Intelligent material property prediction system based on graph neural network

Abstract

The invention discloses an intelligent material property prediction system based on a graph neural network, and the system comprises the following modules: a structure graph construction module which is used for analyzing material structure data and constructing a three-order tensor graph; the tensor graph coding module is used for inputting the third-order tensor graph into a multi-scale tensor graph neural network model to generate high-dimensional tensor node embedding representation; the asymmetric propagation module is used for constructing a directional tensor connection structure based on the high-dimensional node embedding representation and generating an updated feature representation; the model optimization module is used for inputting the updated feature representation into an improved eagle swarm search algorithm to generate optimal configuration; the multi-target prediction module is used for loading optimal configuration and performing multi-target material property prediction; and the continuous learning module is used for identifying a deviation sample based on the prediction error, feeding back and updating the multi-scale tensor graph neural network model, and writing the model into a learning buffer area. According to the invention, multi-channel graph modeling and an improved eagle swarm search algorithm are fused, and an intelligent material property prediction system is constructed.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and material science and technology, and in particular to an intelligent material property prediction system based on 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 that rely on experiments or rule-driven methods have gradually failed to meet the needs of high-throughput material design and performance evaluation. In order to achieve high-precision modeling of material properties, academia and industry have introduced intelligent prediction technologies based on artificial intelligence, especially graph neural networks (GNN) and other graph structure learning models. Because they can characterize the topological connection and physical coupling relationship between atoms inside the material, they have become a key direction in material intelligent prediction research in recent years.

[0003] Existing methods for predicting material properties based on graph neural networks generally regard materials as atomic structure graphs with nodes as atoms and edges as bonds, and use the GNN model to learn the graph structure representation between nodes, and then predict target physical properties such as thermal conductivity, electrical conductivity, Young's modulus, lattice constant, etc. Typical public technologies such as CGCNN (Crystal Graph Convolutional Neural Network) and MEGNet (Materials Graph Network) can accurately fit the physical properties of crystal structures in existing databases. However, most of these technologies are limited to the processing of static topological structures between atoms, and lack the ability to model dynamic information such as complex defect structures and physical field distribution inside materials, resulting in a significant decrease in the performance of the model when processing complex materials with multi-scale coupling characteristics or non-ideal structures. At the same time, existing methods are often trained based on a single graph channel, which cannot fully integrate the interaction between the microstructure of the material and the 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, lack the ability to jointly model defect perturbation graphs and physical field coupling graphs, and have difficulty in characterizing the heterogeneous characteristic 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 distribution, thereby affecting the overall performance of the material, and traditional methods often find it difficult to introduce such local perturbations into a unified graph structure. At the same time, for material systems under different boundary conditions or under the action of multiple physical fields, existing graph neural networks have significant deficiencies in constructing physically sensitive graph structures and processing dynamic changes in edge weight information.

[0005] In terms of model design, current GNN methods mostly use fixed structures or empirically set hyperparameter configurations, and lack targeted tuning paths in terms of model depth, node dimension, fusion mechanism, etc., which leads to slow convergence, overfitting, or poor generalization when facing material systems with inconsistent structural complexity. Some work attempts to introduce evolutionary algorithms or meta-learning methods to search for model parameters, but there is still a lack of optimization mechanisms that are deeply integrated with the graph neural network structure, and it is impossible to 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 with different physical distribution laws such as thermal properties, mechanical properties, electrical properties, and structural stability, gradient conflicts and training instability often occur. In addition, due to the complexity of the material itself and the uneven distribution of data, the prediction accuracy of existing technologies is significantly reduced 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 feedback learning mechanisms. Even when there are systematic errors in model predictions, it is often impossible to proactively identify the source of prediction deviations and conduct continuous training corrections. Especially in the actual deployment stage, the prediction results of materials may be offset due to factors such as changes in boundary conditions and disturbances of defective samples. Traditional models cannot be dynamically updated or fine-tuned, resulting in insufficient model accuracy and stability. Therefore, the existing technology still lacks a graph neural network prediction system with continuous learning, adaptive fine-tuning and enhanced data generation capabilities, especially in terms of technical integration combining multi-source graph structures, multi-property outputs and model optimization feedback loops.

[0008] Therefore, how to provide an intelligent material property prediction system based on graph neural networks is an urgent problem that technical personnel in this field need to solve. Summary of the invention

[0009] One purpose of the present invention is to propose an intelligent material property prediction system based on graph neural network. The present invention integrates multi-channel graph modeling and improved Osprey swarm search algorithm to construct an intelligent material property prediction system based on graph neural network. By introducing the third-order tensor structure of atomic structure graph, defect perturbation graph and physical field coupling graph, the model's structural perception ability of complex materials is improved. At the same time, the adaptive optimization of the prediction model is realized through a continuous learning mechanism. 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 types of materials.

[0010] According to an embodiment of the present invention, a smart material property prediction system based on a graph neural network includes the following modules: a structure graph construction module, which is used to parse material structure data, construct an atomic structure graph, a defect perturbation graph, and a physical field coupling graph, and map them into a third-order tensor graph; a tensor graph encoding module, which is used to input the third-order tensor graph into a multi-scale tensor graph neural network model to generate a high-dimensional tensor node embedding representation, wherein the multi-scale tensor graph 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, which is used to construct a method based on the high-dimensional node embedding representation. A directional tensor connection structure is built, and the incoming propagation path and the outgoing propagation path are designed to perform asymmetric coupled propagation and generate updated feature representations; a model optimization module is used to input the updated feature representation characteristics into the improved Osprey swarm search algorithm, optimize the multi-scale tensor graph neural network model, and generate the optimal configuration; a multi-objective prediction module is used to load the optimal configuration and perform multi-objective material property prediction, and output the prediction results of thermal properties, mechanical properties, electrical properties and structural stability; a continuous learning module is used to identify deviation samples based on prediction errors, feedback update the multi-scale tensor graph neural network model, and write it into the learning buffer.

[0011] According to an embodiment of the present invention, a method for predicting properties of intelligent materials based on a graph neural network comprises the following steps: S1, analyzing the structural data of the material, constructing three types of graph channels including an atomic structure graph, a defect perturbation graph and a physical field coupling graph, and mapping them into a third-order tensor graph according to a node alignment rule; S2, importing the third-order tensor graph into a multi-scale tensor graph neural network model, performing feature encoding on the three types of graph channels respectively, and generating a high-dimensional tensor node embedding representation through a tensor attention fusion module; S3, constructing an asymmetric coupling propagation mechanism between nodes in the tensor space based on the high-dimensional tensor node embedding representation, and generating an updated feature table S4. Based on the updated feature representation, the improved Osprey swarm search algorithm is embedded to optimize the multi-scale tensor graph neural network model, and the fitness function is used to comprehensively evaluate the training loss and accuracy of each configuration to select the optimal configuration; S5. The optimal configuration is used for the material property prediction task, and the multi-objective prediction results are output. The multi-objective prediction results include thermal properties, mechanical properties, electrical properties and structural stability; S6. Based on the actual error feedback of the multi-objective prediction results, a continuous learning mechanism is introduced to construct the deviation sample set as an enhanced sample, and the multi-scale tensor graph neural network model is updated by feedback.

[0012] Optionally, the S1 specifically includes: S11, parsing the structural data of the material, extracting the three-dimensional spatial coordinate value of each atom and the type code of the atom, and constructing an atomic structure diagram, where each atom corresponds to an atomic structure diagram node. When the Euclidean distance between any two atomic nodes satisfies , establish an edge connection between two atomic nodes, where Indicates Atomic nodes, Indicates Atomic nodes, Indicates the set bonding threshold; S12, automatically identify the types of vacancies, dislocations, interstitial atoms and impurity defects in the atomic structure diagram, and mark each detected defect center as a defect node ,in Indicates A defect center is selected, and a subgraph within the local disturbance area is constructed with the defect node as the core. The defect node and the adjacent atoms are represented as defect disturbance graph nodes, and the influence paths between the defect and the surrounding atoms are established as edges to form a defect disturbance graph; S13, multi-physical field data of the material is collected, and 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 graph, a physical field coupling graph is constructed, and physical field difference weights are introduced in the edge construction method: ;in, Indicates Atomic nodes and The physical field difference weights between atomic nodes, , and represents the weight coefficient, Denotes the stress tensor in Atomic nodes and The gradient between atomic nodes, Indicates The electric field strength vector of each atomic node, Indicates The electric field strength vector of each atomic node, Indicates The magnetic flux density vector of the atomic node, Indicates The magnetic flux density vector of the atomic node, Represents the Euclidean norm; S14, coordinate alignment of all node sets in the atomic structure diagram, defect perturbation diagram and physical field coupling diagram, merge nodes with spatial position differences less than a preset error threshold into unified structural position nodes, and finally generate a unified node set; S15, retain the edge connection relationship of the three types of graph channels respectively, and construct a third-order tensor graph, in which the first dimension and the second dimension represent node pairs, and the third dimension represents the graph channel number, and each position in the third-order tensor is defined as: ;in, represents the third-order tensor graph structure, represents the first dimension, represents the second dimension, represents the third dimension, Indicates Atomic nodes, Indicates Atomic nodes, represents the atomic structure diagram channel, represents the defect perturbation map channel, Represents a physics coupling plot channel.

[0013] Optionally, the S2 specifically includes: S21, constructing a multi-scale tensor graph neural network model, the multi-scale tensor graph neural network model 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 graph encoder, a defect perturbation graph encoder and a physical field coupling graph encoder, which respectively correspond to the three channel inputs of the third-order tensor graph; S22, inputting 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, setting a jump connection mechanism in the tensor attention fusion module to respectively connect the original node features and the intermediate node features to generate fused node features; S24, normalizing the fused node features in the output embedding generation module, and finally outputting a high-dimensional tensor node embedding representation.

[0014] Optionally, 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 in a fusion operation, specifically including retaining the original node features after the output of each channel graph encoder, and superimposing the original node features and the intermediate node features after channel fusion after dimensional alignment.

[0015] Optionally, the S3 specifically includes: S31, taking the high-dimensional tensor node embedding representation as input, constructing a tensor connection structure between nodes, and assigning a directional mark to each edge in the tensor connection structure to form a node adjacency relationship with direction distinction; S32, constructing an asymmetric coupling propagation mechanism based on the node adjacency relationship, the asymmetric coupling propagation mechanism includes an inbound feature propagation path and an outbound feature propagation path, and defining the feature update formula of the node during the propagation process as: ;in, Indicates The node in The updated features of the layer, represents the activation function, Indicates The node in The updated features of the layer, Indicates The node in The updated features of the layer, represents the weight matrix of the incoming propagation path, Represents the weight matrix of the propagation path, Indicates The set of incoming neighbor nodes of a node, Indicates The set of outgoing neighbor nodes of a node, represents the attention coefficient of the incoming path, Represents the attention coefficient of the outgoing path; S33, performs asymmetric coupling propagation on the high-dimensional tensor node embedding representation, and transmits it through the incoming path and the outgoing path respectively to generate a local feature propagation tensor at the node level; S34, performs channel-adaptive normalization on the local feature propagation tensor, and uses a weighted sum operation to integrate the local asymmetric propagation features to update the overall node representation; S35, outputs the updated feature representation after being updated by the asymmetric coupling propagation mechanism.

[0016] Optionally, the S4 specifically includes: S41, converting the updated feature representation into a configuration vector, with an initialization size of Osprey population , where each configuration vector includes node embedding dimension, number of graph convolution layers, attention weight coefficient, learning rate and message propagation coefficient; S42, in the In the round iteration, the individual position vector of the osprey is updated according to the global gliding search strategy, and the centroid of the osprey population is updated according to the centroid of the osprey population. And the global optimal osprey individual : ;in, Indicates In the round iteration The individual position vectors of ospreys, Indicates In the round iteration The individual position vectors of ospreys, represents the maximum number of iterations, and Represents a uniform distribution S43, calling the multi-scale tensor graph neural network model training sub-process for each updated osprey individual position vector, and calculating the training loss on the validation set Verification accuracy , calculate the fitness score of each osprey individual: ;in, Indicates The fitness score of each osprey individual, and Represents a preset non-negative weight coefficient; S44, when the iteration round reaches the gliding stage threshold After that, , switch to the dive refinement stage, and guide the dive strategy execution through the gradient sign: ;in, represents the step size coefficient, represents the symbolic function, Represents the gradient operator of the configuration vector; S45, maintaining a global memory channel data structure, the global memory channel data structure stores the historical optimal solution ,in Indicates The fitness score of the best osprey individual in the round iteration, when the best individual in the current iteration round is better than the worst record in memory, the memory channel is updated; S46, when the number of iterations reaches the maximum number of iterations Or the global optimal fitness changes Less than the convergence threshold terminated when Indicates The fitness score of the best osprey individual in each round of iteration is calculated, and the current osprey population solution set is output as the optimal configuration.

[0017] Optionally, the S5 specifically includes: S51, loading the obtained optimal configuration parameter set into the multi-scale tensor graph neural network model; S52, decoupling modeling of multi-objective prediction tasks, establishing four independent task subspaces including thermal properties, mechanical properties, electrical properties and structural stability, each task subspace constructs an attribute-specific prediction head structure respectively, and connects with the shared node embedding representation; S53, constructing a parallel decoder structure, setting a separation path between the shared representation input channel and the four independent task subspace output channels, each output channel is composed of two layers of linear mapping units and nonlinear conversion units, and is connected to various physical property indicator nodes at the output end; S54, setting two prediction dimensions of thermal conductivity and thermal diffusivity for thermal properties, setting two prediction dimensions of Young's modulus and Poisson's ratio for mechanical properties, setting two prediction dimensions of dielectric constant and conductivity for electrical properties, and setting two prediction dimensions of binding energy and lattice constant for structural stability; S55, outputting multi-objective prediction results of materials including thermal properties, mechanical properties, electrical properties and structural stability.

[0018] Optionally, the S6 specifically includes: S61, performing error comparison between the multi-objective prediction results and the real physical property parameters of known materials, extracting all samples whose errors exceed a 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, and generating 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 attribute decoder submodule related to the current deviation sample to perform fine-tuning training; S64, after completing the fine-tuning training, updating the parameter set involved in the activation in the multi-scale tensor graph neural network model, and caching all enhanced samples and disturbance source paths, label correction residuals and error response indicators in the continuous learning buffer.

[0019] 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 node to constitute a spatial position perturbation; multiplying any edge weight by a perturbation factor to form an 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 a node attribute perturbation.

[0020] The beneficial effects of the present invention are as follows: First, the present invention breaks through the limitation of traditional GNN models that only model based on static atomic graphs by constructing three types of heterogeneous graph channels, namely atomic structure graph, defect perturbation graph and physical field coupling graph, and uniformly maps them into a third-order tensor graph structure using node alignment rules. It realizes comprehensive integrated modeling of material structure hierarchical information, local defect interference and external physical field influence, greatly improving the model's expression ability and perception depth for non-ideal complex material systems.

[0021] Secondly, 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, so that each graph channel maintains feature decoupling in the encoding stage and achieves weighted aggregation in the fusion stage. This not only enhances the model's ability to distinguish information between heterogeneous channels, but also avoids redundant feature interference and gradient conflicts, ensuring the physical consistency and expression stability of node representation. In the asymmetric propagation stage, the introduction of the inbound and outbound path separation mechanism effectively improves the model's propagation efficiency and update accuracy in processing directional complex topological structures.

[0022] In addition, the present invention introduces an improved Osprey swarm search algorithm, encodes the structural parameters and training hyperparameters of the multi-scale tensor graph neural network into optimizable vectors, and uses a strategy that combines gliding search with dive refinement to achieve dynamic adaptive optimization in a high-dimensional search space, which significantly improves the structural adaptability and training convergence of the model in different mission scenarios, and solves the problem that the traditional model structure relies on artificial experience settings and is difficult to globally tune.

[0023] Furthermore, a parallel decoding structure and task decoupling path were constructed in the multi-objective prediction module to predict thermal properties, mechanical properties, electrical characteristics and structural stability respectively, supporting attribute-level structural design and differentiated tuning. At the same time, a soft sharing mechanism was used to balance the gradient competition relationship among various physical property tasks, achieving multi-indicator collaborative output under a unified model architecture, and enhancing the generalization and engineering application capabilities of the model.

[0024] Finally, the present invention constructs a continuous learning mechanism based on prediction error feedback, which can automatically trigger the enhanced sample generation process when the model output deviation exceeds the threshold, and generates local enhanced samples by combining spatial perturbation, edge weight perturbation and node attribute perturbation strategies. It also dynamically updates 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 shortcomings of the traditional static GNN model in the actual deployment stage. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0026] In the attached picture: Figure 1 This is a schematic diagram of the structure of an intelligent material property prediction system based on graph neural network proposed in the present invention.

[0027] Figure 2 This is an overall flow chart of an intelligent material property prediction method based on graph neural network proposed in the present invention. DETAILED DESCRIPTION

[0028] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0029] refer to Figure 1, an intelligent material property prediction system based on graph neural network, comprising the following modules: a structure graph construction module, used to parse material structure data, construct atomic structure graph, defect perturbation graph and physical field coupling graph, and map them into a third-order tensor graph; a tensor graph encoding module, used to input the third-order tensor graph into a multi-scale tensor graph neural network model to generate a high-dimensional tensor node embedding representation, the multi-scale tensor graph 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, used to construct a directional tensor connection structure based on the high-dimensional node embedding representation, and design an inbound propagation path and an outbound propagation path, perform asymmetric coupling propagation, and generate an updated feature representation; a model optimization module, used to input the updated feature representation into an improved osprey swarm search algorithm, perform multi-scale tensor graph neural network model optimization, and generate an optimal configuration; a multi-target prediction module, used to load the optimal configuration and perform multi-target material property prediction, and output the prediction results of thermal properties, mechanical properties, electrical properties and structural stability; a continuous learning module, used to identify deviation samples based on prediction errors, feedback update the multi-scale tensor graph neural network model, and write it into a learning buffer.

[0030] The intelligent material property prediction system provided by the present invention has an overall architecture based on a graph neural network. For the first time, the atomic-level topology, defect perturbations, and physical field information of the material structure are uniformly embedded in the third-order tensor graph modeling paradigm, and an 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, structural optimization, performance prediction, and continuous learning. The modules work together to retain the multi-scale structural characteristics of the material's ontology information and have highly flexible and dynamically adjustable prediction capabilities. The system can effectively enhance the model's perception of complex materials, defect structures, and changes in the physical environment, significantly improve the prediction accuracy and the model's adaptive evolution capabilities, and is suitable for intelligent performance analysis and high-throughput screening scenarios of multi-objective and multi-type materials.

[0031] refer to Figure 2, a method for predicting properties of intelligent materials based on graph neural networks, comprising the following steps: S1, parsing the structural data of the material, constructing three types of graph channels including atomic structure graph, defect perturbation graph and physical field coupling graph, and mapping them into third-order tensor graphs according to node alignment rules; S2, importing the third-order tensor graph into a multi-scale tensor graph neural network model, encoding the features of the three types of graph channels respectively, and generating a high-dimensional tensor node embedding representation through a tensor attention fusion module; S3, based on the high-dimensional tensor node embedding representation, constructing 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 the improved Osprey swarm search algorithm to optimize the multi-scale tensor graph neural network model, use the fitness function to comprehensively evaluate the training loss and accuracy of each configuration, and select the optimal configuration; S5, use the optimal configuration for material property prediction tasks, and output multi-objective prediction results, which include thermal properties, mechanical properties, electrical properties 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 to update the multi-scale tensor graph neural network model.

[0032] The method of the present invention defines the full-process data processing and task execution method driven by the graph neural network between each functional module, ensuring clear data and control flow between structural 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 analysis and deep representation of third-order tensor input by subsequent modules. By standardizing the graph structure processing flow, the system achieves higher computational consistency and scalability, and provides a clear and efficient control path for multi-objective prediction tasks. This method improves the engineering integration and application flexibility of the overall system, and contributes to the construction of a large-scale material data modeling and performance prediction platform.

[0033] In this embodiment, S1 specifically includes: S11, parsing the structural data of the material, extracting the three-dimensional spatial coordinate value of each atom and the type code of the atom, and constructing an atomic structure diagram, where each atom corresponds to an atomic structure diagram node. When the Euclidean distance between any two atomic nodes satisfies , establish an edge connection between two atomic nodes, where Indicates Atomic nodes, Indicates Atomic nodes, Indicates the set bonding threshold; S12, automatically identify the types of vacancies, dislocations, interstitial atoms and impurity defects in the atomic structure diagram, and mark each detected defect center as a defect node ,in Indicates A defect center is selected, and a subgraph within the local disturbance area is constructed with the defect node as the core. The defect node and the adjacent atoms are represented as defect disturbance graph nodes, and the influence paths between the defect and the surrounding atoms are established as edges to form a defect disturbance graph; S13, multi-physical field data of the material is collected, and 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 graph, a physical field coupling graph is constructed, and physical field difference weights are introduced in the edge construction method: ;in, Indicates Atomic nodes and The physical field difference weights between atomic nodes, , and represents the weight coefficient, Denotes the stress tensor in Atomic nodes and The gradient between atomic nodes, Indicates The electric field strength vector of each atomic node, Indicates The electric field strength vector of each atomic node, Indicates The magnetic flux density vector of the atomic node, Indicates The magnetic flux density vector of the atomic node, Represents the Euclidean norm; S14, coordinate alignment of all node sets in the atomic structure diagram, defect perturbation diagram and physical field coupling diagram, merge nodes with spatial position differences less than a preset error threshold into unified structural position nodes, and finally generate a unified node set; S15, retain the edge connection relationship of the three types of graph channels respectively, and construct a third-order tensor graph, in which the first dimension and the second dimension represent node pairs, and the third dimension represents the graph channel number, and each position in the third-order tensor is defined as: ;in, represents the third-order tensor graph structure, represents the first dimension, represents the second dimension, represents the third dimension, Indicates Atomic nodes, Indicates Atomic nodes, represents the atomic structure diagram channel, represents the defect perturbation map channel, Represents a physics coupling plot channel.

[0034] This step specifically defines the construction process of the third-order tensor graph, and explicitly integrates the atomic structure graph, defect perturbation graph, and physical field coupling graph through node alignment and edge weight difference 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 problem of information loss caused by channel separation in traditional graph modeling methods, and enhances the overall expression of internal heterogeneity, defect perturbations, and physical gradients of the material. By precisely controlling the node alignment and edge construction strategies, the constructed third-order tensor graph has the advantages of strong structural completeness and high representation robustness, which significantly improves the stability and generalization ability of the graph model when dealing with non-ideal material structures.

[0035] In this embodiment, the S2 specifically includes: S21, constructing a multi-scale tensor graph neural network model, the multi-scale tensor graph neural network model 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 graph encoder, a defect perturbation graph encoder and a physical field coupling graph encoder, which respectively correspond to the three channel inputs of the third-order tensor graph; S22, inputting 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, setting a jump connection mechanism in the tensor attention fusion module, connecting the original node features and the intermediate node features respectively, and generating fused node features; S24, normalizing the fused node features in the output embedding generation module, and finally outputting a high-dimensional tensor node embedding representation.

[0036] This step proposes a tensor graph encoding structure based on multi-channel graph input. It adopts a three-branch graph encoder architecture combined with a tensor attention fusion module and a jump connection mechanism to effectively solve the problem of multi-graph heterogeneous structure fusion. Each graph channel models atomic topology, defect perturbation and physical field information respectively. The fusion module uses learnable weight coefficients to dynamically adjust the importance of different channels to ensure that the final generated node embedding representation has a more comprehensive physical semantic expression capability. At the same time, the jump connection mechanism enhances information fluidity and feature fidelity, which helps to improve training stability and structural interpretability. This structure significantly improves the expression efficiency and fusion quality of the model for multi-graph channel input.

[0037] In this embodiment, 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 in the fusion operation, specifically including retaining the original node features after the output of each channel graph encoder, and superimposing the features after dimensional alignment of the original node features and the intermediate node features after channel fusion. Through the jump connection mechanism, the original node features of the graph encoder are superimposed and connected with the intermediate node features during the feature fusion process, so that the fusion result retains the original structural information of the channel and the semantic features after learning, and enhances the multi-dimensional consistency of the node representation. This mechanism can effectively alleviate the information attenuation or misleading aggregation problems that may occur in the multi-channel fusion process, ensure that the model does not lose key initial geometric and topological features when the deep structure is propagated, thereby enhancing the expressive power and model stability of the graph representation, and improving the robustness and accuracy of the prediction results under complex material structures.

[0038] In this embodiment, S3 specifically includes: S31, taking the high-dimensional tensor node embedding representation as input, constructing a tensor connection structure between nodes, and assigning a directional mark to each edge in the tensor connection structure to form a node adjacency relationship with direction distinction; S32, constructing an asymmetric coupling propagation mechanism based on the node adjacency relationship, the asymmetric coupling propagation mechanism includes an inbound feature propagation path and an outbound feature propagation path, and defining the feature update formula of the node during the propagation process as: ;in, Indicates The node in The updated features of the layer, represents the activation function, Indicates The node in The updated features of the layer, Indicates The node in The updated features of the layer, represents the weight matrix of the incoming propagation path, Represents the weight matrix of the propagation path, Indicates The set of incoming neighbor nodes of a node, Indicates The set of outgoing neighbor nodes of a node, represents the attention coefficient of the incoming path, Represents the attention coefficient of the outgoing path; S33, performs asymmetric coupling propagation on the high-dimensional tensor node embedding representation, and transmits it through the incoming path and the outgoing path respectively to generate a local feature propagation tensor at the node level; S34, performs channel-adaptive normalization on the local feature propagation tensor, and uses a weighted sum operation to integrate the local asymmetric propagation features to update the overall node representation; S35, outputs the updated feature representation after being updated by the asymmetric coupling propagation mechanism.

[0039] This step proposes an asymmetric propagation mechanism based on high-dimensional node embedding. By constructing two propagation paths, inbound and outbound, and setting the direction-aware weight matrix and attention coefficient respectively, it realizes the effective distinction of the direction of information flow between nodes. This design can accurately simulate the local asymmetry and anisotropy existing in the microstructure of materials, and effectively enhances the learning ability of graph neural networks for direction-sensitive structures. Through the process of directional modeling and channel adaptive normalization, the local consistency and topological responsiveness of the updated feature representation are significantly improved, providing a more physically interpretable graph embedding foundation for subsequent prediction tasks.

[0040] In this implementation, S4 specifically includes: S41, converting the updated feature representation into a configuration vector, the initialization size of which is Osprey population , where each configuration vector includes node embedding dimension, number of graph convolution layers, attention weight coefficient, learning rate and message propagation coefficient; S42, in the In the round iteration, the individual position vector of the osprey is updated according to the global gliding search strategy, and the centroid of the osprey population is updated according to the centroid of the osprey population. And the global optimal osprey individual : ;in, Indicates In the round iteration The individual position vectors of ospreys, Indicates In the round iteration The individual position vectors of ospreys, represents the maximum number of iterations, and Represents a uniform distribution S43, calling the multi-scale tensor graph neural network model training sub-process for each updated osprey individual position vector, and calculating the training loss on the validation set Verification accuracy , calculate the fitness score of each osprey individual: ;in, Indicates The fitness score of each osprey individual, and Represents a preset non-negative weight coefficient; S44, when the iteration round reaches the gliding stage threshold After that, , switch to the dive refinement stage, and guide the dive strategy execution through the gradient sign: ;in, represents the step size coefficient, represents the symbolic function, Represents the gradient operator of the configuration vector; S45, maintaining a global memory channel data structure, the global memory channel data structure stores the historical optimal solution ,in Indicates The fitness score of the best osprey individual in the round iteration, when the best individual in the current iteration round is better than the worst record in memory, the memory channel is updated; S46, when the number of iterations reaches the maximum number of iterations Or the global optimal fitness changes Less than the convergence threshold terminated when Indicates The fitness score of the best osprey individual in each round of iteration is calculated, and the current osprey population solution set is output as the optimal configuration.

[0041] This step proposes to encode the graph neural network structure and hyperparameter configuration into vector form, and constructs an adaptive model optimization framework by improving the Osprey swarm search algorithm to perform joint optimization of the gliding and diving stages, and introducing global memory channels, fitness functions and convergence control strategies. This mechanism can efficiently search for the best structural combination in different material systems, avoid performance fluctuations caused by manual parameter adjustment, 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.

[0042] In this embodiment, the S5 specifically includes: S51, loading the obtained optimal configuration parameter set into the multi-scale tensor graph neural network model; S52, decoupling modeling of multi-objective prediction tasks, establishing four independent task subspaces including thermal properties, mechanical properties, electrical properties 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 shared representation input channel and the four independent task subspace output channels, each output channel is composed of two layers of linear mapping units and nonlinear conversion units, and is connected to various physical property indicator nodes at the output end; S54, setting two prediction dimensions of thermal conductivity and thermal diffusivity for thermal properties, setting two prediction dimensions of Young's modulus and Poisson's ratio for mechanical properties, setting two prediction dimensions of dielectric constant and conductivity for electrical properties, and setting two prediction dimensions of binding energy and lattice constant for structural stability; S55, outputting multi-objective prediction results of materials including thermal properties, mechanical properties, electrical properties and structural stability.

[0043] This step divides the multi-target prediction task into four attribute subspaces through decoupling modeling and parallel decoder design, corresponding to thermal properties, mechanical properties, electrical properties and structural stability, and sets attribute-specific decoding paths after each channel. This structure avoids gradient interference and information confusion between attributes, and improves the prediction accuracy and stability of each type of indicator. The multi-channel decoding output path is combined with the shared representation input to ensure the physical independence of each attribute dimension while improving the system prediction efficiency, achieving dual optimization of task-level separation and representation sharing, and is suitable for multi-property integrated design scenarios.

[0044] In this embodiment, the S6 specifically includes: S61, performing error comparison between the multi-objective prediction results and the real physical property parameters of the known materials, extracting all samples whose errors exceed 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, and generating 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 attribute decoder submodule related to the current deviation sample to perform fine-tuning training; S64, after completing the fine-tuning training, updating the parameter set involved in the activation in the multi-scale tensor graph neural network model, and caching all enhanced samples and disturbance source paths, label correction residuals and error response indicators in the continuous learning buffer.

[0045] This step constructs a continuous learning mechanism based on prediction error feedback, dynamically identifies deviation samples through error comparison, performs perturbation enhancement to generate new samples, and implements local fine-tuning updates by freezing the backbone structure and activating the fusion layer and decoder. This mechanism enables the model to self-correct online and quickly adapt to the distribution of new samples, effectively alleviating the generalization degradation problem of traditional models during the deployment phase. Through continuous updates and buffer recording mechanisms, the model has strong iterativeness and high evolution capabilities, and is suitable for dynamic learning and stable performance output in complex material scenarios.

[0046] 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 the atomic node to form a spatial position perturbation; multiplying any edge weight by a perturbation factor to form an 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 a node attribute perturbation.

[0047] This step clearly defines 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 representation dimensions in material modeling, have structural controllability and semantic interpretability, can efficiently expand the bias sample set, and improve sample diversity and model robustness. The perturbation generation mechanism is seamlessly integrated with the feedback fine-tuning process to ensure that the continuous learning module has the ability to enhance samples at the structural level.

[0048] Example 1: In order to verify the feasibility of the present invention in implementation, 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, structural analysis and physical property prediction are carried out to evaluate their adaptability as thermal protective coating materials in extreme thermal environments. In this application scenario, the traditional method requires the use of a high-temperature thermal conductivity meter, a resistivity test system, and a stress loading platform to test thermal properties, electrical properties, and mechanical strength 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 microscopic defects, the test results are prone to fluctuations, affecting the stability of the evaluation results.

[0049] In the actual operation, the researchers first used high-precision electron beam tomography technology to obtain the internal atomic structure image of the composite material, and combined with the material preparation process records to extract atomic coordinates, element types and defect location information, and at the same time obtained the stress distribution, electric field distribution and magnetic induction intensity information of the material under the boundary load from the synchrotron radiation X-ray and laser speckle strain gauge. The system imports the above data into the graph construction module, generates atomic structure graphs, defect perturbation graphs and physical field coupling graphs 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 separately, extracts structural hierarchical features, defect response features and boundary field gradient features, and generates a unified node embedding representation through weighted attention mechanism in the fusion module. The nodes are then updated with information through an asymmetric propagation mechanism, and finally output high-dimensional tensor embedding.

[0050] 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 iterations and 12 rounds of diving fine-tuning, it automatically determines the number of graph convolution layers to be 4, the node embedding dimension to be 128, the learning rate to be 0.003, and the number of attention heads to be 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 manual measured data completed in the next 6 days. The results show that the prediction mean square errors of thermal conductivity, electrical conductivity and binding energy are 1.42W / m·K, 0.06S / cm and 0.031eV, respectively, which are better than the traditional CGCNN method (MSE is 2.35, 0.11 and 0.065, respectively). More importantly, the system can identify deviation samples and automatically generate enhanced perturbation data. On the 9th day after deployment, two thermal conductivity prediction deviation samples were found. The system generated 4 sets of enhanced graph data through perturbation. After local fine-tuning, the error dropped by 67.3%, realizing the adaptive iterative evolution of the model during the application process.

[0051] The following table shows the comparison data between the actual physical properties and the predicted results of some samples in the above tasks.

[0052] Table 1 Comparison of predicted and measured performance of HfC-SiC based composite materials

[0053] 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.

[0054] 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.

[0055] 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.

[0056] 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: The method comprises the following modules: a structure graph construction module, which is used to parse material structure data, construct atomic structure graphs, defect perturbation graphs and physical field coupling graphs, and map them into third-order tensor graphs; a tensor graph encoding module, which is used to input the third-order tensor graph into a multi-scale tensor graph neural network model to generate a high-dimensional tensor node embedding representation. The multi-scale tensor graph neural network model comprises a three-channel graph encoder module, a tensor attention fusion module and an output embedding generation module; An asymmetric propagation module, used to construct a directional tensor connection structure based on the high-dimensional tensor node embedding representation, and design an inbound propagation path and an outbound propagation path to perform asymmetric coupled propagation to generate an updated feature representation; The model optimization module is used to input the updated feature characteristics into the improved Osprey swarm search algorithm, perform multi-scale tensor graph neural network model optimization, and generate the optimal configuration; the multi-objective prediction module is used to load the optimal configuration and perform multi-objective material property prediction, and output the prediction results of thermal properties, mechanical properties, electrical properties and structural stability; A continuous learning module is used to identify deviation samples based on prediction errors, provide feedback to update the multi-scale tensor graph neural network model, and write to the learning buffer.

2. The intelligent material property prediction system based on graph neural network according to claim 1 is characterized in that: The modules are implemented by the following methods: 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 third-order tensor graphs according to the node alignment rules; S2. Import the third-order tensor graph into the 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 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, an improved Osprey swarm search algorithm is embedded to optimize the multi-scale tensor graph neural network model, and the training loss and accuracy of each configuration are comprehensively evaluated using the fitness function to select the optimal configuration; S5. The optimal configuration is used for the material property prediction task, and a multi-objective prediction result is output. The multi-objective prediction result includes thermal properties, mechanical properties, electrical properties and structural stability; S6. Based on the actual error feedback of the multi-objective prediction result, a continuous learning mechanism is introduced to construct the deviation sample set as an enhanced sample, and the multi-scale tensor graph neural network model is updated through feedback.

3. The intelligent material property prediction system based on graph neural network according to claim 2 is characterized in that: The S1 specifically includes: S11, parsing the structural data of the material, extracting the three-dimensional spatial coordinate value of each atom and the type code of the atom, and constructing an atomic structure diagram, where each atom corresponds to an atomic structure diagram node. When the Euclidean distance between any two atomic nodes satisfies , establish an edge connection between two atomic nodes, where Indicates Atomic nodes, Indicates Atomic nodes, Indicates the set bonding threshold; S12, automatically identify the types of vacancies, dislocations, interstitial atoms and impurity defects in the atomic structure diagram, and mark each detected defect center as a defect node ,in Indicates A defect center is selected, and a subgraph within the local disturbance area is constructed with the defect node as the core. The defect node and the adjacent atoms are represented as defect disturbance graph nodes, and the influence paths between the defect and the surrounding atoms are established as edges to form a defect disturbance graph; S13, multi-physical field data of the material is collected, and 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 graph, a physical field coupling graph is constructed, and physical field difference weights are introduced in the edge construction method: ;in, Indicates Atomic nodes and The physical field difference weights between atomic nodes, , and represents the weight coefficient, Denotes the stress tensor in Atomic nodes and The gradient between atomic nodes, Indicates The electric field strength vector of each atomic node, Indicates The electric field strength vector of each atomic node, Indicates The magnetic flux density vector of the atomic node, Indicates The magnetic flux density vector of the atomic node, Represents the Euclidean norm; S14, coordinate alignment of all node sets in the atomic structure diagram, defect perturbation diagram and physical field coupling diagram, merge nodes with spatial position differences less than a preset error threshold into unified structural position nodes, and finally generate a unified node set; S15, retain the edge connection relationship of the three types of graph channels respectively, and construct a third-order tensor graph, in which the first dimension and the second dimension represent node pairs, and the third dimension represents the graph channel number, and each position in the third-order tensor is defined as: ;in, represents the third-order tensor graph structure, represents the first dimension, represents the second dimension, represents the third dimension, Indicates Atomic nodes, Indicates Atomic nodes, represents the atomic structure diagram channel, represents the defect perturbation map channel, Represents a physics coupling plot channel.

4. The intelligent material property prediction system based on graph neural network according to claim 2 is characterized in that: The S2 specifically includes: S21, constructing a multi-scale tensor graph neural network model, the multi-scale tensor graph neural network model 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 graph encoder, a defect perturbation graph encoder and a physical field coupling graph encoder, which respectively correspond to the three channel inputs of the third-order tensor graph; S22, inputting 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, setting a jump connection mechanism in the tensor attention fusion module, connecting the original node features and the intermediate node features respectively, and generating fused node features; S24, normalizing the fused node features in the output embedding generation module, and finally outputting a high-dimensional tensor node embedding representation.

5. The intelligent material property prediction system based on graph neural network according to claim 4 is 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 in the fusion operation, specifically including retaining the original node features after the output of each channel graph encoder, and superimposing the original node features and the intermediate node features after channel fusion after dimensional alignment.

6. The intelligent material property prediction system based on graph neural network according to claim 2 is characterized in that: The S3 specifically includes: S31, taking the high-dimensional tensor node embedding representation as input, constructing a tensor connection structure between nodes, and assigning a directional mark to each edge in the tensor connection structure to form a node adjacency relationship with direction distinction; S32, constructing an asymmetric coupling propagation mechanism based on the node adjacency relationship, the asymmetric coupling propagation mechanism includes an inbound feature propagation path and an outbound feature propagation path, and defining the feature update formula of the node during the propagation process as: ;in, Indicates The node in The updated features of the layer, represents the activation function, Indicates The node in The updated features of the layer, Indicates The node in The updated features of the layer, represents the weight matrix of the incoming propagation path, Represents the weight matrix of the propagation path, Indicates The set of incoming neighbor nodes of a node, Indicates The set of outgoing neighbor nodes of a node, represents the attention coefficient of the incoming path, Represents the attention coefficient of the outgoing path; S33, performs asymmetric coupling propagation on the high-dimensional tensor node embedding representation, and transmits it through the incoming path and the outgoing path respectively to generate a local feature propagation tensor at the node level; S34, performs channel-adaptive normalization on the local feature propagation tensor, and uses a weighted sum operation to integrate the local asymmetric propagation features to update the overall node representation; S35, outputs the updated feature representation after being updated by the asymmetric coupling propagation mechanism.

7. The intelligent material property prediction system based on graph neural network according to claim 2 is characterized in that: The S4 specifically includes: S41, converting the updated feature representation into a configuration vector, the initialization size of which is Osprey population , where each configuration vector includes node embedding dimension, number of graph convolution layers, attention weight coefficient, learning rate and message propagation coefficient; S42, in the In the round iteration, the individual position vector of the osprey is updated according to the global gliding search strategy, and the centroid of the osprey population is updated according to the centroid of the osprey population. And the global optimal osprey individual : ;in, Indicates In the round iteration The individual position vectors of ospreys, Indicates In the round iteration The individual position vectors of ospreys, represents the maximum number of iterations, and Represents a uniform distribution S43, calling the multi-scale tensor graph neural network model training sub-process for each updated osprey individual position vector, and calculating the training loss on the validation set Verification accuracy , calculate the fitness score of each osprey individual: ;in, Indicates The fitness score of each osprey individual, and Represents a preset non-negative weight coefficient; S44, when the iteration round reaches the gliding stage threshold After that, , switch to the dive refinement stage, and guide the dive strategy execution through the gradient sign: ;in, represents the step size coefficient, represents the symbolic function, Represents the gradient operator of the configuration vector; S45, maintaining a global memory channel data structure, the global memory channel data structure stores the historical optimal solution ,in Indicates The fitness score of the best osprey individual in the round iteration, when the best individual in the current iteration round is better than the worst record in memory, the memory channel is updated; S46, when the number of iterations reaches the maximum number of iterations Or the global optimal fitness changes Less than the convergence threshold terminated when Indicates The fitness score of the best osprey individual in each round of iteration is calculated, and the current osprey population solution set is output as the optimal configuration.

8. The intelligent material property prediction system based on graph neural network according to claim 2, characterized in that: The S5 specifically includes: S51, loading the obtained optimal configuration parameter set into the multi-scale tensor graph neural network model; S52, decoupling modeling of multi-objective prediction tasks, establishing four independent task subspaces including thermal properties, mechanical properties, electrical properties and structural stability, each task subspace constructs an attribute-specific prediction head structure respectively, and connects with 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 nonlinear conversion units, and is connected to various physical property indicator nodes at the output end; S54, setting two prediction dimensions of thermal conductivity and thermal diffusivity for thermal properties, setting two prediction dimensions of Young's modulus and Poisson's ratio for mechanical properties, setting two prediction dimensions of dielectric constant and conductivity for electrical properties, and setting two prediction dimensions of binding energy and lattice constant for structural stability; S55, outputting multi-objective prediction results of materials including thermal properties, mechanical properties, electrical properties and structural stability.

9. The intelligent material property prediction system based on graph neural network according to claim 2 is characterized in that: The S6 specifically includes: S61, performing error comparison between the multi-objective prediction results and the real physical property parameters of known materials, extracting all samples whose errors exceed a 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, and generating 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 attribute decoder submodule related to the current deviation sample to perform fine-tuning training; S64, after completing the fine-tuning training, updating the parameter set involved in the activation in the multi-scale tensor graph neural network model, and caching all enhanced samples and perturbation source paths, label correction residuals and error response indicators in the continuous learning buffer.

10. The intelligent material property prediction system based on graph neural network according to claim 9, characterized in that: The enhanced perturbation operation specifically includes: adding a random offset with an amplitude not exceeding 3% to the three-dimensional coordinates of the atomic node to form a spatial position perturbation; multiplying any edge weight by a perturbation factor to form an 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 a node attribute perturbation.

Images