Layered multi-granularity material descriptor management method
Through the layered multi-grained material descriptor management method, the information loss and redundancy problems in material descriptor selection and management are solved, efficient multiplexing and management of descriptors are realized, and the efficiency of materials science research is improved.
Patent Information
- Application Number
- CN202411125629.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-16
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-08-16
AI Technical Summary
In the prior art, the selection of material descriptors depends on domain knowledge, there is information loss and redundancy, and there is a lack of unified standards and systemicity, which leads to difficulties in reuse and management of data and affects the efficiency of machine learning models.
A layered multi-grained material descriptor management method is proposed. By defining the descriptor tree, obtaining and fusing descriptors obtained in different ways, using importance score calculation to select the final descriptor, reducing subjectivity and optimizing the structure of the descriptor tree .
The fusion tree of comprehensive information and concise form of descriptors is realized, which improves the reusability and management efficiency of descriptors, reduces the subjectivity of the selection process, and improves the efficiency of materials science research.
Smart Images

Figure CN118939750A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of material technology, and in particular relates to a hierarchical multi-granularity material descriptor management method. Background Art
[0002] With the development of materials science, the amount of material data accumulated is becoming increasingly huge. Extracting useful information from hundreds of millions of material data and analyzing and sorting out the structure-activity relationship of materials' "composition-structure-process-performance" have become the core and key to the research and development of new materials. Machine learning methods can quickly mine potential laws from data, help researchers efficiently predict the performance of materials, and then reverse design new materials according to needs. However, the performance of machine learning models is restricted by data quality. Therefore, scientifically and rationally selecting descriptor features is still the key to establishing high-precision machine learning models. At present, the selection of descriptors depends largely on domain knowledge. The descriptors selected by experts in different fields often differ, which makes it extremely difficult to reuse descriptors in different material systems and seriously limits the efficiency of machine learning methods. The problems of traditional methods for mining material descriptors from literature include: 1) Descriptors mined from literature are prone to information missing and redundancy (there are many words with the same meaning but different meanings in material literature sentences, such as chemical formulas, their abbreviations, and their English names in material literature); 2) The lack of standardized expression, storage, and maintenance of descriptors makes it difficult for researchers to reuse them. Among them, there is a lack of unified standards for the expression of descriptors, and different researchers and teams may use different formats and terminology, resulting in poor data interoperability. Secondly, the storage method of descriptors is not standardized and may be scattered in various documents, databases, and personal records, making them difficult to centrally manage and access. In addition, the maintenance of descriptors lacks systematicity, old data and methods may not be updated in a timely manner, and new discoveries and improvements are difficult to effectively integrate. Therefore, it is urgent to propose a new descriptor selection and management method to guide the selection and reuse of descriptors, so that the entire process has a basis to follow, in order to accelerate the efficiency of machine learning. Summary of the invention
[0003] In order to solve the above technical problems, the present invention proposes a hierarchical and multi-granular material descriptor management method to solve the above problems in the prior art.
[0004] To achieve the above object, the present invention provides a hierarchical and multi-granular material descriptor management method, comprising:
[0005] Define a descriptor tree and get descriptors;
[0006] Constructing a plurality of hierarchical and multi-granular descriptor trees based on the descriptors and the definitions of the descriptors;
[0007] Fusing several of the descriptor trees to obtain a descriptor fusion tree;
[0008] A comprehensive score of each descriptor in the descriptor fusion tree is calculated based on the importance score, and a final descriptor is selected based on the comprehensive score.
[0009] Optionally, the process of defining the descriptor tree includes: defining the descriptor tree based on a nested and structured data structure.
[0010] Optionally, based on different descriptor acquisition approaches, several hierarchical and multi-granular descriptor trees constructed using the definition of a descriptor tree include: an expert experience tree, a text mining tree, and a feature tree.
[0011] Optionally, the process of constructing a descriptor tree based on the nested and structured data structure includes:
[0012] Preprocessing the data file to obtain a preprocessing file;
[0013] Extracting data information of the preprocessing file to construct a nested defined descriptor tree;
[0014] Each node in the nested defined descriptor tree is converted into JSON format data to obtain a structured descriptor tree.
[0015] Optionally, the process of fusing several of the descriptor trees to obtain a descriptor fusion tree includes merging descriptors with the same meaning in the expert experience tree, the text mining tree and the feature tree into one item, retaining descriptors with different meanings at the same time, and constructing a descriptor fusion tree;
[0016] The fusion strategy for fusion of several descriptor trees is:
[0017] R U =R E ∪R T ∪R Ta
[0018] In the formula, R E is the descriptor set of the expert experience tree, E n is the total number of descriptors in the expert experience tree; R T is the descriptor set of the text mining tree, T n is the total number of descriptors in the text mining tree; R Ta is the descriptor set of the feature tree, Ta n is the total number of descriptors in the feature tree; R U is the set of descriptors in the descriptor fusion tree.
[0019] Optionally, after constructing the descriptor fusion tree, the method further includes removing redundant information from the descriptor fusion tree; wherein the rules for removing redundant information include: eliminating redundant concepts and eliminating redundant formulas;
[0020] The expressions for eliminating redundant concepts and redundant formulas are:
[0021]
[0022] Where R is the descriptor set, R = {f 1 ,f 2 ,...,f n}, n is the total number of descriptors, f i (i∈(1,2,...,n)) is a descriptor, Conpt={C 1 ,C 2 ,...,C k} is a set of physical concepts, C K ={C k,1 ,C k,2 ,...,C k,l}(l=1,...,n) is a collection of different objects of the same physical concept.
[0023] Optionally, the process of calculating the comprehensive score of each descriptor in the descriptor fusion tree includes: respectively using word frequency-inverse file frequency, fusion tree, formula and availability to calculate the importance of descriptor nodes to obtain four dimensional scores, and comprehensively calculating the four dimensional scores based on weighted rules to obtain the comprehensive score of each descriptor.
[0024] Optionally, the expression for calculating the importance score of the descriptor based on the TF-IDF method is:
[0025]
[0026] In the formula, S T is the importance score of the descriptor based on the TF-IDF method, T wn is the importance score of each word in the descriptor phrase, n∈[1,2,3,...], where n is the number of words in the descriptor phrase;
[0027] The expression for calculating the importance score of the descriptor based on the fusion tree is:
[0028]
[0029] In the formula, S E is the importance score of the descriptor based on the fusion tree, N i For each node i∈{1,2,...,n} in the descriptor tree, f EFor text mining trees, f T is the expert experience tree, f F is a feature tree.
[0030] Optionally, the importance score of the descriptor based on the formula library method is:
[0031]
[0032] In the formula, S F is the importance score of the descriptor based on the formula, R F represents the formula library, f EE represents a fusion tree or a fusion tree subtree, N i is each node i∈{1,2,...,n} in the descriptor tree, Root represents the root node of the fusion tree or fusion tree subtree, if N i in R F and N i =Root means that if the descriptor is the root node of the fusion tree or fusion tree subtree, and it is in the formula library, then the fusion tree or fusion tree subtree f EE All descriptor nodes of have a score of 1 based on the importance of the formula;
[0033] The importance scores of the descriptors based on availability are:
[0034]
[0035] Among them, S A is the importance score of the descriptor based on availability, N i is each node i∈{1,2,...,n} in the descriptor tree.
[0036] Compared with the prior art, the present invention has the following advantages and technical effects:
[0037] The present invention proposes a hierarchical and multi-granular material descriptor management method. This method aims to integrate descriptors obtained by different methods and store them in the form of a descriptor tree to better express and manage descriptors. This method comprehensively considers the problem of missing information and redundancy that may exist when mining descriptors from literature in different ways. The descriptors obtained by different methods are constructed into the form of a descriptor tree, and then the fusion strategy and redundancy elimination rules are used to fuse and optimize multiple descriptor trees, and finally a descriptor fusion tree with comprehensive information and concise form is obtained. In addition, in order to solve the problem that the current descriptor selection method relies on expert experience, resulting in a strong subjectivity in the selection process, we also adopted a descriptor selection method based on importance scoring to reduce the subjectivity of the selection process from the source, so as to guide domain experts to efficiently select descriptors and construct high-quality data. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:
[0039] Figure 1 The specific process of the multi-granularity descriptor selection method based on the descriptor tree according to the embodiment of the present invention;
[0040] Figure 2 It is a flowchart of the overall solution implementation of an embodiment of the present invention;
[0041] Figure 3 A formalized representation of a descriptor tree according to an embodiment of the present invention;
[0042] Figure 4 A descriptor tree representing a nested definition of an embodiment of the present invention;
[0043] Figure 5 A descriptor tree representing a structured definition of an embodiment of the present invention;
[0044] Figure 6 Constructing a descriptor tree from an Excel file in an embodiment of the present invention;
[0045] Figure 7 Partial display of the expert experience tree based on nested definition in an embodiment of the present invention;
[0046] Figure 8 Decomposing and constructing a structured descriptor tree for the nested descriptor tree of the embodiment of the present invention;
[0047] Fig. 9 It is a part of nodes of an expert tree based on structured definition according to an embodiment of the present invention;
[0048] Fig.10 Constructing a nested descriptor for the structured descriptor tree of the embodiment of the present invention;
[0049] Fig.11 The fusion process of the fusion tree of the embodiment of the present invention;
[0050] Fig.12 is an example of an expert experience tree of an embodiment of the present invention, wherein the leaves represent fine-grained descriptors;
[0051] Fig.13 This is an example of a text mining tree in an embodiment of the present invention, where leaves represent fine-grained descriptors and circles represent nodes with different descriptions of the same concept;
[0052] Fig.14 is an example of a feature tree in an embodiment of the present invention, wherein leaves represent fine-grained descriptors, and circles represent nodes with different descriptions of the same concept;
[0053] Fig.15 is an example of a fusion tree in an embodiment of the present invention, wherein the leaves represent the existence of fine-grained nodes;
[0054] Fig.16 The present invention is a method flow for selecting a descriptor based on importance calculation according to an embodiment of the present invention. DETAILED DESCRIPTION
[0055] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.
[0056] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.
[0057] Embodiment 1
[0058] At present, the acquisition of descriptors in the field of materials is heavily dependent on the domain knowledge of experts. The subjectivity and limitations of expert knowledge will seriously affect the selection of descriptors, thereby affecting the study of material structure-activity relationship. Although natural language processing (NLP) technology has been applied to extract information from materials science literature, there is no research report on how to use NLP methods to obtain specific performance-driven descriptors from literature. In addition, the descriptors mined from the literature may have information missing and redundancy, and how to store and express the obtained descriptors for reference by different researchers is still an urgent problem to be solved.
[0059] Therefore, in order to quickly and reasonably acquire and manage potential descriptors from materials science literature, a hierarchical and multi-granular material descriptor management method is proposed. By filling the NASICON solid electrolyte descriptors acquired in different ways into the descriptor tree, the multiple descriptor trees are fused and optimized according to the fusion strategy and redundancy elimination rules. Finally, a descriptor fusion tree with comprehensive information and concise form is obtained, which greatly facilitates material experts in selecting and managing descriptors during material design and structure-activity relationship research.
[0060] The present invention proposes a hierarchical and multi-granular material descriptor management method. First, the concept of a descriptor tree is proposed. Then, descriptors obtained in different ways are integrated and stored in the form of a descriptor tree (according to the understanding of ion transport mechanism by materials experts, a descriptor tree is constructed based on the inclusion and subordination relationship between materials concepts, and the obtained tree is called an expert experience tree; a conditional text data enhancement method and a descriptor automatic identifier that integrate material field knowledge are used to mine descriptors related to ion transport performance from literature, and the descriptor tree constructed by this method is called a text mining tree; available descriptor information is collected from material research using machine learning methods and filled into the descriptor tree to obtain a feature tree), multiple descriptor trees are fused and optimized according to a fusion strategy and a redundancy elimination rule, and finally a descriptor fusion tree with comprehensive information and concise form is obtained. On this basis, a descriptor selection strategy based on comprehensive scoring was proposed. By eliminating coarse-grained nodes and selecting fine-grained nodes, the importance of descriptor nodes was calculated from four dimensions: TF-IDF, fusion tree, formula and availability, and the comprehensive score was obtained according to the weighted rule. The comprehensive score was used to guide and complete the descriptor selection, reducing the subjectivity of the selection process from the source, so as to guide domain experts to efficiently select descriptors and construct high-quality data.
[0061] like Figure 1 and Figure 2 As shown, this embodiment provides a hierarchical multi-granularity material descriptor management method, including the following steps: defining a descriptor tree and obtaining descriptors; constructing a number of hierarchical multi-granularity descriptor trees based on the descriptors and the definitions of the descriptors; fusing a number of the descriptor trees to obtain a descriptor fusion tree; calculating a comprehensive score for each descriptor in the descriptor fusion tree based on an importance score, and selecting a final descriptor based on the comprehensive score.
[0062] The specific implementation process includes:
[0063] 1. Definition of descriptor tree: In order to facilitate the reading, maintenance and reuse of descriptors, they are integrated and stored in the form of descriptor trees. The form of descriptor trees varies and needs to be determined by the research object. 2. Construction of descriptor tree: fill the descriptors obtained in different ways into the defined descriptor tree. According to the structure of the descriptor tree, the descriptor tree can be divided into nested descriptor trees (mainly used for expression and display) and structured descriptor trees (mainly used for storage and management). 3. Fusion of multiple descriptor trees: merge the descriptors with the same meaning in the multiple descriptor trees obtained in 2) into one item, retain the descriptors with different meanings and remove redundant information to construct a fused tree. 4. Descriptor selection based on importance score calculation: calculate the importance of the descriptor nodes from four dimensions of the descriptor tree constructed in 3), and obtain its comprehensive score according to the weighted rule. The comprehensive score is used to guide and complete the descriptor selection. The process of this method is as follows: Figure 1 The detailed schematic diagram is shown in Figure 2 shown.
[0064] Among them, 1. Definition of descriptor tree: In order to facilitate the reading, maintenance and reuse of descriptors, we integrate and store them in the form of descriptor trees, such as Figure 3 The formal representation of the descriptor tree is shown in formula (1).
[0065] DescriptorsTree(N,E,H,S)(1)
[0066] Where N represents a node, N = {N 1 ,N 2 ,...,N n},n∈[1,m]. E represents edge, E={E 1 ,E 2 ,...,E n}, n∈[1,m]. H represents the hierarchy, H={H 1 ,H 2 ,...,H n}, n∈[1,m]. S represents the attribute (Slot), S={S 1 ,S 2 ,...,S n},n∈[1,m].
[0067] Node: A descriptor in a descriptor tree is called a node, and each node can contain multiple child nodes. For example, a component is a node, and information related to atoms is a child node of the component.
[0068] Edge: The connection between two nodes is called an edge, which represents the relationship between the nodes. By default, an edge represents the relationship between two nodes from abstract to concrete. For example, the descriptors "channel" and "bottleneck" are connected by an edge, indicating that "channel" is a more abstract concept of "bottleneck". In addition, edges can also represent other custom relationships. For example, there may be a potential impact relationship between "edge length" and "volume".
[0069] Attributes: used to store various functions of nodes and edges, such as calculation formulas of descriptors, etc.
[0070] Level: measures the abstractness of the descriptor. Descriptors with higher levels have coarser granularity, more abstract concepts, stronger universality, and contain more fine-grained descriptors. Descriptors with lower levels have finer granularity, more specific concepts, and are more computable and accessible.
[0071] The shape of the descriptor tree is determined by the research object. The hierarchical representation of nodes in the descriptor tree helps to select appropriate descriptors in different application scenarios. When the research system is broad, more universal coarse-grained descriptors should be selected, and when the research object is more specific (such as a specific material), more targeted fine-grained descriptors can be selected.
[0072] The "attribute" component in the descriptor tree is used to store other information about nodes or edges. The specific content can be customized according to actual conditions. In order to facilitate different researchers to quickly understand the content of each descriptor, the node attribute should contain the following information:
[0073] Introduction: The detailed text content of the descriptor corresponding to this node.
[0074] Source: Record the original source of the descriptor to facilitate the traceability of the descriptor. It should be noted that the original source is not limited to a certain article. If the descriptor comes from a table or a calculation software, it should also be recorded.
[0075] Importance score: Score each node according to specific rules to evaluate the importance of the descriptor. This parameter is used to guide the selection of descriptors.
[0076] Calculation formula: record the quantitative characterization formula of the descriptor. If there are multiple characterization methods, you can record them all and select them according to specific scenarios later.
[0077] It is worth noting that the content of the attributes in the descriptor tree is not fixed and can be dynamically adjusted according to different application scenarios. For example, if it is difficult to score each descriptor due to lack of literature or methods, the attribute of importance score can be deleted.
[0078] The above is a theoretical definition of the descriptor tree, which is relatively abstract and not convenient for direct application to system storage. Therefore, when implementing the algorithm, a nested and structured data structure is used to define the descriptor tree. The nested definition is mainly used for the expression and display of the descriptor tree, and the structured definition is mainly used for the storage and management of the descriptor tree.
[0079] Furthermore, nested definition: The data structure of the nested definition is shown in the following formula (2). The nested definition displays the descriptor tree in a hierarchical manner. Each descriptor node contains the nested structure of its child nodes, providing an intuitive visual representation capability, such as Figure 4 shown.
[0080] DescriptorsTree(N),as f N (2)
[0081] N is the node Node, which is initially the root node. The node definition contains the following information: 1) Unique ID: It is the unique identifier of a node. 2) Node Name: The name of the node. 3) Children Array Children: Node Set {N}, recording all child nodes. 4) Type TreeType: It records whether it belongs to the expert experience tree, text mining tree, feature tree, fusion tree or other tree.
[0082] The above are the required properties of a descriptor tree. The descriptor tree can also have the following optional properties to add more information to the descriptor tree: 1) Concept Hierarchy: divided into three categories: coarse-grained layer, fine-grained layer, and concept layer. 2) Horizontal Hierarchy: divided according to the depth of the tree, the root node is 1, the next layer is 2, and so on. 3) Introduction: Introduces the detailed text content of the descriptor corresponding to the node. 4) Source: Records the original source of the descriptor. 5) Importance score: Score each node according to specific rules (optional, depending on the actual situation). 6) Calculation formula: Records the quantitative characterization formula of the descriptor (optional, depending on the actual situation).
[0083] Furthermore, structured definition: The data structure of structured definition is shown in the following formula (3). Compared with nested definition, structured definition loses the nested Children attribute and adds hierarchical information such as ParentId attribute, ChildrenName attribute, and ChildrenId attribute to represent its hierarchical structure and create a descriptor tree, which is convenient for the storage and management of the descriptor tree. Its structure is as follows: Figure 5 shown.
[0084] DescriptorsTree(N),as f S (3)
[0085] N is the node Node, which is initially the root node. The node definition contains the following information: 1) Unique identifier Id: It is a unique identifier of a node. 2) Node name NodeName. 3) Parent node unique identifier ParentId: It records the identifier of the parent node of the current node. 4) Children array name ChildrenName: It records the names of all child nodes. 5) Children array identifier ChildrenId: It records the IDs of all child nodes. 6) Type TreeType: It records whether it belongs to expert experience tree, text mining tree, feature tree, fusion tree or other tree.
[0086] 2. Construction of Descriptor Tree
[0087] The construction of a descriptor tree refers to the process of obtaining descriptors and filling them into the descriptor tree. There are three ways to obtain descriptors: (1) According to the understanding of the ion transport mechanism by materials experts, a descriptor tree is constructed based on the inclusion and subordination between materials concepts. The resulting tree is called an expert experience tree; (2) Using the conditional text data enhancement method and descriptor automatic identifier that integrates material field knowledge proposed in the present invention, descriptors related to ion transport performance are mined from the literature. The descriptor tree constructed by this method is called a text mining tree; (3) Available descriptor information is collected from material research using machine learning methods and filled into the descriptor tree to obtain a feature tree. According to the different definition methods used by the descriptor tree, the construction of the descriptor tree is divided into the construction of a nested descriptor tree and the construction of a structured descriptor tree. Expert experience trees, text mining trees, and feature trees can all be constructed using these two definition methods.
[0088] Furthermore, the construction of nested descriptor trees: nested descriptor trees often have two file formats, one is a JSON file, which can directly construct a nested descriptor tree, but is not suitable for manual operation. The other is an Excel file, which cannot be directly converted into a nested descriptor tree. Therefore, the corresponding algorithm program can be designed to simplify the generation of nested descriptor trees from Excel, which will be conducive to the completion of the construction of descriptor trees based on nested definitions. The specific process is as follows: Figure 6 shown.
[0089] Building a descriptor tree from an Excel file is mainly divided into two steps: file preprocessing and extracting information to build a nested descriptor tree.
[0090] The specific implementation process includes:
[0091] (1) File preprocessing
[0092] File preprocessing mainly completes merged cell detection and merged cell splitting. Since the program only assigns a value to the first cell when reading merged cells, and the values of other cells are all empty, it is necessary to split the merged type cells into non-merged type cells for reading.
[0093] (2) Extract information to build a nested descriptor tree
[0094] Constructing a nested defined descriptor tree from an Excel file is mainly divided into three parts: 1) Reading by row to establish branches: Traverse all data in Excel by row, and according to the nested data represented by merged cells in Excel, each row of data can construct a branch of the descriptor tree from the root node to the leaf node. 2) Branch establishment rules: When constructing a descriptor branch, traverse each descriptor in the current row. If the descriptor does not exist in the descriptor tree, fill the descriptor into the descriptor tree and add a child array to complete all the information of the descriptor. If the descriptor already exists, do not add the descriptor to the descriptor tree, but find the position of the descriptor and point it to the child array of the descriptor to generate a nested structure. 3) End mark: Establish a branch each time a row of data is read. When the last descriptor data of the last row is read, all branches are established, the descriptor tree is built, and the descriptor tree is output, such as Figure 7 shown.
[0095] Furthermore, the construction of the structured descriptor tree: The structured descriptor tree is mainly used for data storage and management. It converts each node of the nested descriptor tree into JSON format data to form a structured descriptor tree. It is impossible to directly construct a structured descriptor tree from the original data. Figure 8 Shows the nested descriptor tree f N Decomposed into a structured descriptor tree f S Algorithm flow. Nested descriptor tree f N Decompose and access from the root node, create two arrays ChildrenId and ChildrenName to record the unique ID and node name of all children respectively, and store these two arrays and the node information itself in f S Then traverse its child array and recursively perform the above operation on each child. After all recursions are popped out of the stack, the stack is empty, and the structured descriptor tree f S The construction is completed, the structured descriptor tree instance is as follows Fig. 9 shown.
[0096] The stored structured descriptor tree can also be converted into a nested descriptor tree, the algorithm is as follows Fig.10 As shown. The algorithm first traverses the structured descriptor tree f S, find the node whose ParentId attribute is empty. This node has no parent node, that is, this node is the root node. Starting from the root node, create a nested array Children, get all the IDs of its child nodes according to ChildrenId, and traverse f according to ID S Find the node, add it to Children and perform recursive operation with it as the root node. Each time the recursive operation points the top pointer of the stack to the child node, the nested array Children is completed for the child node. After the recursion is completed, the nested descriptor tree f is constructed. S .
[0097] 3. Fusion of multiple descriptor trees
[0098] The "trunk" of the descriptor tree can sort out the relationship between descriptors of different granularities, organize domain knowledge from abstract to concrete in an orderly manner, and provide guidance for different research objectives. The "branches and leaves" of the descriptor tree can be specific descriptors, which can record all information related to the node, facilitating the selection and reuse of descriptors. The descriptor trees obtained by different methods have their own characteristics. In terms of content, the text mining tree is the most comprehensive, followed by the expert experience tree, and finally the feature tree. The characteristics of different descriptor trees are shown in Table 1. In order to absorb the advantages of the three trees, complete the descriptors, remove redundant information in the descriptor tree, and facilitate researchers to consult, it is necessary to merge the three trees into a more comprehensive descriptor fusion tree. The above three trees are all constructed using nested definitions.
[0099] Table 1
[0100]
[0101] Fusion strategy: merge the descriptions with the same meaning in the expert experience tree, text mining tree and feature tree into one item, retain the descriptors with different meanings at the same time, and construct a fusion tree. The details are as follows: E is the descriptor set of the expert experience tree, E n is the total number of descriptors in the expert experience tree; R T is the descriptor set of the text mining tree, T n is the total number of descriptors in the text mining tree; R Ta is the descriptor set of the feature tree, Ta n is the total number of descriptors in the feature tree; R U is the set of descriptors in the descriptor fusion tree.
[0102] The fusion strategy is:
[0103] R U =RE ∪R T ∪R Ta (4)
[0104] It should be noted that some descriptors with the same meaning may be expressed differently. For example, "a" and "Latticeparameter a" both represent lattice parameter a, so it is necessary to perform redundancy removal on these two nodes. In addition, if the formula is directly used as the node of the descriptor, the product of the two formulas will be 1, which also requires redundancy removal. Based on the above analysis, rules 1 and 2 are defined to eliminate redundant information.
[0105] Rule 1 Redundant concept elimination rule: Due to different sources of descriptors, there may be different descriptions of the same physical concept, so it is necessary to eliminate redundant information. R is a descriptor set, R = {f 1 ,f 2 ,...,f n}, n is the total number of descriptors. i (i∈(1,2,...,n)) is a descriptor. Conpt={C 1 ,C 2 ,...,C k} is a set of physical concepts, where C K ={C k,1 ,C k,2 ,...,C k,l}(l=1,...,n) is a set of different objects of the same physical concept. The specific rules are shown in formula (5).
[0106]
[0107] Rule 2 Redundant formula elimination rule: In addition to phrases, descriptors can also be expressed directly in formulas. If there are redundant formulas, they also need to be merged. R = {f 1 ,f 2 ,...,f n}, n is the total number of descriptors, f i (i∈(1,2,...,n)) is a descriptor. If two descriptors are reciprocals of each other, the specific rules are as shown in formula (6).
[0108]
[0109] Based on the above fusion strategy and redundancy elimination rules, we will perform the strategy-based multi-descriptor tree fusion as shown in Algorithm 1 to achieve the fusion of multiple descriptor trees. The specific process is as follows Fig.11 shown.
[0110]
[0111] By summarizing the expert experience, the coarse-grained layer, fine-grained layer and concept layer are constructed from top to bottom, and a hierarchical expert experience tree is obtained, such as Fig.12 As shown in the figure. The leaves at the bottom layer represent the finest-grained descriptors (only some fine-grained descriptors are shown). The expert experience tree has clear hierarchies and concepts, with many coarse-grained nodes and few fine-grained nodes. It has no redundant information and provides a basic "skeleton" for the descriptor tree. However, the descriptors in the concept layer are unevenly distributed.
[0112] Text mining tree Fig.13 As shown in the figure, the leaves represent the finest-grained descriptors, and the green spheres represent different expressions of the concept of the node. The results show that the granularity of the descriptors varies. For example, there are coarse-grained descriptors such as "concentration" and fine-grained descriptors such as "M vacancy". The descriptors obtained by the text mining method come from a large number of different articles and authors. On the one hand, there is a lot of redundant information in the descriptors (different expressions of the same concept), such as the lattice constant descriptors "Parameter c" and "Cell parameter c". On the other hand, it also helps domain experts obtain content that they may have overlooked, making the descriptor information more comprehensive. For example, in the research on solid electrolytes at the simulation calculation level, there is a lack of experimental content, and the descriptor automatic identifier can supplement this information to a certain extent, mining experimental descriptors such as "Annealing temperature" and "Cooling rate". In summary, there are many nodes in the coarse-grained layer, fine-grained layer, and concept layer in the text mining tree, and the descriptors are distributed relatively evenly, covering each fine-grained node. At the same time, compared with the expert experience tree, the text mining tree has greatly improved both depth and breadth, and can find important information that can be ignored. However, it has more redundant information and the hierarchy between concepts is not clear.
[0113] The descriptors in the feature tree are standardized, with clear concepts and distinct levels, such as Fig.14 The distribution characteristics of the feature tree are similar to those of the expert experience tree, but the number of coarse-grained nodes is insufficient, the influencing factors considered are not extensive enough, and there is also some redundant information.
[0114] The fusion tree obtained by fusing the above three trees is as follows Fig.15 As shown in the figure, by comparison, it can be found that the fusion tree contains all the information of the expert experience tree, text mining tree and feature tree. Compared with the three initial descriptor trees, the depth and breadth of the fusion tree have been greatly expanded. The problem of redundant descriptor information in the fusion tree has been basically solved.
[0115] 4. Descriptor Selection Based on Importance Score Calculation
[0116] After building the fusion tree, in order to reduce the subjectivity of descriptor selection from the source (selecting descriptors for machine learning tasks) and ensure that the selected descriptors have certain physical connotations, a descriptor selection method based on importance score calculation is proposed. The importance of descriptor nodes is calculated from four dimensions: TF-IDF, fusion tree, formula, and availability. The comprehensive score is obtained according to the weighted rule, and the descriptor selection is guided and completed by the comprehensive score. The process of the descriptor selection method based on importance calculation is as follows: Fig.16 Through this method, the comprehensive importance score can be used to select more accurate, comprehensive and effective descriptors, laying a solid high-quality data foundation for machine learning tasks.
[0117] Furthermore, the importance score calculation based on TF-IDF is as follows: TF-IDF is a statistical method that can be used to evaluate the importance of a word in a document / corpus. The core idea of this method is that the importance of a word is positively correlated with its frequency of occurrence in the text and negatively correlated with its frequency of occurrence in the corpus. The TF-IDF method consists of two parts. TF represents the frequency of occurrence of a word in the text. In order to prevent it from being biased towards longer texts, it is usually normalized. As shown in formula (7).
[0118]
[0119] Among them, TF i,j Represents the word t i In the document d j The frequency in n i,j Represents the word t i In the document d j The number of times it appears in k n k,j This means that in document d j The sum of all word counts in the text. It is worth noting that in actual situations, the frequency of a word is not necessarily positively correlated with its importance. On the contrary, some words with lower frequency of occurrence are particularly important in the text. Therefore, it is not enough to consider only TF, but also IDF.
[0120] IDF indicates the criticality of a word. i The fewer the documents, the larger the IDF, and vice versa, as shown in formula (8).
[0121]
[0122] Where |D| represents the number of all documents, |j:t i ∈dj | indicates that the word t is included i The number of documents in the word list is 1. The reason for adding 1 here is to prevent the denominator from being 0, which may cause calculation errors. Finally, the importance of each word is obtained based on the TF and IDF values, as shown in formula (9).
[0123] TF-IDF=TF×IDF (9)
[0124] The final TF-IDF importance result of a descriptor is the weighted sum of the TF-IDF importance of the descriptor to all documents and the average value. When the descriptor is a phrase, its TD-IDF importance is the weighted sum of the importance of each word in the phrase and the average value, as shown in formula (10).
[0125]
[0126] Among them, S T is the importance score of the descriptor based on the TF-IDF method, T wn is the importance score of each word in the descriptor phrase, n∈[1,2,3,...], n is the number of words in the descriptor phrase. In addition, all words are converted to lowercase letters to eliminate case interference, and some meaningless particles and modal particles such as "in" and "to" are deleted using regular matching methods.
[0127] After obtaining the importance score of the descriptor, in order to unify it with the scores of other dimensions in the following text, it is necessary to perform standardization. First, all scores are normalized to between 0 and 1, and then sorted in descending order. The top third of the descriptors are scored as 1, the bottom third are scored as 0, and the remaining middle third are scored as 0.5.
[0128] Further, the importance score calculation based on the fusion tree: The importance score calculation rule based on the fusion tree is: when the descriptor tree is fused, the number of times each descriptor appears in the text mining tree, the expert experience tree, and the feature tree is used as its importance index. Appearing 3 times (i.e., appearing once in each of the three trees) is 1 point, appearing 2 times (i.e., appearing once in two of the trees) is 0.5 points, and appearing once (appearing once in only one of the trees) is 0 points. The initial value of the importance score based on the fusion tree is -0.5. Every time the descriptor node appears in a tree, the score increases by 0.5 points, as shown in formula (11):
[0129]
[0130] Among them, S E is the importance score of the descriptor based on the fusion tree, N iFor each node i∈{1,2,...,n} in the descriptor tree, f E For text mining trees, f T is the expert experience tree, f F is a feature tree. The fusion tree composed of nodes with scores greater than or equal to 1 is called an intersection tree; the fusion tree composed of nodes with scores greater than or equal to 0.5 is called a middle tree; the fusion tree composed of nodes with scores greater than or equal to 0 is called a union tree.
[0131] Furthermore, the importance score based on the formula is calculated: if a descriptor has appeared in the material formula, its importance score based on the formula is 1, otherwise it is 0. From this rule, it can be seen that this indicator is related to the collected formula library. The more comprehensive the formula library is, the more accurate the associated descriptors are, and the better the effect of this importance indicator is.
[0132] According to Table 2, which is a partial formula library, if the descriptor appears as a leaf node, it itself has a score of 1. If it is the root node of a subtree, all its corresponding subtrees have a score of 1, as shown in formula (12).
[0133]
[0134] Among them, S F is the importance score of the descriptor based on the formula, R F represents the formula library, f EE represents the fusion tree or its subtree, N i is each node i∈{1,2,...,n} in the descriptor tree, Root represents the root node of the fusion tree or subtree. i in R F and N i =Root means that if the descriptor is the root node of the fusion tree or its subtree, and it is in the formula library, then the fusion tree or subtree f EE All descriptor nodes of have a score of 1 based on the importance of the formula.
[0135] Table 2
[0136]
[0137] Furthermore, the importance score based on availability is calculated as follows: if the descriptor is directly available, its importance score is 1; if the descriptor is highly abstract and difficult to quantitatively characterize, its score is 0. Based on these rules, it can be seen that the finer the descriptor, the more specific it is, and the greater the probability that its importance score of availability will be 1. Therefore, this score can make the descriptor selection tend to be finer-grained descriptors, as shown in formula (13).
[0138]
[0139] Among them, S A is the importance score of the descriptor based on availability, N i is each node i∈{1,2,...,n} in the descriptor tree.
[0140] Further, the comprehensive score of importance: The importance scores of descriptors based on TF-IDF, fusion tree, formula and availability may have different effects on their comprehensive scores and should be differentiated according to specific weights. Therefore, the formula for the final score is shown in (14):
[0141] Score=ω 1 S T +ω 2 S E +ω 3 S F +ω 4 S A (14)
[0142] Among them, Score is the final comprehensive score of importance, S A is the importance score of the descriptor based on availability, S T is the importance score obtained based on the word frequency-inverse document frequency method, S E is the score based on the fusion tree, S F is the importance score based on the formula, ω 1 ,ω 2 ,ω 3 ,ω 4 They are weights on different dimensions, and their default value is 1.
[0143] Furthermore, selection of descriptors based on importance scores: After obtaining the comprehensive importance scores of the descriptors, appropriate descriptors can be selected under the guidance of the comprehensive importance scores of the descriptors. With the increase of data and the deepening of research, the comprehensive importance scores of the descriptors will change dynamically according to different weights and the increase or decrease of the number of samples, so it is necessary to select according to the actual situation. The selection strategy is formulated by the researchers, and the descriptors can be checked from high to low scores to select them. You can also set a threshold for selection, such as selecting the tertile, quartile or the mean of all importance scores as the threshold, or the lowest comprehensive importance score of the descriptor in the expert experience tree as the threshold, and select descriptors with a comprehensive importance score greater than or equal to this value. As shown in Table 3, the top 10 descriptors are selected from high to low according to the comprehensive scores. The top descriptors, such as "temperature", "concentration", and 'a' (lattice constant), are relatively specific and easy to understand, and there are research methods to calculate their corresponding descriptor values. They are all descriptors that can be applied to material machine learning. This verifies the feasibility and effectiveness of our descriptor selection method based on importance calculation and the corresponding algorithm program implemented.
[0144] Table 3
[0145]
[0146] Fig.16 The importance comprehensive score of a coarse-grained node is the sum of the importance comprehensive scores of its child nodes. The higher the value, the more child nodes it has, and the more important and easier it is to be selected. Among them, radius, occupancy, electronegativity, and valence state rank high, indicating that these descriptors are easy to be selected. The reasons for the high ranking of these descriptors may include two aspects: first, from the material perspective, the above four descriptors have a greater impact on the ion transport performance, so they are easy to be selected; second, from the perspective of the difficulty of obtaining descriptors, the above descriptors can be directly obtained by reading tables or files, so they rank high. Among the descriptors with importance scores of 3.5 and 3. "temperature" and "concentration" represent temperature and concentration, respectively. These information are very important influencing factors in material experiments. In summary, the importance comprehensive score can effectively screen out important descriptors, making the selection of descriptors reasonable and justifiable.
[0147] In the dataset constructed based on the descriptor tree selected by the descriptor tree, the performance of different machine learning models on the original dataset (45 descriptors and 90 samples) and the modified dataset (45 descriptors and 85 samples) is shown in Table 4. It can be seen that the model always predicts activation energy better on the modified dataset than on the original dataset. On the modified dataset, it is found that the generalization performance of other models except KNN is relatively good, and its R 2 The model with the worst generalization performance is KNN, but its RMSE still reaches 0.079, MAPE reaches 0.051, and R 2 It reaches 0.894. The experimental results prove that the descriptors selected by the descriptor tree can well predict the activation energy of NASICON-type compounds.
[0148] Table 4
[0149]
[0150] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.
Claims
1. A hierarchical and multi-granular material descriptor management method, characterized in that: The following steps are involved: Define a descriptor tree and get descriptors; Constructing a plurality of hierarchical and multi-granular descriptor trees based on the descriptors and definitions of the descriptors; Fusing several of the descriptor trees to obtain a descriptor fusion tree; A comprehensive score of each descriptor in the descriptor fusion tree is calculated based on the importance score, and a final descriptor is selected based on the comprehensive score.
2. The hierarchical multi-granularity material descriptor management method according to claim 1, characterized in that: The process of defining a descriptor tree includes: defining the descriptor tree based on a nested and structured data structure.
3. The hierarchical multi-granularity material descriptor management method according to claim 2, characterized in that: Based on the different ways of obtaining descriptors, several hierarchical and multi-granular descriptor trees are constructed using the definition of descriptor tree, including: expert experience tree, text mining tree and feature tree.
4. The hierarchical multi-granularity material descriptor management method according to claim 3, characterized in that: The process of building a descriptor tree based on nested and structured data structures includes: Preprocessing the data file to obtain a preprocessing file; Extracting data information of the preprocessing file to construct a nested defined descriptor tree; Each node in the nested defined descriptor tree is converted into JSON format data to obtain a structured descriptor tree.
5. The hierarchical multi-granular material descriptor management method according to claim 3, characterized in that: The process of fusing several of the descriptor trees to obtain a descriptor fusion tree includes merging descriptors with the same meaning in the expert experience tree, the text mining tree and the feature tree into one item, retaining descriptors with different meanings at the same time, and constructing a descriptor fusion tree; The fusion strategy for fusion of several descriptor trees is: R U =R E ∪R T ∪R Ta In the formula, R E is the descriptor set of the expert experience tree, E n is the total number of descriptors in the expert experience tree; R T is the descriptor set of the text mining tree, T n is the total number of descriptors in the text mining tree; R Ta is the descriptor set of the feature tree, Ta n is the total number of descriptors in the feature tree; R U is the set of descriptors in the descriptor fusion tree.
6. The hierarchical multi-granular material descriptor management method according to claim 5, characterized in that: After constructing the descriptor fusion tree, the method further includes removing redundant information from the descriptor fusion tree; wherein the rules for removing redundant information include: eliminating redundant concepts and eliminating redundant formulas; The expressions for eliminating redundant concepts and redundant formulas are: Where R is the descriptor set, R = {f1,f2,...,f n }, n is the total number of descriptors, f i (i∈(1,2,...,n)) is a descriptor, Conpt={C1,C2,...,C k } is a set of physical concepts, C K ={C k,1 ,C k,2 ,...,C k,l }(l=1,...,n) is a collection of different objects of the same physical concept.
7. The hierarchical multi-granularity material descriptor management method according to claim 1, characterized in that: The process of calculating the comprehensive score of each descriptor in the descriptor fusion tree includes: respectively using word frequency-inverse file frequency, fusion tree, formula and availability to calculate the importance of descriptor nodes to obtain four dimensional scores, and comprehensively calculating the four dimensional scores based on weighted rules to obtain the comprehensive score of each descriptor.
8. The hierarchical multi-granular material descriptor management method according to claim 7, characterized in that: The expression for calculating the importance score of the descriptor based on the TF-IDF method is: In the formula, S T is the importance score of the descriptor based on the TF-IDF method, T wn is the importance score of each word in the descriptor phrase, n∈[1,2,3,...], where n is the number of words in the descriptor phrase; The expression for calculating the importance score of the descriptor based on the fusion tree is: In the formula, S E is the importance score of the descriptor based on the fusion tree, N i For each node i∈{1,2,...,n} in the descriptor tree, f E For text mining trees, f T is the expert experience tree, f F is a feature tree.
9. The hierarchical multi-granularity material descriptor management method according to claim 7, characterized in that: The importance score of the descriptor based on the formula library method is: In the formula, S F is the importance score of the descriptor based on the formula, R F represents the formula library, f EE represents a fusion tree or a fusion tree subtree, N i is each node i∈{1,2,...,n} in the descriptor tree, Root represents the root node of the fusion tree or fusion tree subtree, if N i in R F and N i =Root means that if the descriptor is the root node of the fusion tree or fusion tree subtree, and it is in the formula library, then the fusion tree or fusion tree subtree f EE All descriptor nodes of have a score of 1 based on the importance of the formula; The importance scores of the descriptors based on availability are: Among them, S A is the importance score of the descriptor based on availability, N i is each node i∈{1,2,...,n} in the descriptor tree.
Citation Information
Patent Citations
Text data descriptor identification method and device and medium
CN114997176A
