A metamaterial structure design method based on an artificial intelligence generation model

Abstract

The application discloses a metamaterial structure design method based on an artificial intelligence generation model, comprising the following steps: constructing a multi-granularity text description to physical field mapping framework, injecting physical constraints into the diffusion generation network through LoRA fine-tuning, and converting the text description into a physical field; high-order interaction between the divided domains is modeled by using a hypergraph neural network; the optimized physical field representation is input into the structure generation network as a boundary condition; and a metamaterial structure which is strictly consistent with the target physical field is generated. In addition to being able to meet the input of multi-granularity text description and generate a metamaterial structure design, the cascading characteristics of the framework can directly receive a physical field as input, thereby meeting the application scenario of structure replacement, such as replacing a structure design composed of materials that are difficult to process or have high processing costs with a structure design composed of materials that are easy to process and have lower costs.

Description

Technical Field

[0001] This invention relates to the field of metamaterial structure inverse design and generation, and in particular to a metamaterial structure design method based on an artificial intelligence generation model. Background Technology

[0002] Metamaterial inverse design achieves programmable mechanical properties through artificially designed substructures, showing great potential in complex geometric fields such as aerospace components and biomedical implants. However, existing technologies face severe challenges: traditional topology optimization and finite element analysis rely on fragmented numerical inputs, which fail under nonlinear high-dimensional states (plasticity, buckling) due to nonconvex optimization problems. Their computational complexity increases exponentially with topological parameters and cannot resolve complex semantic intentions. Existing methods lack explicit physical generalization mechanisms, and the generated structures often violate constitutive laws. Although modern data-driven paradigms can handle complex design spaces, they are still limited by single-objective numerical specifications. When faced with multi-granularity requirements, the generated structures cannot simultaneously meet physical consistency requirements, which seriously restricts the innovative exploration of metamaterials in cross-domain applications. Summary of the Invention

[0003] The purpose of this invention is to address the shortcomings of existing technologies by providing a metamaterial structure design method based on an artificial intelligence generative model.

[0004] The objective of this invention is achieved through the following technical solution: A metamaterial structure design method based on an artificial intelligence generative model, comprising:

[0005] For structural design application scenarios with unknown structures, a multi-granularity text description is constructed and encoded according to the application requirements of metamaterials. The encoded description is then transformed into a spatiotemporal physical field through a PDE-constrained physical field diffusion generation network.

[0006] For structural replacement applications with known structures, the corresponding spatiotemporal physical fields are obtained through finite element simulation or actual measurement.

[0007] The spatiotemporal physical field is divided into unit partitions and a hypergraph structure is constructed, wherein each unit partition corresponds to a hypergraph vertex, and physically similar unit partitions are connected by hyperedges.

[0008] By iteratively optimizing vertex features through a hypergraph neural network, higher-order interactions between physical field partitions are captured.

[0009] The optimized vertex features are used as conditional inputs to the metamaterial structure diffusion generation network to generate metamaterial structures that maintain strict consistency with the target physical field.

[0010] Furthermore, the multi-granularity text description includes:

[0011] Scalar parameters: Poisson's ratio, filler ratio, elastic modulus threshold;

[0012] Time series curves: strain-stress curves, energy evolution paths;

[0013] Semantic description: stress redistribution pattern, description of strain concentration regions;

[0014] Text encoding unifies the processing of numerical parameters, curve data, and natural language descriptions.

[0015] Furthermore, the PDE-constrained physical field diffusion generation network specifically implements PDE constraints by fine-tuning a pre-trained video diffusion model using LoRA. This includes: establishing a differentiable mapping from the vertical displacement field to the vertical stress field; calculating the predicted stress field; comparing it with the output of the diffusion model; and forcing physical consistency through physical regularization loss, the physical consistency loss function of which is:

[0016]

[0017] in For the actual stress field, For predicting the stress field, F represents the Frobenius norm calculation;

[0018] The physical consistency loss function is incorporated into the loss function during the fine-tuning process of the physical field diffusion generation network, forcing the generated field to satisfy the constitutive relation:

[0019] .

[0020] Furthermore, the hypergraph construction specifically involves:

[0021] The physical field mesh is divided into vertices, each vertex corresponding to a metamaterial unit partition. The four-channel physical field features of each partition constitute the feature vector X:

[0022]

[0023] in For horizontal displacement, For vertical displacement, U represents the vertical stress, and U represents the strain energy field.

[0024] Generating hyperedges using K-Means clustering:

[0025]

[0026] Vertex v corresponds to a unit partition, and hyperedge e connects units with similar physical evolution.

[0027] Furthermore, the iterative optimization of vertex features using a hypergraph neural network specifically involves:

[0028] Constructing the Laplacian matrix of the hypergraph:

[0029]

[0030] in Let be the vertex degree matrix. Let W be the hyperedge degree matrix, W be the hyperedge weight matrix, and H be the vertex-hyperedge incidence matrix;

[0031] Vertex feature update:

[0032]

[0033] in For the updated vertex features, For learnable parameter matrix, This is the LeakyReLU activation function.

[0034] Furthermore, the metamaterial structure diffusion generation network specifically comprises:

[0035] The features optimized by the hypergraph are transformed into vectors in the conditional space of the diffusion model through a mapping network. These conditional vectors are used as keys and values, and together with the noise latent representation used as a query, they are input into the cross-attention layer for feature fusion, thereby guiding the diffusion network to generate metamaterial structures that match the physical field boundary conditions.

[0036] Furthermore, the loss of the metamaterial structure diffusion generation network is specifically as follows:

[0037] When generating metamaterial structures that match the physical field boundary conditions using diffusion networks, the latent space structure is extracted based on a simplified version of the Hooke quasi-static simulator. Rapid stress reconstruction :

[0038]

[0039] in, For temporary extraction of structures in latent space, Porosity The relevant equivalent Young's modulus. Utilizing... Construct physical loss term Defined as:

[0040]

[0041] in For the actual stress field, F represents the calculated Frobenius norm.

[0042] Furthermore, both the physical field diffusion generation network and the metamaterial structure diffusion generation network are conditionally controlled. The denoising mean calculation process under conditional control is as follows:

[0043]

[0044] in is the noise value predicted by the 3D U-Net denoising network; c is the input condition. To preserve weights, it is defined as follows: , indicating in step t The proportion of original data retained; To retain the cumulative weight, it is defined as follows: , indicating from step 1 to step 2 t The cumulative retention effect; The noise variance scheduling parameter is a predefined parameter that varies with time steps. t The added sequence.

[0045] The present invention also provides a metamaterial structure design device based on an artificial intelligence generative model, including a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it implements the metamaterial structure design method based on an artificial intelligence generative model.

[0046] The present invention also provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the aforementioned metamaterial structure design method based on an artificial intelligence generative model.

[0047] The beneficial effects of this invention are:

[0048] 1. This invention introduces an explicit physical enhancement module into the diffusion generation model. By constructing a differentiable mapping between the vertical displacement field and the stress field, and by optimizing the denoising path in conjunction with physical consistency loss, it significantly improves the consistency between the generated structure and the target stress distribution at the global and local levels, effectively avoiding phenomena such as unreasonable structural deformation and false stress concentration.

[0049] 2. This invention employs a hypergraph neural network to model higher-order interactions between spatial and temporal partitions of the physical field, thereby enhancing the ability to model complex physical distribution characteristics (such as stress redistribution and energy pathways) and achieving dual constraints of regional consistency and global coordination of the target physical field.

[0050] 3. This invention is the first to propose mapping multi-granularity text descriptions (including scalar parameters, time-series curves, and semantic intent) to continuous physical fields, and then using a cascaded diffusion generation framework driven by the physical fields to generate structures. This achieves end-to-end modeling from abstract design intent to physically feasible structures, overcoming the limitations of existing methods that rely solely on non-physical modeling based on images or semantic descriptions. Furthermore, in addition to meeting structural design requirements, the cascaded framework design can directly receive physical fields to meet material replacement needs. Attached Figure Description

[0051] Figure 1 A flowchart illustrating the metamaterial structure design method based on an artificial intelligence generative model provided in an embodiment of the present invention;

[0052] Figure 2 A network structure diagram of the metamaterial structure design method based on an artificial intelligence generative model provided in an embodiment of the present invention;

[0053] Figure 3 This is a schematic diagram of a metamaterial structure design device based on an artificial intelligence generative model, provided in an embodiment of the present invention. Detailed Implementation

[0054] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0055] Figure 1 This is a flowchart illustrating a metamaterial structure design method based on an artificial intelligence generative model, as provided in an embodiment of the present invention. Figure 1 As shown, the metamaterial structure design method based on an artificial intelligence generative model in this embodiment of the invention may include the following steps:

[0056] (1) Pre-processing for metamaterial application requirements:

[0057] (1.1) For structural design application scenarios with unknown structures, construct multi-granularity text descriptions based on the application requirements of metamaterials;

[0058] Specifically, this refers to a textual description that presents a qualitative intent based on actual needs. This includes:

[0059] Scalar parameters: Poisson's ratio, filler ratio, elastic modulus threshold;

[0060] Time series curves: strain-stress curves, energy evolution paths;

[0061] Semantic description: stress redistribution pattern, description of strain concentration regions;

[0062] Text encoding unifies the processing of numerical parameters, curve data, and natural language descriptions.

[0063] (1.2) For structural replacement applications of known structures, such as replacing a structural design composed of materials that are difficult to process or have high processing costs with a structural design composed of materials that are easy to process and have lower costs, the corresponding physical field is obtained through finite element simulation / experimental measurement, specifically:

[0064] Finite element method (FEM) simulations were performed on metamaterial structures using plane strain conditions to simulate the two-dimensional behavior of real extruded structures, avoiding out-of-plane buckling caused by compression. The structure was virtually placed between two rigid horizontal indenters, with nodes at the top and bottom boundaries fixed to the indenters. To simplify the process, the indenter surfaces were assumed to be lubricated, allowing nodes to slide horizontally. Frictional self-contact was considered within the element structure, with a friction coefficient of 0.4. Due to the large deformations (including buckling and contact), an implicit dynamic solver was used to ensure numerical stability. To maintain a quasi-static process, the material density was set to 10⁻⁸, and a smooth displacement amplitude was applied from φ = 0 to φ = 1, ensuring that the kinetic energy remained below 1% of the internal energy. Furthermore, the artificial energy term was also limited to less than 1%. All lengths were dimensionless (due to the system's dimensional invariance), and stresses were output in megapascals (MPa).

[0065] During the simulation, horizontal displacement, vertical displacement, vertical stress components, and strain energy were recorded. Eleven uniformly distributed strain steps were selected on a 96×96 pixel grid, from the initial state to the maximum vertical strain. To avoid a zero-field due to an initial step size of zero, only the stress field starting from 0.2% strain was analyzed. Furthermore, the overall effective stress response (i.e., the vertical resultant force per unit initial area) under each strain step was calculated by recording the vertical reaction force (RF2) at the contact nodes on the upper surface.

[0066] (2) For structural design application scenarios with unknown structures, the new multi-granularity text description and the latent space feature sequence obtained from random sampling of Gaussian noise are jointly input into the trained physical field diffusion generation network to obtain the target physical field corresponding to the new multi-granularity text description. In this embodiment, the physical field is a video containing four channels. The specific process is as follows:

[0067] The network architecture is based on a pre-trained video diffusion model (EasyAnimate v5.1), employing a Diffusion Transformer (DiT) architecture and fine-tuned using Low-Rank Adaptation (LoRA). The text description is first processed by a text encoder and then input into the DiT along with latent space features. During fine-tuning, a physically enforced operator injection module is integrated, which imposes physical constraints during diffusion generation: after the new multi-granularity text description is input along with a sequence of latent space features randomly sampled from Gaussian noise, physical constraints are established at each layer of the Diffusion Transformer. Specifically, within the framework of continuum mechanics, stress-strain constraints based on Hooke's Law (σ = C ⋅ ε) are established, using a PDE-based approach.

[0068]

[0069] in Let C be the stress tensor under compressive deformation, C be the material stiffness tensor, and ϵ be the strain tensor. This law is transformed into a differentiable operator embedded in the video generation model. Specifically, the predicted stress field is calculated using a differentiable operator and compared with the output of a diffusion model, through physical regularization loss (such as...). Forced physical consistency: Define a learnable interaction matrix The material stiffness tensor is represented by an equivalent matrix, and the vertical displacement field generated in the latent space is extracted through a three-layer CNN adapter. Features:

[0070] The stress field is predicted using Einstein's summation convention:

[0071] in This indicates that element-wise multiplication and addition operations are performed according to the summation rules.

[0072] In each layer of the diffusion Transformer, the above explicit difference operator is established using Hooke's law, followed by the construction of the physical loss:

[0073]

[0074] We divide the physical field spatially into There are 24 regions in actual practice. twenty four, express The first in the region Each region For the actual stress field, For predicting the stress field, F represents the Frobenius norm calculation;

[0075] In one embodiment, the diffusion model for the physical field generation network includes a forward noise addition process and a reverse noise reduction process;

[0076] The forward noise addition process specifically involves injecting Gaussian-distributed noise into a Markov chain at time steps t∈[1,2,...,T]. ;in t This is the time step index, and its value range is... t ∈[1,2,…, T ] indicates the current step in adding noise. T The total step size is usually set to... T =256, representing the total number of steps in the noise-adding process. This represents the initial latent variables, which are encoded from the real physical field sequence by the encoder of the Video Variational Autoencoder (VAE). The noise variance scheduling parameter is represented at time step. t The amount of injected noise is a predefined parameter (usually obtained through linear or cosine scheduling) that controls the noise injection rate. I Let be the identity matrix, representing the covariance matrix of the noise, with the same dimensions as the latent variables; calculate the t-step noise addition result by accumulating the noise weights: . To preserve weights, it is defined as follows: , indicating in step t The proportion of original data retained. To retain the cumulative weight, it is defined as follows: , indicating from step 1 to step 2 t The cumulative retention effect. ϵ It is Gaussian noise. ϵ ~N(0, I ), obtained by sampling from a standard Gaussian distribution.

[0077] The reverse denoising process specifically involves using a diffusion Transformer (DiT) denoising network to remove noise... Data is gradually restored. The conditionally controlled denoised mean is calculated as follows:

[0078]

[0079] Where condition c is the feature vector obtained by the text encoder processing multi-granular text descriptions, determined by the DiT network parameters. θ predict, The noise value is predicted by the Di network. The denoising process follows a distribution:

[0080]

[0081] in The variance scheduling parameter is usually set to... = To control the randomness in the noise reduction process.

[0082] During training, physical constraints are applied through the physical loss function. The applied loss, combined with the denoising loss, is jointly optimized to ensure that the generated physics field conforms to the constitutive relation, where Here are the coefficients of the regularization term:

[0083]

[0084] In this way, the physical constraint terms affect the weight update through automatic differentiation, forcing the generated field to satisfy the constitutive relation.

[0085] (3) Use a hypergraph neural network to optimize the feature representation of the physical field to obtain the optimized vertex features;

[0086] Specifically:

[0087] The 24×24 physics grid is divided into 576 vertices, each vertex corresponding to a metamaterial unit partition. The feature vector X is composed of 11 frames of four-channel physics features from each partition.

[0088]

[0089] in These are the horizontal displacement field, vertical displacement field, vertical stress field, and strain energy field in the four-channel physical field.

[0090] K-Means clustering (selecting the number of hyperedges E based on the silhouette coefficient) was used to divide the 576 cells into 15 clusters:

[0091]

[0092] in The cluster center is used for iterative optimization until convergence or the maximum number of iterations is reached.

[0093] Construct the vertex-hyperedge incidence matrix H. The incidence matrix is ​​defined as follows:

[0094]

[0095] Vertex v corresponds to the current cell partition in the 576 cells of the physical field grid, and hyperedge e connects cells with similar physical evolution.

[0096] Constructing the Laplacian matrix of the hypergraph based on the correlation matrix H:

[0097]

[0098] in Let be the vertex degree matrix. Let W be the hyperedge degree matrix and W be the hyperedge weight matrix.

[0099] Vertex feature update:

[0100]

[0101] in For learnable parameter matrix, This is the LeakyReLU activation function.

[0102] (4) The optimized vertex features are input into the metamaterial structure diffusion generation network, and X̃ is transformed into a vector c in the diffusion model conditional space through a lightweight mapping network MLP. In the reverse denoising process of the metamaterial structure diffusion generation network (whose denoising network is 3DU-Net), this conditional feature c serves as the key and value, and is related to the noise latent representation as the query. The features are fed together into the cross-attention layer for feature fusion, thereby guiding the diffusion network to generate metamaterial structures that match the physical field boundary conditions.

[0103]

[0104] in , , V = , , , It is a learnable projection matrix.

[0105] When generating metamaterial structures that match the physical field boundary conditions using diffusion networks, the latent space structure is extracted based on a simplified version of the Hooke quasi-static simulator. Rapid stress reconstruction :

[0106]

[0107] in, For temporary extraction of structures in latent space, Porosity The relevant equivalent Young's modulus. Utilizing... Construct physical loss term Defined as:

[0108]

[0109] in For the actual stress field, F represents the calculated Frobenius norm. The loss term is incorporated as part of the diffusion model training process and jointly optimized with the denoising loss. The complete loss term of the metamaterial-generated diffusion model can be expressed as:

[0110]

[0111] In one embodiment, the diffusion model for generating metamaterial structures includes a forward noise addition process and a reverse noise reduction process;

[0112] The forward noise addition process specifically involves injecting Gaussian-distributed noise into a Markov chain at time steps t∈[1,2,...,T]. ;in t This is the time step index, and its value range is... t ∈[1,2,…, T ] indicates the current step in adding noise. T The total step size is usually set to... T =256, representing the total number of steps in the noise-adding process. This represents the initial latent variables, which are encoded from the real physical field sequence by the encoder of the Video Variational Autoencoder (VAE). The noise variance scheduling parameter is represented at time step. t The amount of injected noise is a predefined parameter (usually obtained through linear or cosine scheduling) that controls the noise injection rate. I Let be the identity matrix, representing the covariance matrix of the noise, with the same dimensions as the latent variables; calculate the t-step noise addition result by accumulating the noise weights: . To preserve weights, it is defined as follows: , indicating in step t The proportion of original data retained. To retain the cumulative weight, it is defined as follows: , indicating from step 1 to step 2 t The cumulative retention effect. ϵ It is Gaussian noise. ϵ ~N(0, I ), obtained by sampling from a standard Gaussian distribution.

[0113] The goal of the reverse denoising process is to reconstruct a physically conforming metamaterial structure from random noise, conditioned on the physical field features generated in the previous stage and optimized by a Hypergraph Neural Network (HGNN). Its core lies in a conditionally controlled 3D U-Net denoising network. This process follows the distribution below:

[0114]

[0115] in Indicates that the parameter is θThe conditional probability distribution defined by the neural network; Indicates at time step t -1 latent variable (the denoised state); Indicates at time step t The latent noise variables; c : Represents conditional control information, derived from the physical field features X̃ optimized by the hypergraph neural network through a lightweight mapping network (MLP). This conditional information carries the boundary constraints of the physical field and guides the entire generation process; denoising mean. It was predicted by the 3DU-Net network; The variance scheduling parameter is usually set to... = To control the randomness in the noise reduction process. I Let be the identity matrix, and let represent the covariance.

[0116] The denoised mean value under conditional control is calculated as follows:

[0117]

[0118] in The noise value is predicted by the 3D U-Net denoising network, and its parameters are: θ ; To preserve weights, it is defined as follows: , indicating in step t The proportion of original data retained. To retain the cumulative weight, it is defined as follows: , indicating from step 1 to step 2 t The cumulative retention effect. The noise variance scheduling parameter is a predefined parameter that varies with time steps. t The added sequence.

[0119] The 3D U-Net network effectively fuses conditional information through its encoder-decoder structure and cross-scale skip connections. c With noise latent representation Gradually output the denoised latent variables .go through T After the denoising process, the final denoising result is obtained. Ultimately, the noise reduction result is... The generated target metamaterial structure is obtained through the VAE decoder:

[0120]

[0121] in This refers to the design of the generated metamaterial structure.

[0122] In one embodiment, a new multi-granularity text description and a latent space feature sequence randomly sampled from Gaussian noise are jointly input into a trained physical field diffusion generation network to obtain the target physical field corresponding to the new multi-granularity text description. The feature representation of the physical field is optimized using a hypergraph neural network to obtain optimized vertex features. These optimized vertex features and the latent space feature sequence randomly sampled from Gaussian noise are then jointly input into a trained metamaterial structure diffusion generation network to obtain a metamaterial structure corresponding to the application requirements of the metamaterial, such as... Figure 2 As shown, specifically:

[0123] Encode multi-granularity text descriptions; add physical constraints during the physical generation process using a LoRA-fine-tuned diffusion generation network, transforming text descriptions into spatiotemporal physical fields;

[0124] The feature representation of the physical field is optimized by using a hypergraph neural network. The optimized vertex features are then input into a metamaterial structure diffusion generation network to generate a metamaterial structure that maintains strict consistency with the target physical field.

[0125] Corresponding to the aforementioned embodiment of a metamaterial structure design method based on an artificial intelligence generative model, the present invention also provides an embodiment of a metamaterial structure design device based on an artificial intelligence generative model.

[0126] See Figure 3 The present invention provides a metamaterial structure design device based on an artificial intelligence generative model, comprising a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it is used to implement a metamaterial structure design method based on an artificial intelligence generative model as described in the above embodiment.

[0127] The present invention provides an embodiment of a metamaterial structure design device based on an artificial intelligence generative model, which can be applied to any device with data processing capabilities, such as a computer. The device embodiment can be implemented through software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 3 The diagram shown is a hardware structure diagram of any device with data processing capabilities, which is the metamaterial structure design device based on an artificial intelligence generative model provided by this invention. (Except for...) Figure 3In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0128] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0129] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0130] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements a metamaterial structure design method based on an artificial intelligence generative model as described in the above embodiments.

[0131] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0132] The present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements the aforementioned metamaterial structure design method based on an artificial intelligence generative model.

[0133] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this application are indicated by the claims.

[0134] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this application. This application is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A metamaterial structure design method based on an artificial intelligence generative model, characterized in that, include: For structural design applications with unknown structures, a multi-granularity text description is constructed and encoded based on the application requirements of metamaterials. This encoded description is then used as a condition to transform the physical field into a spatiotemporal physical field through a PDE-constrained physical field diffusion generation network. Specifically, the PDE-constrained physical field diffusion generation network is based on a pre-trained video diffusion model, fine-tuned using LoRA to achieve PDE constraints. This includes: establishing a differentiable mapping from the vertical displacement field to the vertical stress field; calculating the predicted stress field; comparing it with the output of the diffusion model; and forcing physical consistency through physical regularization loss, the physical consistency loss function of which is: in For the actual stress field, For predicting the stress field, F represents the Frobenius norm calculation; The physical consistency loss function is incorporated into the loss function during the fine-tuning process of the physical field diffusion generation network, forcing the generated field to satisfy the constitutive relation: ； For structural replacement applications with known structures, the corresponding spatiotemporal physical fields are obtained through finite element simulation or actual measurement. The spatiotemporal physical field is divided into unit partitions and a hypergraph structure is constructed, wherein each unit partition corresponds to a hypergraph vertex, and physically similar unit partitions are connected by hyperedges. By iteratively optimizing vertex features using a hypergraph neural network, higher-order interactions between physical field partitions are captured; specifically, the iterative optimization of vertex features using a hypergraph neural network involves: Constructing the Laplacian matrix of the hypergraph: in Let be the vertex degree matrix. Let W be the hyperedge degree matrix, W be the hyperedge weight matrix, and H be the vertex-hyperedge incidence matrix; Vertex feature update: in For the updated vertex features, For learnable parameter matrix, Use the LeakyReLU activation function; The optimized vertex features are used as conditional inputs to the metamaterial structure diffusion generation network to generate metamaterial structures that maintain strict consistency with the target physical field.

2. The metamaterial structure design method based on an artificial intelligence generative model according to claim 1, characterized in that, The multi-granularity text description includes: Scalar parameters: Poisson's ratio, filler ratio, elastic modulus threshold; Time series curves: strain-stress curves, energy evolution paths; Semantic description: stress redistribution pattern, description of strain concentration regions; Text encoding unifies the processing of numerical parameters, curve data, and natural language descriptions.

3. The metamaterial structure design method based on an artificial intelligence generative model according to claim 1, characterized in that, The hypergraph construction specifically involves: The physical field mesh is divided into vertices, each vertex corresponding to a metamaterial unit partition. The four-channel physical field features of each partition constitute the feature vector X: in For horizontal displacement, For vertical displacement, U represents the vertical stress, and U represents the strain energy field. Generating hyperedges using K-Means clustering: Vertex v corresponds to a unit partition, and hyperedge e connects units with similar physical evolution.

4. The metamaterial structure design method based on an artificial intelligence generative model according to claim 1, characterized in that, The metamaterial structure diffusion generation network is specifically as follows: The features optimized by the hypergraph are transformed into vectors in the conditional space of the diffusion model through a mapping network. These conditional vectors are used as keys and values, and together with the noise latent representation used as a query, they are input into the cross-attention layer for feature fusion, thereby guiding the diffusion network to generate metamaterial structures that match the physical field boundary conditions.

5. The metamaterial structure design method based on an artificial intelligence generative model according to claim 4, characterized in that, The loss of the diffusion generation network of the metamaterial structure is specifically as follows: When generating metamaterial structures that match the physical field boundary conditions using diffusion networks, the latent space structure is extracted based on a simplified version of the Hooke quasi-static simulator. Rapid stress reconstruction : in, For temporary extraction of structures in latent space, Porosity The relevant equivalent Young's modulus; using Construct physical loss term Defined as: in For the actual stress field, F represents the calculated Frobenius norm.

6. The metamaterial structure design method based on an artificial intelligence generative model according to claim 1, characterized in that, Both the physical field diffusion generation network and the metamaterial structure diffusion generation network are conditionally controlled. The denoising mean calculation process under conditional control is as follows: in is the noise value predicted by the 3D U-Net denoising network; c is the input condition. To preserve weights, it is defined as follows: , indicating the proportion of original data retained at step t; To retain the cumulative weight, it is defined as follows: , representing the cumulative retention effect from step 1 to step t; The noise variance scheduling parameter is a predefined sequence that increases with time step t.

7. A metamaterial structure design device based on an artificial intelligence generative model, comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that, When the processor executes the executable code, it implements a metamaterial structure design method based on an artificial intelligence generative model as described in any one of claims 1-6.

8. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements a metamaterial structure design method based on an artificial intelligence generative model as described in any one of claims 1-6.

Images