Intelligent form-finding design method for large-span spatial curved special-shaped structure

Abstract

The application discloses a kind of large-span space curved surface special-shaped structure intelligent form finding design method, and the method comprises the following steps: constructing the initial three-dimensional parameterized model of curved surface special-shaped structure;Establish a hybrid loss function with structural mechanics performance index as the core;Intelligent form finding model is constructed based on conditional generative adversarial network, wherein the generator is responsible for the intelligent adjustment and variation of the initial model in curved surface form, and the discriminator synchronously judges the mechanical rationality, geometric feasibility and architectural expressiveness of the generated structure based on the hybrid loss function;Through the adversarial iterative training of generator and discriminator, the structure form is driven to be optimized based on mechanical properties, architectural form and other targets, and finally the ideal structure form is output.The application deeply integrates intelligent adversarial generation and reinforcement learning in structure form finding, and realizes the automatic, efficient and intelligent generation and performance-driven self-evolution of structure form under complex architectural constraints.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of building structure design and computer-aided design, and particularly relates to an intelligent form finding method for large-span spatial curved special-shaped structures by fusing artificial intelligence, generative adversarial network and reinforcement learning technology. BACKGROUND

[0002] Large-span spatial curved special-shaped structures, such as airport terminals, stadiums, cultural centers and the like, are typical forms of modern landmark buildings. In order to pursue unique architectural expressiveness, such structures often present complex spatial curved shapes such as continuous and smooth, hyperbolic, twisted and the like. For example, a double-layer net shell structure with a complex curved shape supported by a giant arch shell is adopted in a certain international airport terminal. The design core of such a "curved special-shaped structure" is "form finding", that is, finding a three-dimensional shape that is accurate to architectural creativity and has high mechanical performance and economic efficiency.

[0003] Traditional form finding methods, such as the force density method and the dynamic relaxation method, have obvious limitations: 1) the process is highly dependent on the experience of engineers, and a large number of manual iterations are required, which is low in efficiency and long in design cycle; 2) it is essentially "analysis-oriented" rather than "generation-oriented", and it is difficult to find a global optimal solution in a large shape solution space, and it is easy to fall into a local optimum; 3) for conditions with fixed support points (such as giant arch shells) and complex curved surface constraints, it is difficult for traditional methods to systematically explore the complex mapping relationship between shape and performance, and often sacrifices structural efficiency in order to meet architectural modeling.

[0004] In recent years, generative artificial intelligence has made breakthrough progress in the field of image and video generation. Through the adversarial game between the generator and the discriminator, the generative adversarial network can learn the distribution from the data and create highly realistic and diverse new samples. This provides a new paradigm for structure form finding, which regards the generation of structure shape as a "creative" process that meets the laws of mechanics, and automatically explores the shape with high mechanical performance and architectural aesthetics through adversarial learning of AI.

[0005] The information disclosed in this BACKGROUND section is only intended to enhance the understanding of the general background of the application, and should not be construed as recognition or any form of suggestion that this information constitutes prior art known to those of ordinary skill in the art. SUMMARY

[0006] The present application aims to overcome the shortcomings of the prior art and provide an efficient, intelligent and performance-driven form finding method for large-span spatial curved special-shaped structures, which can deeply integrate high-dimensional mechanical performance requirements and intelligent adversarial generation technology to automatically generate a structure shape with near-optimal mechanical performance under complex architectural and geometric constraints.

[0007] In order to achieve the above object, the present application adopts the following technical solutions:

[0008] The present application provides an intelligent form finding design method for large-span spatial curved surface special-shaped structures, comprising the following steps:

[0009] S1: parameterized model construction, based on the architectural concept scheme, extracting the key control parameters of the structure, constructing a parameterized initial three-dimensional structure model, the initial three-dimensional structure model containing a mathematical control network describing a complex curved surface;

[0010] S2: mixed loss function establishment, establishing a comprehensive mixed loss function L total , which at least contains a mechanical performance loss term L mech , a geometric constraint loss term L geo and an architectural effect loss term L aes , i.e. total =α·L mech +β·L geo +γ·L aes , wherein α, β, γ are weight coefficients;

[0011] S3: intelligent form finding model construction, constructing an intelligent form finding model based on a conditional generative adversarial network, the model comprising:

[0012] Generator G: taking a random noise vector and the parameters of the initial three-dimensional structure model as conditional input, generating a new structure model with morphological variation on the curved surface through a deep learning network;

[0013] Discriminator D: taking the structure model output by the generator or the real optimal structure case as input, outputting a comprehensive score based on the mixed loss function, for judging the advantages and disadvantages of the structure in mechanical performance, geometric feasibility and architectural effect;

[0014] S4: adversarial iterative training and performance-driven evolution, inputting the initial three-dimensional structure model into the intelligent form finding model, through multiple adversarial training of the generator and the discriminator, and introducing a reinforcement learning mechanism based on policy gradient, taking the score of the discriminator as a reward signal to drive the generator to explore the morphological space with better mechanical performance, and optimize the discrimination ability of the discriminator;

[0015] S5: optimal form output, when the training process converges, applying the optimized generator to the initial model to output the final structure form with the best comprehensive mechanical performance and architectural aesthetics under the given constraint conditions.

[0016] Further, the mechanical performance loss term L mechThrough a pre-trained agent model, the prediction value of the key mechanical performance index is directly output, the key mechanical performance index including the maximum equivalent stress, the maximum displacement, the buckling critical load and the structural strain energy of the structure.

[0017] Further, the generator in the step S3 adopts a convolutional neural network or a graph neural network with a U-Net architecture, the input of the generator being a voxelized or point clouded structure model, and the output being a new structure model data after morphological adjustment.

[0018] Further, a reinforcement learning mechanism is introduced in the step S4, specifically, the generator is regarded as an agent executing a policy, the comprehensive score of the generated structure given by the discriminator is regarded as a reward given by the environment, and the parameters of the generator are updated by a policy gradient method to maximize the cumulative expected reward, so that the directional exploration to the high-performance morphological space is realized.

[0019] Further, in the step S1, the key control parameters include the control point coordinates and weights of the NURBS surface, the topological connection relationship of the bifurcated column, the section size library of the rod, and the boundary conditions of the structure.

[0020] By adopting the above technical solutions, the present application has the following beneficial effects:

[0021] Leading performance-driven design, the mechanical performance index is directly taken as the optimization target of AI through the agent model and the loss function, ensuring the structural efficiency of the generated result, and realizing the leap from "constructable" to "optimal construction".

[0022] Intelligent morphological exploration, the combination of the adversarial mechanism of GAN and the directional exploration of reinforcement learning can efficiently search the morphological solution space far beyond the range of human experience, and easily find innovative and excellent "super-normal forms".

[0023] Multi-objective coordination, through the hybrid loss function, the structure performance, material consumption, construction feasibility and architectural effect and other multiple objectives are naturally unified in one framework for optimization, and the best balance point is automatically found.

[0024] High-efficiency design, which liberates the designer from the tedious "modeling-analysis-adjustment" cycle, realizes algorithm-driven design, and greatly shortens the engineering design cycle.

[0025] Strong pertinence: especially suitable for solving the form-finding problems of modern space structures with complex curved surfaces, special shapes, large spans and other characteristics, such as a certain coastal international airport. BRIEF DESCRIPTION OF DRAWINGS

[0026] In order to more clearly illustrate the technical solutions in the specific embodiments of the present application or the prior art, the accompanying drawings needed to be used in the description of the specific embodiments or the prior art will be briefly introduced. Obviously, the accompanying drawings in the following description are some embodiments of the present application, and all other embodiments obtained by a person of ordinary skill in the art without creative work based on these drawings also belong to the protection scope of the present application.

[0027] Figure 1 is a flowchart of the intelligent form-finding method for large-span spatial curved surface special-shaped structures based on mechanical properties provided by the present application;

[0028] Figure 2 is a flowchart of step S2 of the intelligent form-finding method for large-span spatial curved surface special-shaped structures based on mechanical properties provided by the present application;

[0029] Figure 3 is a flowchart of step S3 of the intelligent form-finding method for large-span spatial curved surface special-shaped structures based on mechanical properties provided by the present application;

[0030] Figure 4 is a flowchart of step S4 of the intelligent form-finding method for large-span spatial curved surface special-shaped structures based on mechanical properties provided by the present application;

[0031] Figure 5 is a flowchart of step S5 of the intelligent form-finding method for large-span spatial curved surface special-shaped structures based on mechanical properties provided by the present application;

[0032] Figure 6 is a form-finding demonstration example of the intelligent form-finding method for large-span spatial curved surface special-shaped structures based on mechanical properties provided by the present application in the structural system of Tianjin Airport T3 terminal building. DETAILED DESCRIPTION

[0033] The technical solutions of the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work also belong to the protection scope of the present application.

[0034] The specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application.

[0035] The application provides an intelligent form finding method for large-span spatial curved surface special-shaped structures based on mechanical properties, which comprises the following steps: constructing an initial three-dimensional parametric model containing a curved surface special-shaped structure; establishing a hybrid loss function with structural mechanical performance indicators as the core; constructing an intelligent form finding model based on a conditional generative adversarial network, wherein the generator is responsible for intelligent mutation of the curved surface form of the initial model, and the discriminator simultaneously evaluates the mechanical rationality, geometric feasibility and architectural expressiveness of the generated structure based on the hybrid loss function; through adversarial iterative training of the generator and the discriminator, the structure form is driven to evolve towards the Pareto frontier with optimal mechanical properties and beautiful architectural form, and finally the optimal structure form is output. The application deeply integrates intelligent adversarial generation and reinforcement learning into structure form finding, realizes automatic, efficient and intelligent generation and performance-driven self-evolution of the structure form under complex architectural constraints, and is particularly suitable for the design of large-span spatial curved surface special-shaped structures such as airport terminals and stadiums.

[0036] The intelligent form finding design method for large-span spatial curved surface special-shaped structures provided by the application is characterized in that an intelligent generative adversarial network driven by mechanical properties is constructed, and the process is as shown in Figure 1 , which comprises the following steps:

[0037] S1: Parametric model construction:

[0038] Key control parameters of the structure are defined according to the architectural scheme by using a parametric design platform (such as Rhino / Grasshopper). For a double-layer curved surface reticulated shell structure supported by bifurcated columns, the parameters include: coordinates (Px, Py, Pz) and weights of NURBS surface control points, a topological connection matrix of bifurcated columns, thicknesses and spacings of upper and lower reticulated shells, a library of sectional sizes of members, and boundary support conditions of the structure. By adjusting these parameters, a series of three-dimensional models with different forms but conforming to the basic architectural concept can be generated as the basic data for intelligent form finding.

[0039] S2: Hybrid loss function establishment:

[0040] A hybrid loss function L total is established, as shown in Figure 2 . It is a navigation system driving the entire AI model to evolve towards the optimal solution space. The function is a weighted sum of multiple losses with clear physical meanings:

[0041] Mechanical performance loss L mech : This is the core of the application. A lightweight neural network (i.e., a surrogate model) is trained offline to replace the cumbersome finite element analysis, quickly predicting the maximum equivalent stress σ_max, the maximum displacement d_max and the buckling critical load coefficient λ of the structure. L mech can be designed as: L mech= w1 * max(0, (σ_max / σ_allow -1)) + w2 * max(0, (d_max / d_allow - 1)) - w3 * λ, where σ_allow and d_allow are material allowable stress and code allowable displacement, w1, w2, w3 are weights, this design encourages the structure to pursue higher stability while meeting the strength and stiffness requirements.

[0042] Geometric constraint loss L geo : Ensure that the generated structure meets the geometric and construction requirements, such as the minimum clearance between members, the constructability of node domains, the fixed position constraints of arch shell support points, and the smoothness of curved surfaces (measured by the rate of curvature change).

[0043] Architectural effect loss L aes : By calculating the similarity (such as Chamfer Distance) between the generated structure and the target reference surface (such as the initial concept model) provided by the architect in the control point coordinates or point cloud level, and combining the Gaussian curvature distribution characteristics of its surface, it is defined to ensure that the result does not deviate from the architectural intent.

[0044] S3: Intelligent form finding model construction (based on cGAN):

[0045] Construct a conditional generative adversarial network model, whose structure is shown in Figure 3 .

[0046] Generator G: Convolutional Neural Network or Graph Neural Network with U-Net structure. It takes the voxelized data or point cloud data of the initial structure model and a random noise vector z as input. Noise z provides randomness and diversity of morphological variation. The generator outputs a new structure model G(z|initial_model) with intelligent adjustment of surface morphology through a series of encoding and decoding operations.

[0047] Discriminator D: Convolutional Neural Network or Graph Neural Network. It receives a structure model (from the generator or a pre-set excellent structure case library) and outputs a scalar score D(X). This score not only represents the probability that the structure "looks like a real excellent structure", but more importantly, one of its training goals is to accurately evaluate and reflect the L total > value of the input structure, that is, it is a "expert discriminator" who understands mechanics and codes.

[0048] S4: Adversarial iterative training and performance-driven evolution:

[0049] This process is an AI "self-game" and competitive evolution process, as shown in Figure 4 .

[0050] Generation: The generator intelligently mutates the initial model based on the current strategy to generate a batch of new structural configurations.

[0051] Evaluation: The discriminator evaluates the batch of new structures and the pre-set excellent structure cases, giving a comprehensive score

[0052] Adversarial and evolutionary update:

[0053] (1) Update the discriminator D: The goal is to maximize its ability to distinguish between "excellent cases" and "generated structures", while associating the discriminator's score with lower L total values for excellent cases.

[0054] (2) Update the generator G: The invention introduces a policy gradient method in reinforcement learning. The generator is regarded as an agent, and its generation behavior is a policy. The comprehensive score given by the discriminator is the reward obtained by the behavior. The goal of the generator update is to adjust its parameters to maximize the probability of generating high-reward (i.e. high-performance) structural configurations. This approach can effectively guide the generator to conduct directional exploration in a wide configuration space.

[0055] Iterate steps 1-3 until the generator produces stable structural configurations that meet or exceed the pre-set convergence criteria in terms of mechanical and aesthetic performance.

[0056] S5: Optimal configuration output:

[0057] After training convergence, the optimal generator model is applied to the initial conceptual model, which can output the final structural configuration, as shown in Figure 5 . This configuration is a high-mechanical-efficiency and aesthetically-pleasing structural configuration obtained through massive intelligent exploration and performance-driven evolution under given multiple constraints.

[0058] The invention will be described in detail below in conjunction with an example of a double-layer curved surface net shell structure of a giant arch shell support of a certain airport T3 terminal.

[0059] Implementation preparation: Establish a parametric model in the Rhino / Grasshopper environment. Use NURBS surface to define the roof shape, and adjust the control points and their weights to control the surface modeling. Define the base point position, branching logic and angle of the branching column. Output the model as point cloud data or convert it to voxel grid as input data for the AI model.

[0060] Model training:

[0061] In the Python environment, use the PyTorch deep learning framework to build a cGAN model.

[0062] Generator G: A 3D U-Net is adopted, taking 100-dimensional noise z and 128x128x128 voxels data of initial model as input, and outputting the same size of voxels data representing the new form.

[0063] Discriminator D: A 3D CNN is adopted, using spectral normalization to stabilize the training process, and finally outputting a comprehensive score.

[0064] Surrogate model: A separate 3D CNN regression model is trained in advance using 50,000 structure samples generated by parameterization method and their mechanical indicators (σ_max, d_max, λ) analyzed by finite element method. This model is frozen during the training of cGAN and used to quickly calculate L mech .

[0065] Training process: Set the batch size to 8, use the Wasserstein GAN with Gradient Penalty (WGAN-GP) loss to enhance the stability of training. Train the discriminator and generator alternately. Use the Adam optimizer. Train for 20,000 epochs. For the update of the generator, update the generator once using the policy gradient method for every 5 times of training the discriminator.

[0066] Form finding process: input the prepared initial model of the airport terminal into the trained system. The system runs the processes of S1 to S4 in the background. Engineers can monitor the evolution of the generated form and the change curve of its key mechanical indicators (such as maximum stress) in real time through the visualization interface.

[0067] Result output and verification: After training, output the final form as shown in Figure 6 Import the form result into the commercial finite element software MIDAS for high-precision verification analysis. Practice shows that the structure form generated by the method has better mechanical performance than the form obtained by traditional methods under the premise of meeting all constraint conditions, and has an optimization potential of 5%-15% in material usage.

[0068] In summary, the present application provides a new, efficient and intelligent solution for the design of large-span spatial curved special-shaped structures by deep coupling of AI adversarial generation, reinforcement learning and structural mechanical performance, which has broad engineering application prospects and important industrial value.

[0069] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions recorded in the above embodiments can be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. An intelligent form-finding design method for large-span spatial curved surface special-shaped structures, characterized in that, The method comprises the following steps: S1: parameterized model construction, based on the architectural conceptual scheme, extracting the key control parameters of the structure, constructing a parameterized initial three-dimensional structure model, and the initial three-dimensional structure model comprises a mathematical control network describing a complex curved surface; S2: mixed loss function establishment, a comprehensive mixed loss function is established L total, the function at least contains a mechanical performance loss term L mech, a geometric constraint loss term L geo and a building effect loss term L aes, that is L total =α· L mech+β· L geo+ γ· L aes, wherein α, β, γ are weight coefficients; S3: intelligent form-finding model construction, constructing an intelligent form-finding model based on a conditional generative adversarial network, the model comprising: a generator G: taking a random noise vector and the parameters of the initial three-dimensional structure model as conditional input, generating a new structure model with morphological variation through a deep learning network; a discriminator D: taking the structure model output by the generator or a real optimal structure case as input, outputting a comprehensive score based on the hybrid loss function, for judging the advantages and disadvantages of the structure in terms of mechanical performance, geometric feasibility and architectural effect; S4: adversarial iterative training and performance-driven evolution, inputting the initial three-dimensional structure model into the intelligent form-finding model, through multiple adversarial training of the generator and the discriminator, and introducing a reinforcement learning mechanism based on policy gradient, taking the score of the discriminator as a reward signal to drive the generator to explore the morphological space with better mechanical performance, and to optimize the discrimination ability of the discriminator; S5: optimal form output, when the training process converges, applying the optimized generator to the initial model to output the final structure form with the best comprehensive mechanical performance and architectural aesthetics under the given constraint conditions; Establishing a hybrid loss function L total, which is a weighted sum of multiple physically meaningful losses: Loss of mechanical properties L mech: replace the complicated finite element analysis with a lightweight neural network trained offline to quickly predict the maximum equivalent stress σ max , maximum displacement d max and buckling critical load factor λ of a structure; L mech is designed as: L mech = w1 * max(0, (σ max / σ allow - 1)) + w2 * max(0, (d max / d allow - 1)) - w3 * λ, where σ allow and d allow are the material allowable stress and code allowable displacement, w1, w2, w3 are weights, this design encourages the structure to pursue higher stability while meeting the strength and stiffness requirements; Geometric constraint loss L geo : ensures that the generated structure meets the geometric and construction requirements; Loss of architectural effect L aes : The similarity between the generated structure and the target reference surface provided by the architect at the control point coordinates or point cloud level is calculated, and the Gaussian curvature distribution characteristics of the surface are combined to define it, ensuring that the result does not violate the architectural intent.

2. The intelligent form finding design method of large-span spatial curved surface special-shaped structure according to claim 1, characterized in that, The mechanical property loss term L mech The key mechanical performance indexes including the maximum equivalent stress, the maximum displacement, the buckling critical load and the structural strain energy are directly outputted by a pre-trained agent model which takes the control point coordinates of the structural form and the cross section of the rod as inputs.

3. The intelligent form finding design method of large-span spatial curved surface special-shaped structure according to claim 1, characterized in that, The generator in step S3 adopts a convolutional neural network or a graph neural network with a U-Net architecture, the input of the generator is a voxelized or point clouded structure model, and the output is a new structure model data after morphological adjustment.

4. The intelligent form finding design method of large-span spatial curved surface special-shaped structure according to claim 1, characterized in that, In step S4, a reinforcement learning mechanism is introduced, specifically: the generator is regarded as an intelligent agent executing a policy, the comprehensive score of the discriminator on the generated structure is regarded as the reward given by the environment, the parameters of the generator are updated through a policy gradient method to maximize the cumulative expected reward, so as to realize directional exploration to the high-performance morphological space.

5. The intelligent form finding design method of large-span spatial curved surface special-shaped structure according to claim 1, characterized in that, In step S1, the key control parameters include the control point coordinates and weights of the NURBS curved surface, the topological connection relationship of the bifurcated column, the section size library of the rod, and the boundary conditions of the structure.

Images