Building design method and system based on generative adversarial network

Abstract

The invention discloses a building design method and system based on a generative adversarial network, and the method comprises the following steps: S1, obtaining the initial input data of building design, including function requirements, design styles, site conditions, regulation requirements and reference drawings; s2, extracting key features in the architectural drawing by using a self-organizing mapping network; s3, performing multi-generation optimization on the feature mapping mode through the evolutionary self-organizing mapping network; s4, the optimized feature mapping is input into a generative adversarial network generator to generate a new design scheme, a discriminator evaluates the generation scheme and feeds back the generation scheme to the generator, and a generation strategy is continuously optimized; through multi-round adversarial training and initialized evolutionary optimization of the self-organizing mapping network, the quality of the design scheme is gradually improved; and selecting an optimal scheme and outputting the optimal scheme in a visual form. The self-organizing mapping network is used for extracting key features from a reference architectural drawing, optimization is carried out through a genetic algorithm, a new design scheme is generated, and the efficiency and quality of architectural design are improved.

Description

Technical Field

[0001] The present invention relates to the field of neural network technology, and in particular to an architectural design method and system based on a generative adversarial network. Background Art

[0002] With the continuous development of the construction industry, the architectural design process has become more and more complex and diverse. Traditional architectural design methods rely on the expertise and experience of architects and usually require a lot of time and resources to complete. Although this method has certain advantages in precision and creative expression, its efficiency is low, especially in the face of large-scale projects or when it is necessary to respond quickly to market changes. Traditional design methods seem to be unable to cope with it. In addition, there is a lot of manual intervention in the traditional architectural design process, and the quality and consistency of the design scheme are difficult to guarantee. Therefore, the field of modern architectural design urgently needs an intelligent method that can automatically generate high-quality design schemes.

[0003] In recent years, with the development of artificial intelligence technology, generative adversarial networks have become a powerful data generation tool and have achieved remarkable results in the fields of image generation, style transfer, and data enhancement. Generative adversarial networks can generate realistic data distributions through mutual adversarial learning between generators and discriminators. This technology has also been gradually applied to the field of architectural design to generate diverse design solutions. However, existing architectural design methods based on generative adversarial networks still have many shortcomings and challenges.

[0004] Defects of the prior art:

[0005] 1. Insufficient diversity and innovation of design solutions: Existing architectural design methods based on generative adversarial networks are often limited to the feature distribution of training data sets when generating design solutions, which leads to a lack of sufficient diversity and innovation in the generated architectural design solutions. This method is difficult to generate breakthrough ideas beyond existing design styles and cannot meet the growing diversification and personalization needs of the construction industry. Especially when faced with emerging architectural design trends and complex design requirements, the limitations of existing methods are particularly obvious.

[0006] 2. Unstable quality of generated design solutions: During the training process of the generative adversarial network, the adversarial learning between the generator and the discriminator may lead to unstable quality of the design solutions generated by the generator, resulting in mode collapse or inconsistent generation results. This phenomenon is particularly unfavorable in the field of architectural design, because architectural design has strict requirements on the rationality of the structure and the unity of aesthetics, and any unstable generation solution will affect practical applications.

[0007] 3. Lack of optimization for complex geometric structures and spatial layouts: Although the existing generative adversarial network models have certain advantages in generating two-dimensional floor plans and simple three-dimensional models, they have limited optimization capabilities for complex geometric structures and spatial layouts. Architectural design often involves complex spatial organization, multi-dimensional structural relationships, and fine geometric structures. Existing generative adversarial network models often show inadequacies in dealing with these problems, and the generated solutions are difficult to meet the requirements of practical applications.

[0008] 4. Weak responsiveness to design goals and requirements: Existing generative adversarial network models in architectural design often lack the ability to flexibly respond to specific design goals and requirements. Due to the lack of sufficient control mechanisms, the generated design solutions are difficult to fully meet the specific architectural design requirements, specific functional requirements, site restrictions, and regulatory compliance. The generation of design solutions lacks effective docking with actual application needs, which seriously limits the effectiveness of generative adversarial network technology in practical applications.

[0009] 5. Insufficient evolutionary optimization process: Although some methods try to combine evolutionary algorithms to optimize the design solutions generated by generative adversarial networks, these methods are usually limited to simple parameter adjustments and lack systematic evolutionary strategies and complex optimization tools. The potential of evolutionary algorithms has not been fully utilized in the field of architectural design, resulting in the quality of the generated design solutions being difficult to meet expectations. Especially when facing large-scale architectural design projects, the existing optimization methods seem to be unable to fully utilize the advantages of evolutionary optimization.

[0010] 6. Convergence problem of generative adversarial networks: During the training process of generative adversarial networks, the confrontation between the generator and the discriminator often makes it difficult for the training process to converge, especially when dealing with high-dimensional architectural design problems. The application of existing generative adversarial network models in architectural design often faces the challenge of generating solutions that are difficult to meet the expected quality standards, resulting in delays in the design process or failure to achieve the expected results.

[0011] 7. The practicality and feasibility of the generated solutions are insufficient: Although the existing architectural design solutions generated by generative adversarial networks are innovative in concept, they often face the problem of insufficient feasibility in practical applications. Due to the lack of comprehensive consideration of building physics, material properties, and construction feasibility factors, the generated solutions are difficult to implement in practical applications, resulting in a disconnect between design and actual needs.

[0012] Therefore, how to provide an architectural design method based on generative adversarial networks is an urgent problem that technicians in this field need to solve. Summary of the invention

[0013] One purpose of the present invention is to propose an architectural design method based on a generative adversarial network, which makes full use of the combination of a generative adversarial network and a self-organizing map network to generate a high-quality architectural design solution through genetic algorithm optimization and evolution, with a high level of automation and strong practicality.

[0014] The present invention is achieved through the following technical solutions.

[0015] A building design method based on a generative adversarial network comprises the following steps:

[0016] Obtain architectural design data, input the architectural design data into the trained generative adversarial network, screen multiple generated architectural design schemes based on design goals and requirements, and output the design scheme;

[0017] The training of a generative adversarial network consists of the following steps:

[0018] S1. Obtaining initial input data for architectural design, the initial input data including reference architectural drawing data;

[0019] S2, using genetic algorithm to initialize the network structure, weight vector, learning rate and neighborhood function of the self-organizing map network, and extracting features from the reference building drawing data through the self-organizing map network;

[0020] S3. Evolve the features for multiple generations by evolving the self-organizing map network, perform genetic selection, mutation, and crossover operations, and generate and gradually optimize the feature map of the self-organizing map network:

[0021] S4, inputting the optimized feature map into the generator of the generative adversarial network for multiple rounds of adversarial training, screening multiple generated architectural design schemes based on design goals and requirements, and outputting the design scheme, including the following steps:

[0022] The generator combines complex geometry-time crystal mapping, instanton effect modulation, and algebraic topology-spinor field perturbation algorithm to generate preliminary architectural design solutions;

[0023] During the adversarial training process of the generative adversarial network, the discriminator evaluates the generated architectural design solutions and feeds back the evaluation results to the generator;

[0024] In the process of multiple rounds of adversarial training and evolutionary optimization, the generator continuously optimizes the architectural design scheme based on the feedback information of the discriminator, and further optimizes the quality of each generation of design schemes by combining high-dimensional hypersurface functionals, non-commutative algebraic geometry, dynamical systems and symplectic geometry transformation algorithms;

[0025] S5. Output the final selected design solution into a visual form.

[0026] Preferably, step S2 comprises the following steps:

[0027] S21. Create a self-organizing map network: Use hyperelliptic curve cluster decomposition to set the initial self-organizing map network structure of the self-organizing map network, and generate an initial weight vector W by hyperelliptic curve cluster decomposition ij (0), the weight vector reflects the connection mode and initial distribution characteristics between the nodes of the self-organizing map network;

[0028] S22, determine the initial weight vector W ij The dimension d of (0) weight and the dimension d of the input feature space input ,The weight vector of each node in the self-organizing map network is initialized as a vector in a high-dimensional space;

[0029] S23, set the learning rate η(t) and neighborhood function h BMU,i (t), the learning rate η(t) is adaptively adjusted through nonlinear fractional calculus and then gradually converges. The neighborhood function h BMU,i (t) represents the distance between node i and the network node that best represents the input data features during the training process at time t;

[0030] The self-organizing map network adaptively adjusts the learning rate and neighborhood influence range during training to optimize the network's ability to respond to complex features;

[0031] S24, input the reference architectural drawing data into the self-organizing map network, and gradually adjust the weight vector W of each node through multiple rounds of training ij (t), the update of the weight vector combines the moduli space theory and the principle of quantum state superposition to adapt to the complex features of the reference architectural drawings and extract key features:

[0032]

[0033] in, represents the manifold mapping function on the moduli space, which is used to adjust the representation of the weight vector in the moduli space. The weight vector represents the spatial layout, structural form, facade design and architectural style characteristics in the architectural drawings. m is the quantum state superposition coefficient, ψ m (t) is the quantum state, W ij (Q) represents the weight vector of node i and input feature j at time step Q, represents the manifold mapping function on the moduli space, which is used to adjust the representation of the weight vector in the moduli space. M represents the number of nodes, which is the feature dimension used for optimization in the self-organizing map network.

[0034] S25, in each generation of evolution from step S21 to step S25, the network is globally optimized and mutated using the Riemann-Theta function, and the weight vector generated and adjusted from step S21 to step S25 is used to optimize the weight vector and the self-organizing map network structure through the following fitness function to obtain the extracted feature F optimal :

[0035] F optimal =Θ(z|Ω)·exp(-γ·∑ i,j ∥W ij (t)-W ij (0)∥);

[0036] Where Θ(z|Ω) is the Riemann-Theta function, z is the current state vector, Ω is the Riemann matrix, and γ is the attenuation coefficient; ∥W ij (t)-W ij (0)∥ is the norm of the weight vector, which is used to represent the degree of change of the weight vector of node i and input feature j between time step t and the initial time step 0, W ij (t) represents the weight vector of node i and input feature j at time step t, W ij (0) represents the weight vector of node i and input feature j at time step 0.

[0037] Preferably, step S3 comprises the following steps:

[0038] S31, based on the self-organizing map network optimized in step S2, the extracted features are evolved for multiple generations, and the weight vector W is adjusted by combining the non-stationary spatiotemporal model ij (t) Perform multi-scale analysis and optimization:

[0039]

[0040] in, represents the local loss function under the non-stationary space-time model, T is the instability parameter on the time scale, σ is the control parameter, is the weight vector model after multi-scale feature decomposition, Ω is the integral area of ​​the weight space, and the overall fitness is evaluated by integrating the area. generation represents the objective function generated after the multi-scale analysis and optimization process, which is used to evaluate the overall fitness of the self-organizing map network;

[0041] S32. In each generation of evolution, use variational Bayesian inference to infer the weight vector W ij (t) is selected, and variational Bayesian inference selects the optimal weight vector by maximizing the following variational lower bound:

[0042]

[0043] Among them, ELBO represents the lower bound of evidence, which is used to approximate the true posterior distribution in variational inference. ELBO minimizes the gap between the true distribution and the variational distribution, and plays a role in optimizing the weight vector. Denotes the variational distribution q(W ij ) takes the expectation, which is used to evaluate the assumed weight vector W ij Obey the variational distribution q(W ij ), the expected value of the log-likelihood, p(W ij ) indicates W ij Prior distribution; in Bayesian inference, the prior distribution is used to describe the prior knowledge about the model parameters before observing the data, W ij represents the weight vector between node i and input feature j in the self-organizing map network, D KL is the Kullback-Leibler divergence, which achieves the optimal selection of the weight vector by maximizing the following variational lower bound. X represents the input data used to train and adjust the parameters of the model;

[0044] S33. After the selection operation is completed, random sampling and mutation of the weight vector are performed. The mutation operation is described by the following formula:

[0045] W ij (Q ’ )=W ij (t)+ΔW ij (t);

[0046] Where, ΔW ij (t) is the random perturbation generated by Gaussian distribution. It optimizes the generalization ability of the network by exploring new feature mapping methods through moderate mutation. ij (t) represents the weight vector of node i and input feature j at time step t, W ij (Q ’ ) indicates that at time step Q ’ The weight vector at , represents the new state of the connection strength between node i and input feature j after a mutation operation;

[0047] S34. After the mutation operation is completed, a crossover operation based on topological data analysis is performed, and the crossover process of the weight vector is guided by the topological invariant:

[0048] W ij ′ (Q ’ )=α·TDA(W ij (t))+(1-α)·TDA(W ik (t));

[0049] Among them, TDA(Wij (t)) represents the weight vector W ij (t) The result after topological data analysis, α is the cross coefficient, and a new weight vector is generated by combining the topological characteristics of the two weight vectors, W ij (Q ’ ) indicates that at time step Q ’ The weight vector at the time represents the new state of the connection strength between node i and input feature j after a mutation operation, W ik (t) is a random perturbation generated by a Gaussian distribution, which is used to introduce a certain amount of randomness to help the network explore new feature mapping methods; W ij ′ (Q ’ ) represents the weight vector after the crossover process;

[0050] S35. After multiple generations of evolutionary optimization, the final optimized feature map is generated

[0051] Preferably, the generator generates a preliminary architectural design scheme by combining complex geometry-time crystal mapping, instanton effect modulation and algebraic topology-spinor field perturbation algorithm, specifically including the following steps:

[0052] S411, the final feature map optimized in step S3 The input is sent to the generator G of the generative adversarial network, which generates a preliminary architectural design plan through complex geometry-time crystal mapping.

[0053]

[0054] in It is a complex geometric mapping, acting on the feature mapping exp(.) is the phase evolution based on the time crystal structure. is the input random noise vector; b represents the imaginary unit, z represents the input random noise vector; S and T are both time periods;

[0055] S412. During the generation process, the generator G modulates the feature map and noise by combining the instanton effect in quantum field theory:

[0056]

[0057] Among them, z ′ represents the time crystal phase noise, represents the result of the instanton effect, λ n is the action coefficient of the instanton, is the action quantity related to the feature map and noise, z represents the input random noise vector, and n is the index variable in the summation symbol;

[0058] S413, the generator G modulates the feature map through the instanton effect The phase noise of the time crystal z ′ Combined with algebraic topology-spinor field perturbation to generate architectural design solutions x gen :

[0059]

[0060] Among them, S AT represents the joint perturbation function of algebraic topology and spinor field, TC represents the time crystal operation, is the tensor product operation, represents the phase noise z modulated by the instanton effect of the generator G ′ and feature map weights As input, generate the final output x of the architectural design solution gen ;

[0061] S414, in the generated preliminary design plan x gen The D-brane dynamics in superstring theory is applied to optimize its spatial dimension and stability through the interaction between the D-brane and the generative design scheme:

[0062]

[0063] Among them, D brane (x gen ) represents the D-brane dynamics optimization result, Σ is the world volume of the D-brane, represents the interaction term with the generated design solution, σ is the control parameter, and h αβ is the induction metric, d p σ is the volume differential element in the integral, representing a volume unit on the world volume of the D-brane;

[0064] S415, generator G generates architectural design solutions after D-membrane dynamics optimization Perform quantum tunneling state correction, use quantum tunneling effect to globally correct the local minimum in the design scheme, and finally output the optimized architectural design scheme

[0065] Preferably, during the adversarial training process of the generative adversarial network, the discriminator evaluates the generated architectural design scheme and feeds back the evaluation result to the generator, which specifically includes the following steps:

[0066] S421, in the adversarial training process of the generative adversarial network, the architectural design generated by the generator G in step S4 The input is sent to the discriminator D for evaluation. The discriminator D compares the generated design with the real architectural design data to determine its authenticity and rationality.

[0067] S422, the discriminator D will evaluate the results Feedback is given to the generator G to guide the generator G to adjust its generated architectural design plan. The generator minimizes the following loss function based on the feedback from the discriminator. The way to update its parameters:

[0068]

[0069] in, is the overall loss function of the generative adversarial network, p z (z) is the probability distribution of the noise vector z, p data (x) is the probability distribution of the real data x, where x refers to the existing architectural design scheme that meets the requirements. The generator gradually adjusts the generated architectural design scheme by minimizing this loss function, and the optimized generated architectural design scheme gradually approaches the real data. It indicates that a preliminary architectural design solution is generated through complex geometry-time crystal mapping;

[0070] In order to optimize the parameters of the generator, adversarial gradient descent is used to update the parameters of the generator in each adversarial training iteration; the adversarial training process of the generator and the discriminator is repeated until the loss function Converging to a preset threshold or the generated architectural design scheme meets the design requirements, as the training progresses, the generator G continuously improves its generated architectural design scheme.

[0071] Preferably, during multiple rounds of adversarial training and evolutionary optimization, the generator continuously optimizes the architectural design scheme based on the feedback information of the discriminator, and further optimizes the quality of each generation of design schemes by combining high-dimensional hypersurface functionals, non-commutative algebraic geometry, dynamical systems, and symplectic geometry transformation algorithms, including:

[0072] In the process of multiple rounds of adversarial training and evolutionary optimization, the generator G and the discriminator D gradually improve the quality of the architectural design plan through interaction. Based on the feedback of the discriminator D, the generator G uses high-dimensional hypersurface functional combined with mirror symmetry transformation to optimize the generated architectural design plan:

[0073]

[0074] in, is a mirror symmetry transformation, is a high-dimensional hypersurface functional, represents the boundary conditions of the architectural design in the feature space, is a high-dimensional hyperspace, μ is the measure, θ G Usually represents the parameters of the generator G, is the parameter update result of the generator after the t+1th iteration, represents the optimized feature map;

[0075] In each generation of adversarial training, the architectural design solutions generated by the generator are evolutionarily optimized by combining non-commutative algebraic geometry and deformed quantum field theory in non-commutative algebraic geometry;

[0076] In the process of multiple rounds of adversarial training and evolutionary optimization, the generator's generated architectural design scheme is further optimized by combining the evaluation results of the discriminator D and using the dynamic system on the manifold of the multi-modal space. The discriminator D evaluates the architectural design scheme generated in each generation and feeds the evaluation results back to the generator through the following formula to adjust the generated architectural design scheme:

[0077]

[0078] in, is the dynamical system function on the moduli space manifold, ω mod is the moduli space form, Represents along the path γ k Parallel move group operations, is the state of the feature mapping matrix at the t+1th generation after multiple rounds of adversarial training and optimization. (L) ) represents the output of the discriminator D for the input data X at the Lth iteration, ψ k It is the control variable that adjusts the generator parameter optimization process, affects the process of parallel moving group operation, and thus affects the direction and amplitude of the generator parameter optimization;

[0079] After multiple rounds of adversarial training and optimization, the generated architectural design Approaching the design goal and passing the convergence test, the symplectic geometric transformation on the infinite-dimensional Lie algebra is used to adjust and correct the final output of the generator:

[0080]

[0081] Among them, Ω Lie is a symplectic form on a Lie algebra, is an infinite-dimensional Lie algebraic space, p i and q jThey represent a pair of regular variables in symplectic geometry, called generalized momentum and generalized coordinates. The generalized coordinates correspond to the geometric features of the building. The generalized momentum represents the weight associated with the feature change. The generalized momentum and generalized coordinates jointly describe the state of the generated solution in the design space. H represents the Hamiltonian, which is used to describe the energy wedge product of the system. ∧ is used to combine the relationship between different variables. The variables include materials and the characteristics of the building. ξ is in the form of Lie algebra. The symplectic geometric transformation ensures that the final generated solution has consistency and stability in the geometry and topological structure in the design space. After the adversarial training process is completed, the final generator G outputs the optimized architectural design solution x * .

[0082] A system for implementing the architectural design method based on a generative adversarial network comprises:

[0083] The data preprocessing module organizes and formats the input data to ensure that the input data can be used effectively later;

[0084] The feature extraction module uses a genetic algorithm to initialize a self-organizing map network to extract key architectural design features from reference architectural drawings. The extracted features include spatial layout, structural form, facade design, and architectural style.

[0085] The evolutionary optimization module uses the evolutionary self-organizing map network to perform multi-generation evolutionary optimization on the extracted features, and gradually optimizes the features through selection, crossover and mutation operations to make the target design scheme meet the requirements;

[0086] Generative adversarial network module, including generator and discriminator. The generator generates a preliminary architectural design plan based on the optimized features. The discriminator evaluates the generated architectural design plan, compares the difference between the generated plan and the real architectural data, and feeds the evaluation results back to the generator. The generator continuously optimizes the generated architectural design plan based on the feedback. In this process, the generator and the discriminator undergo multiple rounds of adversarial training until the generated plan meets the design requirements.

[0087] The visualization module outputs the scheme that meets the project requirements in a visual form, including two-dimensional floor plans and three-dimensional models.

[0088] A computer device of the present invention comprises: a memory and a processor and a computer program stored in the memory. When the computer program is executed on the processor, the architectural design method based on a generative adversarial network is implemented.

[0089] Compared with the prior art, the present invention has the following beneficial effects:

[0090] 1. The Generative Adversarial Network of the Present Invention Improves the Automation Level of Architectural Design Generative Adversarial Network: By combining the techniques of Generative Adversarial Network and Self-Organizing Map Network, the present invention effectively automates the architectural design process. The Self-Organizing Map Network is used to extract key features from reference architectural drawings and optimize them through genetic algorithms, while the Generative Adversarial Network generates new design solutions. The entire process significantly reduces the reliance on manual design and improves design efficiency.

[0091] 2. Generative adversarial networks enhance the innovation and diversity of design solutions Generative adversarial networks: This invention gradually improves the quality of generated architectural design solutions through multiple rounds of adversarial training and evolutionary optimization. Combining complex mathematical tools, high-dimensional hypersurface functionals, non-commutative algebraic geometry and dynamic systems, the generator is able to explore innovative design paths in complex multi-dimensional spaces, and the generated design solutions have high geometric complexity and innovation.

[0092] 3. Generative adversarial networks ensure the quality and practicality of design solutions Generative adversarial networks: In multiple rounds of adversarial training and evolutionary optimization, the discriminator evaluates the generated solutions and further adjusts and corrects the output of the generator through complex optimization algorithms. The final output architectural design solution is not only highly innovative in theory, but also has excellent spatial configuration, structural rationality and design aesthetics in practical applications, ensuring the feasibility and practical value of the solution.

[0093] 4. Generative adversarial networks improve the stability and consistency of design solutions. Generative adversarial networks: Through the high-dimensional geometric tools of symplectic geometry transformation on infinite-dimensional Lie algebra, the present invention can make global geometric and topological adjustments to the generated architectural design solutions to ensure the stability and consistency of the final solution in the design space, avoiding the design quality fluctuation problem caused by parameter instability in traditional methods.

[0094] 5. Generative adversarial networks optimize the response speed and flexibility of architectural design Generative adversarial networks: Through multiple rounds of adversarial training and evolutionary optimization, the generated architectural design solutions can quickly respond to specific design requirements and goals. The systematic screening and optimization process ensures that the optimal solution that meets the actual application requirements can always be generated under various design requirements, greatly improving the flexibility and response speed of architectural design. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0096] Figure 1 A flowchart of a building design method based on a generative adversarial network proposed by the present invention;

[0097] Figure 2 A flowchart for generating a preliminary architectural design solution for an embodiment;

[0098] Figure 3 Flowchart of optimizing the design scheme for the embodiment. DETAILED DESCRIPTION

[0099] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0100] like Figure 1-Figure 3 As shown, a building design method based on a generative adversarial network includes the following steps:

[0101] S1. Obtaining initial input data for architectural design, the initial input data including reference architectural drawing data;

[0102] S2, using genetic algorithm to initialize the network structure, weight vector, learning rate and neighborhood function of the self-organizing map network, and extracting features from the reference building drawing data through the self-organizing map network;

[0103] S3. Evolve the features for multiple generations by evolving the self-organizing map network, perform genetic selection, mutation, and crossover operations, and generate and gradually optimize the feature map of the self-organizing map network:

[0104] S4, inputting the optimized feature map into the generator of the generative adversarial network for multiple rounds of adversarial training, screening multiple generated architectural design schemes based on design goals and requirements, and outputting the design scheme, including the following steps:

[0105] The generator combines complex geometry-time crystal mapping, instanton effect modulation, and algebraic topology-spinor field perturbation algorithm to generate preliminary architectural design solutions;

[0106] During the adversarial training process of the generative adversarial network, the discriminator evaluates the generated architectural design solutions and feeds back the evaluation results to the generator;

[0107] In the process of multiple rounds of adversarial training and evolutionary optimization, the generator continuously optimizes the architectural design scheme based on the feedback information of the discriminator, and further optimizes the quality of each generation of design schemes by combining high-dimensional hypersurface functionals, non-commutative algebraic geometry, dynamical systems and symplectic geometry transformation algorithms;

[0108] S5. Output the final selected design solution into a visual form.

[0109] In this implementation, step S2 includes the following steps:

[0110] S21. Create a self-organizing map network: Use hyperelliptic curve cluster decomposition to set the initial self-organizing map network structure of the self-organizing map network, and generate an initial weight vector W by hyperelliptic curve cluster decomposition ij (0), the weight vector reflects the connection mode and initial distribution characteristics between the nodes of the self-organizing map network;

[0111] S22, determine the initial weight vector W ij The dimension d of (0) weight and the dimension d of the input feature space input ,The weight vector of each node in the self-organizing map network is initialized as a vector in a high-dimensional space;

[0112] S23, set the learning rate η(t) and neighborhood function h BMU,i (t), the learning rate η(t) is adaptively adjusted through nonlinear fractional calculus and then gradually converges. The neighborhood function h BMU,i (t) represents the distance between node i and the network node that best represents the input data features during the current training process;

[0113] The self-organizing map network adaptively adjusts the learning rate and neighborhood influence range during training to optimize the network's ability to respond to complex features;

[0114] S24, input the reference architectural drawing data into the self-organizing map network, and gradually adjust the weight vector W of each node through multiple rounds of training ij (t), the update of the weight vector combines the moduli space theory and the principle of quantum state superposition to adapt to the complex features of the reference architectural drawings and extract key features:

[0115]

[0116] in, represents the manifold mapping function on the moduli space, which is used to adjust the representation of the weight vector in the moduli space. The weight vector represents the spatial layout, structural form, facade design and architectural style characteristics in the architectural drawings. m is the quantum state superposition coefficient, |ψ m (t)> is the quantum state, W ij (Q) represents the weight vector of node i and input feature j at time step Q, represents the manifold mapping function on the moduli space, which is used to adjust the representation of the weight vector in the moduli space. M represents the number of nodes, which is the feature dimension used for optimization in the self-organizing map network.

[0117] S25, in each generation of evolution from step S21 to step S25, the network is globally optimized and mutated using the Riemann-Theta function, and the weight vector generated and adjusted from step S21 to step S25 is used to optimize the weight vector and the self-organizing map network structure through the following fitness function to obtain the extracted feature F optimal :

[0118] F optimal =Θ(z|Ω)·exp(-γ·∑ i,j ∥W ij (t)-W ij (0)∥);

[0119] Where Θ(z|Ω) is the Riemann-Theta function, z is the current state vector, Ω is the Riemann matrix, γ is the attenuation coefficient, ∥W ij (t)-W ij (0)∥ represents the norm of the weight vector, which indicates the degree of change of the weight vector of node i and input feature j between time step t and the initial time step 0.

[0120] In this implementation, step S3 includes the following steps:

[0121] S31, based on the self-organizing map network optimized in step S2, the extracted features are evolved for multiple generations, and the weight vector W is adjusted by combining the non-stationary spatiotemporal model ij (t) Perform multi-scale analysis and optimization:

[0122]

[0123] in, represents the local loss function under the non-stationary space-time model, T is the instability parameter on the time scale, σ is the control parameter, is the weight vector model after multi-scale feature decomposition, Ω is the integral area of ​​the weight space, and the overall fitness is evaluated by integrating the area. generation represents the objective function generated after the multi-scale analysis and optimization process, which is used to evaluate the overall fitness of the self-organizing map network;

[0124] S32. In each generation of evolution, use variational Bayesian inference to infer the weight vector W ij (t) is selected, and variational Bayesian inference selects the optimal weight vector by maximizing the following variational lower bound:

[0125]

[0126] Among them, ELBO represents the lower bound of evidence, which is used to approximate the true posterior distribution in variational inference. ELBO minimizes the gap between the true distribution and the variational distribution, and plays a role in optimizing the weight vector. Denotes the variational distribution q(W ij ) takes the expectation, which is used to evaluate the assumed weight vector W ij Obey the variational distribution q(W ij ), the expected value of the log-likelihood, p(W ij ) means (W ij ) prior distribution; in Bayesian inference, the prior distribution is used to describe the prior knowledge about the model parameters before observing the data, (W ij ) represents the weight vector between node i and input feature j in the self-organizing map network, D KL is the Kullback-Leibler divergence, which achieves the optimal selection of the weight vector by maximizing the following variational lower bound. X represents the input data used to train and adjust the parameters of the model;

[0127] S33. After the selection operation is completed, random sampling and mutation of the weight vector are performed. The mutation operation is described by the following formula:

[0128] W ij (Q ’ )=W ij (t)+ΔW ij (t);

[0129] Where, ΔW ij (t) is the random perturbation generated by Gaussian distribution. It optimizes the generalization ability of the network by exploring new feature mapping methods through moderate mutation. ij (t) represents the weight vector of node i and input feature j at time step t, W ij (Q ’ ) indicates that at time step Q ’ The weight vector at , represents the new state of the connection strength between node i and input feature j after a mutation operation;

[0130] S34. After the mutation operation is completed, a crossover operation based on topological data analysis is performed, and the crossover process of the weight vector is guided by the topological invariant:

[0131] W ij ′ (Q ’ )=α·TDA(W ij (t))+(1-α)·TDA(W ik (t));

[0132] Among them, TDA(W ij (t)) represents the weight vector Wij (t) The result after topological data analysis, α is the cross coefficient, and a new weight vector is generated by combining the topological characteristics of two weight vectors, W ij (Q ’ ) indicates that at time step Q ’ The weight vector at the time represents the new state of the connection strength between node i and input feature j after a mutation operation, W ik (t) is a random perturbation generated by a Gaussian distribution, which is used to introduce a certain amount of randomness to help the network explore new feature mapping methods; W ij ′ (Q ’ ) represents the weight vector after the crossover process;

[0133] S35. After multiple generations of evolutionary optimization, the final optimized feature map is generated

[0134] like Figure 2 As shown, in this embodiment, the generator combines complex geometry-time crystal mapping, instanton effect modulation and algebraic topology-spinor field perturbation algorithm to generate a preliminary architectural design scheme, which specifically includes the following steps:

[0135] S411, the final feature map optimized in step S3 The input is sent to the generator G of the generative adversarial network, which generates a preliminary architectural design plan through complex geometry-time crystal mapping.

[0136]

[0137] in, It is a complex geometric mapping, acting on the feature mapping exp(.) is the phase evolution based on the time crystal structure. is the input random noise vector, S is the time period; b represents the imaginary unit, z represents the input random noise vector; T represents the time period;

[0138] S412. During the generation process, the generator G modulates the feature map and noise by combining the instanton effect in quantum field theory:

[0139]

[0140] Among them, z ′ represents the time crystal phase noise, represents the result of the instanton effect, λ n is the action coefficient of the instanton, is the action quantity related to the feature map and noise, z represents the input random noise vector, and n is the index variable in the summation symbol;

[0141] S413, the generator G modulates the feature map through the instanton effect The phase noise of the time crystal z ′ Combined with algebraic topology-spinor field perturbation to generate architectural design solutions x gen :

[0142]

[0143] Among them, S AT represents the joint perturbation function of algebraic topology and spinor field, TC represents the time crystal operation, is the tensor product operation, represents the phase noise z modulated by the instanton effect of the generator G ′ and feature map weights As input, generate the final output x of the architectural design solution gen ;

[0144] S414, in the generated preliminary design plan x gen The D-brane dynamics in superstring theory is applied to optimize its spatial dimension and stability through the interaction between the D-brane and the generative design scheme:

[0145]

[0146] Among them, D brane (x gen ) represents the D-brane dynamics optimization result, Σ is the world volume of the D-brane, represents the interaction term with the generated design solution, σ is the control parameter, and h αβ is the induction metric, d p σ is the volume differential element in the integral, representing a volume unit on the world volume of the D-brane;

[0147] S415, generator G generates architectural design solutions after D-membrane dynamics optimization Perform quantum tunneling state correction, use quantum tunneling effect to globally correct the local minimum in the design scheme, and finally output the optimized architectural design scheme

[0148] During the adversarial training of the generative adversarial network, the discriminator evaluates the generated architectural design and feeds back the evaluation results to the generator, which specifically includes the following steps:

[0149] S421, in the adversarial training process of the generative adversarial network, the architectural design generated by the generator G in step S4 The input is sent to the discriminator D for evaluation. The discriminator D compares the generated design with the real architectural design data to determine its authenticity and rationality.

[0150] S422, the discriminator D will evaluate the results Feedback is given to the generator G to guide the generator G to adjust its generated architectural design plan. The generator minimizes the following loss function based on the feedback from the discriminator. The way to update its parameters:

[0151]

[0152] in, is the overall loss function of the generative adversarial network, p z (z) is the probability distribution of the noise vector z, p data (x) is the probability distribution of the real data x, where x refers to the existing architectural design scheme that meets the requirements. The generator gradually adjusts the generated architectural design scheme by minimizing this loss function, and the optimized generated architectural design scheme gradually approaches the real data. It indicates that a preliminary architectural design solution is generated through complex geometry-time crystal mapping;

[0153] In order to optimize the parameters of the generator, adversarial gradient descent is used to update the parameters of the generator in each adversarial training iteration; the adversarial training process of the generator and the discriminator is repeated until the loss function Converging to a preset threshold or the generated architectural design scheme meets the design requirements, as the training progresses, the generator G continuously improves its generated architectural design scheme.

[0154] During multiple rounds of adversarial training and evolutionary optimization, the generator continuously optimizes the architectural design scheme based on the feedback information of the discriminator, and further optimizes the quality of each generation of design schemes by combining high-dimensional hypersurface functionals, non-commutative algebraic geometry, dynamical systems, and symplectic geometry transformation algorithms, including:

[0155] In the process of multiple rounds of adversarial training and evolutionary optimization, the generator G and the discriminator D gradually improve the quality of the architectural design plan through interaction. Based on the feedback of the discriminator D, the generator G uses high-dimensional hypersurface functional combined with mirror symmetry transformation to optimize the generated architectural design plan:

[0156]

[0157] in, is a mirror symmetry transformation, is a high-dimensional hypersurface functional, represents the boundary conditions of the architectural design in the feature space, is a high-dimensional hyperspace, μ is the measure, θ G Usually represents the parameters of the generator G, is the parameter update result of the generator after the t+1th iteration, represents the optimized feature map;

[0158] In each generation of adversarial training, the architectural design solutions generated by the generator are evolutionarily optimized by combining non-commutative algebraic geometry and deformed quantum field theory in non-commutative algebraic geometry;

[0159] In the process of multiple rounds of adversarial training and evolutionary optimization, the generator's generated architectural design scheme is further optimized by combining the evaluation results of the discriminator D and using the dynamic system on the manifold of the multi-modal space. The discriminator D evaluates the architectural design scheme generated in each generation and feeds the evaluation results back to the generator through the following formula to adjust the generated architectural design scheme. :

[0160]

[0161] in, is the dynamical system function on the moduli space manifold, ω mod is the moduli space form, Represents along the path γ k Parallel move group operations, is the state of the feature mapping matrix at the t+1th generation after multiple rounds of adversarial training and optimization. (L) ) represents the discriminator D at the Lth iteration for the input data x (t) The output of k It is the control variable that adjusts the generator parameter optimization process, affects the process of parallel moving group operation, and thus affects the direction and amplitude of the generator parameter optimization;

[0162] After multiple rounds of adversarial training and optimization, the generated architectural design Approaching the design goal and passing the convergence test, the symplectic geometric transformation on the infinite-dimensional Lie algebra is used to adjust and correct the final output of the generator:

[0163]

[0164] Among them, Ω Lie is a symplectic form on a Lie algebra, is an infinite-dimensional Lie algebraic space, p i and q jThey represent a pair of regular variables in symplectic geometry, called generalized momentum and generalized coordinates. The generalized coordinates correspond to the geometric features of the building. The generalized momentum represents the weight associated with the feature change. The generalized momentum and generalized coordinates jointly describe the state of the generated solution in the design space. H represents the Hamiltonian, which is used to describe the energy wedge product of the system. ∧ is used to combine the relationship between different variables. The variables include materials and the characteristics of the building. ξ is in the form of Lie algebra. The symplectic geometric transformation ensures that the final generated solution has consistency and stability in the geometry and topological structure in the design space. After the adversarial training process is completed, the final generator G outputs the optimized architectural design solution x * .

[0165] The present invention also provides a system for implementing a building design method based on a generative adversarial network, comprising:

[0166] The data preprocessing module organizes and formats the input data to ensure that the input data can be used effectively later;

[0167] The feature extraction module uses a genetic algorithm to initialize a self-organizing map network to extract key architectural design features from reference architectural drawings. The extracted features include spatial layout, structural form, facade design, and architectural style.

[0168] The evolutionary optimization module uses the evolutionary self-organizing map network to perform multi-generation evolutionary optimization on the extracted features, and gradually optimizes the features through selection, crossover and mutation operations to make the target design scheme meet the requirements;

[0169] Generative adversarial network module, including generator and discriminator. The generator generates a preliminary architectural design plan based on the optimized features. The discriminator evaluates the generated architectural design plan, compares the difference between the generated plan and the real architectural data, and feeds the evaluation results back to the generator. The generator continuously optimizes the generated architectural design plan based on the feedback. In this process, the generator and the discriminator undergo multiple rounds of adversarial training until the generated plan meets the design requirements.

[0170] The visualization module outputs the scheme that meets the project requirements in a visual form, including two-dimensional floor plans and three-dimensional models.

[0171] In order to verify the feasibility of the present invention in implementation, this embodiment is applied to a super high-rise office building planned to be built in the CBD core area of ​​a large city. The building needs to meet the needs of efficient office space and have a modern design style to meet the requirements of an international business environment. The construction project requires the design to be completed within a compact construction period and to optimize the building function and aesthetic effect to the greatest extent. However, traditional architectural design methods face many challenges: long design cycle, high labor cost, difficult to innovate design solutions, and difficult to meet complex functional and aesthetic requirements.

[0172] The architectural design method based on the generative adversarial network of the present invention aims to generate high-quality architectural design solutions through automation and intelligent means, shorten the design cycle, and improve the innovation and practicality of the design solutions.

[0173] First, the design requirements and relevant data of the super high-rise building were collected and sorted, including but not limited to: functional requirements of the building, site conditions, design style preferences, building regulations and reference architectural drawings.

[0174] The above data is input into the system of the method of the present invention. First, the input data is sorted and formatted through the data preprocessing module to ensure that the subsequent algorithm can effectively process this information. Then, the self-organizing map network is initialized by the genetic algorithm to extract key architectural design features from the reference architectural drawings, including but not limited to spatial layout, structural form, facade design and architectural style; after processing, the self-organizing map network can effectively extract and display the core features of similar architectural design, laying the foundation for generating new design schemes. After the feature extraction is completed, the extracted features are optimized through multiple generations of evolution through the evolutionary self-organizing map network. The genetic algorithm plays an important role in this process. These features are gradually optimized through selection, crossover and mutation operations to make them more in line with the requirements of the target design scheme. The optimized feature map is input into the generator of the generative adversarial network, and the generator generates a preliminary architectural design scheme based on these features. The generated scheme contains a variety of possibilities, reflecting the combination of different spatial layouts and architectural styles.

[0175] The generated architectural design is then input into the discriminator of the generative adversarial network for evaluation. The discriminator compares the difference between the generated solution and the real building data, and feeds the evaluation results back to the generator, which continuously optimizes the generation strategy based on the feedback. In this process, the generator and the discriminator gradually improve the quality of the architectural design through multiple rounds of adversarial training until the generated solution meets the design requirements.

[0176] In the process of multiple rounds of adversarial training and evolutionary optimization, combined with the evolutionary optimization function of the self-organizing map network, the quality of the design scheme is improved generation by generation to ensure that the scheme generated in each generation is more in line with the requirements of architectural design. Through complex mathematical tools and optimization algorithms, the generated scheme gradually tends to be optimal in terms of spatial layout, structural stability and design aesthetics.

[0177] Finally, multiple high-quality architectural design plans are generated. These plans are screened based on design goals and requirements, and the plan that best meets the project requirements is selected and output in a visual form, including a two-dimensional floor plan and a three-dimensional model. This plan can be directly used in subsequent architectural design and construction.

[0178] Compared with the traditional design method, it can be seen from Table 1 and Table 2 that the design cycle of super high-rise office buildings is shortened from 2-3 months of the traditional method to 3 weeks, saving about 70% of the time and greatly improving the design efficiency. The system generates more than 30 different design schemes, and 85% of the schemes show significant innovation in spatial layout, structural form and facade design. Compared with the existing designs on the market, they are unique and visually impactful.

[0179] Table 1 Design cycle comparison

[0180]

[0181] Table 2 Evaluation of design diversity and innovation

[0182]

[0183] As shown in Table 3 and Table 4, in terms of design quality, the final generated solution met the design requirements in terms of space utilization, structural stability and aesthetic effect, which were improved by 15%, 20% and 25% respectively; the user satisfaction survey showed that 90% of the architects and developer representatives were very satisfied with the final solution, and the satisfaction rate increased by 30%. At the same time, as shown in Table 5, the shortening of the design cycle and the improvement of the solution quality also reduced the project design cost by 10%, further reflecting the economic advantages and practical value of the present invention.

[0184] Table 3 Design quality comparison

[0185] Evaluation Metrics Traditional methods Method of the present invention Improvement ratio Space Utilization 80％ 92％ +15％ Structural stability score 8 / 10 9.6 / 10 +20％ Architectural aesthetics rating 7.5 / 10 9.4 / 10 +25％

[0186] Table 4 User satisfaction survey

[0187] Participant categories Traditional method satisfaction Satisfaction of the method of the present invention Satisfaction improvement ratio Architect (10 persons) 70％ 90％ +29％ Developer Representative (3 persons) 60％ 90％ +50％

[0188] Table 5 Project cost comparison

[0189] Project cost type Traditional design method cost The cost of the method of the present invention Cost saving ratio Design Cost 100％ 90％ -10％

[0190] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A building design method based on generative adversarial networks, characterized in that: The following steps are involved: Obtain architectural design data, input the architectural design data into the trained generative adversarial network, screen multiple generated architectural design schemes based on design goals and requirements, and output the design scheme; The training of a generative adversarial network consists of the following steps: S1. Obtaining initial input data for architectural design, the initial input data including reference architectural drawing data; S2, using genetic algorithm to initialize the network structure, weight vector, learning rate and neighborhood function of the self-organizing map network, and extracting features from the reference building drawing data through the self-organizing map network; S3. Evolve the features for multiple generations by evolving the self-organizing map network, perform genetic selection, mutation, and crossover operations, and generate and gradually optimize the feature map of the self-organizing map network: S4, inputting the optimized feature map into the generator of the generative adversarial network for multiple rounds of adversarial training, screening multiple generated architectural design schemes based on design goals and requirements, and outputting the design scheme, including the following steps: The generator combines complex geometry-time crystal mapping, instanton effect modulation, and algebraic topology-spinor field perturbation algorithm to generate preliminary architectural design solutions; During the adversarial training process of the generative adversarial network, the discriminator evaluates the generated architectural design solutions and feeds back the evaluation results to the generator; In the process of multiple rounds of adversarial training and evolutionary optimization, the generator continuously optimizes the architectural design scheme based on the feedback information of the discriminator, and further optimizes the quality of each generation of design schemes by combining high-dimensional hypersurface functionals, non-commutative algebraic geometry, dynamical systems and symplectic geometry transformation algorithms; S5. Output the final selected design solution into a visual form.

2. The architectural design method based on generative adversarial network according to claim 1, characterized in that: Step S2 includes the following steps: S21. Create a self-organizing map network: Use hyperelliptic curve cluster decomposition to set the initial self-organizing map network structure of the self-organizing map network, and generate an initial weight vector W by hyperelliptic curve cluster decomposition ij (0), the weight vector reflects the connection mode and initial distribution characteristics between the nodes of the self-organizing map network; S22, determine the initial weight vector W ij The dimension d of (0) weight and the dimension d of the input feature space input ,The weight vector of each node in the self-organizing map network is initialized as a vector in a high-dimensional space; S23, set the learning rate η(t) and neighborhood function h BMU,i (t), the learning rate η(t) is adaptively adjusted through nonlinear fractional calculus and then gradually converges. The neighborhood function h BMU,i (t) represents the distance between node i and the network node that best represents the input data features during the training process at time t; The self-organizing map network adaptively adjusts the learning rate and neighborhood influence range during training to optimize the network's ability to respond to complex features; S24, input the reference architectural drawing data into the self-organizing map network, and gradually adjust the weight vector W of each node through multiple rounds of training ij (t), the update of the weight vector combines the moduli space theory and the principle of quantum state superposition to adapt to the complex features of the reference architectural drawings and extract key features: in, represents the manifold mapping function on the moduli space, which is used to adjust the representation of the weight vector in the moduli space. The weight vector represents the spatial layout, structural form, facade design and architectural style characteristics in the architectural drawings. m is the quantum state superposition coefficient, ψ m (t) is the quantum state, W ij (Q) represents the weight vector of node i and input feature j at time step Q, represents the manifold mapping function on the moduli space, which is used to adjust the representation of the weight vector in the moduli space. M represents the number of nodes, which is the feature dimension used for optimization in the self-organizing map network. S25, in each generation of evolution from step S21 to step S25, the network is globally optimized and mutated using the Riemann-Theta function, and the weight vector generated and adjusted from step S21 to step S25 is used to optimize the weight vector and the self-organizing map network structure through the following fitness function to obtain the extracted feature F optimal : F optimal =Θ(z|Ω)·exp(-γ·∑ i,j ∥W ij (t)-W ij (0)∥); Where Θ(z|Ω) is the Riemann-Theta function, z is the current state vector, Ω is the Riemann matrix, and γ is the attenuation coefficient; ∥W ij (t)-W ij (0)∥ is the norm of the weight vector, which is used to represent the degree of change of the weight vector of node i and input feature j between time step t and the initial time step 0, W ij (t) represents the weight vector of node i and input feature j at time step t, W ij (0) represents the weight vector of node i and input feature j at time step 0.

3. The architectural design method based on generative adversarial network according to claim 1, characterized in that: Step S3 includes the following steps: S31, based on the self-organizing map network optimized in step S2, the extracted features are evolved for multiple generations, and the weight vector W is combined with the non-stationary spatiotemporal model ij (t) Perform multi-scale analysis and optimization: in, represents the local loss function under the non-stationary space-time model, T is the instability parameter on the time scale, σ is the control parameter, is the weight vector model after multi-scale feature decomposition, Ω is the integral area of ​​the weight space domain, the overall fitness is evaluated by integrating the region, and Fgeneration represents the objective function generated after the multi-scale analysis and optimization process, which is used to evaluate the overall fitness of the self-organizing map network; S32. In each generation of evolution, the weight vector Wij(t) is selected using variational Bayesian inference. Variational Bayesian inference selects the optimal weight vector by maximizing the following variational lower bound: Among them, ELBO represents the evidence lower bound, which is used to approximate the true posterior distribution in variational inference. ELBO minimizes the gap between the true distribution and the variational distribution, and plays a role in optimizing the weight vector. represents the expectation of the variational distribution q(Wij), which is used to evaluate the expected value of the log-likelihood under the assumption that the weight vector Wij obeys the variational distribution q(Wij). p(Wij) represents the prior distribution of Wij. In Bayesian inference, the prior distribution is used to describe the prior knowledge of the model parameters before the data is observed. Wij represents the weight vector between node i and input feature j in the self-organizing map network. DKL is the Kullback-Leibler divergence. The optimal selection of the weight vector is achieved by maximizing the following variational lower bound. X represents the input data, which is used to train and adjust the parameters of the model. S33. After the selection operation is completed, random sampling and mutation of the weight vector are performed. The mutation operation is described by the following formula: Wij(Q')=Wij(t)+ΔWij(t); Among them, ΔWij(t) is the random perturbation generated by Gaussian distribution. It optimizes the generalization ability of the network by exploring new feature mapping methods through moderate mutation. Wij(t) represents the weight vector of node i and input feature j at time step t. Wij(Q') represents the weight vector at time step Q', which represents the new state of the connection strength between node i and input feature j after a mutation operation. S34. After the mutation operation is completed, a crossover operation based on topological data analysis is performed, and the crossover process of the weight vector is guided by the topological invariant: Wij'(Q')=α·TDA(Wij(t))+(1-α)·TDA(Wik(t)); Among them, TDA(Wij(t)) represents the result of topological data analysis on the weight vector Wij(t), α is the cross coefficient, and a new weight vector is generated by combining the topological characteristics of the two weight vectors. Wij(Q') represents the weight vector at time step Q', which represents the new state of the connection strength between node i and input feature j after a mutation operation. Wik(t) is a random perturbation amount generated by a Gaussian distribution, which is used to introduce a certain degree of randomness to help the network explore new feature mapping methods; Wij′(Q') represents the weight vector after the cross process; S35. After multiple generations of evolutionary optimization, the final optimized feature map is generated 4. The architectural design method based on generative adversarial network according to claim 1, characterized in that: The generator combines complex geometry-time crystal mapping, instanton effect modulation and algebraic topology-spinor field perturbation algorithm to generate a preliminary architectural design plan, which includes the following steps: S411, the final feature map optimized in step S3 The input is sent to the generator G of the generative adversarial network, which generates a preliminary architectural design plan through complex geometry-time crystal mapping. in It is a complex geometric mapping, acting on the feature mapping exp(.) is the phase evolution based on the time crystal structure. is the input random noise vector; b represents the imaginary unit, z represents the input random noise vector; S and T are both time periods; S412. During the generation process, the generator G modulates the feature map and noise by combining the instanton effect in quantum field theory: Among them, z ′ represents the time crystal phase noise, represents the result of the instanton effect, λ n is the action coefficient of the instanton, is the action quantity related to the feature map and noise, z represents the input random noise vector, and n is the index variable in the summation symbol; S413, the generator G modulates the feature map through the instanton effect The phase noise of the time crystal z ′ Combined with algebraic topology-spinor field perturbation to generate architectural design solutions x gen : Among them, S AT represents the joint perturbation function of algebraic topology and spinor field, TC represents the time crystal operation, is the tensor product operation, represents the phase noise z modulated by the instanton effect of the generator G ′ and feature map weights As input, generate the final output x of the architectural design solution gen ; S414, in the generated preliminary design plan x gen The D-brane dynamics in superstring theory is applied to optimize its spatial dimension and stability through the interaction between the D-brane and the generative design scheme: Among them, D brane (x gen ) represents the D-brane dynamics optimization result, Σ is the world volume of the D-brane, represents the interaction term with the generated design solution, σ is the control parameter, and h αβ is the induction metric, d p σ is the volume differential element in the integral, representing a volume unit on the world volume of the D-brane; S415, generator G generates architectural design solutions after D-membrane dynamics optimization Perform quantum tunneling state correction, use quantum tunneling effect to globally correct the local minimum in the design scheme, and finally output the optimized architectural design scheme 5. The architectural design method based on generative adversarial network according to claim 4, characterized in that: During the adversarial training of the generative adversarial network, the discriminator evaluates the generated architectural design and feeds back the evaluation results to the generator, which specifically includes the following steps: S421, in the adversarial training process of the generative adversarial network, the architectural design generated by the generator G in step S4 The input is sent to the discriminator D for evaluation. The discriminator D compares the generated design with the real architectural design data to determine its authenticity and rationality. S422, the discriminator D will evaluate the results Feedback is given to the generator G to guide the generator G to adjust its generated architectural design plan. The generator minimizes the following loss function based on the feedback from the discriminator. The way to update its parameters: in, is the overall loss function of the generative adversarial network, p z (z) is the probability distribution of the noise vector z, p data (x) is the probability distribution of the real data x, where x refers to the existing architectural design scheme that meets the requirements. The generator gradually adjusts the generated architectural design scheme by minimizing this loss function, and the optimized generated architectural design scheme gradually approaches the real data. It indicates that a preliminary architectural design solution is generated through complex geometry-time crystal mapping; In order to optimize the parameters of the generator, adversarial gradient descent is used to update the parameters of the generator in each adversarial training iteration; the adversarial training process of the generator and the discriminator is repeated until the loss function Converging to a preset threshold or the generated architectural design scheme meets the design requirements, as the training progresses, the generator G continuously improves its generated architectural design scheme.

6. The architectural design method based on generative adversarial network according to claim 5, characterized in that: During multiple rounds of adversarial training and evolutionary optimization, the generator continuously optimizes the architectural design scheme based on the feedback information of the discriminator, and further optimizes the quality of each generation of design schemes by combining high-dimensional hypersurface functionals, non-commutative algebraic geometry, dynamical systems, and symplectic geometry transformation algorithms, including: In the process of multiple rounds of adversarial training and evolutionary optimization, the generator G and the discriminator D gradually improve the quality of the architectural design plan through interaction. Based on the feedback of the discriminator D, the generator G uses high-dimensional hypersurface functional combined with mirror symmetry transformation to optimize the generated architectural design plan: in, is a mirror symmetry transformation, is a high-dimensional hypersurface functional, represents the boundary conditions of the architectural design in the feature space, is a high-dimensional hyperspace, μ is the measure, θ G Usually represents the parameters of the generator G, is the parameter update result of the generator after the t+1th iteration, represents the optimized feature map; In each generation of adversarial training, the architectural design solutions generated by the generator are evolutionarily optimized by combining non-commutative algebraic geometry and deformed quantum field theory in non-commutative algebraic geometry; In the process of multiple rounds of adversarial training and evolutionary optimization, the generator's generated architectural design scheme is further optimized by combining the evaluation results of the discriminator D and using the dynamic system on the manifold of the multi-modal space. The discriminator D evaluates the architectural design scheme generated in each generation and feeds the evaluation results back to the generator through the following formula to adjust the generated architectural design scheme: in, is the dynamical system function on the moduli space manifold, ω mod is the moduli space Form, Hol γk Represents along the path γ k Parallel move group operations, is the state of the feature mapping matrix at the t+1th generation after multiple rounds of adversarial training and optimization. (L) ) represents the output of the discriminator D for the input data X at the Lth iteration, ψ k It is the control variable that adjusts the generator parameter optimization process, affects the process of parallel moving group operation, and thus affects the direction and amplitude of the generator parameter optimization; After multiple rounds of adversarial training and optimization, the generated architectural design Approaching the design goal and passing the convergence test, the symplectic geometric transformation on the infinite-dimensional Lie algebra is used to adjust and correct the final output of the generator: Among them, Ω Lie is a symplectic form on a Lie algebra, is an infinite-dimensional Lie algebraic space, p i and q j They represent a pair of regular variables in symplectic geometry, called generalized momentum and generalized coordinates. The generalized coordinates correspond to the geometric features of the building. The generalized momentum represents the weight associated with the feature change. The generalized momentum and generalized coordinates jointly describe the state of the generated solution in the design space. H represents the Hamiltonian, which is used to describe the energy wedge product of the system. ∧ is used to combine the relationship between different variables. The variables include materials and the characteristics of the building. ξ is in the form of Lie algebra. The symplectic geometric transformation ensures that the final generated solution has consistency and stability in the geometry and topological structure in the design space. After the adversarial training process is completed, the final generator G outputs the optimized architectural design solution x * .

7. A system for implementing the architectural design method based on generative adversarial network as claimed in claim 1, characterized in that: include: The data preprocessing module organizes and formats the input data to ensure that the input data can be used effectively later; The feature extraction module uses a genetic algorithm to initialize a self-organizing map network to extract key architectural design features from reference architectural drawings. The extracted features include spatial layout, structural form, facade design, and architectural style. The evolutionary optimization module uses the evolutionary self-organizing map network to perform multi-generation evolutionary optimization on the extracted features, and gradually optimizes the features through selection, crossover and mutation operations to make the target design scheme meet the requirements; Generative adversarial network module, including generator and discriminator. The generator generates a preliminary architectural design plan based on the optimized features. The discriminator evaluates the generated architectural design plan, compares the difference between the generated plan and the real architectural data, and feeds the evaluation results back to the generator. The generator continuously optimizes the generated architectural design plan based on the feedback. In this process, the generator and the discriminator undergo multiple rounds of adversarial training until the generated plan meets the design requirements. The visualization module outputs the scheme in a visualized form that meets the project requirements, including two-dimensional floor plans and three-dimensional models.

8. A computer device, characterized in that: The invention comprises: a memory and a processor and a computer program stored in the memory. When the computer program is executed on the processor, an architectural design method based on a generative adversarial network as described in any one of claims 1 to 7 is implemented.

Images