Blockchain association and aggregation method and system for large-scale material data

Abstract

This invention belongs to the field of materials big data technology and discloses a blockchain-based method and system for the aggregation and association of large-scale materials data. The method includes: constructing a heterogeneous graph of material entities using material description text, encoding and weighted pooling the graph to obtain semantic fingerprint vectors to characterize the comprehensive features of the materials; calculating the cosine similarity of the semantic fingerprint vectors and comparing it with an association threshold to determine implicit associations between the accessed material data, outputting association relationship records; calculating the hash value of the association relationship records, constructing a Merkle tree based on the hash values, storing the association relationship records on the blockchain network, and simultaneously uploading the material data to an off-chain storage medium to achieve reliable storage of association relationships and material data. This invention solves the problem of difficult association by introducing semantic fingerprints to automatically discover and establish reliable associations between data.

Description

Technical Field

[0001] This invention relates to the field of materials big data technology, and in particular to a blockchain-based method and system for the aggregation and association of large-scale materials data. Background Technology

[0002] With the deepening of the Materials Genome Initiative and the popularization of high-throughput experimental technologies, materials science research is undergoing a profound transformation from experience-based trial and error to a data-driven paradigm. The rapid development of large-scale efficient computing, automated experimental technologies, and artificial intelligence has led to an exponential increase in the generation speed of materials data. Many well-known materials data platforms have emerged internationally, accumulating millions of structural and performance data of materials.

[0003] However, materials data are characterized by being multi-source, heterogeneous, and dispersed: data sources include experimental measurements, computational simulations, literature, and patents; data formats include structured databases, semi-structured text, and unstructured charts; and data holders are scattered across different research institutions and enterprises, forming serious data silos. The correlation, fusion, and integration technologies for multi-source heterogeneous materials data are one of the key technological challenges in constructing a reliable data space for new materials. Existing technologies in the field of materials data management are evolving in two directions: Centralized data platforms, exemplified by the Materials Project, aggregate computational data through a unified data model and API interface for researchers to query and use. However, these platforms are primarily geared towards computational data, have limited coverage of experimental data, and cannot resolve the issue of the credibility of data sources.

[0004] Data storage based on blockchain: Recently, there have been attempts to apply blockchain technology to the protection of data intellectual property rights. For example, some patents have proposed to calculate the hash value of the original data and store it on the blockchain to realize the trusted verification of the data sharing process, which provides a technical basis for the tamper-proof storage of data. However, the main focus is on the ownership verification of the data itself, and it has failed to solve the problem of discovering and reliably recording the relationship between data.

[0005] Existing technologies propose a material network construction method, which includes data processing and cleaning: systematically collecting and organizing amorphous alloy data developed experimentally over the past 60 years to form a basic dataset; constructing a material network: unlike traditional tabular storage methods, this scheme represents binary and ternary amorphous alloy data as a graph structure (material network), where nodes represent alloy composition and edges represent the relationships between materials (such as compositional similarity); network topology mining: through in-depth analysis of the material network, it is discovered that fully connected subgraphs in the network can provide design ideas for multi-component alloy materials; simultaneously, a dynamic material network is constructed to analyze the evolution information of material development; and experimental design is guided: the network analysis results are used to predict new material systems, reduce the spatial complexity of materials, and provide new ideas for the development of complex alloy systems. The innovation of this method lies in transforming material data into a graph structure and discovering hidden material relationships through topology analysis.

[0006] However, this scheme has the following limitations: First, the construction of the association mainly relies on manually defined structured features, lacking automatic understanding of the deep semantics of composition-process-performance; second, the association results are stored in a centralized system, lacking anti-tampering mechanisms, and the credibility of the association cannot be guaranteed; third, it does not solve the problem of real-time updating of association when multiple sources of data are dynamically accessed.

[0007] Furthermore, in the field of materials science, data exhibits typical characteristics of being multi-source, heterogeneous, and dispersed. Due to the lack of unified data standards and correlation mechanisms, existing technologies mainly suffer from the following problems caused by causal chains: Data silos and lack of deep connections: Data from different entities (laboratories, enterprises) are isolated from each other. Traditional methods rely solely on keyword matching or simple tags for retrieval, failing to uncover the deep semantic relationships between material data in terms of composition, process, microstructure, and properties. This results in a large amount of potential cross-domain knowledge (such as the performance substitution relationships of materials in different systems) being submerged in massive amounts of data, making it difficult to form a knowledge graph that can support the research and development of new materials, thus limiting the release of data value.

[0008] Data credibility and traceability challenges: Traditional centralized database management models are at risk of data tampering, and the traceability records of experimental conditions are incomplete, resulting in low data reuse rates. Although existing blockchain technology attempts to solve the evidence preservation problem, directly uploading massive amounts of raw data to the chain can lead to storage explosion and performance bottlenecks. Furthermore, it cannot solve the source trust problem of whether the data has been tampered with before being uploaded to the chain, and lacks a reliable recording mechanism for the dynamic relationships between data.

[0009] Therefore, how to provide a blockchain-based method and system for the aggregation of large-scale material data is an urgent problem to be solved. Summary of the Invention

[0010] This invention provides a blockchain-based method and system for converging large-scale material data to address the aforementioned problems in the prior art.

[0011] According to a first aspect of the present invention, a blockchain-based method for the aggregation and association of large-scale material data is provided.

[0012] In one embodiment, the blockchain-based aggregation method for large-scale materials data includes: Material data is accessed based on data templates and information extraction rules, and the material data is processed to unify the format and align entities. The material description text is obtained based on the processing results. Material entity heterogeneous graphs are constructed using material description texts, and the graph entity heterogeneous graphs are encoded and weighted pooled based on graph neural networks to obtain semantic fingerprint vectors to characterize the comprehensive features of the materials. The cosine similarity of semantic fingerprint vectors is calculated based on density clustering algorithm, and the cosine similarity is compared with the association threshold to determine the implicit association between access material data, and the association relationship record is output. The hash value of the association record is calculated, and a Merkle tree is constructed based on the hash value. The association record is then stored in the blockchain network, while the material data is uploaded to the off-chain storage medium to achieve reliable storage of the association and material data.

[0013] In one embodiment, a material entity heterogeneous graph is constructed using material description text, and the graph is encoded and weighted pooled based on a graph neural network to obtain a semantic fingerprint vector, which characterizes the comprehensive features of the material, including: The sequence labels corresponding to the material description text are labeled and input into the entity recognition model. A classification layer is added on top of the entity recognition model to learn the semantic features of the material domain and output the recognized entity. The identified entities are mapped to standard concepts in the materials science ontology library, and the semantic code corresponding to each standard concept is output. Based on the semantic code and the material description text, a heterogeneous graph of material entities containing nodes and edges is constructed. The nodes include composition nodes, process nodes, structural nodes and performance nodes, and the edges represent the co-occurrence relationship or causal relationship between the identified entities. The material entity heterogeneous graph is encoded and the neighbor node information is aggregated by graph neural network. Based on the aggregation result, the embedding vector representing the node is output. The embedding vector is then weighted and pooled to obtain the semantic fingerprint vector. The semantic fingerprint vector is used to characterize the comprehensive features of the material.

[0014] In one embodiment, a graph neural network is used to encode and aggregate neighbor node information of the heterogeneous graph of the material entity. Based on the aggregation result, an embedding vector representing the node is output, and a weighted pooling process is applied to the embedding vector to obtain a semantic fingerprint vector. The semantic fingerprint vector is used to characterize the comprehensive features of the material, including: Initialize the feature vectors of each type of node in the heterogeneous graph of the material entity, and design specific transformation matrices for different types of nodes based on the initialization results to map the node features to a unified space for graph convolution operation; The graph convolution operation results are aggregated with the features of neighboring nodes according to edge type. After weighted summation, the node's own features are fused together with the node's features and then updated by nonlinear activation. The neighboring node information is aggregated through three layers of graph convolution operation, and the output is the embedding vector representing the node. The neighboring node information represents the features of the nodes that are directly or indirectly connected to the target node. Weighted pooling is performed on the embedding vectors of all nodes in the heterogeneous graph of the material entity to obtain graph-level semantic fingerprint vectors. The semantic fingerprint vectors are then used to analyze the distance of the material in the fingerprint space to obtain the comprehensive characteristics of the material.

[0015] In one embodiment, weighted pooling is performed on the embedding vectors of all nodes in the heterogeneous graph of the material entity to obtain a graph-level semantic fingerprint vector, including: The embedding vectors of all nodes in the heterogeneous graph of the material entity are used as input, and the matching degree between each embedding vector and the learnable context vector is calculated to obtain the original attention score of the node. Normalize the original attention scores of all nodes in the heterogeneous graph of the material entity to obtain the normalized attention weight coefficients for each node. Then, perform element-wise weighted multiplication of the normalized attention weight coefficients of any node with the embedding vector to obtain the weighted node embedding vector for any node. For the weighted node embedding vectors of all nodes in the heterogeneous graph of the material entity, perform element-wise summation to obtain a graph-level semantic fingerprint vector that can represent the topological structure and global semantic information of the entire graph.

[0016] In one embodiment, the cosine similarity of semantic fingerprint vectors is calculated based on a density clustering algorithm, and the cosine similarity is compared with an association threshold to determine the implicit associations between the accessed material data. The output association record includes: A high-dimensional mapping process is performed on the semantic fingerprint vector to obtain a set of material feature vectors in a high-dimensional feature space. The directional consistency of the set of material feature vectors is judged to capture the composition ratio and feature change trend between different materials. The cosine similarity of the semantic fingerprint vectors corresponding to any material is calculated. The cosine similarity is compared with the association threshold. When the cosine similarity is greater than or equal to the association threshold, the two materials are determined to be strongly associated. When the cosine similarity is less than the association threshold, any two materials are determined to be weakly associated, the preliminary association result is output, and a material association graph is constructed for the weakly associated materials with material records as nodes and preliminary association results as edges. The embedding representation of each node in the material association graph is output using an inductive graph neural network. The potential association edges between nodes are predicted by aggregating the feature information of neighboring nodes. The weak association strengthening result is determined after weak association enhancement. The weak association strengthening result is combined with the strong association result to output the association relationship record.

[0017] In one embodiment, a high-dimensional mapping process is performed on the semantic fingerprint vector to obtain a set of material feature vectors in a high-dimensional feature space. The directional consistency of the set of material feature vectors is then determined to capture the composition ratio and feature variation trends among different materials. The cosine similarity of the semantic fingerprint vectors corresponding to any material is calculated, including: Discretize the material semantic fingerprint vector, map the material semantic fingerprint vector to a high-dimensional space, and combine the mapping result with the density threshold to perform cluster analysis, and construct clusters based on the analysis results; The process of constructing clusters is repeated until the cluster division of material semantic fingerprint vectors is completed. The consistency of the directional distribution of material feature vectors in high-dimensional space is quantified by calculating the cosine of the angle between each material feature vector in each cluster. Based on the results of directional distribution consistency, we can capture the variation patterns of component ratios and characteristic indexes among different materials within clusters. The dimensional weights of the component ratio variation pattern and the feature index variation trend are assigned, and the cosine similarity of the semantic fingerprint vector corresponding to any material is calculated based on the assignment results using the cosine similarity calculation formula.

[0018] In one embodiment, the material semantic fingerprint vector is discretized, mapped to a high-dimensional space, and the mapping result is combined with a density threshold for cluster analysis. Based on the analysis results, clusters are constructed, including: Each feature dimension of the material semantic fingerprint vector is discretized and encoded to construct an initial dimension vector of the semantic fingerprint. Then, a high-dimensional mapping process is performed on the initial dimension vector of the semantic fingerprint through a linear transformation matrix to obtain a set of material feature vectors with unified dimensions. Material feature vectors with missing feature dimensions and abnormal vector magnitudes are filtered out from the material feature vector set to form an effective feature vector dataset in a high-dimensional feature space. At the same time, the neighborhood radius and density threshold of the high-dimensional feature space are set. Traverse the effective feature vector dataset to find the core vector that satisfies the neighborhood radius and density threshold, and extract all feature vectors in the high-dimensional feature space where the core vector is located. Filter all density-connected feature vectors starting from the core vector until there are no new expandable feature vectors, thus completing the construction of a single cluster.

[0019] In one embodiment, the hash value of the association record is calculated, a Merkle tree is constructed based on the hash value, the association record is stored on the blockchain network, and the material data is uploaded to an off-chain storage medium to achieve trusted storage of the association and material data, including: The hash value of each related record is calculated using hashing technology. The hash values ​​generated within the target time window are used as leaf nodes. Leaf nodes are paired up to obtain parent nodes, and the process continues until all leaf nodes are paired up to obtain the root node. The timestamps of the root node and the target window are merged and packaged into a blockchain transaction and submitted to the blockchain network. The blockchain network nodes verify the blockchain transaction through the consensus mechanism and write it into a block after verification, so as to realize the on-chain storage of the relationship record. Material data and material entity heterogeneous diagrams are stored as raw data in a distributed file storage, and hash addresses are generated for the distributed files. At the same time, a field pointing to the storage address of the raw data is added to the relationship record to achieve reliable and associated storage of the relationship and material data.

[0020] In one embodiment, verifying blockchain transactions using a consensus mechanism via blockchain network nodes includes: Based on blockchain transactions, a set of storage data feature descriptors is generated. Consensus decision-making operations are performed in combination with the set of storage data feature descriptors according to the embedded consensus verification space and dynamic confidence constraints. The initial verification decision scheme is output based on the operation results. The stored data feature descriptor set is segmented into time windows. By predicting the feature evolution path within the future time window, a predicted feature descriptor set is generated. Based on the predicted feature descriptor set, the initial verification decision scheme is calculated to compensate for deviations and generate an optimized verification decision scheme. Based on the optimized verification decision scheme, an execution instruction is sent to the blockchain network nodes, triggering the distributed nodes to perform blockchain transaction verification operations.

[0021] According to a second aspect of the present invention, a blockchain-based data aggregation system for large-scale materials data is provided.

[0022] In one embodiment, the blockchain-based aggregation system for large-scale materials data includes: The data access module is used to access material data based on data templates and information extraction rules, and to perform format unification and entity alignment processing on the material data, and to obtain material description text based on the processing results; The semantic analysis module is used to construct a material entity heterogeneous graph using material description text, and to encode and weight pool the material entity heterogeneous graph based on graph neural network to obtain semantic fingerprint vectors to characterize the comprehensive features of the material. The association calculation module is used to calculate the cosine similarity of semantic fingerprint vectors based on density clustering algorithm, compare the cosine similarity with the association threshold to determine the implicit association between the accessed material data, and output the association relationship record; The blockchain evidence storage module is used to calculate the hash value of the relationship record, construct a Merkle tree based on the hash value, store the relationship record to the blockchain network, and upload the material data to the off-chain storage medium to achieve reliable storage of the relationship and material data.

[0023] According to a third aspect of the present invention, a computer device is provided.

[0024] In some embodiments, the computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method described above.

[0025] According to a fourth aspect of the present invention, a computer-readable storage medium is provided.

[0026] In one embodiment, a computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the steps of the above method.

[0027] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: 1. This invention solves the problem of difficult association by introducing semantic fingerprints to automatically discover and establish credible associations between data. At the same time, it solves the problems of difficult trust and difficult storage by storing the proof of association (rather than the original data) on the blockchain. Ultimately, it achieves secure, efficient, and reliable aggregation and intelligent association of material data.

[0028] 2. This invention extracts four types of entities—composition, process, structure, and performance—through a material entity recognition model, constructs a material entity heterogeneous graph, and uses a graph neural network (GCN) to encode the heterogeneous graph, generating a semantic fingerprint vector that can quantitatively represent the unique position of the material in a multidimensional space. This achieves a mathematical expression of the deep semantic features of the material, providing a foundation for subsequent clustering and association, which is different from traditional association methods based on keyword matching or structured labels.

[0029] 3. This invention performs density clustering on semantic fingerprints in high-dimensional space, automatically identifies strong and weak correlations between data, and uses graph neural networks for correlation verification and enhanced prediction for weakly correlated data. By comparing and learning loss functions, the correlation accuracy is optimized, solving the problem that traditional single clustering algorithms are difficult to discover hidden correlations, and realizing the automatic mining of complex relationships between large-scale material data.

[0030] 4. The present invention separates the calculation results of the association relationship from the blockchain evidence storage. It performs hash calculation on the association record and constructs a Merkle tree. Only the Merkle root is written to the blockchain, while the original data is stored in a distributed file system. Only the IPFS address pointer is stored in the association record. This can both use the blockchain to ensure the immutability and traceability of the association relationship and avoid the storage explosion caused by the massive amount of original data on the chain, thus achieving a balance between data sovereignty protection and trusted sharing.

[0031] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0032] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0033] Figure 1 This is a flowchart illustrating a blockchain-based method for converging and linking large-scale material data, according to an exemplary embodiment. Figure 2 This is a schematic diagram illustrating the principle of a blockchain-based aggregation system for large-scale material data, according to an exemplary embodiment. Figure 3 This is a schematic diagram of the structure of a computer device according to an exemplary embodiment; Figure 4 This is a flowchart illustrating a blockchain-based method for converging and aggregating large-scale material data, according to an exemplary embodiment. Detailed Implementation

[0034] The following description and accompanying drawings fully illustrate specific embodiments described herein to enable those skilled in the art to practice them. Some portions and features of certain embodiments may be included in or replace portions and features of other embodiments. The scope of the embodiments herein includes the entire scope of the claims and all available equivalents thereof. The various embodiments described herein are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments; similar or identical parts between embodiments can be referred to interchangeably.

[0035] The modules in the apparatus or system of this application can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0036] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0037] Figure 1 An embodiment of a blockchain-based method for converging large-scale material data according to the present invention is shown.

[0038] In this optional embodiment, the blockchain-based aggregation method for large-scale material data includes: Step S101: Access material data based on data templates and information extraction rules, perform format unification and entity alignment processing on the material data, and obtain material description text based on the processing results; Step S102: Construct a material entity heterogeneous graph using the material description text, and encode and weighted pool the material entity heterogeneous graph based on a graph neural network to obtain a semantic fingerprint vector to characterize the comprehensive features of the material. Step S103: Calculate the cosine similarity of semantic fingerprint vectors based on density clustering algorithm, compare the cosine similarity with the association threshold to determine the implicit association between the accessed material data, and output the association record; Step S104: Calculate the hash value of the association record, construct a Merkle tree based on the hash value, store the association record in the blockchain network, and upload the material data to the off-chain storage medium to achieve reliable storage of the association and material data.

[0039] In this optional embodiment, a material entity heterogeneous graph is constructed using material description text, and the graph is encoded and weighted pooled based on a graph neural network to obtain a semantic fingerprint vector to characterize the comprehensive features of the material. This includes: labeling the sequence tags corresponding to the material description text and inputting the sequence tags into the entity recognition model; adding a classification layer at the top of the entity recognition model to learn semantic features in the material domain and outputting the identified entities; mapping the identified entities to standard concepts in the materials science ontology library and outputting the semantic code corresponding to each standard concept; constructing a material entity heterogeneous graph containing nodes and edges based on the semantic code and the material description text; nodes include composition nodes, process nodes, structural nodes, and performance nodes, and edges represent co-occurrence relationships or causal relationships between identified entities; encoding and aggregating neighbor node information in the material entity heterogeneous graph using a graph neural network; outputting embedding vectors representing nodes based on the aggregation results; and performing weighted pooling on the embedding vectors to obtain a semantic fingerprint vector, which is then used to characterize the comprehensive features of the material.

[0040] In this optional embodiment, a graph neural network is used to encode and aggregate neighbor node information in the heterogeneous graph of the material entity. Based on the aggregation result, an embedding vector representing the node is output, and a weighted pooling process is applied to the embedding vector to obtain a semantic fingerprint vector. The semantic fingerprint vector is used to characterize the comprehensive features of the material, including: initializing the feature vectors of each type of node in the heterogeneous graph of the material entity; designing specific transformation matrices for different types of nodes based on the initialization result to map the node features to a unified space for graph convolution operation; aggregating neighbor node features according to edge type based on the graph convolution operation result; after weighted summation, fusing the node's own features and updating the node through nonlinear activation, so as to aggregate neighbor node information through a three-layer graph convolution operation and output the embedding vector representing the node; the neighbor node information represents the features of nodes directly or indirectly connected to the target node; weighted pooling is applied to the embedding vectors of all nodes in the heterogeneous graph of the material entity to obtain a graph-level semantic fingerprint vector; the semantic fingerprint vector is used to analyze the distance of the material in the fingerprint space to obtain the comprehensive features of the material.

[0041] In this optional embodiment, weighted pooling of the embedding vectors of all nodes in the heterogeneous graph of the material entity to obtain a graph-level semantic fingerprint vector includes: taking the embedding vectors of all nodes in the heterogeneous graph of the material entity as input and calculating the matching degree between each embedding vector and the learnable context vector to obtain the original attention score corresponding to the node; performing normalization processing on the original attention scores of all nodes in the heterogeneous graph of the material entity to obtain the normalized attention weight coefficient corresponding to each node, and performing element-wise weighted multiplication of the normalized attention weight coefficient of any node with the embedding vector to obtain the weighted node embedding vector corresponding to any node; and performing element-wise summation processing on the weighted node embedding vectors of all nodes in the heterogeneous graph of the material entity to obtain a graph-level semantic fingerprint vector that can represent the topological structure and global semantic information of the entire graph.

[0042] It should be explained that the semantic fingerprint vector does not use graph neural networks (GCNs) for automatic embedding learning. Instead, it can be obtained by constructing an ontology library in the field of materials science, defining semantic distance rules (such as Euclidean distance and Mahalanobis distance) between components, processes, structures, and properties, and directly calculating the similarity between material entities as fingerprints; or by directly inputting the material description text into a pre-trained Transformer model (such as SciBERT or MatBERT), and taking the labeled output vector as the semantic fingerprint of the material, replacing the graph neural network encoding process; for material data containing microstructure images (such as SEM photos), convolutional neural networks (CNNs) are introduced to extract image features, which are then concatenated with text features to form a multimodal semantic fingerprint, suitable for material databases with rich image data, such as spectral libraries and metallographic image libraries.

[0043] In this optional embodiment, the cosine similarity of semantic fingerprint vectors is calculated based on a density clustering algorithm, and the cosine similarity is compared with an association threshold to determine the implicit association between the accessed material data. The output association record includes: performing high-dimensional mapping processing on the semantic fingerprint vectors to obtain a set of material feature vectors in a high-dimensional feature space, and judging the directional consistency of the set of material feature vectors to capture the component ratio and feature change trend between different materials; calculating the cosine similarity of the semantic fingerprint vectors corresponding to any material; comparing the difference between the cosine similarity and the association threshold; when the cosine similarity is greater than or equal to the association threshold, it is determined that any two materials are strongly associated; when the cosine similarity is less than the association threshold, it is determined that any two materials are weakly associated, and the preliminary association result is output. For the weakly associated materials, a material association graph is constructed with material records as nodes and preliminary association results as edges; the embedding representation results of each node in the material association graph are output using an inductive graph neural network, and the potential association edges between nodes are predicted by aggregating the feature information of neighboring nodes to determine the weak association strengthening result after weak association enhancement; and the weak association strengthening result is combined with the strong association result to output the association record.

[0044] In the process of realizing association aggregation computation, hierarchical clustering algorithm (AGNES) can also be used to merge similar material data from bottom to top to form a tree-like association structure. By setting different similarity thresholds, the association relationship between materials can be observed at different granularities, so that a hierarchical association view can be generated without presetting the neighborhood radius parameter. Alternatively, association enhancement based on knowledge graph reasoning can be used instead of graph neural network GraphSAGE. Instead, knowledge graph reasoning methods such as TransE and TransH based on translation models, or PRA algorithm based on path ranking, can be used to infer implicit associations by mining existing entity relationship paths in the knowledge graph. Cross-domain association computation based on federated learning can also be applied. When the data holder is unwilling to share the original data, a federated learning framework can be used. Each participant calculates the semantic fingerprint locally, and only the encrypted fingerprint gradient is uploaded to the central server for aggregation and clustering, realizing cross-domain association computation where the data does not move but the model moves.

[0045] In this optional embodiment, high-dimensional mapping processing is performed on the semantic fingerprint vectors to obtain a set of material feature vectors in a high-dimensional feature space. The directional consistency of the set of material feature vectors is then judged to capture the composition ratio and feature variation trend among different materials. The cosine similarity of the semantic fingerprint vectors corresponding to any material is calculated as follows: the material semantic fingerprint vectors are discretized and encoded; the material semantic fingerprint vectors are mapped to a high-dimensional space; the mapping result is combined with a density threshold for cluster analysis; and clusters are constructed based on the analysis results. The cluster construction process is repeated until the cluster division of the material semantic fingerprint vectors is completed. After all cluster divisions are completed, for each material feature vector within each cluster, the cosine value of the angle between the vectors is calculated to quantify the directional distribution consistency of the material feature vectors in the high-dimensional space. Thus, the similarity of material features within the same cluster can be measured from the perspective of vector direction, realizing the digital quantification of directional consistency and reflecting the material's... The overall direction and distribution pattern of feature changes are analyzed. Based on the consistency results of the directional distribution obtained by quantification, the variation patterns of component ratios and feature indexes among different materials within each cluster are mined and captured. This allows for the extraction of core variation features of materials from the clustering results and directional consistency data, transforming abstract vector data into interpretable material composition and trend information. This intuitively captures the implicit component ratio correlations and feature change logics between materials, providing a core basis for similarity calculation that aligns with the actual properties of the materials. Based on the extracted component ratio variation patterns and feature index change trends, differentiated dimension weights are assigned to different feature dimensions of the semantic fingerprint vector. Component dimensions and trend dimensions with high material feature discrimination and strong change sensitivity are given high weights, while those with low weights are given low weights. Then, based on the vectors after weight assignment, the cosine similarity between the semantic fingerprint vectors of any two materials is calculated using the standard cosine similarity calculation formula.

[0046] In this optional embodiment, the material semantic fingerprint vector is discretized and encoded, mapped to a high-dimensional space, and the mapping result is combined with a density threshold for cluster analysis. Based on the analysis results, clusters are constructed by: performing discretization encoding on the material semantic fingerprint vector; independently discretizing and encoding each feature dimension of the material semantic fingerprint vector; converting continuous or categorical features such as material composition types, component proportions, physicochemical indicators, and feature change gradients into standardized coded values ​​to construct an initial dimension vector of the semantic fingerprint; and performing high-dimensional mapping processing on the constructed initial dimension vector of the semantic fingerprint using a preset linear transformation matrix, mapping the low-dimensional, low-information-density initial vector to a preset high-dimensional feature space, ultimately obtaining a set of material feature vectors with completely unified dimensions. This allows for the effective separation of material features that are difficult to distinguish in low-dimensional space in high-dimensional space, fully amplifying subtle feature differences between materials; and then... The material feature vector set undergoes data filtering to remove invalid material feature vectors with missing feature dimensions or abnormal vector magnitudes, thus forming a clean and effective feature vector dataset within a high-dimensional feature space. Simultaneously, based on the distribution characteristics of material features, a neighborhood radius and density threshold are set for the high-dimensional feature space. The processed effective feature vector dataset is then traversed, searching for core vectors that satisfy the set neighborhood radius and density threshold conditions. Once a core vector is found, all feature vectors within the high-dimensional feature space containing that core vector are extracted. Further filtering is performed to identify all density-connected feature vectors originating from that core vector. This process continuously expands the clustering range until no new expandable feature vectors are available, completing the construction of a single cluster. This clustering process is repeated, traversing all effective feature vectors, until all material semantic fingerprint vectors are clustered. Based on density clustering rules, material vectors with similar features can be grouped into the same cluster, achieving automatic classification and aggregation of material features.

[0047] In this optional embodiment, the hash value of the association record is calculated, and a Merkle tree is constructed based on the hash value. The association record is stored on the blockchain network, and the material data is uploaded to the off-chain storage medium to achieve trusted storage of the association and material data. This includes: calculating the hash value of each association record using hash technology, using the hash value generated within the target time window as the leaf node, pairing leaf nodes to obtain the parent node, and so on until all leaf nodes are paired to obtain the root node; merging the root node with the timestamp of the target window into a blockchain transaction and submitting it to the blockchain network; using blockchain network nodes to verify the blockchain transaction through a consensus mechanism, and writing it into a block after verification to achieve on-chain storage of the association record; storing the material data and material entity heterogeneous graph as raw data in a distributed file storage, generating a hash address for the distributed file, and adding a field to the association record pointing to the storage address of the raw data to achieve trusted and associated storage of the association and material data.

[0048] It's important to clarify that, in implementing blockchain-based evidence storage, privacy-preserving evidence storage based on zero-knowledge proofs can also be applied. This avoids directly uploading the hash value of the association to the blockchain; instead, it generates a zero-knowledge proof of the association. Verifiers can verify the existence and compliance with preset rules of the association without knowing the specific content of the association. Alternatively, a DAG-based evidence storage architecture can be used. This architecture doesn't employ the traditional block + chain structure of blockchains but uses a Directed Acyclic Graph (DAG) distributed ledger (such as IOTA). Each association is directly attached to the DAG graph as an independent transaction, eliminating the need to build a Merkle tree or wait for block confirmation. This results in fast transaction confirmation, suitable for real-time evidence storage of high-frequency association data. However, the consensus mechanism is more complex and less mature than traditional blockchains. Furthermore, off-chain evidence storage based on a Trusted Execution Environment (TEE) can be applied. Instead of writing the association to the blockchain, the association computation is performed in a TEE (such as Intel SGX), and the TEE generates a hardware-signed evidence proof. External verifiers can remotely verify the correctness of the association computation process, ensuring the confidentiality and integrity of the computation.

[0049] In this optional embodiment, verifying blockchain transactions using a consensus mechanism via blockchain network nodes includes: generating a set of corresponding stored data feature descriptors based on the blockchain transactions to be processed. This set transforms key features such as the transaction's subject information, transaction content, data format, behavioral attributes, and node signatures into standardized, computable feature description data. Its function is to uniformly convert unstructured, multi-source, and heterogeneous blockchain transaction information into feature carriers that can be directly computed by the consensus mechanism, eliminating verification barriers caused by differences in transaction data formats, improving the compatibility and processing efficiency of consensus operations, and, based on the pre-set consensus verification space and dynamic confidence constraints embedded in the blockchain system, processing the aforementioned stored data feature descriptions... The consensus decision-making process involves substituting a set of symbols into a consensus verification space. This space defines the scope, dimensions, and boundaries of the verification rules. Dynamic confidence constraints adjust the verification strictness and decision threshold in real time based on node historical credibility, transaction risk level, and network status. The resulting calculation outputs an initial verification decision scheme. This scheme, based on preset rules and boundary conditions, allows for the preliminary determination of the legality and validity of transactions in the first round, forming a basic verification conclusion. This ensures that consensus decisions operate within a controllable and standardized framework, avoiding efficiency losses and decision biases caused by unconstrained computation. Furthermore, dynamic confidence constraints adapt to different network environments and transaction risk levels, enhancing the rationality of the initial decision. The generated stored data feature descriptors are then processed. The dataset is segmented into time windows, dividing the complete feature data into multiple feature subsets across consecutive time periods. Based on historical feature change patterns, network state trends, and transaction behavior patterns, the evolution path of features within future time windows is predicted, generating a set of predicted feature descriptors. This set reflects potential changes and shifts in transaction features in subsequent periods, supplementing dynamic information that static features cannot cover from a temporal evolution perspective. This overcomes the limitations of traditional consensus verification, which relies solely on current-moment data, allowing for early detection of potential abnormal fluctuations and risks in transaction features. Furthermore, the set of predicted feature descriptors is used to calculate bias compensation for the initial verification decision, comparing it with the initial decision based on... The difference between current features and predicted features is analyzed to correct the judgment bias, omission, or misjudgment caused by ignoring temporal changes and feature drift in the initial decision-making process. This generates a more accurate and stable optimized verification decision scheme. Furthermore, by calibrating and optimizing the initial verification conclusions at the temporal level, decision errors caused by instantaneous data and local anomalies are eliminated. Based on the obtained optimized verification decision scheme, a unified execution instruction is sent to each distributed node in the blockchain network, triggering all nodes to synchronously execute blockchain transaction verification operations according to the optimized decision rules and judgment standards. This allows the final decision to be implemented as a unified execution action of distributed nodes, ensuring that the verification logic and results of all nodes in the network are consistent, and completing the closed loop of the entire transaction verification process.

[0050] Figure 2An embodiment of a blockchain-based data aggregation system for large-scale materials data according to the present invention is shown.

[0051] In this optional embodiment, the blockchain-based aggregation system for large-scale materials data includes: The data access module 201 is used to access material data based on data templates and information extraction rules, and to perform format unification and entity alignment processing on the material data, and to obtain material description text based on the processing results. The semantic analysis module 202 is used to construct a material entity heterogeneous graph using the material description text, and to encode and weight pool the material entity heterogeneous graph based on a graph neural network to obtain a semantic fingerprint vector to characterize the comprehensive features of the material. The association calculation module 203 is used to calculate the cosine similarity of semantic fingerprint vectors based on density clustering algorithm, compare the cosine similarity with the association threshold to determine the implicit association between the accessed material data, and output the association relationship record. The blockchain evidence storage module 204 is used to calculate the hash value of the association record, construct a Merkle tree based on the hash value, store the association record to the blockchain network, and upload the material data to the off-chain storage medium to achieve reliable storage of the association and material data.

[0052] To facilitate understanding of the above technical solutions of the present invention, the following further describes the above technical solutions of the present invention from the perspectives of architecture and principle, as follows: This invention provides a blockchain-based process for the aggregation and association of large-scale material data, such as... Figure 4 As shown, its architecture consists of a data access layer, a semantic analysis layer, a correlation computing layer, a blockchain evidence storage layer, and an application service layer.

[0053] The data access layer connects multiple material data sources, including experimental databases, computational simulation platforms, and literature and patent databases. It achieves unified access to heterogeneous data through standardized API interfaces. This layer deploys data cleaning functions to perform format conversion, missing value handling, and outlier detection on the raw data. The semantic analysis layer is the core processing layer, containing material entity recognition and semantic fingerprint generation functions. The material entity recognition function is based on a materials science ontology library and uses a BERT-BiLSTM-CRF deep learning model to automatically extract four key entities from unstructured text: material composition, preparation process, microstructure, and macroscopic properties. The semantic fingerprint generation function maps the extracted entity information into high-dimensional semantic feature vectors. The association calculation layer is deployed on a high-performance computing cluster and includes a density clustering engine and a graph neural network calculation module, responsible for similarity calculation and association mining of semantic fingerprints. The blockchain evidence storage layer consists of blockchain network nodes and includes association hash calculation, Merkle tree construction, and consensus evidence storage functions, responsible for storing association proofs on the blockchain. The application service layer provides a graph visualization interface and API query interface, supporting association retrieval and source tracing analysis of material data. The specific aggregation process is as follows: Step 1: Multi-source data access and preprocessing; Standardized data adapters are used to access material data from different sources. For structured data (such as data exported from a database), it is directly mapped to a predefined data model (the interpretation of the predefined semantic model is to formulate data templates according to different material categories. For example, for steel, the composition template is defined as C, Si, Mn, S, P, Nb, V, Ti, Ni, Cu; the performance template is tensile strength, yield strength, elongation after fracture, impact temperature, and impact energy; for high-temperature alloys, the composition template is defined as C, Cr, Ni, Co, W, Mo, Al, Ti, Fe, Nb; the performance template is tensile strength, yield strength, elongation after fracture, high-temperature test temperature, creep time, creep elongation, etc.). For semi-structured data (such as experimental report PDFs), a rule-based information extraction method is used. Structured data is read directly from the database table; semi-structured data is read from the PDF according to the data template rules of the semantic model. For unstructured text (such as scientific research papers), a material entity recognition model is called for extraction.

[0054] Specific data preprocessing includes: format standardization, converting all data into JSON-LD format; entity alignment, mapping synonymous entities from different sources (such as heat treatment temperature and annealing temperature) to standard ontologies; and quality filtering, removing records with entity integrity below 60%.

[0055] Step 2: Semantic fingerprint generation; The core innovation of step two lies in constructing a semantic fingerprint of material data, enabling a unique quantitative representation of materials in a four-dimensional space of composition, process, structure, and performance. Material entity recognition employs a pre-trained BERT model as the basic encoder, fine-tuned on a materials science corpus. Specifically, this involves collecting materials literature and labeling entities with BIO tags for composition, process, structure, and performance. The pre-trained BERT model and word segmenter are then loaded, and a classification layer is added on top of the BERT model, with the output dimension equal to the number of entity tags. The model is trained using labeled data, and its parameters are updated through backpropagation, allowing it to learn the semantic features of the materials domain and ultimately accurately identify entities.

[0056] The model input is material description text, and the output is BIO tags for each word (B indicates the start of an entity, I indicates the interior of an entity, and O indicates a non-entity). The identified entity categories include: composition entities (e.g., Fe-0.3C), process entities (e.g., rolled at 1200℃), structural entities (e.g., martensite), and performance entities (e.g., tensile strength 850MPa).

[0057] Semantic fingerprint generation: Drawing inspiration from the molecular fingerprint construction approach in the SubGrapher method, this invention constructs a material semantic fingerprint vector. F ∈R n , n The generation process is as follows, based on the preset material feature dimensions: The identified entities are mapped to standard concepts in a materials science ontology library. Each concept corresponds to a predefined semantic code. For example, the composition Fe-0.3C is mapped to the concept node Fe-C alloy and 0.3wt% carbon content. A heterogeneous graph of material entities is then constructed. G = ( V , E ), where nodes V Including composition nodes, process nodes, structural nodes, performance nodes, and edges. E It represents the co-occurrence or causal relationship between entities, and the edge weight is calculated based on the co-occurrence frequency of the entities in the text.

[0058] Simultaneously, a graph neural network (GCN) is used to encode the heterogeneous graph. Specifically, feature vectors are initialized for each type of node in the heterogeneous graph. At the same time, type-specific transformation matrices are designed for different types of nodes to map the node features to a unified space and perform graph convolution operations: in each layer, the features of the node are aggregated from the features of its neighbors (considering edge types), weighted and summed, and then combined with its own features. The node representation is then updated by nonlinear activation. After repeating multiple layers, the node embedding with fused structural information is obtained. The neighbor node information is aggregated through three layers of graph convolution operations to generate the node embedding representation.

[0059] The neighbor node information refers to the features of nodes that are directly or indirectly connected to the target node. The three convolutional operations are as follows: the first layer aggregates the features of one-hop neighbors and updates the node representation; the second layer aggregates the two-hop neighbor information based on the updated representation, so that the node contains the structural information within the two hops; the third layer aggregates the three-hop neighbor information. Finally, the node embedding integrates the structural and feature information of the three-order neighborhood. Each layer includes neighbor feature aggregation, such as summation, combination of its own features, and nonlinear transformation.

[0060] The embedding vectors of all nodes in the entire graph are weighted and pooled. Specifically, the attention score between each node embedding and the learnable context vector (or the query vector) is first calculated, then normalized to weights using softmax, and each node embedding is multiplied by its weight. Finally, all weighted embeddings are summed to obtain the graph-level semantic fingerprint vector. F This method enables the model to focus on important nodes, quantifies the comprehensive characteristics of the material, and obtains the final semantic fingerprint vector. F This fingerprint can quantitatively characterize the comprehensive features of materials: materials with similar compositions are closer together in the fingerprint space, while similar materials with different processes produce distinguishable vector offsets in the process dimension.

[0061] Step 3: Density-based correlation aggregation calculation; Step 3 is executed on an off-chain high-performance computing cluster to perform large-scale cluster analysis on the semantic fingerprints, automatically discovering implicit relationships between material data. Density clustering relationships are used to analyze the semantic fingerprints of all material data. F i Mapping to a high-dimensional space, an improved DBSCAN density clustering algorithm is used for clustering. However, conventional DBSCAN typically uses Euclidean distance, which measures the straight-line length between two points in a coordinate system. In materials science, two materials with significantly different compositions may be far apart in high-dimensional space, but this distance is only mathematical and does not directly equate to a strong relationship. Therefore, this invention uses cosine similarity, which measures the consistency of direction rather than the distance. Physically, cosine similarity ignores the magnitude of the fingerprint vector (which may represent some absolute strength or order of magnitude) and focuses on the similarity of component proportions or feature patterns. In materials science, this means that even if a certain performance indicator of material A is an order of magnitude higher than that of material B, as long as the trends or component proportions are similar, cosine similarity can still capture this similar relationship. The core parameters of the algorithm include the neighborhood radius. ε and the minimum number of samples, MinPts. For any two material records... i and j Calculate the cosine similarity Sim( of its fingerprint vector). i , j )= F i · Fj / (|| F i ||×|| F j ||); when Sim( i , j )≥threshold T When 1 (in this embodiment, 0.85) is taken, it is determined to be a strong correlation, representing multiple tests on the same sample or repeated experiments on the same material; when T 2≤Sim( i , j )< T 1 ( T When the correlation coefficient is 0.65, it is considered a weak correlation, representing a performance comparison of different literature reports on similar materials or materials with similar compositions.

[0062] Graph Neural Network Association Enhancement: For weakly correlated data, a graph neural network is further used to verify the associations and construct a material association map. G m =( V m , E m The nodes represent material records, and the edges represent preliminary association results. The GraphSAGE model is used to learn the embedding representations of the nodes. By aggregating the feature information of neighboring nodes, potential association edges between nodes are predicted (constructing a material association graph, where nodes represent materials and edges represent preliminary associations. Neighbors are sampled, features are aggregated to update node embeddings, and based on the learned embeddings, the probability of potential association edges between nodes is predicted). The model training uses a contrastive learning loss function, ensuring that the embedding distance of positive sample pairs (known associations) is less than the distance of negative sample pairs (no associations). The association calculation output is the association relationship record. R ={ r id , source id , target id , relation type , similarity , timestamp},in relation type The value can be either a strong association or a weak association.

[0063] Step 4: Reliable evidence of the relationship; The core innovation of step four lies in putting only the proof of the relationship on the blockchain, while the original data is still stored off-chain. This ensures the immutability of the relationship and avoids the storage explosion caused by putting massive amounts of original data on the blockchain.

[0064] Relationship hash calculation: For each relationship record RCalculate its SHA-256 hash value H ( R )=SHA-256( r id || source id || target id || relation type || similarity || timestamp ), where || represents string concatenation, H ( R ) is the association hash value, and SHA-256 is a hash algorithm. r id It is a unique identifier for the related records. source id It is the ID of the data source. target id It is the target ID. relation type Association type (e.g., strong association / weak association) similarity It is a similarity score. timestamp It's a timestamp.

[0065] Merkle Tree Construction: Using all the hash values ​​of relationships generated within a time window (set to 1 hour in this example) as leaf nodes, a binary Merkle tree is constructed. The tree construction process is as follows: pair leaf nodes together, calculate their hash values, concatenate them, and then hash again to obtain the parent node; this process is recursively repeated until a unique root node, Merkle, is obtained. Root This process draws on the Merkle tree construction method used for blockchain transaction notarization.

[0066] On-chain evidence storage: Merkle Root The transaction is packaged with the timestamp of the specified time window into a single blockchain transaction and submitted to the blockchain network. Blockchain nodes verify the transaction through a consensus mechanism (this embodiment uses the Practical Byzantine Fault-Tolerant (PBFT) algorithm). Once verified, the transaction is written into a block, and the transaction hash Tx is recorded. Hash It is returned to the system as evidence.

[0067] Off-chain raw data storage: Raw material data (including text, graphs, tables, etc.) is stored in the distributed file system IPFS. Each file generates a unique IPFS hash address, and an ipfs field is added to the association record R. hash The address points to the storage address of the original data. When querying, the IPFS address can be obtained through the proof of association on the blockchain, and then the original data can be retrieved from IPFS.

[0068] Step 5: Dynamic updating of the graph and consensus verification; When new data is added to the system, an incremental update process is triggered. Steps one through three are executed on the new data to generate its semantic fingerprint and perform correlation calculations with existing data. For newly discovered correlations, step four is executed to store the data on the blockchain. At the same time, the smart contract performs consistency verification on the newly generated correlations: if the new correlation conflicts with historical correlations (such as the same material record being contained in two different clusters), a manual review process is triggered; if the verification passes, the new correlation is written to the graph database to update the material knowledge graph.

[0069] In the process of implementing graph storage and writing, relational databases (PostgreSQL and pgRouting) or NoSQL databases (MongoDB) can be used to store the relationships. Graph queries can be implemented through adjacency lists or document nesting, thus utilizing existing database infrastructure without introducing new components; however, the performance of complex graph traversal queries may degrade. Semantic storage based on RDF triples can be used, which uses the resource description framework RDF to represent relationships as (subject, predicate, object) triples and stores them in a triple library (such as Apache Jena, Virtuoso). It supports SPARQL semantic queries, making it compliant with the Semantic Web standard and facilitating interconnection with other open databases.

[0070] Taking a steel material database as an example, the implementation process of this invention is described. It is assumed that the database contains 5,000 experimental records of steel materials. Each record includes composition (C%, Mn%, Si%, etc.), process (hot rolling temperature, cooling rate), microstructure (grain size, phase composition), and mechanical properties (yield strength, tensile strength, elongation).

[0071] Example Data Access: 5000 records are imported in batches from a CSV file. The data access module automatically identifies field mappings and completes data cleaning and format conversion.

[0072] Semantic fingerprint generation: The semantic analysis module generates a 2048-dimensional fingerprint vector for each record. For example, the cosine similarity between the fingerprint vectors of record A (composition: Fe-0.2C-1.5Mn, process: hot rolling at 1200℃, microstructure: ferrite and pearlite, property: yield strength 450MPa) and record B (composition: Fe-0.18C-1.4Mn, process: hot rolling at 1180℃, microstructure: ferrite and pearlite, property: yield strength 440MPa) is 0.89.

[0073] Association calculation: The density clustering algorithm clustered 5000 records into 127 clusters. Records A and B were clustered into the same cluster, which was determined to be a strong association (similarity 0.89 > 0.85). At the same time, it was found that the similarity between record C (composition: Fe-0.3C-1.2Mn, process: annealed at 900℃, performance: yield strength 380MPa) and record A was 0.72, which was determined to be a weak association, representing the influence of different processes on the performance of similar materials.

[0074] Related evidence storage: Construct a Merkle tree from the 3562 related relationship records generated within 1 hour, with the root hash value being 0x7a3f...8e9d, write it into the blockchain transaction 0x4b2c...f1a5, store the original data in IPFS, and the IPFS address of record A is QmX...yZ.

[0075] Atlas application: When querying high-strength steel materials, the atlas returns similar materials under different process conditions that are weakly correlated with it, providing reference clues for process optimization. At the same time, the generation time and immutability of the correlation can be verified through blockchain notarization.

[0076] In summary, by utilizing the above-described technical solution of this invention, compared to existing material network construction methods that leap from manual feature definition to automatic semantic fingerprint representation, this invention addresses the challenge that existing technologies primarily rely on manually defined structured features (such as compositional similarity) to capture the deep semantic connections between composition, process, structure, and performance. For example, materials with similar compositions but different processes may be considered strongly correlated in existing technologies, even though their actual performance differences are significant. This invention, by introducing semantic fingerprint technology and utilizing graph neural networks to automatically learn the embedding representation of materials in a multi-dimensional feature space, achieves a quantitative representation of the deep semantics of materials science, resulting in association results that better align with the inherent laws of materials science.

[0077] Furthermore, existing technologies use a single network topology analysis method to mine associations, which has limited ability to identify weak associations or cross-domain associations. This invention adopts a dual mechanism of density clustering and graph neural networks. Density clustering is used to quickly discover strongly associated clusters, and graph neural networks are used to enhance the verification of weak cross-cluster associations. This hierarchical association strategy not only ensures the processing efficiency of large-scale data, but also improves the ability to discover implicit associations.

[0078] Meanwhile, existing technologies store association results in centralized databases, which are susceptible to tampering, and the credibility of association relationships cannot be guaranteed. This invention adopts an off-chain computing and on-chain evidence storage architecture. Association relationships are generated in an off-chain high-performance computing cluster, but the association proof (Merkelgen) is written to the blockchain, realizing the immutability and traceability of association relationships. This allows the authenticity of association relationships to be verified through the blockchain when designing materials based on association graphs, solving the core pain point of low credibility of data association results in existing technologies. This invention supports real-time access to new data, incremental association calculation, and automatic verification of smart contracts, realizing the dynamic evolution of the material knowledge graph and better adapting to the continuous growth characteristics of material data.

[0079] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores static and dynamic information data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0080] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0081] In addition, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0082] In addition, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0083] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0084] This invention is not limited to the structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this invention is limited only by the appended claims.

Claims

1. A blockchain-linked aggregation method of large-scale material data, characterized in that, The method includes: Material data is accessed based on data templates and information extraction rules, and the material data is processed to unify the format and align entities. The material description text is obtained based on the processing results. Material entity heterogeneous graphs are constructed using material description texts, and the graph entity heterogeneous graphs are encoded and weighted pooled based on graph neural networks to obtain semantic fingerprint vectors to characterize the comprehensive features of the materials. The cosine similarity of semantic fingerprint vectors is calculated based on density clustering algorithm, and the cosine similarity is compared with the association threshold to determine the implicit association between access material data, and the association relationship record is output. The hash value of the association record is calculated, and a Merkle tree is constructed based on the hash value. The association record is then stored in the blockchain network, while the material data is uploaded to the off-chain storage medium to achieve reliable storage of the association and material data.

2. The blockchain-based aggregation method for large-scale material data according to claim 1, characterized in that, The method involves constructing a material entity heterogeneous graph using material description text, and then encoding and weighted pooling the graph using a graph neural network to obtain a semantic fingerprint vector that characterizes the comprehensive features of the material. The sequence labels corresponding to the material description text are labeled and input into the entity recognition model. A classification layer is added on top of the entity recognition model to learn the semantic features of the material domain and output the recognized entity. The identified entities are mapped to standard concepts in the materials science ontology library, and the semantic code corresponding to each standard concept is output. Based on the semantic code and the material description text, a heterogeneous graph of material entities containing nodes and edges is constructed. The nodes include composition nodes, process nodes, structural nodes and performance nodes, and the edges represent the co-occurrence relationship or causal relationship between the identified entities. The material entity heterogeneous graph is encoded and the neighbor node information is aggregated by graph neural network. Based on the aggregation result, the embedding vector representing the node is output. The embedding vector is then weighted and pooled to obtain the semantic fingerprint vector. The semantic fingerprint vector is used to characterize the comprehensive features of the material.

3. The blockchain-based aggregation method for large-scale material data according to claim 2, characterized in that, The process involves encoding and aggregating neighbor node information in a heterogeneous graph of the material entity using a graph neural network. Based on the aggregation result, an embedding vector representing the node is output. Weighted pooling is then applied to the embedding vector to obtain a semantic fingerprint vector. This semantic fingerprint vector is used to characterize the comprehensive features of the material, including: Initialize the feature vectors of each type of node in the heterogeneous graph of the material entity, and design specific transformation matrices for different types of nodes based on the initialization results to map the node features to a unified space for graph convolution operation; The neighbor node features are aggregated according to edge type based on the results of graph convolution operations. After weighted summation, the node's own features are fused together with the node's updated features through nonlinear activation. This process aggregates neighbor node information through three layers of graph convolution operations and outputs an embedding vector representing the node. The neighbor node information represents the features of nodes that are directly or indirectly connected to the target node. Weighted pooling is performed on the embedding vectors of all nodes in the heterogeneous graph of the material entity to obtain graph-level semantic fingerprint vectors. The semantic fingerprint vectors are then used to analyze the distance of the material in the fingerprint space to obtain the comprehensive characteristics of the material.

4. The blockchain-based aggregation method for large-scale material data according to claim 3, characterized in that, The weighted pooling of the embedding vectors of all nodes in the heterogeneous graph of the material entity to obtain the graph-level semantic fingerprint vector includes: The embedding vectors of all nodes in the heterogeneous graph of the material entity are used as input, and the matching degree between each embedding vector and the learnable context vector is calculated to obtain the original attention score of the node. Normalize the original attention scores of all nodes in the heterogeneous graph of the material entity to obtain the normalized attention weight coefficients for each node. Then, perform element-wise weighted multiplication of the normalized attention weight coefficients of any node with the embedding vector to obtain the weighted node embedding vector for any node. For the weighted node embedding vectors of all nodes in the heterogeneous graph of the material entity, perform element-wise summation to obtain a graph-level semantic fingerprint vector that can represent the topological structure and global semantic information of the entire graph.

5. The blockchain-based aggregation method for large-scale material data according to claim 1, characterized in that, The method calculates the cosine similarity of semantic fingerprint vectors based on density clustering algorithm, compares the cosine similarity with the association threshold to determine the implicit association between access material data, and outputs association relationship records including: A high-dimensional mapping process is performed on the semantic fingerprint vector to obtain a set of material feature vectors in a high-dimensional feature space. The directional consistency of the set of material feature vectors is judged to capture the composition ratio and feature change trend between different materials. The cosine similarity of the semantic fingerprint vectors corresponding to any material is calculated. The cosine similarity is compared with the association threshold. When the cosine similarity is greater than or equal to the association threshold, the two materials are determined to be strongly associated. When the cosine similarity is less than the association threshold, any two materials are determined to be weakly associated, the preliminary association result is output, and a material association graph is constructed for the weakly associated materials with material records as nodes and preliminary association results as edges. The embedding representation of each node in the material association graph is output using an inductive graph neural network. The potential association edges between nodes are predicted by aggregating the feature information of neighboring nodes. The weak association strengthening result is determined after weak association enhancement. The weak association strengthening result is combined with the strong association result to output the association relationship record.

6. The blockchain-based aggregation method for large-scale material data according to claim 5, characterized in that, The process of performing high-dimensional mapping on the semantic fingerprint vectors to obtain a set of material feature vectors in a high-dimensional feature space, and judging the directional consistency of the set of material feature vectors to capture the composition ratio and feature change trends between different materials, and calculating the cosine similarity of the semantic fingerprint vectors corresponding to any material includes: Discretize the material semantic fingerprint vector, map the material semantic fingerprint vector to a high-dimensional space, and combine the mapping result with the density threshold to perform cluster analysis, and construct clusters based on the analysis results; The process of constructing clusters is repeated until the cluster division of material semantic fingerprint vectors is completed. The consistency of the directional distribution of material feature vectors in high-dimensional space is quantified by calculating the cosine of the angle between each material feature vector in each cluster. Based on the results of directional distribution consistency, we can capture the variation patterns of component ratios and characteristic indexes among different materials within clusters. The dimensional weights of the component ratio variation pattern and the feature index variation trend are assigned, and the cosine similarity of the semantic fingerprint vector corresponding to any material is calculated based on the assignment results using the cosine similarity calculation formula.

7. The blockchain-based method for linking and aggregating large-scale material data according to claim 6, characterized in that, The discretized material semantic fingerprint vector maps the material semantic fingerprint vector to a high-dimensional space, and combines the mapping result with a density threshold for cluster analysis. Based on the analysis results, clusters are constructed, including: Each feature dimension of the material semantic fingerprint vector is discretized and encoded to construct an initial dimension vector of the semantic fingerprint. Then, a high-dimensional mapping process is performed on the initial dimension vector of the semantic fingerprint through a linear transformation matrix to obtain a set of material feature vectors with unified dimensions. Material feature vectors with missing feature dimensions and abnormal vector magnitudes are filtered out from the material feature vector set to form an effective feature vector dataset in a high-dimensional feature space. At the same time, the neighborhood radius and density threshold of the high-dimensional feature space are set. Traverse the effective feature vector dataset to find the core vector that satisfies the neighborhood radius and density threshold, and extract all feature vectors in the high-dimensional feature space where the core vector is located. Filter all density-connected feature vectors starting from the core vector until there are no new expandable feature vectors, thus completing the construction of a single cluster.

8. The blockchain-based aggregation method for large-scale material data according to claim 1, characterized in that, The process of calculating the hash value of the association record, constructing a Merkle tree based on the hash value, storing the association record on the blockchain network, and simultaneously uploading the material data to an off-chain storage medium to achieve trusted storage of the association and material data includes: The hash value of each related record is calculated using hashing technology. The hash values ​​generated within the target time window are used as leaf nodes. Leaf nodes are paired up to obtain parent nodes, and the process continues until all leaf nodes are paired up to obtain the root node. The timestamps of the root node and the target window are merged and packaged into a blockchain transaction and submitted to the blockchain network. The blockchain network nodes verify the blockchain transaction through the consensus mechanism and write it into a block after verification, so as to realize the on-chain storage of the relationship record. Material data and material entity heterogeneous diagrams are stored as raw data in a distributed file storage, and hash addresses are generated for the distributed files. At the same time, a field pointing to the storage address of the raw data is added to the relationship record to achieve reliable and associated storage of the relationship and material data.

9. The blockchain-based method for linking and converging large-scale material data according to claim 8, characterized in that, The verification of blockchain transactions using blockchain network nodes through a consensus mechanism includes: Based on blockchain transactions, a set of storage data feature descriptors is generated. Consensus decision-making operations are performed in combination with the set of storage data feature descriptors according to the embedded consensus verification space and dynamic confidence constraints. The initial verification decision scheme is output based on the operation results. The stored data feature descriptor set is segmented into time windows. By predicting the feature evolution path within the future time window, a predicted feature descriptor set is generated. Based on the predicted feature descriptor set, the initial verification decision scheme is calculated to compensate for deviations and generate an optimized verification decision scheme. Based on the optimized verification decision scheme, an execution instruction is sent to the blockchain network nodes, triggering the distributed nodes to perform blockchain transaction verification operations.

10. A blockchain-based system for linking and aggregating large-scale materials data, characterized in that, The system includes: The data access module is used to access material data based on data templates and information extraction rules, and to perform format unification and entity alignment processing on the material data, and to obtain material description text based on the processing results; The semantic analysis module is used to construct a material entity heterogeneous graph using material description text, and to encode and weight pool the material entity heterogeneous graph based on graph neural network to obtain semantic fingerprint vectors to characterize the comprehensive features of the material. The association calculation module is used to calculate the cosine similarity of semantic fingerprint vectors based on density clustering algorithm, compare the cosine similarity with the association threshold to determine the implicit association between the accessed material data, and output the association relationship record; The blockchain evidence storage module is used to calculate the hash value of the relationship record, construct a Merkle tree based on the hash value, store the relationship record to the blockchain network, and upload the material data to the off-chain storage medium to achieve reliable storage of the relationship and material data.

Images