpolation method is as follows: based on x... p The CFD mesh element in question uses linear 3D interpolation (trilinear interpolation for hexahedral elements and barycentric coordinate interpolation for tetrahedral elements). The x-value is obtained by weighting the velocity values ​​of the vertex nodes within that element. pThe fluid velocity at a certain point is used to propel the overall transport of the foreign object; when the foreign object approaches the boundary of the structural component, it is further used in conjunction with the CAD collision geometry for subsequent collision response calculations.

[0042] A geometric model of the solid components of the nuclear reactor is constructed; the basic parameters of the foreign object are set, and the initial position and initial velocity of the foreign object are randomly sampled using Monte Carlo; the Lagrange equations of motion are solved using RK4 integral in the frozen background flow field; and the collision detection and restitution coefficient method are performed to determine the velocity response.

[0043] The determination of jamming includes, Rule 1: When the equivalent diameter of the foreign object is greater than the local channel diameter at the current location, it is determined to be a jam. Criterion 2: Within N consecutive time steps, calculate the standard deviation of the oscillation at the particle's center of mass position. When the standard deviation of the oscillation at the particle's center of mass position is less than the convergence threshold r, it is determined that the foreign object is stuck in the region. When the sticking determination is met, record the three-dimensional coordinates of the sticking, and record the sticking type and the corresponding working condition; calculate the sticking probability, obtain the set of high-probability sticking regions, and output the coordinates; take the union of the high-probability sticking region sets for each working condition.

[0044] In Example 2, the whole process includes an offline build phase and an online tracking phase.

[0045] Offline construction phase: A CFD model characterizing the core's main flow resistance using porous media is established, along with a CAD model for collision detection. Steady-state CFD calculations are performed under multiple operating conditions covering the normal operating space, generating a velocity field snapshot dataset. The snapshot data undergoes POD (Positive Displacement) reduction to extract the flow field spatial modal basis, and an ANN is trained to establish the mapping relationship between operating parameters and modal coefficients. Simultaneously, surface triangulation, normal extraction, local path representation, and spatial query organization are performed on the CAD model, and encapsulation is completed under a unified coordinate system and a unified calling interface. The resulting physical AI engine is a fast computation engine based on both CFD and CAD models, possessing both rapid flow field prediction capabilities and realistic structural geometric information.

[0046] Online tracking phase: The physical AI engine is invoked to reconstruct the flow field for representative operating conditions, forming a set of frozen background flow fields consistent with the CFD model; in each frozen background flow field, Monte Carlo random sampling is performed on the initial conditions of the foreign object, the Lagrange particle tracking method is used to solve the motion equation of the foreign object, and the geometric information organization module obtained by the CAD real geometric model is invoked to perform collision detection, velocity response and jamming determination; finally, the jamming three-dimensional coordinates under each operating condition are statistically analyzed and high-probability jamming areas are output.

[0047] The two stages described above are set around the same technical goal: the porous medium CFD model is responsible for generating multi-condition flow field snapshots and predicting background flow, while the CAD real geometric model is responsible for representing local wall positions, minimum diameters, and collision normals, thus balancing multi-condition computational efficiency and the geometric realism required for local collision determination within the same computational system.

[0048] The key technical points of this invention mainly include: 1. Structural constraints of the physics AI engine: The physics AI engine is defined as a "fast calculation engine based on CFD model + CAD model, which combines rapid flow field prediction and real structural geometric information." Internally, it includes a porous media CFD model and its offline snapshot library, POD spatial modal basis, ANN condition mapping, CAD geometric information organization, unified coordinate alignment, and online calling mechanism. Its output includes modal coefficients corresponding to the condition, reconstructed background flow field, and callable real structural geometric information; 2. Collaborative division of labor between the CFD model and the CAD model: The porous media CFD model is used to represent the overall pressure drop, flow distribution, and main transport, while the CAD real geometric model is used to represent local wall boundaries, normals, and minimum diameters; 3. Integrated calling of reconstructed flow field and collision tracking: The frozen background flow field output by the physics AI engine is directly used for Lagrange particle tracking and coupled with CAD real geometric collision detection in the same coordinate system, outputting high-probability jamming regions for multiple conditions.

[0049] Step 1: Building a Physics AI Engine Step 1.1: Establish the porous medium CFD model and the actual CAD geometric model; For the reactor internals of the target nuclear power plant, a porous media CFD model for flow field calculation and a CAD realistic geometric model for collision detection were established. Both models were built under the same equipment coordinate reference and aligned within the common spatial area involved in foreign object tracking.

[0050] (1) Model composition and division of labor This invention employs two types of models that work collaboratively according to their respective roles in the foreign object localization task: Porous media CFD model: The entire fuel assembly region of the reactor core is equivalent to a porous media continuum. Its drag parameters can be set separately in the axial and radial directions as needed, avoiding fine meshing of millions of fuel rods. This model covers the main flow path from the inlet section of the cold tube section to the outlet section of the core, and is responsible for the overall pressure drop, flow redistribution and main transport background characterization. It is used to generate multi-condition flow field snapshots and support subsequent flow field reconstruction. CAD realistic geometric model: Solid components in the lower chamber that significantly affect the analysis of foreign object movement trajectories (control rod guide tubes, core support plates and drainage holes, vortex suppression plates, instrumentation columns, etc.) are modeled using realistic 3D geometry based on the reactor design CAD drawings. Figure 3 As shown, this model represents the collision boundary, bounce normal, and local minimum path, and is used for subsequent collision detection and local path determination.

[0051] Taking a three-loop megawatt-class pressurized water reactor as an example: the solid components of the lower chamber include 61 control rod guide tubes 1-1 (outer diameter of about 250 mm, arranged according to the positions of 61 RCC assemblies), core support plate 1-2 (perforated circular plate, the number of water flow holes corresponds one-to-one with the number of fuel assemblies, each assembly corresponds to 4 water flow holes, totaling about 628), upper vortex suppression thin plate 1-3, lower vortex suppression thin plate 1-4, and secondary support structure 1-5.

[0052] (2) Determination of CFD model parameters for porous media The porous media domain is fully characterized by three parameters: porosity. (Dimensionless) Viscous resistance coefficient (Unit: m) -2 ) and inertial drag coefficient (unit: m) -1 These three parameters are determined through the following steps: Step ①: Calculate porosity

[0053] Porosity is defined as the volume fraction through which fluid can pass. For a 17×17 square fuel assembly, the cross-section of the assembly can be approximated as having a side length of... If the shape is square, the porosity can be estimated using the following formula:

[0054] in, : The total number of fuel rods in the fuel rod assembly (in a 17×17 arrangement, the actual number of fuel rods is 264, with the remaining 25 rods used for guide tubes and instrument tubes). Fuel rod outer diameter (unit: m), taken from reactor design drawings; Fuel rod pitch (the distance between the centers of two adjacent fuel rods, in meters), taken from reactor design drawings; : Approximate cross-sectional area of ​​a 17×17 square array component.

[0055] For a typical pressurized water reactor with a 17×17 arrangement, , Substituting into If the area occupied by guide tubes, instrument tubes and other internal structures is further taken into account, adjustments can be made based on this.

[0056] Step 2: Calibrate the viscous resistance coefficient and inertial drag coefficient

[0057] The flow resistance along the flow direction (axial direction) is described by the Darcy-Forchheimer equation:

[0058] in, The pressure difference (Pa) across the two ends of the computational domain; a positive value indicates the pressure drop along the flow direction. : Length of the computational domain along the flow direction (m); The fluid dynamic viscosity (Pa·s), obtained from the IAPWS-IF97 water property database under rated operating conditions (292°C, 15.5 MPa), is approximately... ; Fluid density (kg / m³), approximately under rated operating conditions ; Apparent flow velocity (m / s). When the total flow rate... When using the commonly used engineering unit m³ / h, first convert it to... (m³ / s), then ,in This represents the total cross-sectional area of ​​the reactor core (m²).

[0059] Calibration process: using at least 3 different total flow rates As a representative operating condition, the corresponding axial pressure drop is calculated by calling the sub-channel program (SUBCHANFLOW or COBRA-IV). ,get For the data pairs, the least squares method is used to fit the above equation to a value about... The quadratic polynomial is obtained by reading the coefficients of the linear terms. The coefficient of the quadratic term is obtained Then divide by the corresponding fluid property to obtain and If anisotropic treatment is required for the lateral (radial) drag coefficient, the above steps can be repeated for the representative lateral working condition.

[0060] (3) CFD solver settings CFD solutions can be obtained using software platforms such as ANSYS Fluent, STAR-CCM+, and OpenFOAM. The preferred governing equations are the steady-state incompressible Reynolds-averaged Navier-Stokes (RANS) equations. The turbulence model can be selected based on specific working conditions and mesh conditions, such as the standard Navier-Stokes (RANS) model. - Model with wall function ( The boundary conditions are set as follows: given mass flow rate at the inlet of the cold pipe section (when...). When expressed in m³ / h, first convert to ,but and uniform temperature Core outlet pressure given outlet boundary (rated pressure) All solid walls are subjected to no-slip boundary conditions. The convergence criterion can be set to the residuals of the principal variables being less than... (The residuals of the continuity equation can be appropriately relaxed to...) ).

[0061] (4) Coordinate alignment and calling relationship between CFD model and CAD model The porous media CFD model and the CAD real geometric model are established using the same device coordinate reference and aligned through a unified spatial coordinate system. Offline CFD snapshots are generated based on the porous media CFD model. During the online tracking phase, the physics AI engine reconstructs the background flow field on the velocity degrees of freedom of the CFD model, providing the background flow information required for velocity interpolation, drag, and added mass terms in the particle motion equations. Collision detection calls the geometric information organization module obtained from the CAD real geometric model, utilizing the organized triangular facet geometry, wall normals, and local path information to process the collision response and jamming determination between foreign objects and walls. Thus, the physics AI engine forms a unified computational system for background flow field reconstruction and real structural geometry invocation under the same device coordinate reference.

[0062] Step 1.2: Multi-condition steady-state velocity field sampling (1) Operating condition variables and sampling space In this embodiment, the total volumetric flow rate of the primary loop is selected. (Unit: m³ / h) and inlet temperature of cold pipe section (Unit: °C) is used as a variable for operating conditions; the system pressure is fixed at the rated value. (Since the system pressure fluctuations during normal nuclear power plant operation are relatively small, they can be considered constant and are not included in the operating condition space of this embodiment.) In other embodiments, other operating parameters may be further introduced as needed.

[0063] The sampling range covers the normal power regulation range and minor over-power range of nuclear power plants:

[0064] in This is the rated total flow rate (m³ / h, given in the nuclear power plant design documents).

[0065] (2) Latin hypercube sampling (LHS) In the aforementioned two-dimensional working space, Latin hypercube sampling is used to generate... A uniformly distributed sample point In one embodiment A value of 200 can be selected. Compared to completely random sampling, LHS can achieve more uniform spatial coverage with fewer samples. The specific steps are as follows: Will The sampling range is divided into Each interval has equal width. A value is randomly selected from each interval to form a sequence. Then, randomly shuffle the order; right Repeat the above steps to form a randomly permuted sequence. ; Pair the two sequences one by one to obtain One working condition sample point.

[0066] (3) CFD snapshot calculation and velocity field extraction For each working condition sample Set CFD boundary conditions (if If expressed in m³ / h, then first convert to Import mass flow rate The inlet temperature is Export pressure is Run steady-state CFD calculations until convergence, and extract all computational domains. The three-dimensional velocity vector of each grid node ( , , Three components).

[0067] The first The velocity fields of the next snapshot are arranged as column vectors:

[0068] The subscript indicates the node number ( The superscript indicates the operating condition number ( ).Will The column vectors are arranged horizontally to form a velocity snapshot matrix:

[0069] This represents the total number of grid nodes in the CFD computation model. After simplification using porous media, the number of model nodes is approximately in the millions (a reduction of 1 to 2 orders of magnitude compared to the hundreds of millions of nodes in similar full-order CFDs).

[0070] (4) Fluid property calculation Fluid density under various operating conditions and dynamic viscosity All calculations were performed according to the IAPWS-IF97 International Standard for Water and Water Vapor Properties, with temperature as the input parameter. and pressure .for , Approximate range of values: , .

[0071] Step 1.3, Snapshot POD order reduction – Extracting flow field spatial modal basis Snapshot matrix of velocity The specific steps for performing snapshot intrinsic orthogonal decomposition (Sirovich snapshot method) are as follows: Step 1: Construct the snapshot covariance matrix

[0072] The element Indicates the first and the The inner product (flow field similarity) between snapshots is used. The covariance matrix can be found in This significantly reduces the computational load.

[0073] Step 2: Eigenvalue decomposition For covariance matrix Solving the standard real symmetric eigenvalue problem:

[0074] get Eigenvalues ​​and eigenvectors ,in eigenvalues ( (A positive semi-definite matrix). Arranged in descending order of eigenvalues: Eigenvalues Indicates the first The flow field "energy" (mean square velocity component) carried by each mode.

[0075] Step 3: Calculate the spatial modal basis vectors By the eigenvectors Construct the corresponding spatial mode basis vectors :

[0076] in It is the Euclidean norm. It is a linear combination of all snapshots, with the physical meaning of the first snapshot. Step flow field spatial model (including) Each velocity component value represents the three-dimensional velocity direction distribution of each node. Constructing an orthogonal basis: (Kronecker symbol).

[0077] Step 4: Modal Truncation and Energy Criterion Before calculation Cumulative energy contribution rate of each mode:

[0078] Based on cumulative energy contribution rate The truncation criterion determines the number of modes to be retained. That is, select the one that satisfies The smallest integer In practice, the flow field inside the reactor is subject to boundary conditions ( , ) control, the flow field change has a low intrinsic dimension, usually (much smaller) ).

[0079] Will be retained The spatial modal basis vectors are arranged horizontally to form the spatial modal basis matrix:

[0080] Step 5: Calculate the modal coefficients of the training samples (ANN training labels) The first Projecting each snapshot onto the spatial modal basis yields its modal coefficient vector. :

[0081] in For the first The snapshot in the The projection amplitude in each modal direction. Group The input-output dataset that constitutes the training of an ANN.

[0082] Step 1.4: ANN Training, Geometric Information Organization and Unified Encapsulation – Forming a Physics AI Engine (1) Network structure Constructing a multilayer perceptron (MLP) neural network The structure is as follows: Input layer: Operating parameters The dimension is 2; normalization is performed before feeding it into the network: , Compress the input values ​​to Interval; normalization boundary is consistent with sampling range: , , , ; Hidden layers: A multi-layer fully connected structure can be used. The number of layers, the number of neurons per layer, and the activation function can be adjusted according to the training error and inference efficiency requirements. For example, 2 to 4 hidden layers can be used, with 64 to 128 neurons per layer, and ReLU or tanh activation functions can be used. Output layer: POD mode coefficients , dimension Linear activation (no activation function).

[0083] (2) Training data preparation and partitioning The result calculated in step ⑤ Group The dataset is randomly divided into a training set and a training set, with 80% and 20% of the data set being used as the training set. Groups) and validation sets ( Group).

[0084] (3) Loss function and optimizer The mean squared error (MSE) loss function is used:

[0085] in This is the network prediction value. This is the true value calculated from the snapshot projection. The Adam optimizer (initial learning rate) is used. , , Network parameters are updated iteratively using mini-batch gradient descent. .

[0086] (4) Training convergence criteria After each training epoch, the L2 relative error of the velocity field reconstruction is calculated on the validation set:

[0087] Among them, the reconstructed velocity field , for Vie Euclidean norm. When When training is deemed convergent, training is stopped and the model parameters are saved. .

[0088] (5) CAD geometric information organization and unified engine encapsulation After the ANN training is completed, the structural information of the CAD real geometric model is organized and processed. Specifically, the CAD model surface is triangulated, and the spatial position and normal information of each facet are extracted. Local path representation and spatial query structures are established for collision-sensitive locations such as the gaps between control rod guide tube bundles, the drainage holes in the core support plate, and the adjacent areas of the vortex suppression plate. Subsequently, the processed CAD real geometric information is encapsulated into a geometric information organization module. And with porous media CFD models and offline snapshot libraries POD spatial mode basis Mapping network with ANN The association is completed within the same device coordinate system, forming a unified data organization and calling interface. The mathematical representation of the physics AI engine is:

[0089] Among them, the online reconstruction operator Depend on and Together, they generate a reconstructed velocity field. It provides queryable structural boundary information; the trajectory of the foreign object, collision results, and jamming conclusions are completed by subsequent steps through the coupled invocation of both. The training phase information for the physics AI engine is as follows:

[0090] Step 2: Online flow field reconstruction and provision of frozen background flow field for collision tracing Step 2.1: Selection of representative operating points From the working space , Selected from Representative steady-state operating conditions Typically includes full power ( , ), 75% power ( ), 50% power ( Typical operating states include (e.g., ...). In one implementation, 3 to 5 are acceptable.

[0091] Step 2.2: Call the physics AI engine to reconstruct the flow field. For each operating point First, the input operating parameters are normalized. Then, the modal coefficients are obtained by calling the ANN forward inference. Finally, linear reconstruction is performed to obtain the reconstructed velocity field corresponding to the porous medium CFD model. And organize it into a nodal velocity vector field. .

[0092] Step 2.3: Freeze the background flow field set Will The reconstructed velocity field is denoted as the frozen background flow field set:

[0093] "Freezing" means that during the subsequent foreign object tracking phase, the time derivative of the flow field... The spatial distribution of the flow field remains unchanged (steady-state assumption). This assumption applies when the characteristic size of the foreign object is... Much smaller than the characteristic scale of the computational domain, under which the disturbance of the flow field by foreign objects is negligible. This set serves as the background flow field source for the subsequent particle tracking stage; collision boundary, wall normal, and local path information are provided by a geometric information organization module obtained from the processing of the CAD real geometric model.

[0094] Step 2.4, Spatial Interpolation of Flow Field The position of the foreign particle during tracking Typically, the particles do not coincide with CFD mesh nodes; spatial interpolation is required to obtain the fluid velocity at the particle's location. The interpolation method is as follows: based on The CFD mesh element in question uses linear 3D interpolation (trilinear interpolation for hexahedral elements and barycentric coordinate interpolation for tetrahedral elements), and the result is obtained by weighting the velocity values ​​of the vertex nodes of that element. The fluid velocity at that location. This local velocity is used to propel the overall transport of the foreign object; as the foreign object approaches the boundary of the structural component, it is further used in conjunction with the CAD collision geometry for subsequent collision response calculations. This interpolation process is embedded within the time integration loop of step three, at each time step. Executed in China.

[0095] Step 3: Monte Carlo Lagrange Particle Tracking Step 3.1, Geometric Model of Solid Components The surface geometry of solid components such as the inner and outer walls of the control rod guide tube, the upper and lower surfaces of the core support plate, the side of the water inlet, and the outer wall of the instrumentation pipe are extracted from the reactor design CAD model. Each surface is triangulated into a finite number of triangular facets to form a collision detection geometric model. ,in From the coordinates of three vertices and unit normal vector describe, This represents the total number of triangular faces. This geometric model corresponds to the actual CAD geometric model and is created once before tracking begins, for use in collision detection at subsequent time steps.

[0096] Step 3.2, Equations of Motion of the Foreign Object (1) Definition of basic parameters Before starting the tracking, first determine the basic parameters of the foreign object and the fluid:

[0097] (2) Equations of motion High-temperature and high-pressure water has a higher density. (Approximately 600 times the density of air), the added mass effect is not negligible. The velocity of the foreign matter core. The changes are controlled by the following formula:

[0098] In the formula Effective mass is the sum of the mass of the foreign object and the added mass. The mass derivative of fluid velocity degenerates into the following value in a frozen steady-state flow field: It is calculated numerically from the velocity gradient of the flow field.

[0099] Calculation methods for each force: Fluid drag (direction along relative velocity) ):

[0100] Relative velocity , Obtained by the aforementioned spatial interpolation.

[0101] drag coefficient The Schiller-Naumann correlation is used for piecewise calculations based on particle Reynolds number:

[0102] Combined buoyancy gravity (Vertical direction):

[0103] in Buoyancy is a unit vector pointing vertically upwards. ,gravity The sum of the two is For metallic foreign objects ( In high-pressure water ( During exercise, The resultant force is downward, with gravity dominating; for low-density aluminum foreign objects, the buoyancy effect is more significant.

[0104] Collision force When there is no collision When a collision occurs, it is handled by the collision detection module, which uses the velocity jump method to directly correct the velocity after the collision. The collision force only acts instantaneously at the moment the collision occurs.

[0105] (3) Fourth-order Runge-Kutta (RK4) time integral The equations of motion are rewritten as a system of first-order ordinary differential equations, with the particle position as the state variable. and speed :

[0106] In the time step Internally, calculated according to RK4 format:

[0107]

[0108]

[0109]

[0110] Displacement update introduces the corresponding displacement slope vector (i.e., the speed value of each sub-step):

[0111]

[0112] (4) Time step selection (CFL condition) Time step The CFL stability condition must be met:

[0113] in The smallest local feature scale that needs to be distinguished within the tracking area can be determined based on the size of the foreign object and the size of the local channel. Conservatively taking the peak flow rate at the core inlet (approximately) Safety factor In engineering practice, the time step that meets the stability requirements can be selected by combining the local geometric scale and the peak flow velocity, with a typical value on the sub-millisecond level.

[0114] Step 3.3, Collision Detection and Velocity Response (1) Collision detection At the end of each time step, it is checked whether the foreign object has collided with the solid component: using a geometric information organization module generated from the CAD real geometric model, the foreign object core is analyzed. With geometric model Perform a nearest neighbor query on the relevant triangular facets and calculate the signed distance to the nearest candidate facet; if the nearest distance is less than the equivalent radius of the foreign object... If a collision occurs, the position of the foreign object is corrected to the collision surface, and the unit normal vector of the collision surface is denoted as... (Towards the foreign object).

[0115] (2) Velocity decomposition Pre-collision velocity Decomposed into normal and tangential components:

[0116] (3) Recovery coefficient method for speed update Post-collision velocity is updated using the restitution coefficient method:

[0117] in: - Normal restitution coefficient, describing normal energy dissipation; under conditions of stainless steel foreign object and stainless steel wall surface, high temperature and high pressure water lubrication, one embodiment may take [value missing]. ; - Tangential restitution coefficient, describing tangential friction dissipation; in one embodiment, it can be taken as... - The above values ​​are empirical engineering values ​​for metal-metal water lubrication collisions, which can be adjusted according to the actual material and lubrication conditions.

[0118] After the collision is handled, Replace the current speed and continue with the next step of time integration.

[0119] Step 3.4, Monte Carlo Statistical Framework For each representative steady-state condition (common (number), execution Sub-independent tracking simulation; in one implementation, A value of 500 can be used. Each simulation is independent, and the initial conditions are randomly sampled as follows: (1) Initial position sampling initial position Foreign object entry area Uniform random sampling within: Represented by the center of the circle Center, radius Circular cross section ( The coordinates are fixed at the inlet section height. When sampling uniformly within a circle, it is necessary to take... To ensure uniform area distribution, angle exist If the top is uniformly randomly generated, then the initial position is... , , .

[0120] (2) Initial velocity sampling The initial velocity amplitude is sampled based on the uniform distribution of the local fluid characteristic kinetic energy, and the direction is consistent with the local fluid velocity direction: Calculate the fluid velocity at the initial position (Obtained by flow field interpolation); Equivalent volume of foreign object The corresponding local fluid characteristic kinetic energy is expressed as: (J); Randomly sampled initial kinetic energy: (in for (uniform random variable on) Converted to initial velocity amplitude: ; The initial velocity direction is the same as the local flow direction: .

[0121] (3) Single tracking process Each tracking from Starting from the designated point, the system uses the RK4 format for propulsion, performing collision detection and velocity response at each step, and checking the jamming criteria in real time until one of the following three termination conditions occurs: the jamming criteria are met (geometric jamming or oscillatory convergence jamming); or the total tracking time reaches the preset upper limit. The movement of the foreign object exceeds the computational domain.

[0122] Step 4: Determining the Jam and Outputting Coordinates Step 4.1, Criteria for Determining Jams At the end of each time step, the current particle position is... Real-time testing of the following two jamming criteria: Rule 1: When the equivalent diameter of the foreign object If the diameter exceeds the minimum local channel diameter at the current location, it is considered stuck.

[0123] in (m) is the minimum diameter of the local flow channel where the particle is currently located. It is given by the local diameter information pre-established by the geometric information organization module based on the CAD real geometric model. In the region where the local area is approximately constrained by opposing walls on both sides, it can be obtained from the distance between the nearest opposing walls.

[0124] Rule Two: In recent consecutive Within each time step, the oscillation standard deviation of the particle's center of mass position is calculated; in one embodiment, 500 can be selected, corresponding to the time window. :

[0125] in The average position (m) within the time window.

[0126] When the oscillation amplitude is below the convergence threshold When the object repeatedly collides in the area but is unable to escape, it is determined to be kineticly stuck.

[0127] Step 4.2: Record the card's coordinates. Record the simulation (number) at the moment when any criterion is met. , The three-dimensional coordinates of the stuck state:

[0128] Simultaneously record the jam type (Criterion 1 / Criterion 2) and the corresponding working condition number. .

[0129] Step 4.3: Probability Distribution Calculation and Coordinate Output exist After the simulation is completed, let The second time a jam occurred ( ), corresponding to the set of stuck coordinates .

[0130] The lower chamber space of the in-core components is divided into volumetric elements (three-dimensional mesh, the element side length can be determined according to...). (Select), count the number of stuck points in each unit, and calculate the local stuck probability:

[0131] Output satisfies The three-dimensional coordinate range of all volumetric units, i.e., the set of high-probability jamming regions, provides spatial guidance for foreign object inspection. In one embodiment, A percentage of 5% to 10% is acceptable.

[0132] Step 4.4, Multi-condition union For all The above process is completed for each representative operating condition. The high-probability jamming areas output by each operating condition are then combined and used as the priority areas for final maintenance and troubleshooting.

[0133] In this third embodiment, a method for locating foreign objects in nuclear reactor internals based on a physical AI engine is used as an application scenario for locating foreign objects stuck in the internals of a three-loop megawatt-class pressurized water reactor.

[0134] A certain three-loop 1,000 MW pressurized water reactor has a rated thermal power of approximately 2900 MWt. The primary loop adopts a three-loop design, with a rated total flow rate of... m³ / h (3 main pumps, approximately 22,500 m³ / h per loop), rated inlet temperature of cold pipe section The temperature of the heat pipe section is approximately 329°C, and the system pressure is... MPa; the core contains 157 fuel assemblies (17×17 square arrangement, active section height 3658 mm, equivalent diameter 3040 mm) and 61 control rod guide tubes.

[0135] Step 1 (a): Construction of porous media CFD model, preparation of CAD model and parameter calibration A porous media CFD model was established: the computational domain covers the area from the inlet section of the cold tube section to the core outlet section. In this embodiment, ANSYS Fluent was used to construct the porous media CFD model. The fuel assembly core region is equivalent to a porous medium, and the porous medium parameters are calculated according to the following formulas: Porosity: ,in (Actual number of fuel rods in 17×17). , Substituting into ; Drag coefficient calibration: Call the SUBCHANFLOW subchannel program to calibrate the drag coefficient. , , Axial pressure drop calculated under three representative working conditions ,right Perform least squares fitting under rated operating conditions ( , , , The subscript is determined , .

[0136] After simplification using porous media, the total number of nodes in the CFD calculation mesh is... Compared to similar heap types, full-order CFD is approximately The number of nodes is reduced by approximately two orders of magnitude. In this embodiment, the turbulence model adopts the standard... - The model, incorporating wall functions, has a convergence criterion of the residuals of all equations being less than [a certain value]. .

[0137] Establish a realistic CAD geometric model: The solid components such as the control rod guide tube, core support plate and its approximately 628 flow holes, vortex suppression plate, and instrumentation column are established as realistic three-dimensional geometric models according to the CAD drawings. Before subsequent collision calculations, these models are triangulated into a collision detection geometric model to provide information on local wall positions, normals, and minimum diameters.

[0138] In this embodiment, the core fuel assembly region is characterized by an overall pressure drop and flow distribution using a porous media CFD model; collision-sensitive areas such as control rod guide tubes and core support plate drainage holes are provided with local wall positions and minimum diameter information using a CAD realistic geometric model. This division of labor allows for the batch generation of multi-condition snapshots at an acceptable computational cost, while preserving necessary realistic structural boundaries for subsequent collision assessments.

[0139] The two types of models are aligned under the same device coordinate reference, and together with the subsequent POD order reduction, ANN mapping and online flow field reconstruction, they constitute the physical AI engine.

[0140] In this embodiment, before entering online tracking, the CAD real geometric model further completes triangulation, wall normal organization, and local path query organization to form a geometric information organization module obtained by the CAD model, which can be called for collision detection and jamming determination.

[0141] Latin hypercube sampling and CFD snapshot: In this embodiment, the sampling parameters are taken as follows: Operating conditions range ( ), For each sample working condition First, convert the volumetric flow rate to (m³ / s), and then set the inlet mass flow rate boundary conditions. (Fluid properties are found in IAPWS-IF97 tables within this temperature and pressure range) Run CFD until convergence, extract the global 3D velocity field, and press... Formatted storage. 200 snapshot column vectors are merged into a velocity snapshot matrix. ( A single CFD computation takes approximately 4 hours (64 cores in parallel).

[0142] Step 1(b): POD reduction and ANN training POD decomposition process: Step 1: Build covariance matrix ; Step 2: For Eigenvalue decomposition yields 200 eigenvalues. ; Step 3: From Calculate the basis vectors of each spatial mode ( ); Step 4: Calculate the cumulative energy contribution rate ,when hour Before truncation Each mode constitutes ; Step 5: Calculate the modal coefficients for each snapshot. 200 sets of training data were generated. (Input has been normalized to) ).

[0143] ANN Training: In this embodiment, the network structure consists of 2 input nodes, 2 hidden layers (ReLU activation) with 64 nodes each, and 18 output nodes, for a total of 3 fully connected layers. The training set contains 160 datasets, and the validation set contains 40 datasets; the Adam optimizer is used. , , After 2000 training iterations, the L2 relative error on the validation set decreased to 8.2%, satisfying the convergence criterion of no more than 10%. The training time was approximately 35 minutes. After training, the ANN was used to predict modal coefficients and, combined with the POD modal basis, to reconstruct the background flow field.

[0144] Step 2: Online flow field reconstruction to generate a frozen background flow field set. In this embodiment, the selected For each representative working condition, the frozen background flow field is generated sequentially through normalization, ANN forward inference, and linear reconstruction:

[0145] Reconstructing the velocity field under various operating conditions Composition of frozen background flow field set This set shares the same device coordinate system with the CAD real geometric model and the geometric information organization module obtained from its processing in step one (a). The subsequent step three calls this flow field and performs collision detection through this geometric information organization module.

[0146] Step 3: Monte Carlo Lagrange Particle Tracking Basic parameters of foreign objects:

[0147] Monte Carlo initial condition sampling: In this embodiment Next, the initial position is at the inlet section (equivalent to a circle) at the bottom of the lower chamber. ,area , Within (as the zero point of the coordinate system) Uniform sampling; initial velocity amplitude , The direction is consistent with the local fluid velocity.

[0148] kinematic equation parameters: fluid density Dynamic viscosity (100% power operation, IAPWS-IF97 table reference); Effective quality ; drag coefficient According to the Schiller-Naumann formula Calculation; take the vertically upward direction as positive. Then the combined buoyancy force The direction is downward, and gravity is dominant.

[0149] Time step: based on the local minimum feature scale Peak particle velocity Safety factor Estimate, get This embodiment takes Maximum duration of a single tracking session (Approximately 125,000 steps).

[0150] Collision response parameters: Both the foreign object and the wall are made of stainless steel. Under water lubrication conditions, the normal coefficient of restitution is used in this embodiment. Tangential coefficient of restitution (Engineering experience value).

[0151] Step 4: Determining the Jam and Outputting Results Rule 1 (Geometric hysteresis): At each step, the geometric information organization module queries the local path corresponding to the particle's current position. ,when Geometric jamming is determined when the diameter exceeds this limit. In this example, the diameter of the flow orifice is approximately... greater than Therefore, geometric jamming is not triggered; the minimum diameter at the gap of the control rod guide tube bundle can be as low as less than Therefore, this will trigger geometric jamming.

[0152] Criterion 2 (Oscillation Convergence Stuck): Time Window in this Example step( Convergence threshold Real-time calculation of the standard deviation of position within the sliding window. ,like Then the dynamics are determined to be stuck.

[0153] Under 100% power conditions, for The set of stagnation coordinates obtained from the simulation was statistically analyzed. (about Divide the element into a 3D mesh based on its side length, and calculate the values ​​of each element. Output The unit is considered a high-probability jamming region: geometric jamming region (criterion 1): concentrated in the gap of the control rod guide tube bundle ( Taking the inlet cross-section as ); Oscillation convergence stagnation region (criterion two): concentrated in the region adjacent to the vortex suppression plate ( Taking the inlet cross-section as ).

[0154] Take the union of the results from the three operating conditions This yields the final priority areas for investigation, which can serve as a reference range for subsequent video inspections.

[0155] like Figure 4 As shown, this invention employs the Monte Carlo simulation method to perform N independent particle trajectory simulations to simulate the random motion of foreign objects in a complex flow field. During the simulation, the system uses two strict "trapping criteria" to identify whether a particle is trapped: one is "geometric trapping," where a particle is considered physically trapped when its diameter is larger than the local flow channel diameter (e.g., a gap of only 5-10 mm between control rod guide tubes); the other is "oscillatory convergence trapping," where the standard deviation of the particle's position within a sliding time window is calculated. If the particle oscillates violently within a very small range (less than 6 mm) without moving with the main flow, it is considered trapped by the eddy current. After obtaining all the simulated trapping coordinates, the researchers divided the lower chamber space into a three-dimensional grid with sides of approximately 20 mm, statistically analyzed the frequency of trapping within each grid cell, calculated the trapping probability P, and finally selected cells with a trapping probability greater than 10%, rendering them as the black areas in the figure, thus providing a visual output of the trapping probability. Figure 4 The diagram shows the high probability of lag distribution.

[0156] from Figure 4 In terms of the effect shown, the calculated "high-risk areas" are displayed as a dark heatmap overlaid on a transparent model of the reactor's lower chamber structure. For example... Figure 4 As shown, two main black high-probability jamming areas are clearly presented: one is located in the area near the upper vortex suppression plate in the lower middle part, and the other is located in the gap of the control rod guide tube bundle further below (near the secondary support structure). This indicates that under simulated conditions, foreign objects are most likely to accumulate or jam at these two locations, while the jamming risk is lower in most other flow channel areas.

[0157] This invention relates to a method for locating foreign objects in nuclear reactor internals based on a physics AI engine. The method includes constructing a CFD model and a realistic CAD model; performing steady-state CFD calculations to form a snapshot dataset; performing POD (Programmable Observation Decomposition) reduction on the snapshot data; and training an ANN to establish a mapping relationship between operating parameters and modal coefficients. The CAD model is then processed for surface triangulation, normal extraction, local path representation, and spatial query organization, with unified coordinates and a unified calling interface used for encapsulation to construct a physics AI engine. The physics AI engine is then used to reconstruct the flow field, forming a frozen background flow field set. Monte Carlo Lagrange particle tracking is performed. The CAD model is then used for collision detection, velocity response, and jamming determination. The jamming 3D coordinates under various operating conditions are statistically analyzed, and high-probability jamming region coordinates are output. This method achieves millisecond-level online flow field reconstruction and unified access to realistic geometric collision information, balancing computational efficiency across multiple operating conditions with the geometric realism of local collision determination.

[0158] The beneficial effects of this invention include: 1. Achieve unified access to millisecond-level online flow field reconstruction and real geometric collision information through a physics AI engine. The physics AI engine encapsulates POD-ANN flow field reconstruction and CAD geometric information organization within a unified calling interface. In the online phase, single-condition flow field reconstruction requires only one ANN forward inference and one matrix-vector multiplication, significantly reducing the computation time compared to full-order CFD (which can take hours). This allows for batch Monte Carlo tracking calculations under multiple representative conditions. Simultaneously, the real CAD geometric information is organized into callable geometric information organization modules under the same coordinate datum, eliminating the need for simplified geometric assumptions in collision detection.

[0159] 2. By leveraging the collaborative division of labor between CFD and CAD models, the computational feasibility of batch snapshot generation is balanced with the geometric accuracy of collision detection. After the porous media CFD model is equivalent to a porous medium, the computational scale is reduced by 1 to 2 orders of magnitude compared to full-order CFD, making it feasible to generate hundreds of working condition snapshots in batches within actual engineering time. Meanwhile, the wall boundaries, normals, and minimum diameters of the collision-sensitive areas are still directly provided by the actual CAD geometric model, without losing the geometric basis for collision determination due to the simplification of the flow field.

[0160] 3. The reconstructed flow field and collision detection are completed in the same coordinate system, which is applicable to complex three-dimensional regions such as control rod guide tubes. The frozen background flow field output in the online stage shares the same coordinate system with the CAD collision detection geometry, avoiding coordinate transformation errors when stitching multiple models. Local velocity interpolation, collision response, and jamming determination in particle motion are all completed under the same device coordinate reference, thus enabling this method to be applied to control rod guide tube regions with vertical circular tube arrays, complex flow hole geometry, and multiple three-dimensional collision interfaces, rather than being limited to simple geometries such as planar tube sheets.

[0161] 4. Can output multi-condition stuck areas for maintenance and troubleshooting. After conducting tracking calculations under multiple representative steady-state operating conditions, the union output of the high-probability stuck areas under each operating condition can be directly used to guide the priority inspection scope during nuclear power plant overhauls, reducing the workload of full-area video flaw detection and radiation exposure.

[0162] In summary, this embodiment fully verifies that the present invention has the significant effect of achieving millisecond-level online flow field reconstruction and unified access to real geometric collision information, while addressing the lack of pre-physical prediction capability for foreign object jamming in foreign object handling schemes, and balancing multi-condition calculation efficiency with geometric realism of local collision determination.

[0163] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention; therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein, and the data collection in this invention is carried out without violating laws, social ethics, or harming public interests.

Claims

1. A method for locating foreign objects in nuclear reactor internals based on a physics AI engine, characterized in that, include: The offline build phase and the online tracking phase; the offline build phase includes, S1. Construct a CFD model based on porous media to characterize the flow resistance of the reactor core, perform steady-state CFD model calculations to form a velocity field snapshot dataset, perform POD order reduction on the snapshot data and perform ANN training to establish the mapping relationship between operating parameters and modal coefficients. S2. Simultaneously construct a real CAD model, perform surface triangulation, normal extraction, local path expression, and spatial query organization on the real CAD model, and encapsulate it under a unified coordinate and calling interface. A physical AI engine is constructed based on the CFD model and the CAD real model; the CAD real model is a model of solid components that affect the movement of foreign objects in the lower chamber, including one or more of the following: control rod guide tube, core support plate and water flow hole, vortex suppression plate, and instrument column. The online tracking phase includes, S3. Call the physical AI engine, select representative working conditions, perform online flow field reconstruction on the representative working conditions, obtain a set of frozen background flow fields consistent with the CFD model, and perform Monte Carlo Lagrange particle tracking on each frozen background flow field in the set of frozen background flow fields. S4. Call the actual CAD model to perform collision detection, speed response and jamming judgment; The three-dimensional coordinates of the jamming under each of the aforementioned working conditions are statistically analyzed, and the coordinates of the high-probability jamming area are output.

2. The method for locating foreign objects in nuclear reactor internals based on a physics AI engine according to claim 1, characterized in that, Step S1 includes: determining the parameters of the porous medium CFD model, which consists of porosity, viscous drag coefficient and inertial drag coefficient, and characterizes the porous medium domain; calculating the porosity and calibrating the viscous drag coefficient and inertial drag coefficient; and setting the solver of the CFD model.

3. The method for locating foreign objects in nuclear reactor internals based on a physics AI engine according to claim 1, characterized in that, Step S1 further includes: selecting operating condition variables and sampling space, performing Latin hypercube sampling, setting boundary conditions for the CFD model for the samples of the operating condition, running the steady-state CFD model to convergence and extracting the velocity field; and calculating fluid properties.

4. The method for locating foreign objects in nuclear reactor internals based on a physics AI engine according to claim 1, characterized in that, Step S1 further includes: using the Sirovich snapshot method to perform snapshot intrinsic orthogonal decomposition on the snapshot data and extracting the flow field spatial mode basis.

5. The method for locating foreign objects in nuclear reactor internals based on a physics AI engine, as described in claim 4, is characterized in that... The step of performing snapshot intrinsic orthogonal decomposition on the snapshot data using the Sirovich snapshot method includes: constructing the covariance matrix C of the snapshot data; performing eigenvalue decomposition on the covariance matrix C to calculate the spatial mode basis vectors; and truncating the spatial modes according to the energy contribution rate. Calculate the modal coefficients of the training samples.

6. The method for locating foreign objects in nuclear reactor internals based on a physics AI engine according to claim 1, characterized in that, Step S2 includes: constructing a multilayer perceptron neural network, which consists of an input layer, a hidden layer, and an output layer; preparing training data and randomly dividing the training data; setting a loss function and an optimizer; and organizing and processing the geometric structure information of the CAD real model to form a unified data organization and calling interface.

7. The method for locating foreign objects in nuclear reactor internals based on a physics AI engine according to claim 1, characterized in that, Step S3, which involves calling the physics AI engine to select representative working conditions and reconstructing the flow field online, includes: normalizing the input working condition parameters for each working condition, calling ANN forward inference to obtain modal coefficients, performing linear reconstruction, obtaining the reconstructed velocity field of the CFD model and organizing it into a nodal velocity vector field; recording the reconstructed velocity field as a frozen background flow field set; using spatial interpolation to obtain the fluid velocity at the particle location for flow field spatial interpolation; and then performing Monte Carlo Lagrange particle tracking.

8. A method for locating foreign objects in nuclear reactor internals based on a physics AI engine, as described in claim 1 or 7, characterized in that, The Monte Carlo Lagrange particle tracking process includes: calling the CAD real model in the physical AI engine; setting the basic parameters of the foreign object; performing Monte Carlo random sampling of the initial position and initial velocity of the foreign object; solving the Lagrange equations of motion using RK4 integration in the frozen background flow field; and performing collision detection and velocity response using the restitution coefficient method.

9. A method for locating foreign objects in nuclear reactor internals based on a physics AI engine, as described in claim 1, is characterized in that... The jamming determination in step S4 includes... Rule 1: When the equivalent diameter of the foreign object is greater than the local channel diameter at the current location, it is determined to be a jam. Criterion 2: Calculate the oscillation standard deviation of the particle's centroid position over N consecutive time steps. When the oscillation standard deviation of the particle's centroid position is less than the convergence threshold r, it is determined that the foreign object is stuck in the region.

10. The method for locating foreign objects in nuclear reactor internals based on a physics AI engine according to claim 1, characterized in that, The step S4 of statistically analyzing the three-dimensional coordinates of the jamming under each working condition also includes recording the three-dimensional coordinates of the jamming when the jamming determination is met, and recording the jamming type and the corresponding working condition. Calculate the jamming probability, obtain the set of high-probability jamming regions, and output their coordinates; take the union of the high-probability jamming region sets for each of the stated working conditions.

Images