Interactive evolution graph intelligent design method based on GNN and LLM

Abstract

The invention belongs to the technical field of AI design tools, and particularly relates to an intelligent design method for an interactive evolution graph based on GNN and LLM. According to the method, the design process of a mature designer is quantified through a graph neural network (GNN), the relation and dynamic change between design steps are captured, more comprehensive design thought reference is provided for a green designer, compared with a traditional linear method, the method is more flexible and accurate, the interaction ability of a large language model (LLM) is combined, and the design efficiency is improved. A novice designer can interact with the system through a natural language, adjust an interactive design drawing in real time and perform self-adaptive optimization according to feedback, so that the guide design process is more efficient and meets requirements, an interactive tool is constructed by utilizing a design data set of a mature designer, the novice designer can directly refer to and learn the design thought, and the design efficiency is improved. And the design efficiency and quality are obviously improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of AI design tools, and specifically relates to an interactive evolution graph intelligent design method based on GNN and LLM. Background Art

[0002] The creative process of designers usually relies on rich experience and complex thinking evolution, especially in design innovation. The complexity and subjectivity of these design processes often make novice designers face challenges in imitation and learning. Traditional design auxiliary tools mainly provide static templates or basic graphic editing functions, lacking effective dynamic feedback and process guidance. Therefore, how to extract and quantify the design thinking process of mature designers to help novices gradually innovate has become an important research topic.

[0003] Existing studies have used long short-term memory (LSTM) networks to model the decision-making behavior of designers in the system design process. This method is based on the function-behavior-structure (FBS) model and predicts their future decisions in design tasks by modeling the decision sequence of designers. Although this method has shown some effectiveness in simulating the decision-making process of designers, it mainly focuses on the design process of linear coding chains and fails to fully consider the interaction and processing capabilities between multi-level design information. Therefore, how to further expand this method to adapt to more diverse decision paths in complex design environments is still an urgent problem to be solved. A design optimization method based on reinforcement learning has been proposed in the prior art. Although it has advantages in some areas, it mainly focuses on the optimization tasks of a single objective and lacks the ability to interact and process multi-level design information. Compared with these linear design path methods, this scheme regards the design process as a gradually interactively evolving graph structure. By quantifying the evolution of the designer's thinking process, it captures the relationship and mutual influence between the design steps. This graph quantification method can more comprehensively reflect the complexity and dynamic changes in the design process, and provide designers with more flexible support and guidance.

[0004] On the other hand, in recent years, natural language processing technology based on large language models (LLM) has been widely used in image generation, design assistance and other fields. For example, large language models can generate rich design suggestions through contextual semantic understanding. At the same time, the advantages of graph neural networks (GNN) in processing graph structured data have also been gradually applied to fields such as data mining and pattern recognition. GNN can capture the relationship between design steps by treating the designer's design process as graph structured data. For example, graph convolutional networks (GCN) have demonstrated excellent performance in node classification tasks, which provides a basis for the structured expression of complex design thinking. In addition, generative models such as generative adversarial networks (GAN) and variational autoencoders (VAE) provide powerful tools for the generation and modeling of complex data.

[0005] Therefore, the present invention further combines GNN and LLM, uses GNN to quantify and learn the graph structure obtained in the design process, and introduces a large language model for intelligent feedback and guidance. Summary of the invention

[0006] The purpose of the present invention is to provide an interactive evolutionary graph intelligent design method based on GNN and LLM, which can quantify the design process of mature designers with interactive evolutionary graphs, further use GNN and LLM and other related AI technologies to learn these graph structures, and then build an interactive design platform to help novice designers complete the design, thereby realizing the original design process of a given mature designer and guiding novice designers to complete the complete design.

[0007] The technical solution adopted by the present invention is as follows: An interactive evolutionary graph intelligent design method based on GNN and LLM, comprising: Obtain a design task and an original text corresponding to the design task, and then use LLM text embedding and design classification to quantify the original text into a design drawing dataset of the design task; Build a VAE-based GNN generative neural network with point sequence as the step-by-step construction process, fit the design diagram dataset, and generate candidate solutions for the design diagram; Evaluate the generated design drawing candidate solutions, determine the matching degree between the design drawing candidate solutions and the design task, and perform iterative optimization based on the matching results to obtain an interactive evolution model; Acquire newly added tasks, match design tasks that match the newly added tasks, and output a guide design diagram candidate solution through the interactive evolution model.

[0008] In a preferred embodiment, the step of quantifying the original text into a design drawing dataset of a design task includes: Obtaining original text corresponding to the design task; Based on the node classification method of design science, the original text is segmented into multiple text categories; The segmented text category nodes are connected into edges according to the text logic or spatial relationship between the nodes to construct an original design diagram.

[0009] In a preferred embodiment, the text categories include: function-based analysis, behavior-based analysis, structure-based analysis, function-based precedents, behavior-based precedents, structure-based precedents, function-based evaluation, behavior-based evaluation, structure-based evaluation, concept synthesis-function, concept synthesis-behavior, concept synthesis-structure.

[0010] In a preferred solution, after the design drawing data set is output, data preprocessing is performed to process the design drawing data set into a preliminary embedded graph data set.

[0011] In a preferred embodiment, the construction of the VAE-based GNN generation network includes the following steps: Build a graph neural network generation method based on variational autoencoder (VAE), and map the embedded graph dataset into latent variables through the encoder; The decoder gradually restores the sampled latent variables to the original graph dataset in a point-by-point manner; Among them, the loss function of the training process is the KL divergence loss of the encoder and the reconstruction distribution loss of the decoder.

[0012] In a preferred solution, after the original text is collected, the original text is expanded, and each of the original texts is expanded into several copies; The expanded text data is embedded using the M3E Chinese LLM model to generate a 768-dimensional language feature vector; The 768-dimensional language embedding vector is reduced to 24 dimensions through a dimensionality reduction network, and the node category embedding dimension is adjusted to 24 dimensions through dimensionality up-coding to generate the final design node feature encoding.

[0013] In a preferred solution, the GNN encoder encodes the input graph data through a message passing mechanism, extracts global graph structure information, and generates latent variable representation. The specific steps are: Initially embed the node features and edge features of the input graph and map them to a high-dimensional space; Through multi-layer graph neural network, message passing is performed to aggregate node and edge information and obtain a global representation of the graph. Use graph pooling technology to compress global graph features and obtain the global embedding vector of the graph; The global embedding vector is mapped to the latent space through a fully connected layer, the mean and variance of the latent variable are output, and then the mean and variance are sampled to generate the latent variable representation of the latent space.

[0014] In a preferred solution, the GNN decoder generates a graph structure from latent variables and recursively restores the original graph data step by step. The specific steps are as follows: The latent variables sampled from the latent space are input into the decoder, and the decoder first generates the starting node of the graph; Use a recursive generation network to gradually generate nodes of the graph and predict the type of each newly generated node in turn; Based on the graph structure of the currently generated nodes and their edges, predict the edge connections between the newly generated nodes and the existing nodes; The global representation of the graph is updated using the message passing mechanism until the nodes and edges of the entire graph are generated, gradually restoring the original graph structure.

[0015] The present invention also provides an interactive evolutionary graph intelligent design system based on GNN and LLM, using the above-mentioned interactive evolutionary graph intelligent design method based on GNN and LLM, including: An initialization module, the initialization module is used to obtain a design task and an original text corresponding to the design task, and then quantify the original text into a design drawing dataset of the design task by using LLM text embedding and design classification; A building module, wherein the building module is used to build a VAE-based GNN generative neural network with a point sequence as a step-by-step construction process, fit the design drawing data set, and generate a candidate solution for the design drawing; An optimization module, wherein the optimization module is used to evaluate the generated design drawing candidate solutions, determine the matching degree between the design drawing candidate solutions and the design task, and perform iterative optimization according to the matching degree result to obtain an interactive evolution model; A guiding module is used to obtain a newly added task, match a design task that matches the newly added task, and output a guiding design diagram candidate solution through the interactive evolution model.

[0016] And, an electronic device, the electronic device comprising: at least one processor; and a memory communicatively coupled to the at least one processor; Among them, the memory stores a computer program that can be executed by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the above-mentioned interactive evolution graph intelligent design method based on GNN and LLM.

[0017] The technical effects achieved by the present invention are: The present invention quantifies the design process of mature designers through graph neural networks (GNNs), captures the relationships and dynamic changes between design steps, and provides novice designers with a more comprehensive reference for design ideas. Compared with traditional linear methods, it is more flexible and accurate. Combined with the interactive capabilities of large language models (LLMs), novice designers can interact with the system through natural language, adjust the interactive design diagram in real time, and adaptively optimize based on feedback, making the guided design process more efficient and in line with demand. By using the design data sets of mature designers to build interactive tools, novice designers can directly refer to and learn their design ideas, significantly improving design efficiency and quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a schematic flow chart of the method of the present invention; Figure 2 It is a schematic diagram of the design tool module of the present invention. DETAILED DESCRIPTION

[0019] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0020] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0021] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure or characteristic that may be included in at least one implementation of the present invention. The phrase "in a preferred embodiment" that appears in different places in this specification does not refer to the same embodiment, nor is it a separate or selective embodiment that is mutually exclusive with other embodiments.

[0022] Secondly, the present invention is described in detail with reference to the schematic diagram. When describing the embodiments of the present invention in detail, for the sake of convenience, the cross-sectional diagrams showing the device structure will not be partially enlarged according to the general scale, and the schematic diagrams are only examples, which should not limit the scope of protection of the present invention. In addition, in actual production, the three-dimensional dimensions of length, width and depth should be included.

[0023] See also Figure 1 As shown, the present invention provides an interactive evolution graph intelligent design method based on GNN and LLM, comprising: S1. Obtain the design task and the original text corresponding to the design task, and then use LLM text embedding and design classification to quantify the original text into a design diagram dataset for the design task; S2. Build a VAE-based GNN generative neural network with point sequence as the step-by-step construction process, fit the design drawing data set, and generate candidate solutions for the design drawing; S3, evaluating the generated design drawing candidate solutions, determining the matching degree between the design drawing candidate solutions and the design task, and iteratively optimizing according to the matching degree results to obtain an interactive evolution model; S4. Obtain the newly added tasks, match the design tasks that match the newly added tasks, and guide the candidate solutions of the design diagram through the interactive evolution model output.

[0024] As described in the above steps S1-S4, as the quantitative research of the design process gradually attracts attention, people realize the importance of converting the design task into a computable data structure. In this embodiment, the design task and its related original text information are first obtained, and then the LLM is used to perform text embedding processing to convert the text information into a numerical form for subsequent processing. Then, the numerical text data is further converted into the design drawing data set required for the design task through the design classification method. Secondly, a graph generation neural network based on variational autoencoder (VAE) is constructed. The neural network takes the point sequence as the core of the step-by-step construction process. In this way, the design drawing data set can be fitted and generated. A series of candidate solutions of design drawings, and then a detailed evaluation of the generated candidate solutions of design drawings is required. The evaluation process includes determining the matching degree of each candidate solution with the design task, which usually involves multiple design indicators and standards. According to the evaluation results, the candidate solutions of the design drawings are iteratively optimized to improve their matching degree with the design task. Through continuous iteration, an interactive evolutionary model is finally obtained to guide the generation and optimization process of candidate solutions of design drawings. Finally, when the user (beginner) inputs a new task, the relevant information of the new task will be obtained and the matching design task will be matched. Through the interactive evolutionary model, the guided design drawing candidate solutions can be output to guide the designer in design work.

[0025] In a preferred embodiment, the step of quantifying the original text into a design drawing dataset of the design task includes: Obtain the original text corresponding to the design task; Based on the node classification method of design science, the original text is segmented into multiple text categories; The segmented text category nodes are connected into edges according to the text logic or spatial relationship between the nodes to construct the original design graph.

[0026] In this implementation, in order to quantify the original text into a design drawing dataset for a design task, first, it is necessary to obtain the original text corresponding to the design task. The original text can be obtained in a variety of ways, such as extracting it from existing design documents, project reports or related technical papers, or by interviewing mature designers to obtain the original text, ensuring that the obtained text is closely related to the design task and providing sufficient information for subsequent analysis and quantification. Then, based on the node classification method of design science, the original text is segmented and divided into multiple text categories. The text categories include: function-based analysis, behavior-based analysis, structure-based analysis, function-based precedents, behavior-based precedents, structure-based precedents, function-based evaluation, behavior-based evaluation, structure-based evaluation, concept synthesis-function, concept synthesis-behavior, concept synthesis-structure, wherein the text data described by the designer is manually classified into 12 different categories based on the node classification method of design science. , these 12 categories include: function-based analysis (focusing on user needs and product design goals), behavior-based analysis (focusing on usage and human-computer interaction), structure-based analysis: focusing on the physical characteristics of the product and existing products on the market, function-based precedents: recalling function-related precedents in design experience, behavior-based precedents: recalling behavior and interaction precedents in design experience, structure-based precedents: recalling structure-related precedents in design experience, function-based evaluation: evaluating whether the solution meets user needs and design goals, behavior-based evaluation: evaluating the feasibility of interaction design, structure-based evaluation: evaluating the rationality and feasibility of the solution structure, concept synthesis-function: generating or improving function solutions, concept synthesis-behavior: generating or improving behavior and interaction solutions, concept synthesis-structure: generating or improving structure solutions. Finally, the segmented text category nodes are connected into edges according to the text logic or spatial relationship between the nodes to construct the original design diagram. In this process, we need to identify the relationships between different nodes. These relationships can be causal relationships, parallel relationships, hierarchical relationships, etc. By representing these relationships in the form of edges, a complete and structured design diagram can be constructed. Each node in the design diagram is connected to other nodes through edges, forming a complex relationship network, which intuitively shows the mutual connection between the information and concepts in the original text.

[0027] Some data examples in this embodiment are shown in the following table (including only five nodes): Node Number original Design Behavior Input source node 1 First of all, I think the appearance design of this robot. Structure-based analysis 0 2 The appearance design needs to imitate humans as much as possible. Concept Synthesis-Function 1 3 If it is made too mechanical, it may appear too cold. Structure-based evaluation 2 4 Next, it can tell something about the elderly. Behavior-based analytics 3 5 For example, telling stories to the elderly. Concept Synthesis-Behavior 4 Based on this, the original text can be effectively quantified into a design drawing dataset for the design task, providing strong support for subsequent design analysis and decision-making.

[0028] In a preferred embodiment, when constructing a VAE-based GNN generation network, the following steps are included: Build a graph neural network generation method based on variational autoencoder (VAE), and map the embedded graph dataset into latent variables through the encoder; The decoder gradually restores the sampled latent variables to the original graph dataset in a point-by-point manner; Among them, the loss function of the training process is the KL divergence loss of the encoder and the reconstruction distribution loss of the decoder.

[0029] In this implementation, when constructing a graph neural network generation model based on a variational autoencoder (VAE), we first start by building a graph neural network generation method based on a variational autoencoder. The core is to map the embedded graph data set into latent variables through an encoder. This process involves compressing the complex structure and feature information in the graph data set into a low-dimensional latent space representation, and then using a decoder to gradually restore the sampled latent variables to the original graph data set. This process is carried out in a point-by-point manner, that is, the decoder will gradually generate nodes and edges in the graph, thereby reconstructing the complete graph structure. The key to this step is that the decoder needs to be able to accurately capture the features and structural information in the original graph data set so that high-quality graph data can be generated. During the training process, a specific loss function is used to guide the learning of the model. The loss function consists of two parts: the KL divergence loss of the encoder and the reconstruction distribution loss of the decoder. The KL divergence loss of the encoder is used to measure the difference between the distribution of latent variables and the prior distribution during the encoding process, ensuring that the latent variables can effectively capture the key information in the graph dataset, while the reconstruction distribution loss of the decoder is used to measure the difference between the graph data generated by decoding and the original graph data, ensuring that the generated graph data can be as close to the original data as possible. In this way, a VAE-GNN model that can generate high-quality graph data can be effectively trained, which can not only capture the complex structure and feature information in the graph data, but also perform effective data generation and transformation in the latent space.

[0030] During the training process of the GNN network, the loss function consists of the KL divergence loss of the encoder and the reconstruction distribution loss of the decoder. The KL divergence is used to make the encoding z in the latent space as close as possible to the standard normal distribution N(0,I). In VAE, the encoder outputs the mean of the latent variable and variance , the KL divergence loss is calculated as: ; The reconstruction loss is used to measure the decoder Generated graph And the real input graph In graph neural networks, the graph structure includes nodes and edges. The reconstruction loss can be divided into two parts: node reconstruction loss: measuring the similarity between the node type of the generated graph and the nodes of the real graph, and edge reconstruction loss: measuring the similarity between the edge structure (topology) of the generated graph and the edge structure of the real graph. For node reconstruction, assuming that the real graph have nodes, and each node has a corresponding category onehot vector , the softmax feature of the generated prediction graph is , the node reconstruction loss is: ; For edge reconstruction, assume that the adjacency matrix of the real graph is , the adjacency matrix of the generated graph is ,The binary cross entropy loss function is used to measure whether each edge is generated correctly, and the edge reconstruction loss is: ; The loss function in the overall training process is: ; and It is the weight used to adjust the relative importance of node and edge reconstruction loss. Finally, the back propagation algorithm is used to train and optimize the encoder and decoder of GNN to obtain the original design diagram dataset for mature designers. The trained VAE-based GNN generates a neural network, which can be used to build an interactive evolutionary graph intelligent design tool later.

[0031] In addition, an interactive evolution graph intelligent design tool is built: using the trained GNN network, a tool is built for novice designers to gradually design, including the following modules: Front-end module: used to interact with users, display design paths and provide feedback, and includes the following sub-modules: Interactive design diagram display module: displays the relationship between nodes and edges in the design path, helping users to intuitively understand the current design status; Dialog module: used to pop up the current interactive design task and prompt the user to select the node and edge type for the next design path. The user can choose to accept the prompt or change the node and edge type through the interactive dialog box to update the current interactive evolution graph.

[0032] In this exemplary embodiment, when the user is designing, a dialog box will pop up to display the current design task. The user can select the node or edge type of the next design path through the dialog box. If the user is not satisfied with the system's suggestions, the user can freely change the node or edge through the dialog box, and the current interactive evolution graph will be updated; Prompt module: In combination with the large language model (LLM), according to the current state of the interactive evolution graph, the user is prompted with a reference text input. The text can be a suggestion generated by the GPT model or a prompt content defined according to the current node. In this exemplary embodiment, appropriate suggestions are generated according to the current state of the interactive evolution graph to help the user select a design path; Backend module: Dynamic prediction and optimization of the interactive evolution graph based on the trained VAE-based GNN generation model, including the following submodules: Interaction evolution graph prediction module: backend encapsulation of VAE decoder , receiving user feedback information from the front end , process the subgraph and predict the next node and the corresponding edge Information, namely: ; Design feedback processing module: the backend provides feedback based on the text input by the user and the selected nodes and edge type , using embedded Update to update the state of the interactive evolution graph subgraph, .

[0033] In this exemplary embodiment, the user has described some nodes during the design process: Node Number original Design Behavior Input source node 1 First, we considered a solution for a home companion robot. Function-based analysis 0 2 Introduce the product's positioning. This is a home robot that may include additional functions, but it is mainly aimed at the home market. Concept Synthesis-Function 1 3 Its main function is medical, such as monitoring medical indicators such as heart rate and blood pressure. Function-based precedent 2 4 A sphygmomanometer for measuring blood pressure can be integrated into this robot. Concept Synthesis-Function 3 At this point, the next design node is predicted based on the trained VAE-GNN model. The predicted next design behavior category is concept synthesis-function. Combined with the system prompt module, a suggestion based on the current design status is shown to the user: GPT suggests: "Next, you can also consider including the ability to measure blood sugar and blood oxygen levels, which will further enhance the robot's capabilities in health monitoring."

[0034] Based on the predictions and prompts, the user can choose to accept the prompts or make modifications. If the user chooses to accept, the design diagram will be automatically updated and the next node will be added. The design diagram after this step is as follows: Node Number original Design Behavior Input source node 1 First, we considered a solution for a home companion robot. Function-based analysis 0 2 Introduce the product's positioning. This is a home robot that may include additional functions, but it is mainly aimed at the home market. Concept Synthesis-Function 1 3 Its main function is medical, such as monitoring medical indicators such as heart rate and blood pressure. Function-based precedent 2 4 A sphygmomanometer for measuring blood pressure can be integrated into this robot. Concept Synthesis-Function 3 5 Next, the ability to measure blood sugar and blood oxygen levels could also be included. Concept Synthesis-Function 4 Through the above process, the gradual generation and interactive updating of design drawings are realized, which can provide novice designers with design path guidance based on mature design ideas. Through the combination of GNN and LLM, real-time prediction and dynamic adjustment of design drawings can be carried out according to user feedback information, making the design process more interactive and flexible. At the same time, design suggestions can be generated through large language models to provide users with rich design inspiration.

[0035] In a preferred embodiment, after the original text collection is completed, the original text is expanded, and each original text is expanded into several copies; The expanded text data is embedded using the M3E Chinese LLM model to generate a 768-dimensional language feature vector; The 768-dimensional language embedding vector is reduced to 24 dimensions through a dimensionality reduction network, and the node category embedding dimension is adjusted to 24 dimensions through dimensionality up-coding to generate the final design node feature encoding.

[0036] In this implementation, after completing the collection of the original text, the collected original text needs to be further expanded. Specifically, each original text needs to be expanded into multiple new texts to increase the diversity and richness of the data. This process can be achieved through various text expansion technologies, such as synonym replacement, sentence rearrangement, paragraph reorganization, etc., to ensure that the expanded text has certain changes and innovations while maintaining the original semantics. After the text expansion is completed, the expanded text data is input into the M3E Chinese LLM model. The M3E Chinese LLM model can deeply understand and embed the text, and convert the expanded text data into a 768-dimensional language feature vector. The language feature vector can capture the semantic information and contextual relationship in the text. Subsequent processing provides rich feature representations. In order to further optimize and simplify these feature vectors, this embodiment reduces the 768-dimensional language embedding vector to 24 dimensions through a dimensionality reduction network. The role of the dimensionality reduction network is to extract the most important features and remove redundant information while maintaining the intrinsic structure and semantics of the data. Through dimensionality reduction, a more compact and efficient feature representation can be obtained, which helps to improve the computational efficiency and accuracy of subsequent processing steps. After dimensionality reduction, the node category needs to be embedded, and the node category embedding dimension is adjusted to 24 dimensions to ensure that the node category information can be fully expressed and utilized when generating the final design node feature coding. Through dimensionality increase coding, the node category information can be combined with the language feature vector to generate a more comprehensive and accurate design node feature coding.

[0037] In a preferred implementation, the GNN encoder encodes the input graph data through a message passing mechanism, extracts the global graph structure information, and generates a latent variable representation. The specific steps are: Initially embed the node features and edge features of the input graph and map them to a high-dimensional space; Through multi-layer graph neural network, message passing is performed to aggregate node and edge information and obtain a global representation of the graph. Use graph pooling technology to compress global graph features and obtain the global embedding vector of the graph; The global embedding vector is mapped to the latent space through a fully connected layer, the mean and variance of the latent variable are output, and then the mean and variance are sampled to generate the latent variable representation of the latent space.

[0038] The graph neural network (GNN) encoder extracts the global structural information of the graph and generates a latent variable representation by encoding the input graph data through a message passing mechanism. First, the encoder will initially embed the node features and edge features of the input graph. This process involves mapping the original features of the nodes and edges into a high-dimensional space to better capture the complexity of the graph data, laying the foundation for subsequent message passing. Then, message passing is performed through a multi-layer graph neural network. During this process, each node collects information from its neighboring nodes and combines its own features to aggregate information through a series of aggregation functions. The aggregation function can be a simple sum, average, or a more complex attention mechanism. Through multiple rounds of messages The information of nodes and edges is fully transmitted, so as to obtain the global representation of the graph. Then, the encoder uses graph pooling technology to compress the global graph features. Graph pooling is a technology similar to the pooling operation in convolutional neural networks, which aims to reduce the complexity of the graph while retaining the key information of the graph. Through graph pooling, the global embedding vector of the graph can be obtained, and the global features of the graph can be concentrated. Finally, the global embedding vector is mapped to the latent space through the fully connected layer. In the latent space, the encoder outputs the mean and variance of the latent variables. These parameters describe the probability distribution of the latent variables. By sampling them, the latent variable representation in the latent space can be generated, which captures the global structural information of the graph data and provides a basis for further graph analysis and processing.

[0039] In a preferred embodiment, the GNN decoder generates a graph structure from latent variables and recursively restores the original graph data step by step. The specific steps are as follows: The latent variables sampled from the latent space are input into the decoder, and the decoder first generates the starting node of the graph; Use a recursive generation network to gradually generate nodes of the graph and predict the type of each newly generated node in turn; Based on the graph structure of the currently generated nodes and their edges, predict the edge connections between the newly generated nodes and the existing nodes; The global representation of the graph is updated using the message passing mechanism until the nodes and edges of the entire graph are generated, gradually restoring the original graph structure.

[0040] In this implementation, the graph neural network (GNN) decoder generates a graph structure from latent variables and recursively restores the original graph data step by step. The specific process is as follows: first, the latent variables sampled from the latent space are input into the decoder. After receiving the latent variables, the decoder begins to perform its core task, namely, generating the starting node of the graph. The starting node is the starting point of the entire graph generation process and lays the foundation for the generation of subsequent nodes. Then the decoder uses the recursive generation network to gradually generate each node in the graph. In each step of the generation process, the decoder predicts the type of the newly generated node. The prediction process is based on the currently generated node information and the characteristics of the latent variables, ensuring that the type of the new node is consistent with the structure and semantics of the entire graph. After generating the new node, the decoder will The graph structure of the generated nodes and their edges is used to predict the edge connections between the newly generated nodes and the existing nodes. This prediction process involves an in-depth understanding of the graph structure to ensure that the generated edges can reasonably connect the nodes to form a meaningful graph structure. Finally, the decoder uses a message passing mechanism to update the global representation of the graph. This mechanism allows information exchange between nodes in the graph so that each node can obtain global context information. In this way, the global representation of the graph is continuously updated until all nodes and edges of the entire graph are generated, and a graph structure that is highly similar to the original graph data is gradually restored. The whole process is an iterative and recursive process. By continuously generating nodes and edges and updating the global representation of the graph, a graph structure that is highly consistent with the original graph data can eventually be generated.

[0041] See also Figure 2 , an interactive evolutionary graph intelligent design system based on GNN and LLM, using the above-mentioned interactive evolutionary graph intelligent design method based on GNN and LLM, including: Initialization module: The initialization module is used to obtain the design task and the original text corresponding to the design task, and then use LLM text embedding and design classification to quantify the original text into a design drawing dataset of the design task; The building module is used to build a VAE-based GNN generative neural network with point sequence as the step-by-step construction process, fit the design diagram data set, and generate candidate solutions for the design diagram; The optimization module is used to evaluate the generated candidate solutions of the design drawing, determine the matching degree between the candidate solutions of the design drawing and the design task, and perform iterative optimization according to the matching results to obtain the interactive evolution model; Guidance module,The guidance module is used to obtain new tasks, match design tasks that match the new tasks, and output the candidate solutions of the guidance design graph through the interactive evolution model.

[0042] As mentioned above, the system includes an initialization module, a building module, an optimization module and a guiding module, which work together to realize an interactive evolutionary graph intelligent design system based on GNN and LLM. The initialization module is responsible for obtaining the design task and the original text information corresponding to the design task. By using LLM for text embedding processing, the original text is converted into a numerical feature representation, and then these features are further quantified into the design drawing data set required for the design task through the design classification method. This process ensures that the initial data of the design task can be effectively understood and processed by the system. The core function of the building module is to build a GNN generative neural network based on a variational autoencoder (VAE), which adopts a point-by-point step-by-step construction method to effectively fit the design drawing data set and generate a series of design drawing candidates. Solution selection, candidate solutions are the system's initial attempts when exploring the design space, which provide a basis for subsequent optimization. The optimization module focuses on in-depth evaluation of the generated design drawing candidate solutions. Through a series of evaluation indicators, the evaluation indicator is the preset matching degree, which determines the degree of matching between the design drawing candidate solutions and the design tasks. Iterative optimization is performed based on the evaluation results to gradually improve the quality of the design drawing candidate solutions, and finally an interactive evolution model is obtained. Finally, the role of the guidance module is to obtain new design tasks and find matching design tasks. Through the interactive evolution model, the guidance module can output guided design drawing candidate solutions to provide valuable references and inspiration for novice designers. These guided design drawing candidate solutions can help designers find the entry point and direction of the design faster when facing new tasks.

[0043] The present invention also provides an electronic device, the electronic device comprising: at least one processor; and a memory communicatively coupled to the at least one processor; Among them, the memory stores a computer program that can be executed by at least one processor, and the computer program is executed by at least one processor so that the at least one processor can execute the above-mentioned interactive evolution graph intelligent design method based on GNN and LLM.

[0044] In the above, the electronic device may be a smart phone, a tablet computer, a laptop computer, a desktop computer, a server or other types of computing devices, the processor of the electronic device may be a central processing unit (CPU), a graphics processing unit (GPU), an application specific integrated circuit (ASIC) or other types of processors, and the memory may include a random access memory (RAM), a read-only memory (ROM), a hard disk drive (HDD), a solid-state drive (SSD) or other types of storage media.

[0045] In addition, the electronic device also includes an arithmetic unit and input-output devices. The arithmetic unit is used to perform various mathematical operations and logical operations to ensure that the electronic device can handle complex data processing tasks. The input-output devices include a keyboard, mouse, touch screen, display, printer, etc., which provide users with an interface for interacting with the electronic device, allowing users to input instructions, view results and output data.

[0046] The above is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications should also be considered as the protection scope of the present invention. The structures, devices and operating methods not specifically described and explained in the present invention shall be implemented according to the conventional means in the art unless otherwise specified and limited.

Claims

1. An intelligent design method for interactive evolutionary graphs based on GNN and LLM, characterized by: include: Obtain a design task and an original text corresponding to the design task, and then use LLM text embedding and design classification to quantify the original text into a design drawing dataset of the design task; Build a VAE-based GNN generative neural network with point sequence as the step-by-step construction process, fit the design diagram dataset, and generate candidate solutions for the design diagram; Evaluate the generated design drawing candidate solutions, determine the matching degree between the design drawing candidate solutions and the design task, and perform iterative optimization based on the matching results to obtain an interactive evolution model; Acquire newly added tasks, match design tasks that match the newly added tasks, and output a guide design diagram candidate solution through the interactive evolution model.

2. According to claim 1, a GNN and LLM-based interactive evolution graph intelligent design method is characterized by: The step of quantifying the original text into a design drawing dataset for a design task comprises: Obtaining original text corresponding to the design task; Based on the node classification method of design science, the original text is segmented into multiple text categories; The segmented text category nodes are connected into edges according to the text logic or spatial relationship between the nodes to construct an original design diagram.

3. The interactive evolution graph intelligent design method based on GNN and LLM according to claim 2 is characterized by: The text categories include: function-based analysis, behavior-based analysis, structure-based analysis, function-based precedents, behavior-based precedents, structure-based precedents, function-based evaluation, behavior-based evaluation, structure-based evaluation, conceptual synthesis-function, conceptual synthesis-behavior, conceptual synthesis-structure.

4. The interactive evolution graph intelligent design method based on GNN and LLM according to claim 1 is characterized by: After the design drawing data set is output, data preprocessing is performed to process the design drawing data set into a preliminary embedded graph data set.

5. The interactive evolution graph intelligent design method based on GNN and LLM according to claim 1, characterized in that: The construction of the VAE-based GNN generation network includes the following steps: Build a graph neural network generation method based on variational autoencoders, and map the embedded graph dataset into latent variables through the encoder; The decoder gradually restores the sampled latent variables to the original graph dataset in a point-by-point manner; Among them, the loss function of the training process is the KL divergence loss of the encoder and the reconstruction distribution loss of the decoder.

6. The interactive evolution graph intelligent design method based on GNN and LLM according to claim 4 is characterized by: After the original text collection is completed, the original text is expanded, and each of the original texts is expanded into several copies; The expanded text data is embedded using the M3E Chinese LLM model to generate a 768-dimensional language feature vector; The 768-dimensional language embedding vector is reduced to 24 dimensions through a dimensionality reduction network, and the node category embedding dimension is adjusted to 24 dimensions through dimensionality up-coding to generate the final design node feature encoding.

7. The interactive evolution graph intelligent design method based on GNN and LLM according to claim 5 is characterized by: The GNN encoder encodes the input graph data through a message passing mechanism, extracts global graph structure information, and generates latent variable representation. The specific steps are as follows: Initially embed the node features and edge features of the input graph and map them to a high-dimensional space; Through multi-layer graph neural network, message passing is performed to aggregate node and edge information and obtain a global representation of the graph. Use graph pooling technology to compress global graph features and obtain the global embedding vector of the graph; The global embedding vector is mapped to the latent space through a fully connected layer, the mean and variance of the latent variable are output, and then the mean and variance are sampled to generate the latent variable representation of the latent space.

8. The interactive evolution graph intelligent design method based on GNN and LLM according to claim 5 is characterized by: The GNN decoder generates a graph structure from latent variables and recursively restores the original graph data step by step. The specific steps are as follows: The latent variables sampled from the latent space are input into the decoder, and the decoder first generates the starting node of the graph; Use a recursive generation network to gradually generate nodes of the graph and predict the type of each newly generated node in turn; Based on the graph structure of the currently generated nodes and their edges, predict the edge connections between the newly generated nodes and the existing nodes; The global representation of the graph is updated using the message passing mechanism until the nodes and edges of the entire graph are generated, gradually restoring the original graph structure.

9. An interactive evolutionary graph intelligent design system based on GNN and LLM, characterized by: The interactive evolution graph intelligent design method based on GNN and LLM according to any one of claims 1 to 8 comprises: An initialization module, the initialization module is used to obtain a design task and an original text corresponding to the design task, and then use LLM text embedding and design classification to quantify the original text into a design drawing dataset of the design task; A building module, wherein the building module is used to build a VAE-based GNN generative neural network with a point sequence as a step-by-step construction process, fit the design drawing data set, and generate a candidate solution for the design drawing; An optimization module, wherein the optimization module is used to evaluate the generated design drawing candidate solutions, determine the matching degree between the design drawing candidate solutions and the design task, and perform iterative optimization according to the matching degree result to obtain an interactive evolution model; A guiding module is used to obtain a newly added task, match a design task that matches the newly added task, and output a guiding design diagram candidate solution through the interactive evolution model.

10. An electronic device, characterized in that: The electronic device comprises: at least one processor; and a memory communicatively coupled to the at least one processor; In which, the memory stores a computer program that can be executed by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the interactive evolution graph intelligent design method based on GNN and LLM as described in any one of claims 1 to 8.

Images