Mechanical structure global intelligent design system based on topological evolution optimization

Abstract

The invention relates to the technical field of mechanical design, in particular to a mechanical structure global intelligent design system based on topological evolution optimization. Comprising a global parametric modeling module configured to construct an initial population of a mechanical structure by using an explicit geometric description method based on a preset design domain, a mechanical boundary condition and a manufacturing process constraint, each individual in the initial population being uniquely represented by a group of geometric parameter vectors, according to the method, a closed-loop feedback mechanism of a physical perception agent model and active learning evolution is constructed, so that the inherent contradiction of huge time consumption of traditional evolution optimization calculation and insufficient physical precision of a pure data driving model is effectively overcome; according to the system, the deep neural network is used for replacing high-frequency finite element calculation, the global search efficiency of a complex mechanical structure is improved by two orders of magnitude, meanwhile, real-time error calibration is carried out on an agent model through online high-fidelity verification, and it is ensured that the mechanical property prediction error of an optimal solution is lower than a preset threshold value.

Description

Technical Field

[0001] This invention relates to the field of mechanical design technology, and more specifically to a global intelligent design system for mechanical structures based on topological evolution optimization. Background Technology

[0002] With the increasing demand for lightweight and high performance in the aerospace and precision machinery fields, structural topology optimization technology has become an important tool in mechanical design. Although traditional topology optimization based on the variable density method (SIMP) is widely used, it is essentially a gradient-based local search algorithm, which is prone to getting trapped in local optima. Moreover, the generated structures often contain a large number of gray-scale units with intermediate density, resulting in blurred boundaries and making it difficult to directly meet the process constraints of additive manufacturing (3D printing).

[0003] To overcome the local optima problem, existing technologies attempt to introduce evolutionary algorithms such as Genetic Algorithms (GA) and Particle Swarm Optimization (PSO) for global optimization. However, evolutionary algorithms rely on a large population size and numerous iterations, requiring finite element analysis (FEA) for performance evaluation of each individual in each iteration. For complex mechanical structures, a single FEA calculation is extremely time-consuming, leading to an exponential increase in computational cost for evolution-based global optimization, making it difficult to implement in engineering practice.

[0004] In recent years, data-driven methods based on deep learning have been proposed to accelerate the optimization process. However, existing end-to-end neural network models are often regarded as "black boxes," lacking physical interpretability. Their predictions often fail to meet strict mechanical equilibrium equations and have poor generalization ability for working conditions outside the training set.

[0005] In summary, there is an urgent need in the existing technology for an intelligent design system that can achieve global search using evolutionary algorithms, significantly reduce computational costs through physical sensing mechanisms, and take into account structural manufacturability. Summary of the Invention

[0006] To achieve the above objectives, the present invention provides a global intelligent design system for mechanical structures based on topological evolution optimization, comprising: a global parametric modeling module configured to construct an initial population of mechanical structures using an explicit geometric description method based on a preset design domain, mechanical boundary conditions and manufacturing process constraints, wherein each individual in the initial population is uniquely represented by a set of geometric parameter vectors; The physical perception agent prediction module is configured to pre-set a trained deep neural network model, which is used to receive the geometric parameter vector as input and output the corresponding predicted value of the structural physical field response. A multi-strategy co-evolutionary engine is configured to execute an evolutionary iteration process, calculate the fitness function of individuals in the population using the predicted values ​​of the physical field response, and perform selection, crossover, and mutation operations on the population based on the evolutionary strategy to generate offspring individuals. A high-fidelity verification and feedback calibration module is configured to monitor the prediction confidence of the physical perception agent prediction module during the evolutionary iteration process. When the prediction confidence is lower than a preset threshold or the evolutionary iteration reaches a preset interval of generations, the high-fidelity verification and feedback calibration module calls the finite element analysis solver to perform physical field simulation calculations on the preferred individuals in the current population, obtains the real physical field response value, and feeds back the deviation between the real physical field response value and the predicted physical field response value to the physical perception agent prediction module to update the weight parameters of the deep neural network model online. The post-processing output module receives the final converged optimal geometric parameter vector, transforms it into a triangular mesh model using a level set function, automatically identifies the overhanging region based on the manufacturing process constraints, generates the support structure path, and finally outputs the additive manufacturing slice file.

[0007] Furthermore, the explicit geometric description method adopted by the global parametric modeling module is the moving shape component method or the moving deformable cavity method; wherein, the geometric parameter vector includes the center coordinates, length, width, tilt angle and polynomial shape control parameters of the component; The global parametric modeling module is configured to directly map the minimum feature size constraint and the maximum overhang angle constraint to the geometry of the initial population by limiting the range of values ​​of the geometric parameter vector.

[0008] Furthermore, the deep neural network model pre-installed in the physical perception agent prediction module adopts a fully convolutional neural network architecture or a graph neural network architecture. The deep neural network model includes an encoder path and a decoder path. The encoder path is used to extract deep topological features of the input geometry, and the decoder path is used to map the extracted features back to the physical field space, directly outputting pixel-level or node-level von Mises stress distribution maps and displacement cloud maps.

[0009] Furthermore, the multi-strategy co-evolutionary engine adopts a covariance matrix adaptive evolutionary strategy. The multi-strategy co-evolutionary engine is configured to calculate the fitness function of an individual based on the following formula. : , The structural flexibility predicted by the surrogate model, For structural volume, Penalties for violating manufacturing process constraints and These are the weighting coefficients.

[0010] Furthermore, the high-fidelity verification and feedback calibration module monitors the prediction confidence level by: enabling Monte Carlo dropout or using deep ensemble technology during the inference process of the deep neural network model; performing multiple forward propagation inferences on the geometric parameter vector of the same entity, and calculating the variance or standard deviation of the multiple inference results; using the variance or standard deviation as a quantitative indicator to measure the prediction confidence level, and determining that the prediction confidence level is low when the variance is greater than a preset threshold, thus triggering a finite element analysis call.

[0011] Furthermore, when updating the weight parameters of the deep neural network model online, the high-fidelity verification and feedback calibration module adopts a transfer learning strategy: freezing the weight parameters of the shallow feature extraction layer of the deep neural network model, and using only the error between the real physical field response value and the predicted value, fine-tuning the weight parameters of the deep fully connected layer or output layer of the model through the backpropagation algorithm.

[0012] Furthermore, the global parameterized modeling module also includes a Latin hypercube sampling unit; during the system initialization phase, the Latin hypercube sampling unit is used to uniformly generate a set of sample structures within the design space and call the finite element analysis solver to calculate their label data, which is used to construct an initial training dataset for pre-training the deep neural network model.

[0013] Furthermore, the high-fidelity verification and feedback calibration module is also equipped with an anomaly removal mechanism; when the real physical field response value calculated by the finite element analysis solver shows that the structure has diverged or singular deformation, the individual is determined to be an invalid design, the data of the individual is not included in the update training set of the deep neural network model, and the individual is directly removed from the evolutionary population. The input to the physical sensing agent prediction module also includes an external load condition vector; the deep neural network model is configured to simultaneously receive geometric parameter vectors and external load condition vectors, thereby predicting the physical field response of the structure under different load boundary conditions.

[0014] Beneficial effects Compared with known public technologies, the technical solution provided by this invention has the following beneficial effects: The significant advantages of this invention are as follows: By constructing a physical perception proxy model and a closed-loop feedback mechanism for active learning and evolution, it effectively overcomes the inherent contradiction between the enormous computational time consumption of traditional evolutionary optimization and the insufficient physical accuracy of purely data-driven models. This system utilizes deep neural networks to replace high-frequency finite element calculations, improving the global search efficiency of complex mechanical structures by two orders of magnitude. Simultaneously, online high-fidelity verification performs real-time error calibration on the proxy model, ensuring that the mechanical performance prediction error of the optimal solution is below a preset threshold. Furthermore, combined with parametric modeling technology based on explicit geometric description, this invention can directly embed process constraints such as minimum wall thickness and overhang angle at the design stage, completely eliminating grayscale units and microstructures that are difficult to manufacture in traditional topology optimization. This achieves a seamless "zero-post-processing" connection from conceptual design to additive manufacturing instruction generation, greatly shortening the R&D cycle of high-performance mechanical products. Attached Figure Description

[0015] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a schematic diagram of the computer device of the present invention; Figure 3 This is a flowchart of the active learning optimization method of the present invention. Detailed Implementation

[0016] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0017] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0018] The present invention will now be described in further detail with reference to the accompanying drawings: Example: As shown in the figure, the present invention provides a global intelligent design system for mechanical structures based on topological evolution optimization, including: a global parametric modeling module, configured to construct an initial population of mechanical structures based on a preset design domain, mechanical boundary conditions and manufacturing process constraints, using an explicit geometric description method, wherein each individual in the initial population is uniquely represented by a set of geometric parameter vectors; Furthermore, the operation process of the global parametric modeling module includes: 1. First, the global parametric modeling module reads the initial design request input by the user and determines the design domain. , Define design domain The maximum possible spatial extent of the structure is defined. Simultaneously, non-design domains, such as bolt mounting holes, bearing housing holes, or stress boundary regions, are identified based on mechanical interface requirements. During subsequent optimization, the material distribution within these non-design domains remains constant and does not participate in evolutionary variation.

[0019] 2. Modules in the design domain An initial set of movable and deformable components is arranged within the design domain, assuming the domain contains... Each component has a geometry defined by its corresponding topological description function. To express, Each component is defined as a "hyperellipse" or "beveled rectangle" structure with a certain thickness and length. The geometric features of each component are defined by a set of geometric parameter vectors. The unique representation, its mathematical expression is as follows:

[0020] in, Indicates the first The position coordinates of the center point of each component in the global coordinate system. This represents half the length of the component feature. This represents half the width of the component feature. Indicates the tilt angle of the component's main axis. The polynomial parameter represents the degree of curvature that controls the shape of the component.

[0021] 3. The global parametric modeling module uses Boolean operations to... Individual components are integrated into a unified structure. The global level set function. Defined as the maximum value of all component description functions:

[0022] Among them, when When, it indicates that point Located in the solid material area; When, it indicates that point Located in a hollow area; When, it represents the boundary contour of the structure. In this way, the topological configuration of the mechanical structure is completely mapped to a high-dimensional set of geometric parameter vectors.

[0023] In the subsequent multi-strategy co-evolutionary engine, the object of evolution is precisely this vector group. Rather than a large grid density matrix.

[0024] 4. The global parametric modeling module directly applies explicit constraints to the geometric parameter vectors when generating the initial population and subsequent parameter mutations: Minimum size constraint: Sets the width parameter of the component. The range of values ​​is ,in The minimum wall thickness allowed for a 3D printer.

[0025] Self-supporting constraint: Sets the tilt angle of the component. The range of values ​​for . Assuming the stacking direction in additive manufacturing is the Z-axis, if the critical overhang angle of the printed material is . Then restrictions Avoid The range, or in subsequent optimization objectives, impose a very large penalty value on components that violate this angle range.

[0026] In this embodiment, the global parametric modeling module is primarily responsible for constructing the mathematical description space of the mechanical structure. Unlike traditional implicit topology optimization methods based on finite element mesh density (such as the SIMP method), this embodiment employs an explicit geometric description method based on moving morphological components. This method transforms the topology optimization problem into a shape and size optimization problem for geometric parameters, thereby enabling direct control over the geometric features of the structure.

[0027] The physical perception agent prediction module is configured to pre-set a trained deep neural network model, which is used to receive the geometric parameter vector as input and output the corresponding predicted value of the structural physical field response. Furthermore, the operation process of the physical perception agent prediction module includes: 1. Because the global parametric modeling module outputs a set of geometric parameter vectors. To adapt it to the input format of deep neural networks, this module first performs feature discretization. Specifically, the system divides the design domain into... A regular grid. Utilizing the aforementioned level set function. This maps the geometric parameter vector to a two-dimensional (or three-dimensional) geometric feature tensor. The input tensor... Contains multiple channels: Channel 1 (Structure Channel): Material density value of the corresponding mesh node (0 or 1, or smooth grayscale value at the boundary); Channel 2 (Boundary Condition Channel): Marks the location of displacement fixed constraints (e.g., 1 for fixed areas and 0 for the rest); Channel 3 (Load Channel): Marks the location and magnitude of the applied external force. Through this multi-channel tensor design, the neural network can not only perceive the shape of the structure, but also the changes in the mechanical boundary conditions.

[0028] 2. The physical perception proxy model employs a fully convolutional neural network based on the U-Net architecture. This network structure is U-shaped and includes encoder and decoder paths: Encoder path (shrinkage path): Used to extract deep abstract features of the geometry. It consists of a set of repeatedly applied operations: two... The convolutional layers are followed by a modified linear unit activation function and a stride of 2. Max pooling layer. Each downsampling operation halves the size of the feature map while doubling the number of feature channels.

[0029] Bottleneck layer: Located at the bottom of the U-shaped structure, it contains the deepest convolutional layer and is used to capture global contextual information, such as the topological features of the force transmission path.

[0030] Decoder path (expansion path): Used to restore abstract features to a physical field distribution map consistent with the input size. It contains upsampling layers and introduces skip connections. Skip connections directly concatenate high-resolution feature maps from corresponding layers in the encoder path into the decoder path to preserve high-frequency details such as structural edges.

[0031] 3. The output layer of this neural network uses a linear activation function, and the output tensor... The corresponding predicted physical field cloud map includes the von Mises stress field and displacement field. To reflect the physical perception characteristics, the loss function of this module during the pre-training phase is... The design incorporates both data-driven loss and physical mechanism loss.

[0032] in, For mean square error loss, For physical residual loss, These are the weighting coefficients, which the system uses to predict the displacement field. Based on the equilibrium equation of linear elasticity , where K is the stiffness matrix and F is the external load, calculate the force imbalance residual for each node.

[0033] In this embodiment, the system is equipped with a physical perception proxy prediction module, which is used to construct a fast mapping relationship between design variables and structural physical performance, in order to replace the time-consuming full finite element analysis in the evolutionary optimization process.

[0034] A multi-strategy co-evolutionary engine is configured to execute an evolutionary iteration process, calculate the fitness function of individuals in the population using the predicted values ​​of the physical field response, and perform selection, crossover, and mutation operations on the population based on the evolutionary strategy to generate offspring individuals. Furthermore, the operational process of the multi-strategy co-evolution engine includes: To quantitatively evaluate the performance of each individual, a fitness function is constructed. The engine first defines the fitness function. Unlike traditional objective functions that only consider stiffness, this approach employs a multi-objective weighted aggregation strategy to couple mechanical properties with manufacturing constraints.

[0035] in, The geometric parameter vector output by the global parametric modeling module. For the flexibility of the structure, As the normalization factor, For manufacturing process constraints and penalties; when geometric parameters When the minimum wall thickness or overhang angle constraint is violated, this term increases exponentially, thereby automatically eliminating unproducible individuals during the evolutionary process.

[0036] 2. This engine employs a strategy of combining global coarse search and local fine search to address the slow convergence and premature convergence issues of traditional genetic algorithms: Phase 1: Global Exploration Based on Adaptive Differential Evolution in the Early Stages of Optimization, such as before In this generation, the engine employs an adaptive differential evolution algorithm, utilizing a large mutation factor. and crossover probability The driving population performs a broad random search within the design domain. During this stage, the engine batch inputs the geometric parameter vectors of all generated offspring individuals into the physics-aware surrogate prediction module. The surrogate model returns the predicted physics field within milliseconds, and the engine calculates the fitness based on this. Quickly identify promising seed individuals.

[0037] Phase Two: Adaptive Local Refinement Based on Covariance Matrix When the fitness variance of the population decreases to a preset threshold, or the number of iterations exceeds... Then, the engine automatically switches to the CMA-ES strategy. CMA-ES learns the distribution covariance matrix of dominant individuals in the population and dynamically adjusts the search step size and direction. This strategy can perform high-precision gradient estimation for continuous geometric parameters of MMC (moving morphology components) (such as coordinate fine-tuning and angle fine-tuning), thereby optimizing the shape of the structure edge at the micrometer level while ensuring that the topological configuration remains unchanged.

[0038] 3. During the evolution process, the engine not only focuses on the predicted values ​​output by the proxy model, but also on the uncertainty of the prediction.

[0039] When the engine invokes the physics-aware proxy model, it enables Monte Carlo drop mode, meaning that for the same individual... conduct Next, for example Random reasoning. Calculate this. Variance of the prediction results The engine defines the active learning trigger criteria:

[0040] in, The confidence threshold. This is a mandatory calibration cycle. When the above criteria are met, the engine marks the individual as a "high-value uncertain individual" and sends it to the next module.

[0041] In this embodiment, the system core is configured with a multi-strategy co-evolutionary engine. This engine is not a single optimization algorithm, but a hierarchical hybrid optimization framework designed to leverage the high-speed reasoning capabilities of the physical perception agent prediction module to quickly converge to the global optimum within a vast design space.

[0042] A high-fidelity verification and feedback calibration module is configured to monitor the prediction confidence of the physical perception agent prediction module during the evolutionary iteration process.

[0043] Furthermore, the specific operation procedures for the high-fidelity verification and feedback calibration module include: 1. When this module receives the geometric parameter vectors of "high-value individuals" or "high-uncertainty individuals" selected by the multi-strategy co-evolutionary engine... At this time, the built-in mesh generator is invoked first: Using level set functions to transform geometric parameter vectors The model is converted into a closed STL triangular mesh model, and then tetrahedral or hexahedral meshes are applied to the STL model to establish a high-precision finite element analysis model. During this process, the system automatically identifies excessively fine or sharp geometric features, refines the local mesh, applies preset mechanical boundary conditions (loads and constraints) to the volume mesh, and calls the finite element solver to solve the equilibrium equations. Obtain the true physical field response .

[0044] 2. The system will use the actual response calculated by FEA. Predicted response compared to the previous physical perception agent model Alignment and comparison are performed, and the mean absolute error or root mean square error between the two is calculated. To prevent "catastrophic forgetting" during model updates, this module maintains a dynamic "experience replay pool." Newly generated samples... These samples are stored in this pool. During subsequent training, the system not only uses the latest samples but also randomly selects historical samples from the pool for mixed training.

[0045] 3. The system uses the backpropagation algorithm to incrementally update the neural network in the physical perception proxy prediction module. To ensure update efficiency, this embodiment adopts a transfer learning strategy: freezing the convolutional layers at the front end of the neural network to maintain their ability to extract geometric features; and only updating the weight parameters of the fully connected layers or decoder layers at the back end of the neural network using gradient descent. This strategy allows the model update to be completed with minimal computational resources and sample size, typically enabling the proxy model to learn new physical laws within seconds.

[0046] 4. In practical engineering, some randomly generated topologies may exhibit geometric singularities (such as single-point connections), causing the FEA solver to fail to converge or mesh generation to fail. The module is equipped with an anomaly detection mechanism. If the FEA solver reports an error or the results diverge (e.g., the displacement value tends to infinity), the module determines that the individual is an invalid design. The module sends a very large penalty value to the multi-strategy co-evolutionary engine, forcing the fitness of the individual to be set to the worst. This is equivalent to labeling this type of genetic structure with a lethal tag during the evolutionary process, guiding the population to automatically avoid such geometric configurations that are prone to computational failure in subsequent iterations.

[0047] In this embodiment, the system includes a high-fidelity verification and feedback calibration module. This module not only performs final physical verification of the optimized design, but more importantly, it uses an online active learning mechanism to reverse-correct the physical perception proxy model using real-time generated high-precision data, making it increasingly accurate during the optimization process.

[0048] The post-processing output module receives the final converged optimal geometric parameter vector, transforms it into a triangular mesh model using a level set function, automatically identifies the overhanging region based on the manufacturing process constraints, generates the support structure path, and finally outputs the additive manufacturing slice file.

[0049] Furthermore, the operation flow of the post-processing output module includes: 1. When the optimization iteration meets the termination condition (such as reaching the maximum number of generations or the fitness no longer improving), this module receives the geometric parameter vector of the globally optimal individual. First, call the aforementioned global level set function. The parameterized mathematical description is mapped to a high-resolution three-dimensional scalar field, and then the moving cube algorithm is used to extract... The algorithm transforms the scalar field into a closed surface mesh composed of numerous triangular patches. The module automatically performs a mesh repair procedure, applying Laplacian smoothing to the generated triangular mesh to remove minute jagged edges, and detecting and repairing non-manifold edges and holes to ensure the output model is a watertight solid model.

[0050] 2. Unlike general-purpose slicing software, this module fully utilizes the "process constraint information" embedded in the global parametric modeling stage. Based on the user-preset printing direction (Z-axis), the module traverses all faces of the triangular mesh, calculating the angle between their normal vectors and the Z-axis. Although the optimization algorithm tries to avoid overhangs, for areas where overhangs are unavoidable (i.e., areas where the angle exceeds a critical threshold), the module automatically generates tree-like or grid-like support structures. Because self-support is considered in the design phase, the amount of support generated here is typically only 10%-20% of that in traditional topology optimization structures, significantly saving material.

[0051] 3. Finally, the module formats and encodes the geometric information containing the part body and supporting structure. The system supports outputting standard STL or OBJ files for general interaction. In a further embodiment, the module has a built-in slicing engine that directly generates CLI slicing files or G-code instructions containing the laser scanning path or printhead movement path based on the printer's nozzle diameter and layer height parameters. At this point, the user no longer needs to use third-party slicing software; they can directly import the file into the additive manufacturing equipment to begin production.

[0052] In this embodiment, the system also includes a post-processing output module. This module transforms the mathematical parameter vector finally converged by the multi-strategy co-evolutionary engine into a manufacturing instruction file that can be directly used in industrial production.

[0053] like Figure 2 As shown, a computer device may include: a processor, a communication interface, a memory, and a communication bus.

[0054] The processor, communication interface, and memory communicate with each other via a communication bus. The communication interface is used to communicate with other computer devices or networks. The processor is used to execute programs, specifically the steps described in the above method embodiments.

[0055] Specifically, the program may include program code, which includes computer operation instructions. The processor X01 may be a central processing unit, a specific integrated circuit, or one or more integrated circuits configured to implement embodiments of the present invention. The intelligent design system may include one or more processors for executing computer programs to: acquire the design domain and process constraints of the mechanical structure; input geometric features into a physical perception proxy model for prediction; and perform steps such as multi-strategy evolutionary optimization based on active learning.

[0056] Memory is used to store programs. Memory may include high-speed RAM, and may also include non-volatile memory, such as at least one disk drive.

[0057] In one alternative implementation, the processor executes a computer program stored in memory to perform the following steps: acquiring the design domain, mechanical boundary conditions, and manufacturing process constraints of the mechanical structure, and generating an initial structural population containing multiple individuals using an explicit geometric description method; inputting the geometric features of the individuals in the population into a pre-set physical perception proxy model, and rapidly predicting the physical field response data of each individual through the physical perception proxy model; calculating the fitness of each individual based on the predicted physical field response data, and performing an evolutionary optimization operation to update the population structure; executing an active learning calibration strategy during the evolutionary optimization process, performing high-fidelity finite element analysis on the selected superior individuals, and fine-tuning the physical perception proxy model online based on the error until the optimal mechanical structure design scheme is output.

[0058] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps described in the method embodiments above. The computer-readable storage medium may include various media capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory, random access memory, magnetic disk, or optical disk.

[0059] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A global intelligent design system for mechanical structures based on topological evolution optimization, characterized in that, include: The global parametric modeling module is configured to construct an initial population of mechanical structures based on a preset design domain, mechanical boundary conditions, and manufacturing process constraints using an explicit geometric description method. Each individual in the initial population is uniquely represented by a set of geometric parameter vectors. The physical perception agent prediction module is configured to pre-set a trained deep neural network model, which is used to receive the geometric parameter vector as input and output the corresponding predicted value of the structural physical field response. A multi-strategy co-evolutionary engine is configured to execute an evolutionary iteration process, calculate the fitness function of individuals in the population using the predicted values ​​of the physical field response, and perform selection, crossover, and mutation operations on the population based on the evolutionary strategy to generate offspring individuals. A high-fidelity verification and feedback calibration module is configured to monitor the prediction confidence of the physical perception agent prediction module during the evolutionary iteration process. When the prediction confidence is lower than a preset threshold or the evolutionary iteration reaches a preset interval of generations, the high-fidelity verification and feedback calibration module calls the finite element analysis solver to perform physical field simulation calculations on the preferred individuals in the current population, obtains the real physical field response value, and feeds back the deviation between the real physical field response value and the predicted physical field response value to the physical perception agent prediction module to update the weight parameters of the deep neural network model online. The post-processing output module receives the final converged optimal geometric parameter vector, transforms it into a triangular mesh model using a level set function, automatically identifies the overhanging region based on the manufacturing process constraints, generates the support structure path, and finally outputs the additive manufacturing slice file.

2. The intelligent design system for mechanical structures based on topological evolution optimization according to claim 1, characterized in that, The explicit geometric description method adopted by the global parametric modeling module is either the moving shape component method or the moving deformable cavity method; wherein, the geometric parameter vector includes the component's center coordinates, length, width, tilt angle, and polynomial shape control parameters; The global parametric modeling module is configured to directly map the minimum feature size constraint and the maximum overhang angle constraint to the geometry of the initial population by limiting the range of values ​​of the geometric parameter vector.

3. The intelligent design system for mechanical structures based on topological evolution optimization according to claim 2, characterized in that, The deep neural network model pre-set in the physical perception agent prediction module adopts a fully convolutional neural network architecture or a graph neural network architecture. The deep neural network model includes an encoder path and a decoder path. The encoder path is used to extract deep topological features of the input geometry, and the decoder path is used to map the extracted features back to the physical field space, directly outputting pixel-level or node-level von Mises stress distribution maps and displacement cloud maps.

4. The intelligent design system for mechanical structures based on topological evolution optimization according to claim 3, characterized in that, The multi-strategy co-evolutionary engine adopts a covariance matrix adaptive evolutionary strategy. The multi-strategy co-evolutionary engine is configured to calculate the fitness function of an individual based on the following formula. : , The structural flexibility predicted by the surrogate model, For structural volume, Penalties for violating manufacturing process constraints and These are the weighting coefficients.

5. The intelligent design system for mechanical structures based on topological evolution optimization according to claim 4, characterized in that, The high-fidelity verification and feedback calibration module monitors the prediction confidence level by: enabling Monte Carlo dropout or using deep ensemble technology during the inference process of the deep neural network model; performing multiple forward propagation inferences on the geometric parameter vector of the same entity and calculating the variance or standard deviation of the multiple inference results; using the variance or standard deviation as a quantitative indicator to measure the prediction confidence level; when the variance is greater than a preset threshold, it is determined that the prediction confidence level is low, and finite element analysis is triggered.

6. The intelligent design system for mechanical structures based on topological evolution optimization according to claim 5, characterized in that, When updating the weight parameters of the deep neural network model online, the high-fidelity verification and feedback calibration module adopts a transfer learning strategy: freezing the weight parameters of the shallow feature extraction layer of the deep neural network model, and using only the error between the real physical field response value and the predicted value, fine-tuning the weight parameters of the deep fully connected layer or output layer of the model through the backpropagation algorithm.

7. The global intelligent design system for mechanical structures based on topological evolution optimization according to claim 6, characterized in that, The global parametric modeling module also includes a Latin hypercube sampling unit; During the system initialization phase, the Latin hypercube sampling unit is used to uniformly generate a set of sample structures within the design space and call the finite element analysis solver to calculate their label data, which is used to construct an initial training dataset for pre-training the deep neural network model.

8. The intelligent design system for mechanical structures based on topological evolution optimization according to claim 7, characterized in that, The high-fidelity verification and feedback calibration module is also equipped with an anomaly removal mechanism; when the real physical field response value calculated by the finite element analysis solver shows that the structure has diverged or singular deformation, the individual is determined to be an invalid design, the data of the individual is not included in the update training set of the deep neural network model, and the individual is directly removed from the evolutionary population. The input to the physical sensing agent prediction module also includes an external load condition vector; the deep neural network model is configured to simultaneously receive geometric parameter vectors and external load condition vectors, thereby predicting the physical field response of the structure under different load boundary conditions.

9. A computer device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor; When the processor executes the computer program, it implements the global intelligent design system for mechanical structures based on topological evolution optimization as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the global intelligent design system for mechanical structures based on topological evolution optimization as described in any one of claims 1 to 8.

Images