A complex mechanical structure multidisciplinary design optimization method based on generative AI

Abstract

The application discloses a complex mechanical structure multidisciplinary design optimization method based on generative AI, and belongs to the technical field of engineering design and artificial intelligence. The method comprises the following steps: analyzing the multidisciplinary performance requirements and constraints of a complex mechanical structure during service, and establishing an optimization design model; performing parameterized modeling on the complex mechanical structure, and establishing a finite element model of multi-physical field coupling; sampling in a design space based on the finite element model and performing simulation, and constructing a dataset containing design variables and multidisciplinary performance; constructing and training a generative AI reverse design model embedded with constraints based on the optimization design model and the dataset, the generative AI reverse design model learning reverse mapping from a performance constraint space to a design space; and applying the generative AI reverse design model to generate a candidate design scheme satisfying the constraints. The application considers the coupling effect between multi-physical fields, integrates multidisciplinary constraints into a generative artificial intelligence model, and establishes an intelligent optimization model capable of directly generating a design scheme in reverse from performance requirements, which can systematically and efficiently output design schemes under different constraints, thereby providing an effective data-driven method for innovative design, performance improvement and research and development cycle shortening of complex engineering systems.

Description

Technical Field

[0001] This application belongs to the field of engineering design and artificial intelligence technology, and in particular relates to a multidisciplinary design optimization method for complex mechanical structures based on generative AI. Background Technology

[0002] The design of complex mechanical structures, such as aircraft, vehicles, ships, and large equipment, is essentially an optimization problem involving deep coupling of multiple disciplines, including aerodynamics, structure, materials, and dynamics. Traditional design optimization methods employ a simulation-driven optimization paradigm: first, a parametric model is established; then, high-fidelity multiphysics simulation is used to evaluate the design performance; and finally, an optimizer based on gradient or evolutionary algorithms is used for iterative search. This paradigm faces two major bottlenecks: first, high-fidelity simulation itself is computationally expensive, and the optimization process requires thousands of iterative evaluations, resulting in huge consumption of computational resources and a long design cycle, severely limiting the systematic exploration of high-dimensional, nonlinear design spaces; second, traditional optimization algorithms struggle to efficiently handle complex and interdependent multidisciplinary inequality constraints, often requiring the introduction of complex penalty functions or feasibility rules, which increases the complexity of the problem. Particularly noteworthy is that when design requirements change, the entire optimization process needs to be repeated from scratch, making the initial computational investment unreusable and failing to meet the needs of engineering scenarios where requirements change rapidly in the initial design phase.

[0003] To reduce simulation costs, researchers have introduced surrogate models to approximate the performance response function. While this has accelerated optimization iterations to some extent, it hasn't fundamentally changed the aforementioned paradigm. In recent years, generative artificial intelligence, particularly conditional generative adversarial networks (GANs), has provided new ideas for performance-driven reverse design. However, existing reverse design research based on generative AI largely focuses on achieving precise matching of a single, fixed performance objective, or using multidisciplinary constraints as posterior selection conditions for generated solutions. The core limitation of these methods lies in their failure to embed multidimensional, inequality-based performance constraint boundaries into the model training process. Summary of the Invention

[0004] The purpose of this application is to overcome the shortcomings of existing technologies and provide a multidisciplinary design optimization method for complex mechanical structures based on generative AI. This method constructs an intelligent reverse design model that can be constrained by inequality, thereby achieving efficient and controllable generation from performance requirements to design solutions, and significantly improving the exploration efficiency and quality of complex mechanical structures in the conceptual design stage.

[0005] The objective of this application is achieved through the following technical solution: A multidisciplinary design optimization method for complex mechanical structures based on generative AI, the method comprising: The multidisciplinary performance requirements and constraints of complex mechanical structures during service are analyzed, and an optimization design model is established. Parametric modeling of the complex mechanical structure was performed to establish a multi-physics coupled finite element model; Based on the finite element model, sampling and simulation are performed in the design space to construct a dataset containing design variables and multidisciplinary performance. Based on the dataset, a generative AI reverse design model with embedded constraints is constructed and trained. The generative AI reverse design model learns the reverse mapping from the performance constraint space to the design space. The generative AI reverse design model is applied to generate candidate design schemes that satisfy the constraints.

[0006] Furthermore, the multidisciplinary performance requirements and constraints include aerodynamic performance indicators and constraints, structural static performance indicators and constraints, dynamic characteristic indicators and constraints, and fatigue and durability indicators and constraints.

[0007] Furthermore, the establishment of the optimization design model includes: The performance index and constraints are divided into at least two conflicting optimization objectives and a constraint objective that serves as an inequality constraint, thus obtaining an optimization design model.

[0008] Furthermore, the parametric modeling of the complex mechanical structure and the establishment of a multi-physics coupled finite element model include: The complex mechanical structure is parametrically characterized to establish a unified design space that includes continuous variable vectors and discrete variable vectors; Based on the parametric characterization, a high-fidelity multiphysics coupled simulation analysis process is constructed. The analysis process uses automated solution to couple and calculate the performance parameters of different physical fields, thereby obtaining multidisciplinary performance indicators.

[0009] Furthermore, the step of sampling and simulating in the design space based on the finite element model to construct a dataset containing design variables and multidisciplinary performance includes: Based on a unified design space, a series of sample points covering continuous variable vectors and discrete variable vectors are generated; Based on the high-fidelity multiphysics coupling simulation analysis process, automated simulation calculations are performed on each sample point to obtain the corresponding multidisciplinary performance index data. By integrating the design variable vectors and corresponding multidisciplinary performance index vectors of all sample points, a design-performance mapping dataset is constructed.

[0010] Furthermore, the generative AI reverse design model built and trained based on the dataset and embedding constraints includes: Construct a generative AI reverse design model with embedded inequality constraints, wherein the AI ​​reverse design model includes a generator and a performance predictor; The generator takes as input a multidisciplinary performance constraint boundary vector and trade-off parameters and outputs the generated design variables. The input to the performance predictor is the design variables, and the output is the predicted performance metric. Define a composite loss function for the generator to guide training; Based on the dataset, the generator and performance predictor are jointly trained. Through training optimization, the generator learns the inverse mapping from any given performance constraint boundary and objective trade-off preference to the feasible design space, forming an intelligent core that can directly drive the design.

[0011] Furthermore, the composite loss function is composed of a weighted average of distribution regularization loss, inequality constraint loss, and multi-objective guided loss; The distribution regularization loss uses Sinkhorn divergence based on optimal transport theory to calculate the difference between the generated sample distribution and the real sample distribution in the training dataset under the multidisciplinary performance constraint boundary vector. Inequality constraint loss penalizes the degree to which the generated sample prediction performance violates the corresponding constraint boundaries. Multi-objective-oriented loss uses scalar weighting to balance conflicting optimization objectives through trade-off parameters.

[0012] Furthermore, the joint training of the generator and performance predictor based on the dataset includes: First, the performance predictor is trained to accurately estimate the performance of a given design. After training, real data pairs are sampled from the dataset as constraints, and the generator parameters are updated by minimizing the composite loss function.

[0013] Furthermore, the application of the generative AI reverse design model to generate candidate design solutions that satisfy the constraints includes: The trained generator is used as a design generation tool. The target constraint boundary and the trade-off parameters with different values ​​are input to generate a candidate design set. The Pareto optimal solution set is selected as the output optimal design scheme set.

[0014] The beneficial effects of this application are as follows: By embedding multidisciplinary inequality constraints into the training framework of a generative model, a direct mapping from performance requirements to design solutions is achieved. This method builds a dataset based on prior simulations, and after training, feasible solutions can be quickly generated through conditional sampling, significantly improving design efficiency. The model uses constraint boundaries as input, directly responding to changing design requirements without retraining or re-invoking simulations, demonstrating excellent cross-task reusability. Furthermore, by adjusting tradeoff parameters, the system can systematically generate and present Pareto optimal tradeoffs among multiple objectives, providing efficient and intuitive decision support for the design of complex mechanical structures. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating a multidisciplinary design optimization method for complex mechanical structures based on generative AI, using a complex composite material structure as an example, as an embodiment of this application. Figure 2 This is a schematic diagram of the generative AI model structure with embedded inequality constraints, which is the core of the embodiments of this application. Detailed Implementation

[0016] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0017] Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] Existing technologies cannot build an intelligent approach that can directly understand the boundaries of multidisciplinary constraints and quickly explore the design space.

[0019] To address the aforementioned technical problems, the following embodiments of a multidisciplinary design optimization method for complex mechanical structures based on generative AI are proposed in this application.

[0020] This embodiment uses a complex composite material structure as an example for illustration, referring to... Figure 1 ,like Figure 1 The diagram shown is a flowchart of a multidisciplinary design optimization method for complex mechanical structures based on generative AI, taking a complex composite material structure as an example, according to an embodiment of this application. The specific steps are as follows: S1. Analyze the multidisciplinary performance requirements and constraints of complex composite material structures, and establish an optimization design model.

[0021] By systematically analyzing the multi-physics field coupling environment of complex composite material structures under service conditions, including structure and fatigue, we clarify their key performance indicators and constraint boundaries. We construct an optimization design mathematical model with the core objectives of improving structural stiffness and reducing structural weight, while also satisfying performance constraints from multiple disciplines. This provides a clear mathematical description and optimization framework for exploring the design space.

[0022] S2. Parametric modeling of complex composite material structures is performed to establish a high-fidelity multi-physics coupled finite element model.

[0023] By parametrically characterizing the geometry and material properties of complex composite material structures, a high-fidelity simulation model of its multi-physics coupling is constructed to calculate the corresponding performance indicators.

[0024] S3. Sample and perform large-scale simulations in the design space of complex composite material structures to construct a dataset that includes design variables and multidisciplinary performance.

[0025] Based on the parameterized model established in step S2, the Latin hypercube experimental design method is used to sample within the design space, driving high-fidelity multiphysics coupling simulation for automated calculation, and finally constructing a structured dataset corresponding to design variables and multidisciplinary performance indicators.

[0026] S4. Construct and train a generative AI reverse design model with embedded constraints.

[0027] Based on the optimization design model established in step S1 and the structured dataset constructed in step S3, a generative AI model is constructed with multidisciplinary performance constraint boundaries and objective trade-off parameters of complex composite material structures as input conditions. The training loss function of the model integrates Sinkhorn divergence distribution matching, inequality constraint violation penalty based on the ReLU function, and multi-objective weighted guided loss, enabling it to learn the inverse mapping from the performance constraint space to the design space.

[0028] S5. Generative AI models are used to generate candidate design schemes for complex composite material structures that meet the constraints.

[0029] The generative AI model trained in step S4 is used as the core design algorithm. The target performance constraint boundary is input, and a set of candidate design schemes for complex composite material structures that satisfy all constraints is generated in batches through conditional sampling. The output is then used for design decisions.

[0030] In this embodiment, step S1 is as follows: S11. Determine the multidisciplinary performance indicators and constraints, which specifically cover the following three dimensions of indicators and constraints: (1) Structural static performance indicators and constraints, including structural weight m, maximum structural deformation, etc. Inverse reserve factor of composite material failure strength (2) Dynamic performance indicators and constraints, including first-order natural frequencies (3) Fatigue and durability indicators and constraints, including fatigue life ; S12. Divide the performance index into two conflicting optimization objectives. , and the constraint objective as an inequality constraint Construct a multidisciplinary mathematical model for optimal design: ; in, Here, x represents the constraint boundary value corresponding to the performance index, and x is the design variable vector.

[0031] In this embodiment, step S2 is as follows: S21. Perform parametric characterization on complex mechanical structures and establish a vector containing continuous variables. with discrete variable vector Unified design space Among them, all design variables are Continuous variables are used to characterize the geometric shape of complex mechanical structures, while discrete variables are used to characterize the layup scheme of composite materials.

[0032] S22. Based on the parametric model, a high-fidelity multi-physics coupled simulation analysis process is constructed. This process automatically solves the coupled calculation of performance parameters of different physical fields, and then solves the static and dynamic response of the structure and the damage of composite materials, so as to simultaneously obtain the multi-disciplinary performance indicators of structure, vibration and fatigue.

[0033] In this embodiment, step S3 is as follows: S31. Design space defined based on parametric model The Latin hypercube experimental design method was used to generate a series of sample points covering the range of values ​​for both continuous and discrete variables. S32. Based on the high-fidelity multiphysics field coupling simulation analysis process, perform automated simulation calculations on each sample point to obtain its corresponding multidisciplinary performance index data. S33. Integrate the design variable vectors of all sample points and their corresponding multidisciplinary performance index vectors to construct a structured "design-performance" mapping dataset. It is used for the training and validation of subsequent generative artificial intelligence models.

[0034] In this embodiment, step S4 is as follows: S41. Based on the optimization design model, construct a generative artificial intelligence inverse design model with embedded inequality constraints. The model consists of a generator G and a performance predictor R; the input of the generator G is the boundary vector of multidisciplinary performance constraints. and trade-off parameters The output is the generated design variables. The performance predictor R takes the design variable x as input and outputs the predicted performance metrics. and ; S42. Define the composite loss function for generator G. The composite loss function is used to guide its training. It consists of the following three weighted components: (1) Distribution regularization loss Using Sinkhorn divergence based on optimal transport theory, the distribution of generated samples and the distribution of real samples in the training dataset are calculated under certain conditions. The differences are used to ensure the reasonable distribution of the generated samples; (2) Inequality constraint loss : Prediction performance of generated samples Violation of corresponding constraint boundary The degree of punishment is determined by the expression: ,in It is a linear rectified function; (3) Multi-objective-oriented loss By weighing parameters The optimization objective of the conflict is weighted by a scalar to guide the generator to search in the direction of preference; its expression is: ; The composite loss function is ,in These are the weighting coefficients for each loss term. (Refer to...) Figure 2 ,like Figure 2 The diagram shown is a schematic diagram of the generative AI model structure with embedded inequality constraints, which is the core of this application embodiment.

[0035] S43. Based on a structured dataset, jointly train the generator G and the performance predictor R; prioritize training the performance predictor R to accurately estimate the performance of a given design; after training, sample real data from the dataset to compare the results. As a constraint, by minimizing the composite loss function Update the generator parameters; optimize through training so that the generator G can learn from any given performance constraint boundary. Trade-offs with objectives To feasible design space The robust inverse mapping forms an intelligent core that can directly drive the design.

[0036] In this embodiment, step S5 is as follows: The trained generator G is used as the design generation tool, with the target constraint boundary as input. Trade-off parameters with different values Generate a candidate design set The Pareto front is then selected as the optimal design scheme set for output.

[0037] This embodiment achieves a direct mapping from performance requirements to design solutions by embedding multidisciplinary inequality constraints into the training framework of a generative model. Based on a dataset built from prior simulations, this method can quickly generate feasible solutions through conditional sampling after training, significantly improving design efficiency. The model uses constraint boundaries as input, enabling it to directly respond to changing design requirements without retraining or re-invoking simulations, demonstrating excellent cross-task reusability. Furthermore, by adjusting tradeoff parameters, it can systematically generate and present Pareto optimal tradeoffs among multiple objectives, providing efficient and intuitive decision support for the design of complex mechanical structures.

[0038] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A multidisciplinary design optimization method for complex mechanical structures based on generative AI, characterized in that, The method includes: The multidisciplinary performance requirements and constraints of complex mechanical structures during service are analyzed, and an optimization design model is established. Parametric modeling of the complex mechanical structure was performed to establish a multi-physics coupled finite element model; Based on the finite element model, sampling and simulation are performed in the design space to construct a dataset containing design variables and multidisciplinary performance. Based on the optimization design model and the dataset, a generative AI reverse design model with embedded constraints is constructed and trained. The generative AI reverse design model learns the reverse mapping from the performance constraint space to the design space. The generative AI reverse design model is applied to generate candidate design schemes that satisfy the constraints.

2. The multidisciplinary design optimization method for complex mechanical structures based on generative AI as described in claim 1, characterized in that, The multidisciplinary performance requirements and constraints include aerodynamic performance indicators and constraints, structural static performance indicators and constraints, dynamic characteristic indicators and constraints, and fatigue and durability indicators and constraints.

3. The multidisciplinary design optimization method for complex mechanical structures based on generative AI as described in claim 1, characterized in that, The establishment of the optimization design model includes: The performance index and constraints are divided into at least two conflicting optimization objectives and a constraint objective that serves as an inequality constraint, thus obtaining an optimization design model.

4. The multidisciplinary design optimization method for complex mechanical structures based on generative AI as described in claim 1, characterized in that, The parametric modeling of the complex mechanical structure and the establishment of a multi-physics coupled finite element model include: The complex mechanical structure is parametrically characterized to establish a unified design space that includes continuous variable vectors and discrete variable vectors; Based on the parametric characterization, a high-fidelity multiphysics coupled simulation analysis process is constructed. The analysis process uses automated solution to couple and calculate the performance parameters of different physical fields, thereby obtaining multidisciplinary performance indicators.

5. The multidisciplinary design optimization method for complex mechanical structures based on generative AI as described in claim 4, characterized in that, The process of sampling and simulating in the design space based on the finite element model to construct a dataset containing design variables and multidisciplinary performance includes: Based on a unified design space, a series of sample points covering continuous variable vectors and discrete variable vectors are generated; Based on the high-fidelity multiphysics coupling simulation analysis process, automated simulation calculations are performed on each sample point to obtain the corresponding multidisciplinary performance index data. By integrating the design variable vectors and corresponding multidisciplinary performance index vectors of all sample points, a design-performance mapping dataset is constructed.

6. The multidisciplinary design optimization method for complex mechanical structures based on generative AI as described in claim 1, characterized in that, The generative AI reverse design model built and trained based on the dataset and embedding constraints includes: Construct a generative AI reverse design model with embedded inequality constraints, wherein the AI ​​reverse design model includes a generator and a performance predictor; The generator takes as input a multidisciplinary performance constraint boundary vector and trade-off parameters and outputs the generated design variables. The input to the performance predictor is the design variables, and the output is the predicted performance metric. Define a composite loss function for the generator to guide training; Based on the dataset, the generator and performance predictor are jointly trained. Through training optimization, the generator learns the inverse mapping from any given performance constraint boundary and objective trade-off preference to the feasible design space, forming an intelligent core that can directly drive the design.

7. The multidisciplinary design optimization method for complex mechanical structures based on generative AI as described in claim 6, characterized in that, The composite loss function is composed of a weighted sum of distribution regularization loss, inequality constraint loss, and multi-objective guided loss. The distribution regularization loss uses Sinkhorn divergence based on optimal transport theory to calculate the difference between the generated sample distribution and the real sample distribution in the training dataset under the multidisciplinary performance constraint boundary vector. Inequality constraint loss penalizes the degree to which the generated sample prediction performance violates the corresponding constraint boundaries. Multi-objective-oriented loss uses scalar weighting to balance conflicting optimization objectives through trade-off parameters.

8. The multidisciplinary design optimization method for complex mechanical structures based on generative AI as described in claim 6, characterized in that, The joint training of the generator and performance predictor based on the dataset includes: First, the performance predictor is trained to accurately estimate the performance of a given design. After training, real data pairs are sampled from the dataset as constraints, and the generator parameters are updated by minimizing the composite loss function.

9. The multidisciplinary design optimization method for complex mechanical structures based on generative AI as described in claim 6, characterized in that, The application of the generative AI reverse design model to generate candidate design solutions that satisfy the constraints includes: The trained generator is used as a design generation tool. The target constraint boundary and the trade-off parameters with different values ​​are input to generate a candidate design set. The Pareto optimal solution set is selected as the output optimal design scheme set.

Images