Automatic optimization method based on CAD model

Abstract

The invention discloses an automatic optimization method based on a CAD model, and relates to the technical field of engineering design automation and computer aided design optimization. The method does not depend on manual predefined parameter dimensions any more, directly takes topology and geometric features of the CATIA model as core input, constructs a complete geometric optimization link through automatic extraction, structure mapping and optimization algorithm cooperative processing, ensures synchronous evolution of geometric features and design targets in the optimization iteration process, and improves the optimization efficiency of the CATIA model. And the optimization efficiency and stability are obviously improved. A reinforcement learning or evolutionary algorithm is built in the system to perform efficient search and multi-round adaptive iteration on geometric feature vectors, and global optimization of a complex structure can be completed without manual intervention. According to the method, automatic model generation of an optimization result is realized through a CATIA API and a macro command interface, and manual geometry reconstruction by an engineer is not needed. CATIA can be written back through one key according to optimized feature information, the closed-loop automatic process of input, optimization and output is achieved, the design period is remarkably shortened, and the iteration response efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of engineering design automation and computer-aided design (CAD) optimization technology, specifically to an automatic optimization method based on CAD models, which can be widely applied to structural optimization and performance improvement in fields such as aerospace, intelligent manufacturing, automotive design, and consumer electronics. Background Technology

[0002] In modern engineering design and product development, 3D modeling software, as a core tool for complex structural design, assembly management, and parametric control, has become a crucial foundation in the industrial design process. Among them, CATIA (Computer Aided Three-dimensional Interactive Application), as a leading international high-end CAD / CAE / CAM integrated platform, is widely used in aerospace, automotive manufacturing, shipbuilding, electronic equipment, and flexible structures. Design engineers typically use the CATIA platform to build 3D geometric models, define component structures, constraints, and assembly features, and combine this with parametric modeling techniques to achieve multi-version, multi-condition product structure design and verification.

[0003] However, most mainstream design optimization methods are based on parameter-driven automatic optimization, which involves pre-selecting several geometric or process parameters (such as dimensions, angles, positions, and thicknesses) and using external optimization algorithms to iteratively adjust these parameters to obtain the optimal design solution in terms of performance or objective function. This approach is effective in scenarios with simple structures and clear parameter dependencies, but it reveals significant limitations in complex models and multi-constraint design tasks: First, it requires engineers to invest considerable manpower in parameter definition and variable constraints in the early stages; second, when the design involves curved surfaces, assemblies, or nonlinear interconnected features, the parameter space cannot fully cover geometric changes; and third, traditional parameter optimization cannot directly address the geometric features themselves, leading to discrepancies between the optimization results and the actual manufacturable model, reducing design iteration efficiency and engineering reliability.

[0004] As design tasks become increasingly complex, the relationship between product geometry and performance exhibits highly coupled and multidimensional nonlinear characteristics. Traditional parameter-performance mapping models are inefficient in handling high-degree-of-freedom geometric optimization and struggle to meet the engineering demands for rapid design iteration and high-precision optimization. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes an automated optimization method based on CAD 3D models. This method no longer relies on manually predefined parameter dimensions, but instead directly uses the geometry and topology of the CAD model as its foundation. It combines feature extraction, intelligent optimization algorithms, and automatic model rewriting technology to achieve efficient optimization and automatic model generation for complex structures. This approach significantly reduces parameter setting and maintenance costs, improves the optimization accuracy and automation level of complex structures, and provides a more efficient and universal optimization path for engineering design.

[0006] The specific technical solution of the present invention is as follows:

[0007] An automatic optimization method based on CAD models includes the following steps:

[0008] Step 1: Perform geometric feature extraction and structural description modeling on the CATIA 3D model of the target structure, including feature extraction of control points, lines, surfaces, topology, and geometric attribute information to obtain the geometric features of the control point space; and filter and partition the control points based on geometric and working condition sensitivity.

[0009] Step 2: For the set of control points selected for optimization, perform geometric perturbation on the geometric features of the control points, and then optimize through global search and local correction. After optimization, output the optimal or near-optimal geometric perturbation result that satisfies the optimization objective and constraints.

[0010] Step 3: After each round of geometric perturbation, the structural performance is quickly predicted using a pre-trained AI proxy model. The fitness is calculated based on the objective function, and the top K solutions with the best prediction performance are retained for real simulation verification.

[0011] Step 4: Simulation Verification and Closed-Loop Iteration; Execute the simulation process for the selected optimal candidate solution, with simulation tasks for multiple candidate solutions running in parallel; Set convergence criteria, and determine the convergence of fitness, optimization variable perturbation, and response error for the simulation results. Satisfying any two convergence conditions is considered as reaching a stable convergence state; If convergence is not achieved, add the simulation result samples to the AI ​​proxy model training set, readjust the perturbation step size and search direction, and enter the next round of optimization iteration.

[0012] Furthermore, the geometric features include control points. and based on control points boundary curve curved surfaces curvature Normal vector Local thickness ;

[0013] Furthermore, the selection and partitioning of control points are specifically as follows: The set of control points within the high curvature region and the stress-sensitive region is selected as the set of principal optimization variables. The set of control points in low-sensitivity areas is used as the frozen set. The curvature field and the principal direction of the surface are used to provide directional priors for the disturbance: the maximum and minimum principal curvature directions corresponding to the local principal directions of the surface are uniformized and used as the regional directional basis. The displacement increments generated by subsequent optimization are weighted and decomposed on this directional basis.

[0014] Furthermore, the geometric perturbation is specifically as follows:

[0015] Suppose the three-dimensional geometric model has N control points Composition, for the set of main optimization variables Its position after the disturbance is defined as:

[0016] , ;

[0017]

[0018] in, For the control point displacement increment, the frozen set ,make ; , , The corresponding disturbance coefficient, The principal direction of the surface corresponds to the direction of maximum principal curvature. The principal direction of the surface corresponds to the direction of minimum principal curvature; disturbance amplitude A round-by-round decay mechanism is adopted:

[0019] ;

[0020] in, For the first The threshold parameter at the next iteration The attenuation coefficient is represented by the superscript t, which indicates the number of iterations.

[0021] Apply feasible region constraints: ,in, To control points Related disturbance limits;

[0022] Geometric continuity and compliance constraints: ,in, It is the smoothness coefficient;

[0023] Curvature change rate constraint: ,in, For the first disturbance Curvature values ​​at each control point or its neighborhood corresponding surface sampling point; This refers to the curvature value after disturbance calculated at the same location or corresponding neighborhood after applying displacement increment to the control point; This is the threshold for curvature change.

[0024] Furthermore, the optimization through global search and local correction is as follows:

[0025] The global search uses a genetic algorithm or particle swarm optimization algorithm to generate a set of perturbation candidate solutions in the control point space X; the local correction uses gradient-based or Bayesian optimization strategies to fine-tune excellent individuals and narrow the perturbation range.

[0026] Calculate the fitness function:

[0027] ;

[0028] in: Geometric features The overall fitness value, , , These are the weighting coefficients corresponding to the stress index, weight index, and deformation index, respectively, and satisfy the following conditions: , , ; These are stress evaluation items related to structural strength; Weight evaluation items related to lightweighting; Deformation evaluation items related to stiffness / deformation;

[0029] When the maximum number of iterations is reached or the change in the fitness function meets the convergence threshold, the optimization stops and the optimal or near-optimal geometric design results that satisfy the optimization objective and constraints are output, including the corresponding control point displacement distribution, the updated three-dimensional model, and the simulation evaluation results of each performance index.

[0030] Furthermore, step 3 is detailed as follows:

[0031] Step 3.1: Geometric Disturbance and Response Feature Representation; The input features of the geometric disturbance include the control point displacement increment. Adjacency topology matrix changes Local curvature changes Changes in normal direction Surface principal direction , And geometric scalar features, the above information is uniformly vectorized as Feature extraction is performed.

[0032] ;

[0033] Where z is the design change embedding vector; It is the feature extraction function, where d is the feature dimension;

[0034] Step 3.2: The AI ​​agent model is used to... Predict the corresponding structural performance under these circumstances. Authentic response Based on finite element simulation, the training objective function is mean square error:

[0035] ;

[0036] in, Let the mean squared error loss function be . The sample size is used. In the early stages, the surrogate model is trained offline using an initial simulation sample library; during the optimization process, accuracy is gradually improved through active sampling and incremental training.

[0037] Step 3.3: Perturb the candidate solution set Rapidly predict responses using AI agent models And calculate the fitness based on the objective function:

[0038] ;

[0039] in, This represents the set of control point displacement increments corresponding to the m-th candidate scheme; For the first The weight coefficients corresponding to each evaluation indicator; For the first Each evaluation indicator corresponds to an objective function; To address candidate perturbation schemes The performance response prediction values ​​obtained from the proxy model or simulation model are used to select the top K solutions with the best predicted performance for real simulation verification.

[0040] Furthermore, step 4 is detailed as follows:

[0041] Step 4.1: Perform high-precision finite element simulation or computational fluid dynamics simulation on the selected optimal candidate solution. The simulation tasks of multiple candidate solutions are distributed to the multi-core computing cluster by the scheduler.

[0042] Step 4.2: Closed-loop control and convergence criteria;

[0043] Objective function convergence: The change in the fitness function during iteration is less than or equal to the set convergence threshold;

[0044] Optimization variable perturbation convergence: During iteration, the change in the displacement increment of the control point is less than or equal to the perturbation convergence threshold;

[0045] Response error convergence: The calculated L2 norm of the performance response prediction value and the true response value is less than or equal to the error convergence threshold;

[0046] If any two convergence conditions are met, it is considered to have reached a stable convergence state. If it has not converged, the simulation result samples are added to the AI ​​agent model training set, the perturbation step size and search direction are readjusted, and the next round of optimization iteration is entered.

[0047] Step 4.3: In the closed-loop iteration, an adaptive perturbation shrinkage and surrogate model incremental update mechanism are adopted:

[0048] The perturbation contraction gradually reduces the amplitude of the geometric perturbation to ensure the stability of the final solution.

[0049]

[0050] Incremental learning of proxy models:

[0051]

[0052] in: For the first The parameter set of the AI ​​agent model at the next iteration; This is the set of parameters for the AI ​​agent model updated based on the new samples. The learning rate; For AI agent model parameters The gradient of the loss function is calculated from the loss function corresponding to the new sample. The loss function is constructed based on the newly added training data; the newly added simulation result samples are used to continuously train the surrogate model.

[0053] Implement uncertainty-driven sampling: prioritize simulation of candidate geometries with high prediction uncertainty to avoid convergence and getting trapped in local extrema.

[0054] The beneficial effects of this invention are as follows:

[0055] This invention employs a geometry-driven CATIA model optimization mechanism. Instead of relying on manually predefined parameter dimensions, it directly uses the topology and geometric features of the CATIA model as core input. Through automatic extraction, structural mapping, and collaborative processing with optimization algorithms, a complete geometric optimization chain is constructed. Geometric structures (surfaces, skeletons, assembly relationships) and optimization objectives (mass, stiffness, shape errors, etc.) are uniformly and standardized within the algorithm engine, ensuring the synchronous evolution of geometric features and design objectives during optimization iterations, significantly improving optimization efficiency and stability.

[0056] A multi-objective optimization mechanism based on adaptive evolution: The system incorporates reinforcement learning or evolutionary algorithms to efficiently search and iterate through multiple rounds of geometric feature vectors, achieving global optimization of complex structures without manual intervention. This mechanism effectively overcomes the limitations of traditional parameter optimization in high-dimensional geometric spaces, improving the convergence speed and reliability of the global optimal solution.

[0057] An automated modeling mechanism based on CATIA model write-back: This invention achieves automatic model generation of optimization results through the CATIA API and macro command interface, eliminating the need for engineers to manually reconstruct the geometry. The optimized feature information can be written back to CATIA with a single click, realizing a closed-loop automated process of "input model—optimization calculation—output optimized model," significantly shortening the design cycle and improving iterative response efficiency.

[0058] Design assurance mechanism based on geometric and performance consistency: The system simultaneously constrains geometric features and target performance during the optimization process, realizing real-time consistency control of geometric shape, assembly logic and structural performance, avoiding mismatch between optimization results and manufacturable models, and improving design reliability and engineering usability. Attached Figure Description

[0059] Figure 1 This is a schematic diagram of the overall structure of the method of the present invention.

[0060] Figure 2 This is a flowchart of the geometric perturbation and search strategy. Detailed Implementation

[0061] This invention proposes a direct geometric model optimization closed-loop method for complex structures. This method does not rely on predefined parametric control, but directly optimizes and adjusts the geometric features of the 3D CAD model (such as CATIA's CATPart file). Figure 1 As shown, it mainly includes a geometric topology information extraction module, a geometric feature modification and search module, an AI-assisted response prediction and optimization module, a CATIA model write-back and update module, and a simulation verification and convergence control module. It is suitable for complex engineering structures such as flexible structural components, hinged rotating parts, and aerodynamic / mechanical load-sensitive surfaces, as detailed below:

[0062] Step 1: Extract geometric features and perform structural description modeling of the target 3D structural model based on the geometric topology information extraction module. The purpose of this step is to establish a computable, perturbable, and mappable geometric information representation without introducing explicit parameterization, laying the foundation for subsequent geometric search, simulation prediction, and closed-loop optimization.

[0063] Step 1.1: Geometric Information Extraction Framework; For the CATIA 3D model (.CATPart file) of the target structure, this invention first extracts the following types of information through the CATIA API / COM interface:

[0064] Point information: Coordinates of vertices or control points in a 3D model. ,in Let be the coordinates of the X, Y, and Z axes in the 3D model, and i be the index of the point. The number of points.

[0065] Edge information: consisting of two points , The resulting boundary line segment or curve is represented as: ,in In CATIA 3D model, the first... The face and the first Geometric objects of the edges between faces; Indicates and The corresponding parameterized space curve function; These are the curve parameter variables for the edge curve.

[0066] Surface information: A surface defined by several control points and basis functions:

[0067]

[0068] in, Indicates in parameter pair The first one obtained Coordinates of a spatial point on a surface instance or surface patch; and Surface parameter variables are defined in the parameter space respectively, and are used to locate points on the surface using various surface parameters; Indicates the first Okay, number The three-dimensional spatial coordinate vector of the control points of the NURBS surface in the column; To control points The corresponding weighting coefficients; and These represent the surfaces in the parametric directions, respectively. With parameter direction The order of the B-spline basis functions on; Indicates the direction of the parameter Above, the first Each control point corresponds to B-spline basis functions; Indicates the direction of the parameter Above, the first Each control point corresponds to B-spline basis functions; and These are the index numbers of the control points in the two parameter directions, respectively; and The sub-tables represent the surface in the parametric direction. With parameter direction The number of control points on the device is reduced by one.

[0069] Topological structure information: The connection relationships between points, edges, and faces form a topological adjacency matrix A used to characterize the geometric topological relationships in CAD, and ; This indicates that the size of the adjacency matrix is A real matrix; Adjacency matrix The Line 1 Column elements, used to represent the first column. The point and the first Topological connections between points, where, when When, it indicates the first The point and the first When there are adjacent / connected points, when When, it indicates the first The point and the first The points do not have any adjacent / connected relationships.

[0070] Geometric attribute information: including normal vector curvature Local thickness The principal direction of the surface corresponds to the direction of maximum principal curvature. The principal directions of the surface correspond to the directions of minimum principal curvature. These are used to establish a mapping between geometric perturbations and mechanical responses.

[0071] Step 1.2: Mathematical representation of geometric information; To achieve computational and optimized search of geometry, this invention encodes the above information into a set of vectorized geometric features. :

[0072]

[0073] Where: control point coordinates It is an optimization variable, a boundary curve. Surfaces are used to ensure geometric continuity during disturbances. In response to the computational carrier, curvature With normal vector Used to constrain the range of geometric disturbances, prevent shape distortion, and local thickness. This serves as a reference for local thickness constraints and manufacturing tolerances. In this way, the model is transformed from a graphic file into a mathematically operable object that can be directly searched and modified by optimization algorithms.

[0074] Step 1.3: Correlation Mapping of Geometric and Mechanical Responses; To establish a learnable mapping relationship between geometric perturbations and simulation responses (such as stress and deformation), this invention further defines the relationship between local geometric features and response features:

[0075]

[0076] in: Mapping the responses obtained from finite element simulations (e.g., Mises stress field, deformation field, mass, stiffness, etc.) This mapping can be obtained through AI proxy models or simulation calculations. This mapping provides physical support for subsequent geometry-based optimization searches and rapid AI predictions.

[0077] Step 1.4: Feature Selection and Optimization Dimension Control; Since optimizing all control points simultaneously would lead to excessively high variable dimensionality and low search efficiency, this invention selects and partitions control points based on geometric and working condition sensitivity: the set of control points in high curvature regions and stress-sensitive regions is selected as the set of main optimization variables. Set up control points in low-sensitivity areas As a frozen set, it remains unchanged during optimization iterations.

[0078] Regarding the selection of direction, this invention utilizes the curvature field and the principal direction of the surface to provide a priori direction for the disturbance: for the local principal direction of the surface , After being standardized, the displacement increments generated by subsequent optimization algorithms are weighted and decomposed on this directional basis to make the disturbance more consistent with the surface geometry and maintain the continuity of the surface.

[0079] By employing the aforementioned strategy of "filtering + freezing + directional weighting," the optimization dimensionality is significantly reduced without sacrificing the key shape degrees of freedom, thereby improving search efficiency and convergence stability.

[0080] Step 2: Geometric Perturbation and Search Strategy; The optimization process of this invention does not rely on parameterized variables, but directly performs perturbation and optimization on the control point / surface geometry. To ensure the continuity of the geometric shape, the predictability of the physical response, and the convergence of the computation process, this step defines a complete geometric perturbation model and search control strategy, such as... Figure 2 As shown, the details are as follows:

[0081] Step 2.1: Definition of the perturbation space; Assume the three-dimensional geometric model consists of N control points. Composition, the set of control points involved in optimization Its position after the disturbance is defined as:

[0082] ,

[0083] in, The control point displacement increments (decision variables) generated for the optimization algorithm. For the frozen set. ,make .

[0084] To prevent shape collapse, local distortion, or violation of manufacturing constraints, this invention imposes a feasible region constraint on the displacement increment:

[0085]

[0086] in, To control points The relevant perturbation upper limit is determined by the local curvature, boundary characteristics and manufacturing tolerance, and is used to limit the geometric change range of a single iteration, thereby ensuring geometric effectiveness and engineering manufacturability in the optimization process.

[0087] Step 2.2: Disturbance direction and step size control; In order to control the disturbance direction and make it conform to the intrinsic characteristics of the structural geometry, this invention defines:

[0088]

[0089] in, , , The corresponding disturbance coefficient; disturbance amplitude A round-by-round decay mechanism is adopted:

[0090]

[0091] in, For the first The threshold parameter at the next iteration The value represents the decay coefficient, and the superscript t indicates the number of iterations. This is equivalent to performing a large search in the early stages and refining the process during convergence in the later stages, thus improving global convergence.

[0092] Step 2.3: Geometric Continuity and Smoothness Constraints; Directly perturbing the control points can easily lead to problems such as creases, discontinuities, or manufacturing infeasibility on the surface. This invention constrains the smoothness of the perturbation through the relationship between the curvature field and adjacent points:

[0093]

[0094] This is the smoothness coefficient, which is usually taken as 0.05 to 0.2.

[0095] Furthermore, the curvature change rate constraint ensures that the curvature gradient is not too large after perturbation of adjacent surfaces, thus maintaining the surface quality that is machinable in engineering:

[0096]

[0097] in, For the first disturbance The curvature value at each control point (or its corresponding neighborhood surface sampling point); The curvature value after disturbance is calculated at the same location (or corresponding neighborhood) after applying displacement increment to the control point; This is the threshold for curvature variation (the upper limit allowed).

[0098] Step 2.4: Optimize the search strategy; This invention adopts a search mechanism that combines global and local approaches to balance search efficiency and accuracy.

[0099] Global search uses either a genetic algorithm (GA) or a particle swarm optimization (PSO) algorithm to generate a set of perturbation candidate solutions in the control point space X:

[0100]

[0101] GA encoding for perturbation coefficients combination;

[0102] Local correction uses gradient-based or Bayesian optimization strategies to fine-tune excellent individuals, reduce the perturbation range, and improve convergence speed.

[0103] Fitness function:

[0104]

[0105] in: For candidate designs The overall fitness value, , , These are the weighting coefficients corresponding to the stress index, weight index, and deformation index, respectively, and satisfy the following conditions: , , ; These are stress evaluation items related to structural strength; Weight evaluation items related to lightweighting; This refers to the deformation evaluation item related to stiffness / deformation.

[0106] The optimization objective can be to minimize the maximum equivalent stress, minimize the structural mass, minimize the deformation, or a combination of multiple objectives with weights. After optimization, the optimal or near-optimal geometric design results that satisfy the optimization objectives and constraints are output, including the corresponding control point displacement distribution, the updated 3D model, and simulation evaluation results of each performance index.

[0107] Convergence Criterion:

[0108]

[0109] in: The ratio of the fitness function value to the objective function value at the t-th iteration; This is the convergence threshold; The maximum number of iterations is used to limit the computational budget and termination conditions of the optimization process.

[0110] Step 3: AI Agent Prediction and Search Acceleration;

[0111] Step 3.1: Geometric perturbation and response feature representation

[0112] After each round of geometric perturbation, this invention no longer directly calls the costly finite element simulation, but instead uses a pre-trained AI proxy model to quickly predict the structural performance.

[0113] The input characteristics of the disturbance geometry include control point displacement increments. Adjacency topology matrix changes Local curvature changes Changes in normal direction Surface principal direction , Geometric scalar features such as area, volume, and thickness are uniformly vectorized and represented as... Feature extraction is performed.

[0114]

[0115] Where z is the design change embedding vector; It is the feature extraction function, and d is the feature dimension.

[0116] Step 3.2: AI Agent Model Architecture and Training; The agent model is used to... Predict the corresponding structural performance under these circumstances. Authentic response Based on finite element simulation. This invention supports multiple AI proxy models, including: Multilayer Perceptron (MLP): suitable for low-dimensional perturbations and global features; Graph Neural Network (GNN / GCN): suitable for control points + topological adjacency information, capable of capturing geometric topological changes; Hybrid Model (MLP + GCN): combines global scalars and local geometric relationships, balancing accuracy and generalization.

[0117] Predictive Relationship:

[0118]

[0119] in, For AI agent models, For AI surrogate model parameters; the training objective function is mean squared error:

[0120]

[0121] in, Let the mean squared error loss function be . The sample size is used. In the early stages, the surrogate model is trained offline using an initial simulation sample library; during the optimization process, accuracy is gradually improved through active sampling and incremental training.

[0122] Step 3.3: Coupling of prediction and optimization search; Perturbing candidate solution set Rapidly predict responses using AI agent models And calculate the fitness based on the objective function:

[0123]

[0124] in: This represents the set of optimization variable disturbances / control point displacement increments corresponding to the m-th candidate scheme; For the first The weight coefficients corresponding to each evaluation indicator; For the first Each evaluation indicator corresponds to a target / penalty function; To address candidate perturbation schemes Predicted performance response values ​​obtained from proxy models or simulation models.

[0125] The optimization engine prioritizes retaining the top K solutions with the best prediction performance for verification in real simulation. This step effectively reduces the computational cost of the original full simulation. It has become a computational burden of AI agents + carefully selected simulations. In large-scale perturbation searches, this can achieve a computational speedup of tens of times.

[0126] Step 4: Simulation verification and closed-loop iteration;

[0127] Step 4.1: Simulation task scheduling and automatic execution;

[0128] After initial screening using geometric perturbation and AI proxy models, this invention selects the optimal candidate solution. (i.e., the set of control point disturbances) performs high-precision finite element simulation (FEM) or computational fluid dynamics simulation (CFD).

[0129] The simulation process is implemented using an automated scheduling script, and includes the following steps:

[0130] 1. Geometry Update: The optimization engine uses the CATIA API to... The corresponding perturbation is applied to the CAD geometric model, and a new .CATPart file is saved.

[0131] 2. Automatic Meshing: Meshing is performed using the mesher scripts included with ANSA, HyperMesh, or Abaqus. Mesh quality requirements include a skewness of no more than 0.85 and an aspect ratio of no more than 5 for each mesh element, ensuring that the mesh quality meets the requirements for numerical computation stability and simulation accuracy.

[0132] 3. Simulation Solution: Based on the target task type, Abaqus or Fluent is automatically called; for structural stress / deformation, the Abaqus.inp file is used to call abaqusjob=... batch processing; for fluid-structure interaction or aerodynamic problems, the Fluent batch processing script .jou file is called.

[0133] 4. Task Distribution and Parallelism: Simulation tasks for multiple candidate solutions are distributed to a multi-core / multi-node computing cluster by the scheduler, improving iteration efficiency.

[0134]

[0135] in, The parallel execution time required to complete the same simulation calculation after adopting parallel computing / parallel task distribution; The serial execution time required to complete the same simulation calculation in single-core or serial mode; This refers to the number of computing cores used for parallel computing (number of CPU cores / number of parallel worker processes).

[0136] Step 4.2: Closed-Loop Control and Convergence Criterion: The core of closed-loop control is "simulation feedback → optimization judgment → re-iteration". This invention defines the following convergence criterion:

[0137] The objective function converges:

[0138]

[0139] Optimize variable perturbation convergence:

[0140]

[0141] in, Indicates the first During the nth iteration Displacement increment / disturbance vector of each control point (decision variable); To optimize the variable convergence threshold / perturbation convergence threshold.

[0142] Response error convergence (surrogate model error):

[0143]

[0144] in, The predicted value / predicted response vector of the AI ​​proxy model for the simulation response; This represents the 2-norm (Euclidean norm). The response error threshold / surrogate model error convergence threshold.

[0145] If any two of the above conditions are met, the simulation results are considered to have reached a stable convergence state. If convergence has not occurred, the simulation results samples will be... Add the AI ​​agent model training set, readjust the perturbation step size and search direction, and enter the next round of optimization iteration.

[0146] Step 4.3: Adaptive Iteration and Model Update Mechanism

[0147] In the closed-loop iteration, this invention employs an adaptive perturbation contraction + surrogate model incremental update mechanism:

[0148] The perturbation contraction gradually reduces the amplitude of the geometric perturbation to ensure the stability of the final solution.

[0149]

[0150] Incremental learning of proxy models:

[0151]

[0152] in: For the first The set of parameters of the AI ​​agent model during each iteration (e.g., network weights and biases). This is the set of parameters for the AI ​​agent model updated based on the new samples. The learning rate controls the step size for each parameter update. For model parameters The gradient of the loss function is given by the loss corresponding to the new sample (new simulation result). Calculated; This is the loss function constructed based on the newly added training data.

[0153] The newly added simulation results samples are used to continuously train the surrogate model and improve its accuracy in the local optimal neighborhood.

[0154] Uncertainty-driven sampling: Prioritize simulation of candidate geometries with high prediction uncertainty to avoid convergence and getting trapped in local extrema.

[0155] This embodiment focuses on the antenna array surface structure of a certain type of phased array radar as the optimization object. The goal is to achieve multi-objective design optimization that minimizes antenna array mass and maximizes structural stiffness while meeting radar antenna electrical performance indicators (such as gain, sidelobe level, and beam pointing accuracy) and structural mechanical requirements (such as vibration modes and thermal deformation resistance). This case study employs the proposed geometric model-based optimization closed-loop technology system to automatically complete the entire process of geometric perturbation—AI prediction—simulation verification—convergence iteration.

[0156] Step 1: Geometric feature extraction;

[0157] The initial antenna array CAD model was built using CATIA, with dimensions of approximately 2200mm × 1800mm × 3.2mm. This array, the core supporting structure of the phased array radar, includes array element mounting grooves (192 in total, arranged in a rectangular grid), an edge rigid frame, back reinforcing ribs (in a cross-shaped layout), and reserved holes for signal transmission channels. Geometric topology information was extracted using the CATIA COM interface and HybridShape API, including:

[0158] Number of control points N=512;

[0159] Geometry type: NURBS surface skin + reinforcing ribs (regular solid) + mounting grooves (feature holes);

[0160] Areas of concentrated curvature: the connection between the reinforcing rib and the array skin, and the edge of the mounting groove (areas of stress concentration and sensitivity, which also affect the installation accuracy of the antenna element).

[0161] Topological adjacency matrix A: Automatically generated based on the connectivity of control points (such as the connection points between skin and rib, and the transition points between groove and skin);

[0162] Geometric attribute extraction:

[0163] Initial maximum curvature: Curvature value of the central region of the array skin ≤ 0.002mm -1 The curvature value of the groove edge is ≥0.015mm. -1 .

[0164] Normal distribution: The deviation between the main normal of the array and the radar beam pointing direction is less than or equal to 0.5°, which is used to constrain the direction of disturbance (to avoid affecting electrical performance).

[0165] Initial array mass: 18.6 kg;

[0166] Initial maximum equivalent stress: Under typical airborne vibration loads (10-2000Hz, acceleration 20g), the maximum stress at the groove edge reaches 242.3MPa;

[0167] Initial thermal deformation: Within the operating temperature range (-40℃~65℃), the maximum thermal deformation of the array center relative to the edge is 1.8mm (exceeding the electrical performance allowable threshold of 1.2mm).

[0168] Step 2: Geometric perturbation and search;

[0169] The 512 control points are divided into three categories based on structural importance, electrical performance sensitivity, and manufacturing characteristics: High-sensitivity area (edge ​​of mounting groove + reinforcing rib connection area): 288 control points → allows omnidirectional disturbances; Medium-sensitivity area (non-groove area of ​​array skin): 256 control points → limits normal disturbances (maximum normal displacement ≤ 0.3mm, to avoid changing the phase center of the antenna element), allows slight radial disturbances; Assembly-frozen area (edge ​​rigid frame + signal channel boundary): 96 control points → does not participate in disturbances (virtually ensures assembly accuracy with the radar housing and signal channel sealing).

[0170] Perturbation model:

[0171]

[0172] in The normal to the surface of the array skin (aligned with the beam). , These are the principal curvature directions of the groove edge and the rib connection area, respectively. The disturbance amplitude (initial) =0.25mm), , , Directional weighting coefficient ( =0.2, =0.5, =0.3, prioritizing stability in electrically sensitive directions); Iteration contraction coefficient =0.92 (per iteration). ).

[0173] Search strategy: Use a genetic algorithm (GA) for global search + local Bayesian optimization (BO), with a population size of 80, a maximum number of iterations of 30, a mutation probability of 0.08, and a crossover probability of 0.85;

[0174] Boundary constraints: Control point disturbances must not exceed the radar housing installation space (maximum dimensional deviation ≤ 0.5mm), the inner diameter change rate of the mounting groove ≤ 1.5% (to ensure antenna element plug-in compatibility), and the local curvature change rate... (To avoid stress concentration and exacerbation at the edge of the groove)

[0175] Step 3: AI proxy prediction and search acceleration; Since the antenna array needs to be simulated simultaneously with mechanical simulation (vibration, thermal deformation) and electrical performance simulation (electromagnetic radiation characteristics), a single full-physics field coupled simulation (Abaqus+HFSS joint solution) takes about 4.2 hours and has extremely high computational cost. Therefore, an AI proxy model is introduced to quickly predict the dual-objective performance of the perturbation candidate solution.

[0176] Training sample set: 450 sets of initial perturbation + full physics field simulation samples (covering typical scenarios with different perturbation directions and amplitudes).

[0177] Input feature: Control point displacement increment Variation in curvature at the edge of the groove Skin normal offset Changes in the cross-sectional area of ​​the reinforcing ribs Geometric compensation amount corresponding to local temperature gradient .

[0178] Output response: Maximum equivalent stress Thermal deformation at the center of the array Antenna gain Sidelobe level .

[0179] The surrogate model adopts a hybrid architecture of GCN (handling geometric topological relationships, such as the force transmission path at the connection point between the rib and the skin) + MLP (handling scalar features of electrical performance, such as gain and sidelobe level). The electrical performance-related features are pre-trained separately using an electromagnetic simulation dataset and then fused with mechanical features for fine-tuning.

[0180] Prediction error: RMSE < 1.8% for mechanical performance, RMSE < 0.5dB (gain) and RMSE < 0.3dB (sidelobe level) for electrical performance.

[0181] In each round of optimization, AI dual-objective prediction is performed on 80 candidate perturbation schemes; based on "mechanical performance meeting standards (…)" ≤200MPa ≤1.2mm) + Optimal electrical performance ( ≥32dB The top-10 candidate geometries are selected using the "≤-28dB" filtering rule; high-precision Abaqus+HFSS co-simulation is performed only on these 10 candidate schemes; the new simulation results are fed back to train the surrogate model for incremental updates.

[0182] Computational acceleration effect: The original full simulation computational workload Hourly computational load after using AI agents and carefully selected simulations In 1 hour, the speedup ratio reaches 7.99 times, significantly shortening the optimization cycle.

[0183] Step 4: Simulation verification and closed-loop iteration.

[0184] The top-10 candidate solutions were verified by batch processing using Abaqus+HFSS multiphysics coupling simulation. Simulation conditions:

[0185] Mechanical conditions: random vibration (10-2000Hz, acceleration 20g, according to GJB 150.16-2009 standard); steady-state temperature field (-40℃~65℃, temperature gradient 5℃ / min, considering the thermal load of the array heat dissipation module).

[0186] Electrical performance conditions: Operating frequency (X-band, 8-12GHz); Beam scanning range (azimuth ±60°, elevation ±30°); Input power (50W / unit);

[0187] Material parameters: The main body of the array is made of carbon fiber composite material (T800 / epoxy), with an elastic modulus E 11 =180GPa, E 22 =10GPa, coefficient of thermal expansion , The reinforcing ribs are made of aluminum alloy 7075-T73.

[0188] Extract simulation results:

[0189] Mechanical properties: Maximum equivalent stress Thermal deformation at the center of the array First-order vibration mode frequency (Requires ≥250Hz to avoid resonance)

[0190] Electrical performance: Antenna gain Sidelobe level Beam pointing error (Requires ≤0.1°)

[0191] Convergence Criterion:

[0192] The objective function converges: F is a multi-objective weighting function:

[0193]

[0194] Optimize variable perturbation convergence: The amplitude of the control point disturbance reaches the threshold of the process accuracy.

[0195] Response error convergence: The relative error between the AI ​​agent's predicted value and the simulated measured value is ≤2%. The actual convergence is achieved after 18 iterations. Due to the need to balance the dual objectives of mechanical and electrical performance, the number of iterations is slightly more than that for single mechanical objective optimization.

[0196] The final optimization of the geometry mainly involves adjusting the transition fillets of the mounting grooves, adjusting the topography of the reinforcing rib sections, and correcting the micro-arcs of the array skin.

[0197] Compared with the initial structure:

[0198] Performance indicators Initial structure Optimize structure Improvement range Maximum equivalent stress 242.3MPa 189.7MPa ↓21.7% Total mass of the array 18.6kg 16.1kg ↓13.4% Central thermal deformation 1.8mm 1.1mm ↓38.9% First-order vibration mode frequency 228Hz 265Hz ↓16.2% Antenna gain 31.5dB 32.8dB ↓4.1% Sidelobe level -26.8dB -29.2dB ↑8.9% (better) Beam pointing error 0.15° 0.08° ↓46.7%

[0199] The geometric surface continuity meets the requirements of composite material molding process, the fitting gap between the mounting groove and the antenna unit is controlled within the range of 0.05-0.12mm, the assembly accuracy deviation of the edge frame is ≤0.2mm, and the optimization results can be directly imported into the radar assembly and electrical performance debugging process.

[0200] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. An automatic optimization method based on CAD models, characterized in that, Includes the following steps: Step 1: Perform geometric feature extraction and structural description modeling on the CATIA 3D model of the target structure, including feature extraction of control points, lines, surfaces, topology and geometric attribute information, to obtain the geometric features of the control point space; Control points are selected and partitioned based on geometric and operational condition sensitivity. Step 2: For the set of control points selected for optimization, perform geometric perturbation on the geometric features of the control points, and then optimize through global search and local correction. After optimization, output the optimal or near-optimal geometric perturbation result that satisfies the optimization objective and constraints. Step 3: After each round of geometric perturbation, the structural performance is quickly predicted using a pre-trained AI proxy model. The fitness is calculated based on the objective function, and the top K solutions with the best prediction performance are retained for real simulation verification. Step 4: Simulation verification and closed-loop iteration; Execute the simulation process for the selected optimal candidate solution, and perform simulation tasks for multiple candidate solutions in parallel; Set convergence criteria, and determine the convergence of fitness, optimization variable perturbation, and response error for the simulation results. If any two convergence conditions are met, it is considered to have reached a stable convergence state. If convergence is not achieved, the simulation results samples are added to the AI ​​agent model training set, the perturbation step size and search direction are readjusted, and the next round of optimization iteration is initiated.

2. The automatic optimization method based on a CAD model according to claim 1, characterized in that, The geometric features include control points. and based on control points boundary curve curved surfaces curvature Normal vector Local thickness .

3. The automatic optimization method based on a CAD model according to claim 2, characterized in that, The specific steps for selecting and partitioning control points are as follows: The set of control points within high curvature regions and stress-sensitive regions are selected as the set of principal optimization variables. The set of control points in low-sensitivity areas is used as the frozen set. It remains unchanged during the optimization iteration; Furthermore, the curvature field and the principal directions of the surface are used to provide directional priors for the disturbance: the maximum and minimum principal curvature directions corresponding to the local principal directions of the surface are uniformized and used as regional-level directional bases, and the displacement increments generated by subsequent optimization are weighted and decomposed on this directional base.

4. The automatic optimization method based on a CAD model according to claim 3, characterized in that, The specific geometric perturbation is as follows: Suppose the three-dimensional geometric model has N control points Composition, for the set of main optimization variables Its position after the disturbance is defined as: ， ； ； in, For the control point displacement increment, the frozen set ,make ; , , The corresponding disturbance coefficient, The principal direction of the surface corresponds to the direction of maximum principal curvature. The principal direction of the surface corresponds to the direction of minimum principal curvature; disturbance amplitude A round-by-round decay mechanism is adopted: ； in, For the first The threshold parameter at the next iteration The attenuation coefficient is represented by the superscript t, which indicates the number of iterations. Apply feasible region constraints: ,in, To control points Related disturbance limits; Geometric continuity and compliance constraints: ,in, It is the smoothness coefficient; Curvature change rate constraint: ,in, For the first disturbance Curvature values ​​at each control point or its neighborhood corresponding surface sampling point; This refers to the curvature value after disturbance calculated at the same location or corresponding neighborhood after applying displacement increment to the control point; This is the threshold for curvature change.

5. The automatic optimization method based on a CAD model according to claim 4, characterized in that, The optimization through global search and local correction is as follows: The global search uses a genetic algorithm or particle swarm optimization algorithm to generate a set of perturbation candidate solutions in the control point space X; the local correction uses gradient-based or Bayesian optimization strategies to fine-tune excellent individuals and narrow the perturbation range. Calculate the fitness function: ； in: Geometric features The overall fitness value, , , These are the weighting coefficients corresponding to the stress index, weight index, and deformation index, respectively, and satisfy the following conditions: , , ; These are stress evaluation items related to structural strength; Weight evaluation items related to lightweighting; Deformation evaluation items related to stiffness / deformation; When the maximum number of iterations is reached or the change in the fitness function meets the convergence threshold, the optimization stops and the optimal or near-optimal geometric design results that satisfy the optimization objective and constraints are output, including the corresponding control point displacement distribution, the updated three-dimensional model, and the simulation evaluation results of each performance index.

6. The automatic optimization method based on a CAD model according to claim 5, characterized in that, Step 3 is as follows: Step 3.1: Geometric Disturbance and Response Feature Representation; The input features of the geometric disturbance include the control point displacement increment. Adjacency topology matrix changes Local curvature changes Changes in normal direction Surface principal direction , And geometric scalar features, the above information is uniformly vectorized as Feature extraction is performed. ； Where z is the design change embedding vector; It is the feature extraction function, where d is the feature dimension; Step 3.2: The AI ​​agent model is used to... Predict the corresponding structural performance under these circumstances. Authentic response Based on finite element simulation, the training objective function is mean square error: ； in, Let the mean squared error loss function be . The sample size is used. In the early stages, the surrogate model is trained offline using an initial simulation sample library; during the optimization process, accuracy is gradually improved through active sampling and incremental training. Step 3.3: Perturb the candidate solution set Rapidly predict responses using AI agent models And calculate the fitness based on the objective function: ； in, This represents the set of control point displacement increments corresponding to the m-th candidate scheme; For the first The weight coefficients corresponding to each evaluation indicator; For the first Each evaluation indicator corresponds to an objective function; To address candidate perturbation schemes The performance response prediction values ​​obtained from the proxy model or simulation model are used to select the top K solutions with the best predicted performance for real simulation verification.

7. The automatic optimization method based on a CAD model according to claim 6, characterized in that, Step 4 is as follows: Step 4.1: Perform high-precision finite element simulation or computational fluid dynamics simulation on the selected optimal candidate solution. The simulation tasks of multiple candidate solutions are distributed to the multi-core computing cluster by the scheduler. Step 4.2: Closed-loop control and convergence criteria; Objective function convergence: The change in the fitness function during iteration is less than or equal to the set convergence threshold; Optimization variable perturbation convergence: During iteration, the change in the displacement increment of the control point is less than or equal to the perturbation convergence threshold; Response error convergence: The calculated L2 norm of the performance response prediction value and the true response value is less than or equal to the error convergence threshold; If any two convergence conditions are met, it is considered to have reached a stable convergence state. If it has not converged, the simulation result samples are added to the AI ​​agent model training set, the perturbation step size and search direction are readjusted, and the next round of optimization iteration is entered. Step 4.3: In the closed-loop iteration, an adaptive perturbation shrinkage and surrogate model incremental update mechanism are adopted: The perturbation contraction gradually reduces the amplitude of the geometric perturbation to ensure the stability of the final solution. ； Incremental learning of proxy models: ； in: For the first The parameter set of the AI ​​agent model at the next iteration; This is the set of parameters for the AI ​​agent model updated based on the new samples. The learning rate; For AI agent model parameters The gradient of the loss function is calculated from the loss function corresponding to the new sample. The loss function is constructed based on the newly added training data; the newly added simulation result samples are used to continuously train the surrogate model. Implement uncertainty-driven sampling: prioritize simulation of candidate geometries with high prediction uncertainty to avoid convergence and getting trapped in local extrema.

8. The automatic optimization method based on a CAD model according to claim 7, characterized in that, The AI ​​agent model includes a multilayer perceptron, a graph neural network, or a hybrid model of a multilayer perceptron and a graph neural network.

Images