Material intelligent question-answering and performance prediction method and system based on knowledge graph
Through intelligent question-and-answer method based on knowledge graph and GRID similarity calculation, combined with BERT-CRF technology, the material knowledge graph is constructed, and the problems that have not been revealed in the deep correlation relationship in material data are solved, and important breakthroughs in material performance prediction and new material discovery are achieved. .
Patent Information
- Application Number
- CN202411282050.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-12
- Publication Date
- 2024-12-31
AI Technical Summary
The prior art fails to effectively reveal the deep correlation between materials in material data mining, resulting in the limitation of new materials discovery.
Intelligent Q&A method based on knowledge graph is adopted, and the material knowledge graph is constructed in combination with BERT-CRF technology, and performance prediction is achieved through GRID similarity calculation, and graph construction is accelerated using multi-threaded pooling technology to ensure the readability and information retention of the graph.
Accurate material knowledge Q&A and performance prediction, promote the discovery of new materials and potential information mining, and improve the understanding and data analysis capabilities of materials science research.
Smart Images
Figure CN119230024A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of material prediction and analysis, and more specifically to a material intelligent question-answering and performance prediction method and system based on knowledge graph. Background Art
[0002] At present, with the continuous accumulation of material data and the rapid development of information technologies such as big data, data-driven has become a new research model for the rapid design of new materials. Among them, knowledge graph technology describes physical concepts and their relationships to form a network knowledge structure and deeply explores the relationship between nodes, thereby discovering the laws hidden behind the data.
[0003] Knowledge graphs can integrate knowledge such as the properties and structure of materials and represent them in a triple structured way, thereby helping researchers in the field of materials to efficiently organize material knowledge and capture the deep connections between materials and the associations between material properties. Some researchers used machine learning and a deduplication model of matching dependency algorithms to solve the problem of author name ambiguity, and used the constructed knowledge graph to clearly list the articles and authors belonging to the field, and obtained a milestone map of LiFePO4 research and development through citation information. In addition, a research team has also built a semantic knowledge graph focusing on LIB cathode materials. The nodes represent related materials, the edges represent the correlation between nodes, and the cosine similarity metric is used, which is different from the traditional relationship construction. There are also research teams that extract various entities and relationships from the literature, including seven entities such as materials, performance, and applications, and use statistical metrics to establish relationships between entities. The entities are matched and linked with the corresponding entities in Wikidata and Materials Project to enrich information and achieve cross-domain knowledge discovery. The development of the ontology matching system OTMapOnto can handle semantically similar terms in different ontologies, standardize annotations for material data, and use NLP technology to automatically extract entities and relationships from unstructured literature. A new process for building knowledge graphs was proposed using the above technology. A research team also proposed a method for generating a metal material knowledge graph (MMKG) from DBpedia and Wikipedia, using an algorithm based on direct link subgraph semantic distance (DLSSD) to preliminarily extract metal material entities, and further screen using a combination of semantic distance and string similarity (SDSS) methods, and extract material attribute knowledge from Wikipedia to further enrich MMKG and supplement the missing attribute knowledge of material entities extracted from DBpedia. From the above, it can be seen that the current application of knowledge graphs is mostly concentrated on the accumulation of data in data sets, as well as the comparison and analogy of data in data sets, and no related reports on the acquisition or prediction of new entities have been found.
[0004] Therefore, how to use intelligent question answering and material prediction to reveal the deep-seated correlations between materials, thereby facilitating data mining and analysis and promoting the discovery of new materials is an urgent problem that technical personnel in this field need to solve. Summary of the invention
[0005] In view of this, the present invention provides a material intelligent question answering and performance prediction method and system based on knowledge graph, adopts a new knowledge graph construction method to construct a material knowledge graph, combines BERT-CRF intelligent question answering technology, and uses GRID (the Grouped Representation of Interatomic Distances) similarity calculation to achieve performance prediction, thereby solving the problems in the background technology.
[0006] In order to achieve the above object, the present invention adopts the following technical solution:
[0007] On the one hand, a material intelligent question answering and performance prediction method based on knowledge graph is provided, including:
[0008] Obtain original material data and construct original data set;
[0009] Analyze the original data set and extract data associated with material structure and performance properties as effective information;
[0010] Construct a material knowledge graph based on valid information and set three types of nodes: material structure hash ID node, material measurement ID identifier node, and material attribute node;
[0011] Perform multiple rule templates and synonym replacement processing on the material knowledge graph to build a diverse data set;
[0012] Use diverse data sets to train an entity recognition model based on BERT+CRF to obtain an intelligent question-answering model;
[0013] Use intelligent question-answering models to conduct material knowledge questions and answers;
[0014] It also includes matching similar materials between material entities within the material knowledge graph and newly introduced material entities, building relationships to complete the knowledge of the material knowledge graph, and then realizing material performance prediction.
[0015] Optionally, the data associated with the material structure and performance attributes include site number, molecular formula, volume, calculation parameters, electronic structure, thermodynamic properties, and mechanical properties as material properties, retaining material information and ensuring readability of the spectrum.
[0016] Optionally, the material structure hash id node represents a hash value of the material structure, which is a unique identifier of the structure; the material measurement id identifier node represents data identifiers of different material measurement methods; and the material property node represents different material properties.
[0017] Optionally, constructing the material knowledge graph includes using Neo4j graph database technology to construct and manage material data, and using multi-thread pool technology combined with node uniqueness constraints to construct the knowledge graph.
[0018] Optionally, the method of training an entity recognition model based on BERT+CRF using a diverse dataset specifically includes:
[0019] Use the BERT-CRF entity recognition model, set the loss function to the negative log-likelihood function based on the conditional random field model, the evaluation algorithm to accuracy, recall rate and F1 score, and the optimizer to the Adam algorithm. Further optimize the hyperparameters to obtain the intelligent question-answering model, and save the hyperparameters and weight information corresponding to the model.
[0020] Optionally, the similar material matching specifically includes:
[0021] Through GRID similarity calculation, matching similar material entities are found, the relationship between entities is established in the graph, the knowledge graph is completed, and performance prediction is achieved.
[0022] Optionally, during the calculation process, the GRID descriptors of all entities in the material knowledge graph are calculated in advance, and a grouping strategy is used for optimization to reduce the number of comparisons and achieve fast calculation results.
[0023] On the other hand, a material intelligent question answering and performance prediction system based on knowledge graph is provided, comprising:
[0024] Dataset acquisition module, which acquires the original data of materials and constructs the original data set;
[0025] A data set analysis module analyzes the original data set and extracts data associated with material structure and performance properties as effective information;
[0026] The material knowledge graph construction module builds the material knowledge graph based on the valid information and sets three types of nodes: material structure hash ID node, material measurement ID identifier node, and material attribute node;
[0027] The material knowledge graph processing module performs multiple rule templates and synonym replacement processing on the material knowledge graph to build a diverse data set;
[0028] The intelligent question-answering model building module uses a diverse data set to train an entity recognition model based on BERT+CRF to obtain an intelligent question-answering model;
[0029] The question-answering result output module uses the intelligent question-answering model to conduct material knowledge questions and answers;
[0030] The performance prediction module matches similar materials to the material entities in the material knowledge graph and the newly introduced material entities, builds relationships to complete the knowledge of the material knowledge graph, and then realizes the performance prediction of the material.
[0031] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a material intelligent question-answering and performance prediction method and system based on knowledge graph, which adopts a new knowledge graph construction method to structure the material data, retain the material information to the greatest extent and ensure the readability of the graph. The system framework proposed by the present invention can realize accurate material knowledge question-answering, realize knowledge completion by similarity calculation, and complete material performance prediction, which is of great significance to accelerate the discovery of new materials and mine the potential information of materials. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.
[0033] Figure 1 A flow chart of the method provided by the present invention;
[0034] Figure 2 A flow chart for constructing the material knowledge graph provided by the present invention;
[0035] Figure 3 The framework for designing the intelligent question-answering model based on knowledge graph provided by the present invention;
[0036] Figure 4 The knowledge graph performance prediction framework provided by the present invention;
[0037] Figure 5 The present invention provides a flow chart for implementing the method of the present invention by taking a III-V semiconductor material data set as an example. DETAILED DESCRIPTION
[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0039] On the one hand, the embodiment of the present invention discloses a material intelligent question answering and performance prediction method based on knowledge graph, such as Figure 1 As shown, including:
[0040] Obtain original material data and construct original data set;
[0041] Analyze the original data set and extract data associated with material structure and performance properties as effective information;
[0042] Construct a material knowledge graph based on valid information and set three types of nodes: material structure hash ID node, material measurement ID identifier node, and material attribute node;
[0043] Perform multiple rule templates and synonym replacement processing on the material knowledge graph to build a diverse data set;
[0044] Use diverse data sets to train an entity recognition model based on BERT+CRF to obtain an intelligent question-answering model;
[0045] Use intelligent question-answering models to conduct material knowledge questions and answers;
[0046] It also includes matching similar materials between material entities within the material knowledge graph and newly introduced material entities, building relationships to complete the knowledge of the material knowledge graph, and then realizing material performance prediction.
[0047] In a specific embodiment, the acquisition of raw material data includes: first, data needs to be acquired to construct a data set. The data in the data set can be acquired through literature analysis, experimental measurement, or computational simulation. The data set can be constructed by building a data set according to the material data specification, or by using a public high-quality data set.
[0048] In a specific embodiment, the original data set is analyzed and effective information is extracted. For a material, due to the different measurement methods, the numerical value of the attribute obtained will have a certain error, but the different numerical values generated by the different measurement methods can also reflect the basic properties of the material. Therefore, multiple data cannot be normalized by simple weighting or averaging, which will lose the characteristics of the material and cause information loss. Therefore, the number of sites, molecular formula, volume, calculation parameters, electronic structure (including material properties such as energy gap, band gap type), thermodynamic properties (including material properties such as formation energy, atomic energy), mechanical properties (including elastic components, mechanical quantities, etc.) that are closely related to the material structure and performance attributes are extracted as the material attributes, which can retain the material information to the greatest extent and ensure the readability of the spectrum.
[0049] In a specific embodiment, Figure 2 As shown in the figure, the method of constructing knowledge graph is introduced to enhance the semantic expression ability of graph and optimize the graph structure. Three node settings are adopted: material structure hash id node, material measurement id identifier node, and material attribute node. The structure hash id node represents the hash value of the material structure and is the unique identifier of the structure. The material measurement id identifier node represents the data identifier of different material measurement methods. The material attribute node represents different material attributes. It can distinguish different measurement values caused by different measurement methods of the same structure, and distinguish which data in the database it belongs to.
[0050] In a specific embodiment, the attribute node contains all the information of the material itself, which can be freely customized and applied to different application scenarios. In the process of knowledge graph construction, multi-thread pool technology is used in combination with node uniqueness constraints to speed up the construction of knowledge graph. This construction method can fully retain material information and ensure the readability of the graph, which has rich semantic information.
[0051] In a specific embodiment, using a diverse dataset to train an entity recognition model based on BERT+CRF specifically includes:
[0052] Use the BERT-CRF entity recognition model, set the loss function to the negative log-likelihood function based on the conditional random field model, the evaluation algorithm to accuracy, recall rate and F1 score, and the optimizer to the Adam algorithm. Further optimize the hyperparameters to obtain the intelligent question-answering model, and save the hyperparameters and weight information corresponding to the model.
[0053] like Figure 3 As shown in the figure, using BERT-CRF to perform entity recognition on the question and answer sentences provided by users can obtain more and more accurate entities, thereby obtaining accurate user semantics, thereby improving the accuracy of the question and answer system. Adding different rule templates to build training data sets improves the diversity of data sets and the robustness and generalization of training models. Introducing the BERT pre-trained model and adding the CRF layer for training, we finally obtain a high-quality question and answer model.
[0054] In a specific embodiment, Figure 4 As shown, similar material matching specifically includes:
[0055] Use the GRID method to calculate material similarity, predict material properties and complete knowledge. Use parallel algorithms and other optimization strategies to speed up similarity calculations. Match similar materials for material entities in the graph and newly introduced material entities, build relationships to complete knowledge, and predict material performance. Make the knowledge graph more comprehensive and contain more extensive information. The similarity calculation includes both structural similarity and compositional similarity.
[0056] This includes operations on both new entities and existing entities in the graph. The corresponding knowledge of existing entities is completed, performance prediction is performed on new entities, and graph knowledge is completed.
[0057] Through the above steps, the construction of material knowledge graph and the application of intelligent question answering and performance prediction can be realized. The invention can help us better discover the deep connection between materials, greatly promote the comprehensive understanding and data analysis of the relationship between materials, and promote material science research and the discovery of new materials.
[0058] Specifically, the working principle of the present invention is specifically described by taking a self-extracted III-V semiconductor material data set including the open source material database Materials Project and OQMD as an example.
[0059] According to the present invention, Figure 5 The specific implementation process is as follows:
[0060] The embodiments are derived from data obtained from open source computing datasets, and these data are standardized and constructed into a material dataset that meets the requirements of the research group.
[0061] After extracting, screening and processing the data set, the used data set contains III-V semiconductor material data of different components and structures. The number of data in this database is about 30,000 material data.
[0062] Extract effective basic information from the original material data. This embodiment takes the electronic structure properties as the main research target. The material attribute information includes the number of sites, molecular formula, energy gap, volume, formation energy, atomic energy, material structure hash ID, and material measurement ID. According to the determined node information, the above information is extracted and processed from the database.
[0063] Using the construction method adopted by the present invention, the structure hash id node, material measurement id identifier node, and attribute node are set. Neo4j graph database technology is used to build and manage these data, and multi-thread pool parallel technology is used, combined with Neo4j's unique constraint mechanism, to optimize and accelerate graph construction.
[0064] The data in the material knowledge graph is processed and spliced, different rule templates are added, and synonymous substitution is adopted to allow the model to learn more location information and semantic information. A diverse training data set is constructed, and the BERT-CRF entity recognition model is used. The loss function is set to the negative log-likelihood function based on the conditional random field model, and the evaluation algorithm is the accuracy, recall rate, and F1 score. The optimizer is the Adam algorithm, and the hyperparameters such as the number of single training samples, learning rate, number of training rounds, and weights are further optimized to obtain a more accurate question-answering model. And save the hyperparameters and weights corresponding to the model.
[0065] For material entities in the knowledge graph and newly added material entities, GRID similarity calculation can be used to find matching similar material entities, establish relationships between entities in the graph, complete the knowledge graph, and achieve performance prediction. During the calculation process, the GRID descriptors of all entities in the material knowledge graph are calculated in advance, and the grouping strategy is used for optimization to reduce the number of comparisons and achieve fast calculation results.
[0066] It can be seen that the present invention proposes a material knowledge graph technology that combines an intelligent question-answering system and realizes material performance prediction. A knowledge graph containing a material structure hash id node, a material measurement id identifier node, and a material attribute node is constructed, including the number of sites, molecular formula, volume, electronic structure (including material attributes such as energy gap, band gap type), thermodynamic properties (including material attributes such as formation energy, atomic energy), mechanical properties (including elastic components, mechanical quantities, etc.), etc., and multi-thread pool technology is introduced in the construction of the knowledge graph, which optimizes and accelerates the construction of the graph and saves time to a great extent. The BERT+CRF model is used to perform named entity recognition on the sentence, and the preliminary classification results of the word are obtained by BERT, and then the classification results are constrained by CRF, and finally the obtained entity is spliced and queried. GRID similarity calculation is adopted, and the optimization strategy is used to speed up the calculation speed, and similar material matching is performed to realize material knowledge completion, and it can also be applied to the introduction of new entities.
[0067] On the other hand, a material intelligent question answering and performance prediction system based on knowledge graph is provided, comprising:
[0068] Dataset acquisition module, which acquires the original data of materials and constructs the original data set;
[0069] A data set analysis module analyzes the original data set and extracts data associated with material structure and performance properties as effective information;
[0070] The material knowledge graph construction module builds the material knowledge graph based on the valid information and sets three types of nodes: material structure hash ID node, material measurement ID identifier node, and material attribute node;
[0071] The material knowledge graph processing module performs multiple rule templates and synonym replacement processing on the material knowledge graph to build a diverse data set;
[0072] The intelligent question-answering model building module uses a diverse data set to train an entity recognition model based on BERT+CRF to obtain an intelligent question-answering model;
[0073] The question-answering result output module uses the intelligent question-answering model to conduct material knowledge questions and answers;
[0074] The performance prediction module matches similar materials to the material entities in the material knowledge graph and the newly introduced material entities, builds relationships to complete the knowledge of the material knowledge graph, and then realizes the performance prediction of the material.
[0075] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.
[0076] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A material intelligent question answering and performance prediction method based on knowledge graph, characterized in that: include: Obtain original material data and construct original data set; Analyze the original data set and extract data associated with material structure and performance properties as effective information; Construct a material knowledge graph based on valid information and set three types of nodes: material structure hash ID node, material measurement ID identifier node, and material attribute node; Perform multiple rule templates and synonym replacement processing on the material knowledge graph to build a diverse data set; Use diverse data sets to train an entity recognition model based on BERT+CRF to obtain an intelligent question-answering model; Use intelligent question-answering models to conduct material knowledge questions and answers; It also includes matching similar materials between material entities within the material knowledge graph and newly introduced material entities, building relationships to complete the knowledge of the material knowledge graph, and then realizing material performance prediction.
2. According to claim 1, a material intelligent question-answering and performance prediction method based on knowledge graph is characterized in that: The data associated with the material structure and performance attributes include the number of sites, molecular formula, volume, calculation parameters, electronic structure, thermodynamic properties, and mechanical properties as material attributes, retaining material information and ensuring the readability of the spectrum.
3. The material intelligent question answering and performance prediction method based on knowledge graph according to claim 1 is characterized in that: The material structure hash id node represents the hash value of the material structure, which is the unique identifier of the structure. The material measurement id identifier node represents different material measurement method data identifiers. The material attribute node represents different material attributes.
4. The material intelligent question answering and performance prediction method based on knowledge graph according to claim 1 is characterized in that: The construction of the material knowledge graph includes using Neo4j graph database technology to construct and manage material data, and using multi-thread pool technology combined with node uniqueness constraints to construct the knowledge graph.
5. The material intelligent question answering and performance prediction method based on knowledge graph according to claim 1 is characterized in that: The method of using a diverse dataset to train an entity recognition model based on BERT+CRF specifically includes: Use the BERT-CRF entity recognition model, set the loss function to the negative log-likelihood function based on the conditional random field model, the evaluation algorithm to accuracy, recall rate and F1 score, and the optimizer to the Adam algorithm. Further optimize the hyperparameters to obtain the intelligent question-answering model, and save the hyperparameters and weight information corresponding to the model.
6. The material intelligent question answering and performance prediction method based on knowledge graph according to claim 1 is characterized in that: The similar material matching specifically includes: Through GRID similarity calculation, matching similar material entities are found, the relationship between entities is established in the graph, the knowledge graph is completed, and performance prediction is achieved.
7. The material intelligent question answering and performance prediction method based on knowledge graph according to claim 6 is characterized in that: During the calculation process, the GRID descriptors of all entities in the material knowledge graph are calculated in advance, and the grouping strategy is used for optimization to reduce the number of comparisons and achieve fast calculation results.
8. A material intelligent question-answering and performance prediction system based on knowledge graph, characterized in that: A material intelligent question-answering and performance prediction method based on knowledge graph as described in any one of claims 1 to 7 is applied, comprising: Dataset acquisition module, which acquires the original data of materials and constructs the original data set; A data set analysis module analyzes the original data set and extracts data associated with material structure and performance properties as effective information; The material knowledge graph construction module builds the material knowledge graph based on the valid information and sets three types of nodes: material structure hash ID node, material measurement ID identifier node, and material attribute node; The material knowledge graph processing module performs multiple rule template and synonym replacement processing on the material knowledge graph to construct a diverse data set; The intelligent question-answering model building module uses a diverse data set to train an entity recognition model based on BERT+CRF to obtain an intelligent question-answering model; The question-answering result output module uses the intelligent question-answering model to conduct material knowledge questions and answers; The performance prediction module matches similar materials to the material entities in the material knowledge graph and the newly introduced material entities, builds relationships to complete the knowledge of the material knowledge graph, and then realizes the performance prediction of the material.