An intelligent material property prediction system based on graph neural network

Abstract

The present invention discloses an intelligent material property prediction system based on a graph neural network, including the following modules: a structure diagram construction module for parsing material structure data and constructing a third-order tensor graph; a tensor graph encoding module for inputting the third-order tensor graph into a multi-scale tensor graph neural network model to generate a high-dimensional tensor node embedding representation; an asymmetric propagation module for constructing a directional tensor connection structure based on the high-dimensional node embedding representation to generate an updated feature representation; a model optimization module for inputting the updated feature representation into an improved osprey swarm search algorithm to generate an optimal configuration; a multi-objective prediction module for loading the optimal configuration and performing multi-objective material property prediction; and a continuous learning module for identifying deviation samples based on prediction errors, feedback-updating the multi-scale tensor graph neural network model, and writing it into a learning buffer. The present invention integrates multi-channel graph modeling and an improved osprey swarm search algorithm to construct an intelligent material property prediction system.

Description

Technical Field

[0001] The present invention relates to the technical fields of artificial intelligence and materials science, and particularly relates to an intelligent material property prediction system based on a graph neural network. Background Art

[0002] In the field of materials science, how to efficiently and accurately predict the multi-physical properties of complex material systems has always been a core issue in interdisciplinary research. With the continuous increase in the R & D cycle and experimental costs of new materials, traditional prediction methods relying on experiments or rule-driven approaches have gradually been unable to meet the needs of high-throughput material design and performance evaluation. In order to achieve high-precision modeling of material properties, the academic and industrial communities have introduced intelligent prediction technologies based on artificial intelligence, especially graph structure learning models such as Graph Neural Network (GNN), which have become a key direction in recent material intelligent prediction research because they can depict the topological connections and physical coupling relationships between atoms inside materials.

[0003] Existing material property prediction methods based on graph neural networks generally regard materials as atomic structure graphs with atoms as nodes and bonds as edges, and use GNN models to learn the graph structure representations between nodes, and then predict target physical property parameters such as thermal conductivity, electrical conductivity, Young's modulus, and lattice constant. Typical publicly available technologies such as models like CGCNN (Crystal Graph Convolutional Neural Network) and MEGNet (Materials Graph Network) can perform relatively accurate physical property fitting for crystal structures in existing databases. However, most of these technologies are limited to the processing of the static topological structure between atoms and lack the ability to model dynamic information such as complex defect structures and physical field distributions inside materials, resulting in a significant decline in the performance of the model when dealing with complex materials with multi-scale coupling characteristics or non-ideal structures. At the same time, existing methods often train based on a single graph channel and cannot fully integrate the interaction between the microscopic structure of materials and external field conditions, limiting the physical generalization ability of the model.

[0004] In terms of graph structure modeling, existing technologies often only consider single-channel atomic structure graphs and lack the ability to jointly model defect perturbation graphs and physical field coupling graphs, making it difficult to depict the heterogeneous feature expressions of real materials at different physical levels. For example, in actual materials, microscopic defects such as vacancies, dislocations, and interstitial atoms will significantly change the local electric field and stress distributions, thereby affecting the overall performance of the material, while traditional methods often have difficulty introducing such local perturbations into a unified graph structure. At the same time, for material systems under different boundary conditions or multi-physical field actions, existing graph neural networks have significant deficiencies in constructing physically sensitive graph structures and dealing with dynamic changes in edge weight information.

[0005] In terms of model design, current GNN methods mostly adopt fixed structures or empirically set hyperparameter configurations, lacking targeted tuning paths in aspects such as model depth, node dimension, and fusion mechanism. As a result, when faced with material systems with inconsistent structural complexities, the models exhibit problems such as slow convergence, overfitting, or poor generalization ability. Some works attempt to introduce evolutionary algorithms or meta-learning methods to search for model parameters, but still lack an optimization mechanism that deeply integrates with the graph neural network structure, and cannot achieve an adaptive and efficient model search and tuning process.

[0006] In terms of prediction task modeling, most existing methods use single-objective regression to fit a single physical property, lacking a multi-objective decoding architecture design. Especially when dealing with indicators such as thermal properties, mechanical properties, electrical characteristics, and structural stability, which have different physical distribution laws, problems such as gradient conflict and unstable training often occur. In addition, due to the complexity of the material itself and the imbalance of data distribution, the prediction accuracy of existing technologies drops significantly when facing small-sample material systems, and there is a lack of continuous modeling and adaptive enhancement capabilities for error samples.

[0007] In addition, existing systems generally lack a feedback learning mechanism. Even when systematic errors occur in model predictions, it is often impossible to actively identify the sources of prediction deviations and perform continuous training corrections. Especially in the actual deployment stage, the prediction results of materials may deviate due to factors such as changes in boundary conditions and perturbations of defect samples. Traditional models cannot be dynamically updated or fine-tuned, resulting in insufficient model accuracy and stability. Therefore, in the existing technologies, there is still a lack of a graph neural network prediction system with continuous learning, adaptive fine-tuning, and enhanced data generation capabilities. In particular, the technical integration in combining multi-source graph structures, multi-physical property outputs, and model optimization feedback loops is still blank.

[0008] Therefore, how to provide an intelligent material property prediction system based on graph neural networks is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0009] An object of the present invention is to propose an intelligent material property prediction system based on graph neural networks. The present invention integrates multi-channel graph modeling and an improved osprey swarm search algorithm to construct an intelligent material property prediction system based on graph neural networks. By introducing a third-order tensor structure of atomic structure diagrams, defect perturbation diagrams, and physical field coupling diagrams, the structural perception ability of the model for complex materials is improved. At the same time, through a continuous learning mechanism, the adaptive optimization of the prediction model is realized, and it has the advantages of high prediction accuracy, strong structural adaptability, good generalization ability, and strong online evolution ability, and is suitable for high-throughput performance evaluation and intelligent analysis of various materials.

[0010] An intelligent material property prediction system based on a graph neural network according to an embodiment of the present invention includes the following modules: a structure diagram construction module for parsing material structure data, constructing an atomic structure diagram, a defect perturbation diagram, and a physical field coupling diagram, and mapping them into a third-order tensor diagram; a tensor diagram encoding module for inputting the third-order tensor diagram into a multi-scale tensor diagram neural network model to generate high-dimensional tensor node embedding representations, where the multi-scale tensor diagram neural network model includes a three-channel graph encoder module, a tensor attention fusion module, and an output embedding generation module; an asymmetric propagation module for constructing a directional tensor connection structure based on the high-dimensional node embedding representations, designing an in-ward propagation path and an out-ward propagation path, and performing asymmetric coupling propagation to generate updated feature representations; a model optimization module for inputting the updated feature representations into an improved osprey swarm search algorithm to optimize the multi-scale tensor diagram neural network model and generate an optimal configuration; a multi-objective prediction module for loading the optimal configuration and performing multi-objective material property prediction, and outputting prediction results of thermal properties, mechanical properties, electrical characteristics, and structural stability; a continuous learning module for identifying deviation samples based on prediction errors, feedback-updating the multi-scale tensor diagram neural network model, and writing them into a learning buffer.

[0011] An intelligent material property prediction method based on a graph neural network according to an embodiment of the present invention includes the following steps: S1. Parse the structure data of the material, construct three types of graph channels including an atomic structure diagram, a defect perturbation diagram, and a physical field coupling diagram, and map them into a third-order tensor diagram according to the node alignment rule; S2. Import the third-order tensor diagram into a multi-scale tensor diagram neural network model, perform feature encoding on the three types of graph channels respectively, and generate high-dimensional tensor node embedding representations through a tensor attention fusion module; S3. Based on the high-dimensional tensor node embedding representations, construct an asymmetric coupling propagation mechanism between nodes in the tensor space to generate updated feature representations; S4. Based on the updated feature representations, embed an improved osprey swarm search algorithm to optimize the multi-scale tensor diagram neural network model, comprehensively evaluate the training loss and accuracy of each configuration using a fitness function, and select the optimal configuration; S5. Use the optimal configuration for the material property prediction task and output multi-objective prediction results, where the multi-objective prediction results include thermal properties, mechanical properties, electrical characteristics, and structural stability; S6. Based on the actual error feedback of the multi-objective prediction results, introduce a continuous learning mechanism, construct a set of deviation samples as enhanced samples, and feedback-update the multi-scale tensor diagram neural network model.

[0012] Optionally, S1 specifically includes: S11. Parse the structure data of the material, extract the three-dimensional spatial coordinate values of each atom and the atomic type encoding, and construct an atomic structure diagram. Each atom corresponds to a node in the atomic structure diagram. When the Euclidean distance between any two atomic nodes satisfies , establish an edge connection between the two atomic nodes, where represents the th atomic node, represents the th atomic node, represents the set bonding threshold; S12. Automatically identify vacancies, dislocations, interstitial atoms, and impurity defect types in the atomic structure diagram, and mark each detected defect center as a defect node , where represents the th defect center, construct a subgraph within the range of the local perturbation region with the defect node as the core, represent the defect node and adjacent atoms as defect perturbation graph nodes, establish the influence path between the defect and surrounding atoms as edges, and form a defect perturbation graph; S13. Collect multi-physical field data of the material, the multi-physical field data includes stress tensor, electric field strength vector, and magnetic flux density vector, and construct a physical field coupling graph based on the same atomic node set in the atomic structure diagram. Introduce physical field difference weights in the edge construction method: ; where represents the th atomic node and the th atomic node, the physical field difference weight between them, , and represent weight coefficients, represents the gradient of the stress tensor between the th atomic node and the th atomic node, represents the electric field strength vector of the th atomic node, represents the electric field strength vector of the th atomic node, represents the magnetic flux density vector of the th atomic node, represents the magnetic flux density vector of the th atomic node, represents the Euclidean norm;

[0013] S14. Align the coordinates of all node sets in the atomic structure diagram, the defect perturbation graph, and the physical field coupling graph. For nodes with a spatial position difference less than the preset error threshold, merge them into a unified structural position node, and finally generate a unified node set; S15. Retain the edge connection relationships of the three types of graph channels respectively, and construct a third-order tensor graph, where the first dimension and the second dimension represent node pairs, and the third dimension represents the graph channel number. Each position in the third-order tensor is defined as: ; where represents the third-order tensor graph structure, represents the first dimension, Represents the second dimension, Represents the third dimension, Represents the th atomic node, Represents the th atomic node, Represents the atomic structure diagram channel, Represents the defect perturbation diagram channel, Represents the physical field coupling diagram channel.

[0014] Optionally, the S2 specifically includes: S21. Construct a multi-scale tensor graph neural network model, which includes a three-channel graph encoder module, a tensor attention fusion module, and an output embedding generation module. The three-channel graph encoder module is composed of an atomic structure diagram encoder, a defect perturbation diagram encoder, and a physical field coupling diagram encoder, corresponding to the three-channel inputs of the third-order tensor graph respectively; S22. Input the intermediate node features output by the three-channel graph encoder module into the tensor attention fusion module respectively. The tensor attention fusion module assigns channel weight coefficients to the three types of graph features to achieve dynamic weighted fusion; S23. Set a skip connection mechanism in the tensor attention fusion module to connect the original node features and the intermediate node features respectively to generate the fused node features; S24. Perform normalization processing on the fused node features in the output embedding generation module, and finally output the high-dimensional tensor node embedding representation.

[0015] Optionally, the skip connection mechanism is used to connect the original node features of the three-channel graph encoder and the intermediate node features after channel fusion during the fusion operation, specifically including retaining the original node features after the output of each channel graph encoder, and performing feature superposition after dimension alignment of the original node features and the intermediate node features after channel fusion.

[0016] Optionally, the S3 specifically includes: S31. Taking the high-dimensional tensor node embedding representation as the input, constructing a tensor connection structure between nodes, and assigning a directional label to each edge in the tensor connection structure to form a node adjacency relationship with direction distinction;

[0017] S32. Construct an asymmetric coupling propagation mechanism based on the node adjacency relationship. The asymmetric coupling propagation mechanism includes an in-coming feature propagation path and an out-coming feature propagation path. Define the feature update formula of the node during the propagation process as: ; where, Represents the th node's updated feature at the th layer, Represents the activation function, Represents the th node at the Updated features of the layer Indicates the th node's updated features in the th layer Indicates the weight matrix of the incoming propagation path Indicates the weight matrix of the outgoing propagation path Indicates the set of incoming neighbor nodes of the th node Indicates the set of outgoing neighbor nodes of the th node Indicates the attention coefficient of the incoming path Indicates the attention coefficient of the outgoing path; S33. Perform asymmetric coupled propagation on the high-dimensional tensor node embedding representation, and transmit it through the incoming path and the outgoing path respectively to generate a node-level local feature propagation tensor; S34. Perform channel adaptive normalization processing on the local feature propagation tensor, and use weighted summation operation to integrate local asymmetric propagation features to update the overall node representation; S35. Output the updated feature representation after the asymmetric coupled propagation mechanism

[0018] Optionally, the S4 specifically includes: S41. Convert the updated feature representation into a configuration vector, and initialize an osprey population with a size of , where each configuration vector includes the node embedding dimension, the number of graph convolution layers, the attention weight coefficient, the learning rate, and the message propagation coefficient; S42. In the th round of iteration, update the osprey individual position vector according to the global glide search strategy, and according to the centroid of the osprey population and the global optimal osprey individual : ; where represents the th round of iteration, the th osprey individual position vector, represents the th round of iteration, the th osprey individual position vector, represents the maximum number of iterations, and represent the values randomly sampled from the uniform distribution ; S43. Call the multi-scale tensor graph neural network model training sub-process for each updated osprey individual position vector, calculate the training loss and the validation accuracy on the validation set, and calculate the fitness score of each osprey individual: ; where represents the fitness score of the th osprey individual, and represents a preset non - negative weight coefficient;

[0019] S44. When the iteration round reaches the gliding phase threshold after that, among which , switch to the dive refinement phase, and execute the dive strategy guided by the gradient sign: ; among which, represents the step - size coefficient, represents the sign function, represents the gradient operator for the configuration vector; S45. Maintain the global memory channel data structure, and the global memory channel data structure stores the historical optimal solution pair , among which represents the th round of iteration, the fitness score of the optimal osprey individual. When the optimal individual in the current iteration round is better than the worst record in the memory, update the memory channel; S46. When the number of iterations reaches the maximum number of iterations or the change in the global optimal fitness is less than the convergence threshold , terminate. Among which represents the th round of iteration, the fitness score of the optimal osprey individual, and output the current osprey population solution set as the optimal configuration.

[0020] Optionally, the S5 specifically includes: S51. Load the obtained optimal configuration parameter set into the multi - scale tensor graph neural network model; S52. Decouple and model the multi - objective prediction task, establish four independent task sub - spaces including thermal properties, mechanical properties, electrical properties, and structural stability. Each task sub - space constructs an attribute - specific prediction head structure and connects it to the shared node embedding representation; S53. Construct a parallel decoder structure, set the separation path between the shared representation input channel and the output channels of the four independent task sub - spaces. Each output channel consists of two - layer linear mapping units and non - linear conversion units, and is respectively connected to various physical property index nodes at the output end; S54. Set two prediction dimensions, thermal conductivity and thermal diffusivity, for thermal properties, two prediction dimensions, Young's modulus and Poisson's ratio, for mechanical properties, two prediction dimensions, dielectric constant and conductivity, for electrical properties, and two prediction dimensions, binding energy and lattice constant, for structural stability; S55. Output the multi - objective prediction results of materials including thermal properties, mechanical properties, electrical properties, and structural stability.

[0021] Optionally, S6 specifically includes: S61. Comparing the multi-objective prediction results with the known true physical property parameters of the material, and extracting all samples with errors exceeding the preset threshold to form a deviation sample set; S62. Performing an enhanced perturbation operation on each data sample in the deviation sample set, injecting perturbations into the original third-order tensor graph structure to generate enhanced samples; S63. Constructing a local fine-tuning data set based on the enhanced samples, and freezing the general structure layer of the multi-scale tensor graph neural network model, and only activating the fusion layer and the attribute decoder sub-module related to the current deviation sample to perform fine-tuning training; S64. After completing the fine-tuning training, updating the parameter set participating in activation in the multi-scale tensor graph neural network model, and caching all enhanced samples, perturbation source paths, label correction residuals, and error response metrics into the continuous learning buffer.

[0022] Optionally, the enhanced perturbation operation specifically includes: adding a random offset with an amplitude not exceeding 3% to the three-dimensional coordinates of the atomic nodes to form a spatial position perturbation; multiplying any edge weight by a perturbation factor to form an edge weight perturbation, where the value range of the perturbation factor is [0.9, 1.1]; adding Gaussian noise with a mean of 0 to the node attribute vector as a node attribute perturbation.

[0023] The beneficial effects of the present invention are as follows: First, by constructing three types of heterogeneous graph channels, namely the atomic structure graph, the defect perturbation graph, and the physical field coupling graph, and using the node alignment rule to uniformly map them into a third-order tensor graph structure, the present invention breaks through the limitation of traditional GNN models that only model based on static atomic graphs, realizes the comprehensive fusion modeling of material structure hierarchical information, local defect interference, and external physical field effects, and greatly improves the expression ability and perception depth of the model for non-ideal complex material systems.

[0024] Second, the system designs a multi-scale tensor graph neural network model with a three-channel graph encoder, a tensor attention fusion module, and a skip connection structure, enabling each graph channel to maintain feature decoupling during the encoding stage and achieve weighted aggregation during the fusion stage. This not only enhances the information discrimination ability of the model between heterogeneous channels but also avoids redundant feature interference and gradient conflict problems, ensuring the physical consistency and expression stability of node representations. An in-coming and out-going path separation mechanism is introduced in the asymmetric propagation stage, effectively improving the propagation efficiency and update accuracy of the model in processing directional complex topological structures.

[0025] In addition, by introducing an improved osprey swarm search algorithm, the present invention encodes the structural parameters and training hyperparameters of the multi-scale tensor graph neural network into optimizable vectors, and uses a strategy combining gliding search and diving refinement to achieve dynamic adaptive optimization in a high-dimensional search space, significantly improving the structural adaptability and training convergence of the model in different task scenarios, and solving the problem that the traditional model structure depends on manual experience setting and is difficult to globally optimize.

[0026] Furthermore, a parallel decoding structure and a task decoupling path are constructed in the multi-objective prediction module to predict the thermal properties, mechanical properties, electrical characteristics, and structural stability respectively, support attribute-level structural design and differential tuning, and at the same time balance the gradient competition relationship of each physical property task through a soft sharing mechanism, achieving multi-index collaborative output under a unified model architecture, and enhancing the generalization and engineering application capabilities of the model.

[0027] Finally, by constructing a continuous learning mechanism based on prediction error feedback, the present invention can automatically trigger the enhanced sample generation process when the model output deviation exceeds the threshold, generate local enhanced samples by combining spatial perturbation, edge weight perturbation, and node attribute perturbation strategies, and dynamically update the decoding path and fusion layer parameters based on the fine-tuning mechanism of the frozen backbone layer, truly realizing the online evolution and adaptive enhancement of the model, and making up for the shortcoming of the insufficient response ability of traditional static GNN models in the actual deployment stage. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, and do not constitute a limitation to the present invention.

[0029] In the drawings: Figure 1 is a schematic structural diagram of an intelligent material property prediction system based on a graph neural network proposed by the present invention.

[0030] Figure 2 is an overall flowchart of an intelligent material property prediction method based on a graph neural network proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0031] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0032] Reference Figure 1, An intelligent material property prediction system based on graph neural network, comprising the following modules: a structure diagram construction module, configured to parse material structure data, construct an atomic structure diagram, a defect perturbation diagram and a physical field coupling diagram, and map them into a third-order tensor diagram; a tensor diagram encoding module, configured to input the third-order tensor diagram into a multi-scale tensor diagram neural network model to generate high-dimensional tensor node embedding representations, the multi-scale tensor diagram neural network model including a three-channel graph encoder module, a tensor attention fusion module and an output embedding generation module; an asymmetric propagation module, configured to construct a directional tensor connection structure based on the high-dimensional node embedding representations, and design an in-ward propagation path and an out-ward propagation path for asymmetric coupled propagation to generate updated feature representations; a model optimization module, configured to input the updated feature representations into an improved osprey swarm search algorithm to optimize the multi-scale tensor diagram neural network model and generate an optimal configuration; a multi-objective prediction module, configured to load the optimal configuration and perform multi-objective material property prediction, and output prediction results of thermal properties, mechanical properties, electrical characteristics and structural stability; a continuous learning module, configured to identify deviation samples based on prediction errors, feedback and update the multi-scale tensor diagram neural network model, and write it into a learning buffer.

[0033] The intelligent material property prediction system provided by the present invention has an overall architecture based on graph neural network. For the first time, the atomic-level topology, defect perturbation and physical field information of the material structure are uniformly embedded in the third-order tensor diagram modeling paradigm, and the end-to-end prediction process is realized in a modular manner. The system includes multiple core modules such as graph construction, tensor encoding, asymmetric propagation, structure optimization, performance prediction and continuous learning. Each module operates in coordination, not only retaining the multi-scale structural characteristics of the material body information, but also having a highly flexible and dynamically adjustable prediction ability. The system can effectively improve the model's perception ability of complex materials, defect structures and physical environment changes, significantly improve the prediction accuracy and the model's adaptive evolution ability, and is applicable to intelligent performance analysis and high-throughput screening scenarios of multi-objective and multi-type materials.

[0034] Reference Figure 2, An intelligent material property prediction method based on graph neural network, comprising the following steps: S1. Analyze the structural data of the material, construct three types of graph channels including atomic structure graph, defect perturbation graph and physical field coupling graph, and map them into a third-order tensor graph according to the node alignment rule; S2. Import the third-order tensor graph into a multi-scale tensor graph neural network model, perform feature encoding on the three types of graph channels respectively, and generate high-dimensional tensor node embedding representations through a tensor attention fusion module; S3. Based on the high-dimensional tensor node embedding representations, construct an asymmetric coupling propagation mechanism between nodes in the tensor space to generate updated feature representations; S4. Based on the updated feature representations, embed an improved osprey swarm search algorithm to optimize the multi-scale tensor graph neural network model, comprehensively evaluate the training loss and accuracy of each configuration using a fitness function, and select the optimal configuration; S5. Use the optimal configuration for the material property prediction task, and output multi-objective prediction results, where the multi-objective prediction results include thermal properties, mechanical properties, electrical characteristics and structural stability; S6. Based on the actual error feedback of the multi-objective prediction results, introduce a continuous learning mechanism, construct a deviation sample set as enhanced samples, and feedback and update the multi-scale tensor graph neural network model.

[0035] The method of the present invention defines a full-process data processing and task execution method driven by a graph neural network between each functional module, ensuring clear data and control flows between structure diagram construction, encoding, optimization, prediction and feedback. This method not only retains the complete expression of the material microstructure, but also establishes a logical closed-loop from input to output, supporting the unified parsing and deep representation of the third-order tensor input by subsequent modules. By standardizing the graph structure processing flow, the system obtains higher computational consistency and scalability, providing a clear and efficient control path for the multi-objective prediction task. This method improves the engineering integration and application flexibility of the overall system, and helps to build a large-scale material data modeling and performance prediction platform.

[0036] In this embodiment, S1 specifically includes: S11. Analyze the structural data of the material, extract the three-dimensional spatial coordinate values of each atom and the atomic type encoding, and construct an atomic structure graph. Each atom corresponds to an atomic structure graph node. When the Euclidean distance between any two atomic nodes satisfies , establish an edge connection between the two atomic nodes, where represents the th atomic node, represents the th atomic node, represents the set bonding threshold; S12. Automatically identify the types of vacancies, dislocations, interstitial atoms and impurity defects in the atomic structure graph, and mark each detected defect center as a defect node , where indicating the th defect center, constructing a sub - graph within the local perturbation region with the defect node as the core, representing the defect node and adjacent atoms as nodes of the defect perturbation graph, and establishing the influence path between the defect and surrounding atoms as edges to form the defect perturbation graph; S13. Collecting multi - physical - field data of the material, where the multi - physical - field data includes stress tensor, electric - field intensity vector, and magnetic - flux density vector. Based on the same atomic - node set in the atomic structure diagram, constructing a physical - field coupling graph, and introducing a physical - field difference weight in the way of edge construction: where, indicating the th atomic node and the th atomic node, , and represent weight coefficients, indicating the gradient of the stress tensor between the th atomic node and the th atomic node, indicating the electric - field intensity vector of the th atomic node, indicating the electric - field intensity vector of the th atomic node, indicating the magnetic - flux density vector of the th atomic node, indicating the magnetic - flux density vector of the th atomic node, representing the Euclidean norm; S14. Aligning the coordinates of all node sets in the atomic structure diagram, the defect perturbation graph, and the physical - field coupling graph, and merging those with the spatial - position difference between nodes less than the preset error threshold into unified - structure - position nodes, finally generating a unified node set; S15. Reserving the edge - connection relationships of the three types of graph channels respectively, and constructing a third - order tensor graph, where the first dimension and the second dimension represent node pairs, and the third dimension represents the graph - channel number. Each position in the third - order tensor is defined as: where, represents the third - order tensor - graph structure, represents the first dimension, represents the second dimension, represents the third dimension, indicating the th atomic node, indicating the th atomic node, represents the atomic - structure - diagram channel, represents the defect - perturbation - graph channel, represents the physical - field - coupling - graph channel.

[0037] This step specifically defines the construction process of the third-order tensor graph, clearly integrating the atomic structure graph, defect perturbation graph, and physical field coupling graph through node alignment and edge weight difference methods to form a tensor structure input with unified dimensions, providing multi-channel structural expression capabilities for subsequent graph neural network processing. This design avoids the information loss problem caused by channel separation in traditional graph modeling methods and enhances the overall expression of material internal heterogeneity, defect perturbation, and physical gradients. By precisely controlling the node alignment and edge construction strategies, the constructed third-order tensor graph has advantages such as strong structural completeness and high representation robustness, significantly improving the stability and generalization ability of the graph model when dealing with non-ideal material structures.

[0038] In this embodiment, S2 specifically includes: S21, constructing a multi-scale tensor graph neural network model, which includes a three-channel graph encoder module, a tensor attention fusion module, and an output embedding generation module. The three-channel graph encoder module consists of an atomic structure graph encoder, a defect perturbation graph encoder, and a physical field coupling graph encoder, corresponding to the three-channel inputs of the third-order tensor graph respectively; S22, inputting the intermediate node features output by the three-channel graph encoder module into the tensor attention fusion module respectively, and the tensor attention fusion module assigns channel weight coefficients to the three types of graph features to achieve dynamic weighted fusion; S23, setting a skip connection mechanism in the tensor attention fusion module to connect the original node features and the intermediate node features respectively to generate the fused node features; S24, performing normalization processing on the fused node features in the output embedding generation module, and finally outputting a high-dimensional tensor node embedding representation.

[0039] This step proposes a tensor graph encoding structure based on multi-channel graph input, adopting a three-branch graph encoder architecture combined with a tensor attention fusion module and a skip connection mechanism, effectively solving the problem of multi-graph heterogeneous structure fusion. Each graph channel models atomic topology, defect perturbation, and physical field information respectively. The fusion module dynamically adjusts the importance of different channels using learnable weight coefficients to ensure that the finally generated node embedding representation has more comprehensive physical semantic expression capabilities. At the same time, the skip connection mechanism enhances information mobility and feature fidelity, helping to improve training stability and structural interpretability. This structure significantly improves the expression efficiency and fusion quality of the model for multi-graph channel inputs.

[0040] In this embodiment, the skip connection mechanism is used to connect the original node features of the three-channel graph encoder with the intermediate node features after channel fusion during the fusion operation. Specifically, it includes retaining the original node features after the output of each channel graph encoder, and performing feature superposition on the original node features and the intermediate node features after channel fusion through dimension alignment. Through the skip connection mechanism, the original node features and the intermediate node features of the graph encoder are superimposed and connected during the feature fusion process, so that the fusion result retains the original channel structure information and the learned semantic features, enhancing the multi-dimensional consistency of node representation. This mechanism can effectively alleviate the problems of information attenuation or misleading aggregation that may occur during the multi-channel fusion process, ensure that the model does not lose key initial geometric and topological features during deep structure propagation, thereby enhancing the expressive ability of graph representation and the stability of the model, and improving the robustness and accuracy of the prediction results under complex material structures.

[0041] In this embodiment, S3 specifically includes: S31. Using the high-dimensional tensor node embedding representation as the input, constructing a tensor connection structure between nodes, and assigning a directional label to each edge in the tensor connection structure to form a node adjacency relationship with direction distinction; S32. Based on the node adjacency relationship, constructing an asymmetric coupling propagation mechanism, where the asymmetric coupling propagation mechanism includes an in-coming feature propagation path and an out-coming feature propagation path, and defining the feature update formula for nodes during the propagation process as: ; where represents the updated feature of the th node at the th layer, represents the activation function, represents the updated feature of the th node at the th layer, represents the updated feature of the th node at the th layer, represents the weight matrix of the in-coming propagation path, represents the weight matrix of the out-coming propagation path, represents the set of in-coming neighbor nodes of the th node, represents the set of out-coming neighbor nodes of the th node, represents the attention coefficient of the in-coming path, Represent the attention coefficient of the out-going path; S33. Perform asymmetric coupling propagation on the high-dimensional tensor node embedding representation, and transmit it through the in-coming path and the out-going path respectively to generate a node-level local feature propagation tensor; S34. Perform channel adaptive normalization processing on the local feature propagation tensor, and use weighted summation operation to integrate local asymmetric propagation features to update the overall node representation; S35. Output the updated feature representation after being updated by the asymmetric coupling propagation mechanism.

[0042] This step proposes an asymmetric propagation mechanism based on high-dimensional node embedding. By constructing two propagation paths, the in-coming path and the out-going path, and setting direction-aware weight matrices and attention coefficients respectively, it effectively distinguishes the direction of information flow between nodes. This design can accurately simulate the local asymmetry and anisotropy existing in the microscopic structure of materials, and effectively enhances the learning ability of the graph neural network for direction-sensitive structures. Through the directional modeling and channel adaptive normalization process, the local consistency and topological response ability of the updated feature representation are significantly improved, providing a more physically interpretable graph embedding basis for subsequent prediction tasks.

[0043] In this embodiment, the specific content of S4 includes: S41. Convert the updated feature representation into a configuration vector, and initialize the size to the osprey population , where each configuration vector includes the node embedding dimension, the number of graph convolution layers, the attention weight coefficient, the learning rate, and the message propagation coefficient; S42. In the th round of iteration, update the osprey individual position vector according to the global glide search strategy, and according to the centroid of the osprey population and the global optimal osprey individual : ; where, represents the position vector of the th osprey individual at the th round of iteration, represents the position vector of the th osprey individual at the th round of iteration, represents the maximum number of iterations, and represent the values randomly sampled from the uniform distribution ; S43. Call the multi-scale tensor graph neural network model training sub-process for each updated osprey individual position vector, and calculate the training loss and the validation accuracy on the validation set, and calculate the fitness score of each osprey individual: ; where, represents the fitness score of the th osprey individual, and denotes a preset non - negative weight coefficient; S44. When the iteration round reaches the gliding phase threshold after that, among which , switch to the dive refinement phase, and execute the dive strategy guided by the gradient sign: ; among which, denotes the step - size coefficient, denotes the sign function, denotes the gradient operator for the configuration vector; S45. Maintain the global memory channel data structure, and the global memory channel data structure stores the historical optimal solution pair , among which denotes the th iteration, the fitness score of the optimal osprey individual. When the optimal individual in the current iteration round is better than the worst record in the memory, update the memory channel; S46. When the iteration number reaches the maximum iteration number or the change in the global optimal fitness is less than the convergence threshold , terminate. Among which denotes the th iteration, the fitness score of the optimal osprey individual, and output the current osprey population solution set as the optimal configuration.

[0044] This step proposes to encode the graph neural network structure and hyperparameter configuration into vector form, perform joint optimization of the gliding and diving phases through an improved osprey swarm search algorithm, and introduce a global memory channel, fitness function, and convergence control strategy to construct an adaptive model optimization framework. This mechanism can efficiently search for the optimal structure combination that performs best in different material systems, avoid performance fluctuations caused by manual parameter tuning, and improve the automation level of model structure adjustment. The algorithm design is deeply integrated with the GNN structure, effectively improving the training stability and prediction accuracy, and enhancing the adaptability of the system in complex scenarios.

[0045] In this embodiment, S5 specifically includes: S51. Loading the obtained optimal configuration parameter set into the multi-scale tensor graph neural network model; S52. Decoupling and modeling the multi-objective prediction task, establishing four independent task subspaces including thermal properties, mechanical properties, electrical characteristics, and structural stability. For each task subspace, an attribute-specific prediction head structure is constructed and connected to the shared node embedding representation; S53. Constructing a parallel decoder structure, setting a separation path between the shared representation input channel and the output channels of the four independent task subspaces. Each output channel is composed of two layers of linear mapping units and non-linear conversion units, and is respectively connected to various physical property index nodes at the output end; S54. Setting two prediction dimensions of thermal conductivity and thermal diffusivity for thermal properties, two prediction dimensions of Young's modulus and Poisson's ratio for mechanical properties, two prediction dimensions of dielectric constant and conductivity for electrical characteristics, and two prediction dimensions of binding energy and lattice constant for structural stability; S55. Outputting the multi-objective prediction results of materials including thermal properties, mechanical properties, electrical characteristics, and structural stability.

[0046] In this step, through decoupling modeling and parallel decoder design, the multi-objective prediction task is divided into four types of attribute subspaces, corresponding to thermal properties, mechanical properties, electrical characteristics, and structural stability respectively, and an attribute-specific decoding path is set after each channel. This structure avoids the problems of gradient interference and information confusion between attributes, and improves the prediction accuracy and stability of each type of index. The combination of the multi-channel decoding output path and the shared representation input improves the prediction efficiency of the system while ensuring the physical independence of each attribute dimension, realizing the dual optimization of task-level separation and representation sharing, and is applicable to the multi-physical property integrated design scenario.

[0047] In this embodiment, S6 specifically includes: S61. Comparing the error between the multi-objective prediction result and the known true physical property parameters of the material, and extracting all samples with errors exceeding the preset threshold to form a deviation sample set; S62. Performing an enhanced perturbation operation on each data sample in the deviation sample set, injecting perturbations into the original third-order tensor graph structure to generate enhanced samples; S63. Constructing a local fine-tuning data set based on the enhanced samples, freezing the general structure layer of the multi-scale tensor graph neural network model, and only activating the fusion layer and the attribute decoder sub-module related to the current deviation sample to perform fine-tuning training; S64. After completing the fine-tuning training, updating the parameter set involved in activation in the multi-scale tensor graph neural network model, and caching all enhanced samples, perturbation source paths, label correction residuals, and error response indicators into the continuous learning buffer.

[0048] This step constructs a continuous learning mechanism based on prediction error feedback. It dynamically identifies deviation samples through error comparison, performs perturbation enhancement to generate new samples, and realizes local fine-tuning and update by freezing the backbone structure, activating the fusion layer and the decoder. This mechanism realizes the model's ability to correct itself online and quickly adapt to the distribution of new samples, effectively alleviating the generalization degradation problem of traditional models in the deployment stage. Through continuous update and buffer recording mechanism, the model has strong iterability and high evolution ability, and is suitable for dynamic learning and stable performance output in complex material scenarios.

[0049] In this embodiment, the enhanced perturbation operation specifically includes: adding a random offset with an amplitude not exceeding 3% to the three-dimensional coordinates of atomic nodes to form spatial position perturbation; multiplying any edge weight by a perturbation factor to form edge weight perturbation, and the value range of the perturbation factor is [0.9, 1.1]; adding Gaussian noise with a mean of 0 to the node attribute vector as node attribute perturbation.

[0050] This step clearly stipulates the generation strategies of three enhanced perturbation methods, including spatial position perturbation, edge weight perturbation and node attribute perturbation, which act on the geometric dimension, connection strength and node feature space of the graph structure respectively. The three types of perturbation strategies cover the key graph characterization dimensions in material modeling, have structural controllability and semantic interpretability, can efficiently expand the deviation sample set, improve sample diversity and model robustness. This perturbation generation mechanism is seamlessly integrated with the feedback fine-tuning process to ensure that the continuous learning module has the sample enhancement ability at the structural level.

[0051] Example 1: To verify the feasibility of the present invention in practice, the present invention is applied to the high-throughput carbide ceramic material screening task of the intelligent material development platform of a provincial key laboratory. For a batch of new HfC-SiC-based composite material samples with undisclosed performance indicators, structure analysis and physical property prediction are carried out to evaluate their suitability as thermal protection coating materials in extreme thermal environments. In this application scenario, the traditional method needs to test the thermal properties, electrical properties and mechanical strength through a high-temperature thermal conductivity meter, a resistivity test system and a stress loading platform respectively. The test cycle for each batch of samples is as long as 14 days, the test cost is high, the consumables loss is large, and for samples with micro defects, the test results are prone to fluctuate, affecting the stability of the evaluation results.

[0052] In the actual operation process, researchers first obtain the internal atomic structure image of the composite material through high-precision electron beam tomography technology, and extract the atomic coordinates, element types, and defect location information by combining the records of the material preparation process. At the same time, the stress distribution, electric field distribution, and magnetic induction intensity information of the material under the action of boundary loads are obtained from synchrotron radiation X-rays and laser speckle strain gauges. The system imports the above data into the graph construction module, generates the atomic structure graph, defect perturbation graph, and physical field coupling graph respectively, and forms a unified third-order tensor graph input model through spatial position coordinate alignment. The multi-scale graph neural network model encodes the three types of graph structures respectively, extracts the structural hierarchical features, defect response features, and boundary field gradient features, and generates a unified node embedding representation through weighted generation by the attention mechanism in the fusion module. Information is updated between nodes through the asymmetric propagation mechanism, and finally a high-dimensional tensor embedding is output.

[0053] In the modeling optimization stage, the system automatically starts the improved osprey swarm search algorithm to jointly optimize the structural parameters and training hyperparameters. After 38 rounds of gliding iteration and 12 rounds of diving fine-tuning, it automatically determines that the number of graph convolution layers is 4, the node embedding dimension is 128, the learning rate is 0.003, and the number of attention heads is 4. Finally, the multi-objective loss value on the validation set converges to 0.067. After loading the optimized configuration, the system performs multi-objective physical property prediction on 45 new samples, outputs indicators such as thermal conductivity, electrical conductivity, Young's modulus, Poisson's ratio, dielectric constant, and binding energy, and compares them with the manually measured data completed within the following 6 days. The results show that the mean squared errors of the predictions for the three indicators of thermal conductivity, electrical conductivity, and binding energy are 1.42 W / m·K, 0.06 S / cm, and 0.031 eV respectively, which are all better than the traditional CGCNN method (MSEs are 2.35, 0.11, and 0.065 respectively). More importantly, this system can identify deviation samples and automatically generate enhanced perturbation data. Two thermal conductivity prediction deviation samples were found on the 9th day after deployment. The system generated 4 groups of enhanced graph data through perturbation, and the error decreased by 67.3% after performing local fine-tuning, realizing the adaptive iterative evolution of the model during the application process.

[0054] The following table shows the comparison data of the true physical properties and prediction results of some samples in the above tasks.

[0055] Table 1 Comparison table of predicted and measured properties of HfC-SiC based composite material samples

[0056]

[0057] From the comparison results of the predicted and measured performance of the HfC-SiC-based composite material samples shown in Table 1 above, it can be clearly observed that the prediction accuracy of the system of the present invention has reached a high level in many key physical properties. Taking thermal conductivity as an example, the difference between the predicted values ​​and the measured values ​​of multiple samples is kept within ±1.5W / m·K. Among them, the measured value of the thermal conductivity of the HSC-001 sample is 32.6W / m·K, and the corresponding predicted value is 31.7W / m·K, with an error of only 2.76%. This result fully reflects the modeling accuracy of this system in processing thermal conductivity properties. In terms of electrical conductivity, the system also shows stable performance, and the error of most samples is less than 0.08S / cm, indicating that the model after integrating the physical field coupling diagram channel has a good perception of current-carrying characteristics. As a core indicator of the mechanical properties of materials, the prediction error of Young's modulus is controlled within ±10GPa. For example, the measured value of HSC-014 is 450.8GPa, and the predicted value is 457.6GPa, with an error of about 1.5%, which is a high-precision level in the structural ceramic prediction scenario.

[0058] In terms of the prediction of dielectric constant, the model output results are highly consistent with the measured data. For example, the predicted value of the HSC-022 sample is 7.09, and the measured value is 7.12, with almost no offset, which verifies the effectiveness of the system in modeling the distribution of electrical properties under high-frequency response field conditions. In terms of the binding energy index in the structural stability assessment, the error between the predicted value and the measured value of each sample is controlled within 0.05eV, showing the high reliability of this system in processing deep structural quantum expressions. A comprehensive analysis of the tabular data shows that this system captures the linkage characteristics of materials in multiple physical dimensions by constructing three types of heterogeneous graph channels: atomic structure, defect perturbation, and physical field, and continuously improves the prediction performance through structural self-optimization and continuous learning modules. It has industrial practical value and promotion prospects in multiple key performance dimensions such as heat, force, electricity, and structure.

[0059] The results of this embodiment verify the system integration advantages of the present invention in high-dimensional material structure modeling, multi-graph heterogeneous information fusion, model structure self-optimization and post-prediction feedback enhancement. It is particularly suitable for engineering research and development environments that require a large number of material property predictions, have complex structures or significant defects, and dynamically changing data distribution, providing reliable support for the next generation of intelligent material screening.

[0060] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An intelligent material property prediction system based on graph neural network, characterized in that, It includes the following modules: a structure diagram construction module, which is used to parse material structure data, construct an atomic structure diagram, a defect perturbation diagram, and a physical field coupling diagram, and map them into a third-order tensor diagram; a tensor diagram encoding module, which is used to input the third-order tensor diagram into a multi-scale tensor diagram neural network model to generate a high-dimensional tensor node embedding representation. The multi-scale tensor diagram neural network model includes a three-channel diagram encoder module, a tensor attention fusion module, and an output embedding generation module. The three-channel diagram encoder module is composed of an atomic structure diagram encoder, a defect perturbation diagram encoder, and a physical field coupling diagram encoder, corresponding to the three-channel inputs of the third-order tensor diagram respectively. The intermediate node features output by the three-channel diagram encoder module are respectively input into the tensor attention fusion module. The tensor attention fusion module assigns channel weight coefficients to the three types of diagram features to achieve dynamic weighted fusion. A skip connection mechanism is set in the tensor attention fusion module to connect the original node features and the intermediate node features respectively to generate the fused node features. In the output embedding generation module, the fused node features are normalized, and finally a high-dimensional tensor node embedding representation is output. An asymmetric propagation module, which is used to construct a directional tensor connection structure based on the high-dimensional tensor node embedding representation, design an in-ward propagation path and an out-ward propagation path, and perform asymmetric coupling propagation to generate an updated feature representation. A model optimization module, which is used to input the updated feature representation into an improved osprey swarm search algorithm to optimize the multi-scale tensor diagram neural network model and generate an optimal configuration. A multi-objective prediction module, which is used to load the optimal configuration and perform multi-objective material property prediction, and output prediction results of thermal properties, mechanical properties, electrical characteristics, and structural stability. A continuous learning module, which is used to identify deviation samples based on the prediction error, feedback and update the multi-scale tensor diagram neural network model, and write it into the learning buffer.

2. The intelligent material property prediction system based on a graph neural network according to claim 1, wherein The modules are implemented through the following methods: S1, parse the structure data of the material, construct three types of diagram channels including an atomic structure diagram, a defect perturbation diagram, and a physical field coupling diagram, and map them into a third-order tensor diagram according to the node alignment rule; S2, import the third-order tensor diagram into the multi-scale tensor diagram neural network model, encode the features of the three types of diagram channels respectively, and generate a high-dimensional tensor node embedding representation through the tensor attention fusion module; S3, based on the high-dimensional tensor node embedding representation, construct an asymmetric coupling propagation mechanism between nodes in the tensor space to generate an updated feature representation. S4, based on the updated feature representation, embed an improved osprey swarm search algorithm to optimize the multi-scale tensor diagram neural network model, comprehensively evaluate the training loss and accuracy of each configuration using a fitness function, and select the optimal configuration; S5, use the optimal configuration for the material property prediction task, and output multi-objective prediction results, where the multi-objective prediction results include thermal properties, mechanical properties, electrical characteristics, and structural stability; S6, based on the actual error feedback of the multi-objective prediction results, introduce a continuous learning mechanism, construct the deviation sample set as an enhanced sample, and feedback and update the multi-scale tensor diagram neural network model.

3. The intelligent material property prediction system based on a graph neural network according to claim 2, wherein The specific steps of S1 are as follows: S11: Analyze the structural data of the material, extract the three-dimensional spatial coordinate values of each atom and the atomic type code, and construct an atomic structure diagram. Each atom corresponds to a node in the atomic structure diagram. When the Euclidean distance between any two atomic nodes satisfies , an edge connection is established between the two atomic nodes, where represents the th atomic node, represents the th atomic node, represents the set bond threshold; S12: Automatically identify vacancies, dislocations, interstitial atoms, and impurity defect types in the atomic structure diagram. For each detected defect center, mark it as a defect node , where represents the th defect center. Build a subgraph within the range of the local perturbation region with the defect node as the core. Represent the defect node and neighboring atoms as defect perturbation graph nodes, and establish the influence path between the defect and surrounding atoms as an edge to form a defect perturbation graph; S13: Collect multi-physical field data of the material. The multi-physical field data includes stress tensor, electric field intensity vector, and magnetic flux density vector. Based on the same set of atomic nodes in the atomic structure diagram, construct a physical field coupling graph, and introduce a physical field difference weight in the way of building edges: ; where represents the physical field difference weight between the th atomic node and the th atomic node, , and represent weight coefficients, represents the gradient of the stress tensor between the th atomic node and the th atomic node, represents the electric field intensity vector of the th atomic node, represents the electric field intensity vector of the th atomic node, represents the magnetic flux density vector of the th atomic node, represents the magnetic flux density vector of the th atomic node, denotes the Euclidean norm; S14. Align the coordinates of all node sets in the atomic structure diagram, defect perturbation diagram, and physical field coupling diagram. For nodes with a spatial position difference less than the preset error threshold, merge them into a unified structural position node, and finally generate a unified node set; S15. Retain the edge connection relationships of the three types of diagram channels respectively, and construct a third-order tensor diagram, where the first dimension and the second dimension represent node pairs, and the third dimension represents the diagram channel number. Each position in the third-order tensor is defined as: ; where represents the third-order tensor diagram structure, represents the first dimension, represents the second dimension, represents the third dimension, represents the th atomic node, represents the th atomic node, represents the atomic structure diagram channel, represents the defect perturbation diagram channel, represents the physical field coupling diagram channel.

4. An intelligent material property prediction system based on a graph neural network according to claim 3, characterized in that, The jump connection mechanism is used to connect the original node features of the three-channel graph encoder with the intermediate node features after channel fusion during the fusion operation. Specifically, it includes retaining the original node features after the output of each channel graph encoder, and performing feature superposition after dimension alignment of the original node features and the intermediate node features after channel fusion.

5. An intelligent material property prediction system based on a graph neural network according to claim 2, characterized in that, Specifically, S3 includes: S31. Using the high-dimensional tensor node embedding representation as input, constructing a tensor connection structure between nodes, and assigning directional labels to each edge in the tensor connection structure to form a node adjacency relationship with direction differentiation; S32. Based on the node adjacency relationship, constructing an asymmetric coupling propagation mechanism. The asymmetric coupling propagation mechanism includes an in-coming feature propagation path and an out-coming feature propagation path. Define the feature update formula for nodes during the propagation process as: ; where represents the updated feature of the -th node in the -th layer, represents the activation function, represents the updated feature of the -th node in the -th layer, represents the updated feature of the -th node in the -th layer, represents the weight matrix of the in-coming propagation path, represents the weight matrix of the out-coming propagation path, represents the set of in-coming neighbor nodes of the -th node, represents the set of out-coming neighbor nodes of the -th node, represents the attention coefficient of the in-coming path, represents the attention coefficient of the out-coming path; S33. Performing asymmetric coupling propagation on the high-dimensional tensor node embedding representation, and respectively transmitting through the in-coming path and the out-coming path to generate a node-level local feature propagation tensor; S34. Performing channel adaptive normalization processing on the local feature propagation tensor, and using a weighted summation operation to integrate local asymmetric propagation features to update the overall node representation; S35. Outputting the updated feature representation after being updated by the asymmetric coupling propagation mechanism.

6. The intelligent material property prediction system based on a graph neural network according to claim 2, characterized in that, S4 specifically includes: S41. Convert the updated feature representation into a configuration vector, and initialize the osprey population with a size of , where each configuration vector includes the node embedding dimension, the number of graph convolution layers, the attention weight coefficient, the learning rate, and the message propagation coefficient; S42. In the th iteration, update the osprey individual position vector according to the global gliding search strategy, and according to the centroid of the osprey population and the global optimal osprey individual : ; where, ; among them, represents the position vector of the th osprey individual in the th iteration, represents the position vector of the th osprey individual in the th iteration, represents the maximum number of iterations, and represent the values randomly sampled from the uniform distribution ; S43. Call the multi-scale tensor graph neural network model training sub-process for each updated osprey individual position vector, and calculate the training loss and the validation accuracy on the validation set, and calculate the fitness score of each osprey individual: ; where, represents the fitness score of the th osprey individual, and represent the preset non-negative weight coefficients; S44. When the iteration round reaches the gliding stage threshold , where , switch to the dive refinement stage, and execute through the gradient sign-guided dive strategy: ; where, represents the step size coefficient, represents the sign function, represents the gradient operator for the configuration vector; S45. Maintain the global memory channel data structure, and the global memory channel data structure stores the historical optimal solution pair , where represents the fitness score of the optimal osprey individual in the th iteration. When the optimal individual in the current iteration round is better than the worst record in the memory, update the memory channel; S46. Terminate when the number of iterations reaches the maximum number of iterations or the change in the global optimal fitness is less than the convergence threshold , where represents the The fitness score of the optimal osprey individual during the round iteration, and output the solution set of the current osprey population as the optimal configuration.

7. The intelligent material property prediction system based on a graph neural network according to claim 2, wherein Specifically, S5 includes: S51, loading the obtained optimal configuration parameter set into the multi-scale tensor graph neural network model; S52, decoupling and modeling the multi-object prediction task, establishing four independent task subspaces including thermal properties, mechanical properties, electrical characteristics, and structural stability. Each task subspace constructs an attribute-specific prediction head structure and connects it with the shared node embedding representation; S53, constructing a parallel decoder structure, setting a separation path between the input channel of the shared representation and the output channels of the four independent task subspaces. Each output channel consists of two layers of linear mapping units and a non-linear conversion unit, and is respectively connected to various physical property index nodes at the output end; S54, setting two prediction dimensions of thermal conductivity and thermal diffusivity for thermal properties, two prediction dimensions of Young's modulus and Poisson's ratio for mechanical properties, two prediction dimensions of dielectric constant and conductivity for electrical characteristics, and two prediction dimensions of binding energy and lattice constant for structural stability; S55, outputting the multi-object prediction results of materials including thermal properties, mechanical properties, electrical characteristics, and structural stability.

8. The intelligent material property prediction system based on a graph neural network according to claim 2, characterized in that, Specifically, S6 includes: S61, comparing the error between the multi-object prediction results and the known true physical property parameters of materials, and extracting all samples with errors exceeding the preset threshold to form a deviation sample set; S62, performing an enhanced perturbation operation on each data sample in the deviation sample set, injecting perturbations into the original third-order tensor graph structure to generate enhanced samples; S63, constructing a local fine-tuning data set based on the enhanced samples, freezing the general structure layer of the multi-scale tensor graph neural network model, and only activating the fusion layer and the attribute decoder sub-module related to the current deviation sample to perform fine-tuning training; S64, after completing the fine-tuning training, updating the parameter set participating in activation in the multi-scale tensor graph neural network model, and caching all enhanced samples, perturbation source paths, label correction residuals, and error response indicators into the continuous learning buffer.

9. The intelligent material property prediction system based on a graph neural network according to claim 8, wherein, The enhanced perturbation operation specifically includes: adding a random offset with an amplitude not exceeding 3% to the three-dimensional coordinates of atomic nodes to form a spatial position perturbation; multiplying any edge weight by a perturbation factor to form an edge weight perturbation, where the value range of the perturbation factor is [0.9, 1.1]; adding Gaussian noise with a mean of 0 to the node attribute vector as a node attribute perturbation.

Images