A method for constructing a hydrogen storage material database based on artificial intelligence technology

Abstract

This invention discloses a method for constructing a hydrogen storage material database based on artificial intelligence technology, which is achieved through the following nine steps: 1. Searching and organizing papers; 2. Downloading and archiving papers; 3. Text format conversion; 4. Text preprocessing; 5. Chemical named entity recognition; 6. Text classification; 7. Text relation extraction; 8. Ternary archiving; 9. Multi-terminal storage. Compared with existing technologies, this invention solves the following problems: 1. It solves the problem of the database having a single data source, that is, it achieves data acquisition from published papers; 2. It solves the problem of the database relying on manual data collection, that is, it achieves automated data extraction through artificial intelligence technology; 3. It solves the problem of the database having a single application.

Description

Technical Field

[0001] This invention relates to the field of database maintenance technology, and more specifically to a method for constructing a hydrogen storage material database based on artificial intelligence technology. Background Technology

[0002] With the emergence of the concept of materials genome engineering and the development of modern artificial intelligence technology, combining big data, machine learning, and materials science has become a research hotspot both domestically and internationally in recent years. A large amount of research data can be obtained from scientific papers; however, manual data extraction is time-consuming, labor-intensive, and prone to errors. Therefore, it is necessary to replace manual labor with machines. Thus, we have developed an AI-based automated data extraction tool for scientific papers and proposed a method for automatically supplementing and enriching data to construct and improve a hydrogen storage materials database for subsequent research.

[0003] For example, in existing technology 1 (Baricco M, Palumbo M, Pinatel E, et al. Thermodynamic Database for Hydrogen Storage Materials[J]. Advances in Science & Technology, 2010, 72:213-218.), M Baricco et al. established a relatively complete thermodynamic database for hydrogen storage materials in 2010. However, the data mainly came from within the research group, resulting in a very limited amount of data that could not meet the requirements of machine learning.

[0004] Therefore, to address the aforementioned issues, existing technology 2 (Coppola CM, Tolbatov I, Tranca IC, et al. A database approach for materials selection for hydrogen storage in aerospace technology[J]. Rendiconti Lincei. Scienze Fisiche e Naturali, 2019,30(10).) proposes a database approach for selecting hydrogen storage materials in aerospace in 2019. This approach obtains data from papers and research groups in relevant research directions through manual methods and has achieved good results. However, this technology still has problems: data acquisition relies on traditional manual extraction, which is inefficient and cannot meet the requirements of big data.

[0005] Therefore, to address the aforementioned issues, establishing various materials-related databases for subsequent big data analysis and machine learning is a current research focus. Among these, the Atomly materials science database independently developed by the Institute of Physics, Chinese Academy of Sciences, and the integrated supercomputing, data, and AI platform independently developed by the National Supercomputing Center in Guangzhou in 2021 are representative examples of such applications. These platforms combine databases with supercomputing to improve the efficiency of materials science research. Analysis of the above technologies reveals the following two characteristics:

[0006] 1. Although the data source has been expanded from a single research group to multiple other collaborative research groups, the data sources are still limited and cannot meet the requirements of machine learning.

[0007] 2. Although its database contains a large amount of data, the types of materials it covers are relatively limited and cannot meet the needs of specific application areas.

[0008] Specifically, in the field of hydrogen storage materials involved in this invention, the application of existing database technologies faces the following difficulties:

[0009] 1. Due to the large number of names for hydrogen storage materials and the existence of many technical terms, applying machine learning models that have been applied to the material types in the above database to the field of hydrogen storage materials will not yield the desired results.

[0010] 2. The data sources of the existing work mentioned above all originate from data shared within multiple collaborative research groups. Although the amount of data has been expanded, it only accounts for a small portion of the total data held by all research groups in the entire field of hydrogen storage materials research. Therefore, it is necessary to search and collect data from the entire field of hydrogen storage materials research in order to expand the data sources of the existing hydrogen storage materials database.

[0011] 3. Applying machine learning to the collected hydrogen storage material data also faces challenges such as insufficient data volume, low data quality, and a limited number of variables involved in the prediction, resulting in machine learning prediction models failing to meet the requirements of practical applications.

[0012] To address the aforementioned issues and improve the quantity and quality of acquired data, this invention expands the database by automatically collecting data from publicly available papers on hydrogen storage materials. This expands the data source from existing collaborative research groups to include data from other researchers in the same field published in publicly available papers. Furthermore, this method can obtain high-quality data, thereby achieving high accuracy in subsequent machine learning. Simultaneously, it enables automated data extraction to replace manual data extraction, significantly improving the efficiency of data collection.

[0013] The data collection process has the following problems:

[0014] 1. There are many technical terms for hydrogen storage materials, and traditional chemical named entity recognition models cannot be adapted to them, resulting in poor recognition performance of existing named entity recognition in non-specialized fields in the content of the above-mentioned papers.

[0015] Specifically, taking existing chemical named entity recognition methods, such as Existing Technology 3 (Swain MC, Cole JM. ChemDataExtractor: A toolkit for automated extraction of chemical information from the scientific literature[J]. Journal of Chemical Information & Modeling, 2016:acs.jcim.6b00207.), as an example, although a named entity recognition model was built using Python and can recognize chemical named entities in papers, in actual application, this model has poor recognition performance in papers on hydrogen storage materials. Therefore, establishing a more professional chemical named entity recognition system for hydrogen storage materials is one of the problems that must be solved.

[0016] 2. Since most papers obtained from public channels are in PDF format, the PDF documents must be converted before extracting data.

[0017] The existing common method is to use Python packages to convert PDF to Word or TXT, such as Pdfminer_3k and Python-docx. Although this method can achieve text format conversion, the conversion strategy in the actual conversion process will affect the actual conversion effect.

[0018] The existing conversion strategy of the program package has problems in the process of converting images, resulting in garbled characters and invalid line breaks after the images are converted into text, which increases the workload of subsequent text processing.

[0019] Therefore, PDF format conversion and its conversion strategy need to be adapted to the image conversion strategy, and a skip strategy should be adopted for images that do not need to be processed, so as to reduce the workload of subsequent text processing.

[0020] 3. The converted text may contain irrelevant content, which can reduce the efficiency of subsequent extraction. Therefore, in order to solve this problem, the commonly used method is multi-strategy text preprocessing. This technology is currently mature and widely used, and has low technical difficulty. However, it is necessary to select appropriate preprocessing strategies to improve the execution efficiency of subsequent programs and the efficiency of data extraction.

[0021] 4. Choosing the right data extraction strategy and algorithm is crucial. This part of the work requires the use of part-of-speech analysis in natural language processing combined with matching algorithms. However, how to effectively combine the two determines the final extraction efficiency and accuracy. Summary of the Invention

[0022] The purpose of this invention is to provide a method for constructing a hydrogen storage material database based on artificial intelligence technology.

[0023] By using artificial intelligence technology, valuable high-quality data is extracted from papers on hydrogen storage materials to build a hydrogen storage material database. This addresses the problem that existing hydrogen storage material databases rely on manual collection, which is time-consuming, labor-intensive, and prone to low accuracy due to subjective factors. Furthermore, based on this, big data machine learning methods based on artificial intelligence technology are used to predict the performance parameters of hydrogen storage materials, thereby shortening the research and development cycle of new hydrogen storage materials.

[0024] The technologies involved in realizing this invention include web crawling, text format conversion, natural language processing, text mining, data mining, and database operation and maintenance, so as to achieve automated expansion of the hydrogen storage material database.

[0025] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0026] A method for constructing a hydrogen storage material database based on artificial intelligence technology includes the following nine steps:

[0027] Step 1: Searching for and organizing papers;

[0028] Step 2: Downloading and archiving the paper;

[0029] Step 3: Text format conversion;

[0030] Step 4: Text preprocessing;

[0031] Step 5: Chemical Nominated Entity Recognition;

[0032] Step 6: Text classification;

[0033] Step 7: Text relation extraction;

[0034] Step 8: Tripartite archiving;

[0035] Step 9: Multi-terminal storage.

[0036] The specific implementation method is as follows:

[0037] Step 1: Searching and organizing papers. Use academic search engines to obtain the names of papers related to hydrogen storage materials, and then organize them into the names of papers to be searched for use in Step 2. The search engines used in this embodiment include Baidu Scholar and Google Scholar.

[0038] Step 2, downloading and archiving papers: The papers that need further processing are obtained and archived through automated downloading based on the names of the papers to be searched compiled in Step 1.

[0039] The process for querying and archiving papers in step 2 is as follows: paper title input, web crawling, paper download, and archiving of papers in various formats. The specific steps are as follows:

[0040] Step 2.1: After inputting the paper title, web crawling technology can be used to search for the corresponding paper title. However, when using academic search engines to obtain relevant information about papers, traditional manual operations cannot meet the requirements of big data machine learning. Therefore, this invention uses Python's Selenium automated script to simulate a browser in silent mode to replace traditional manual operations, thereby querying the results of the literature title to be searched and returning the paper information. Since the DOI number is the most useful information in the literature results, this invention selects the DOI number for subsequent processing.

[0041] Step 2.2, saving the above DOI number to the local Excel file, also involves a lot of repetitive operations. Manual operation cannot meet the requirements. Therefore, this invention uses Python's openpyxl to write to the Excel file. Then, the Requests command is used to call the applied Elsevier API to read the DOI number in the above Excel document and download the PDF document of the literature to be searched, so as to achieve simultaneous archiving and organization.

[0042] Step 3, text format conversion: Since the paper downloaded and archived in Step 2 is in PDF format, and the text of the PDF paper cannot be directly processed in subsequent steps, it is necessary to perform text format conversion to convert the PDF paper document into Word and TXT format documents.

[0043] The specific method for text format conversion in step 3 is as follows: use the Python Pdfminer_3k and Python-docx packages to convert PDF documents to Word documents, then use the Win32 package to externally call WPS to convert Word documents into TXT documents and name them Raw_TXT, thereby realizing the organization and archiving of the above-mentioned document formats, and using the system storage command to locally archive the format-converted text.

[0044] Step 4, text preprocessing: Since there are easily confused words and a lot of irrelevant content in the text format conversion process, it is necessary to remove irrelevant content in the Raw_TXT after text format conversion in Step 3 by natural language processing methods.

[0045] The specific steps of text preprocessing in step 4 are as follows:

[0046] Step 4.1: Perform common format conversions, replace easily misspelled words, truncate irrelevant content, remove non-English text, and strip merged words from the Raw_TXT file;

[0047] Step 4.2 requires segmenting the text after the above processing into sentences. Therefore, it is necessary to process the text for common stop words and then archive the segmented text in multiple formats, namely, archive it separately in Excel, Word and TXT formats.

[0048] Step 4.3: Perform word segmentation and content block extraction on the processed TXT document in sequence;

[0049] Step 4.4 involves word segmentation of the text. This is because the accuracy of chemical named entity recognition in subsequent steps depends on the accuracy of the pre-trained word vector model of the hydrogen storage material paper. Therefore, it is necessary to train word vectors on the segmented text to establish a word vector analysis model for subsequent database applications.

[0050] Step 4.5: The extracted content is then divided into sections such as introduction, keywords, abstract, main text, and citations, and these sections are archived for future use.

[0051] Step 5, Chemical Named Entity Recognition: Since there is a lot of useless information in the preprocessed sentence text in Step 4, it will increase the amount of invalid work in subsequent relation extraction and ternary archiving. Relying on manual chemical named entity recognition will affect the accuracy of subsequent steps. Therefore, it is necessary to classify and filter the text by using a named entity recognition model adapted to hydrogen storage materials, and remove text that does not need to be extracted to reduce the invalid extraction in the future.

[0052] The specific steps for chemical nominate entity identification in step 5 are as follows:

[0053] Step 5.1: Use regular expressions to expand and adapt the relevant performance values ​​and named entities of hydrogen storage materials to improve the accuracy of the chemical named entity recognition process for hydrogen storage materials.

[0054] Step 5.2: Use Word2vec and TF-IDF to train a word vector model to generate a high-frequency dictionary of chemical nominal entities for hydrogen storage materials, and apply this to subsequent steps;

[0055] Step 6, text classification: Since the text after text preprocessing in Step 4 contains invalid information, this patent selects sentence segments containing target information from the full text after sentence segmentation based on the named entity recognition results in Step 5, as the processing objects for subsequent relation extraction.

[0056] Step 7, text relation extraction. Although the text processed in Step 6 contains chemical named entities of hydrogen storage materials, it is still impossible to directly obtain the relationship between the various hydrogen storage material matrices, performance names, and performance values ​​in the sentences. Therefore, relation extraction is required. That is, the sentence relationship between various text parts in the text is extracted by natural language processing. The key to relation extraction is to search and match through part-of-speech relations.

[0057] Step 7, text relation extraction, involves Dijkstra's shortest path algorithm, simple sequence matching algorithm, and CoreNLP's part-of-speech distance algorithm. The specific steps of text relation extraction are as follows:

[0058] Step 7.1: Process the sentence segments containing target information after classification in Step 6. Since each word will have a sorting label to represent its position in the sentence after word segmentation, the position of the keyword to be extracted in the current sentence to be processed can be located, and the corresponding keyword label can be attached to it for subsequent processing.

[0059] Step 7.2: Use different search strategies for different keywords to extract to extract the corresponding data values;

[0060] Step 8, Ternary Archiving: Archive data in a ternary format of "keyword + data name + data value". A diagram illustrating this step is shown below. Figure 12 As shown, since database writing operations are required, the hydrogen storage matrix, additives, chemical formula, minimum hydrogen absorption and desorption temperature, minimum hydrogen absorption and desorption pressure, and their performance values ​​and units of the hydrogen storage material extracted in step 7 need to be structured and stored in the hydrogen storage material database in a ternary manner, and then organized and archived using the Python openpyxl package.

[0061] Step 9, Multi-terminal storage: Since the data extracted from the paper needs to be stored on multiple terminals for subsequent big data machine learning of hydrogen storage materials, this patent proposes a multi-terminal storage strategy of "local storage + MySQL + online hydrogen storage material database". The storage strategy is as follows: Figure 13 As shown; based on this strategy, the automatically extracted ternary data such as hydrogen storage matrix, additives, hydrogen absorption / desorption temperature and pressure are organized and archived with the classified texts of other papers at different stages.

[0062] At this point, the entire process is complete.

[0063] The present invention can achieve the following technical effects:

[0064] 1. Improve the efficiency of hydrogen storage material researchers in collecting relevant papers;

[0065] 2. Improve the efficiency of hydrogen storage material developers in collecting relevant data;

[0066] 3. It is conducive to conducting big data machine learning on hydrogen storage materials, thereby improving the overall efficiency of material research and development.

[0067] Therefore, the present invention has the following advantages over the prior art:

[0068] 1. The database data source is innovative. The existing material database data comes from the university research group where the patent applicant is located or the internal data of the company. It has the characteristics of standardized terminology and easy data acquisition, but the amount of data obtained is small and cannot meet the needs of big data machine learning.

[0069] Currently, the extraction of a large amount of high-quality data contained in academic papers in related fields still relies on manual extraction due to inconsistent wording and grammatical styles among different papers. This results in the quality of the extracted data being greatly affected by the subjective factors of the extractor, and the quality of the extracted data is uneven, which increases the workload of subsequent database maintenance for this part of the data.

[0070] Therefore, this patent uses a database data source that is different from the existing technology, which can enrich the hydrogen storage material database under the existing database architecture, change the way of collecting paper data from manual to automated extraction by program, save a lot of time and manpower costs, and can also avoid human data extraction errors to a certain extent, thus improving the quality of data.

[0071] 2. The method of acquiring data is innovative. Existing related technologies have been applied to extract data from papers on high-temperature alloys in materials science, but their practical application in the field of database expansion is not effective. However, the method of constructing a hydrogen storage material database based on artificial intelligence technology proposed in this patent technology realizes the effective application of natural language processing technology in papers on hydrogen storage materials.

[0072] First, the current situation of the single data source of the materials database stems from the limitations of existing natural language processing technology in extracting text data from academic papers; papers from different disciplines have different proper nouns and performance names, so the data extraction methods that are successfully applied in one field do not work well when transferred to the application of other materials.

[0073] Secondly, hydrogen storage materials have few proper nouns and many common nouns with other materials, making them more suitable for this method of building a hydrogen storage material database based on artificial intelligence technology.

[0074] Therefore, this patent innovatively proposes a method for extracting data from papers on hydrogen storage materials to solve the above problems;

[0075] 3. This automated database expansion method has a deeper application in big data and machine learning, and can provide high-quality data for future big data machine learning performance prediction based on hydrogen storage materials.

[0076] The concept of materials genome engineering and the "machine learning + materials" model are currently very popular. Existing research has proven that it is possible to discover the relationship between performance by establishing a materials database. Applying this model to hydrogen storage materials will greatly improve performance.

[0077] Currently, big data and artificial intelligence technologies have played an important role in many research fields of materials science, but their application in hydrogen storage material research is still somewhat insufficient. The applicant's research group intends to make some progress in related fields and change the strategy of predicting the performance of hydrogen storage materials from the original experience-based judgment as the guiding direction and a large number of experiments as the work content to a strategy of machine learning prediction as the guiding direction and fewer but better experiments as the work content.

[0078] 4. This method establishes a portable model for extracting data from academic papers, which has high portability and can be applied to the research of other types of materials, such as multi-component alloys and high-entropy alloys. Attached Figure Description

[0079] Figure 1 This is a flowchart illustrating the overall workflow of the database auto-expansion method in this embodiment.

[0080] Figure 2 This is a flowchart of step 2 in the embodiment.

[0081] Figure 3 This is a detailed schematic diagram of step 2 in the embodiment;

[0082] Figure 4 This is a schematic diagram of the processing flow for step 3 of the embodiment;

[0083] Figure 5 This is a detailed schematic diagram of step 3 in the embodiment;

[0084] Figure 6 This is a flowchart of step 4 in the embodiment.

[0085] Figure 7 This is a detailed schematic diagram of step 4 in the embodiment;

[0086] Figure 8 This is a schematic diagram of Dijkstra's shortest path algorithm in step 7 of the embodiment;

[0087] Figure 9 This is a schematic diagram of the sequential path matching algorithm in step 7 of the embodiment;

[0088] Figure 10 This is a schematic diagram of the CoreNLP part-of-speech distance in step 7 of the embodiment;

[0089] Figure 11 This is a schematic diagram of the bidirectional noun order matching extraction in step 7 of the embodiment;

[0090] Figure 12 This is a detailed schematic diagram of step 8 in the embodiment;

[0091] Figure 13 This is a schematic diagram of the storage strategy in step 9 of the embodiment;

[0092] Figure 14 This is a detailed schematic diagram of step 9 in the embodiment. Detailed Implementation

[0093] The present invention will be further described in detail through embodiments and with reference to the accompanying drawings, but this is not intended to limit the scope of the invention.

[0094] Example 1

[0095] A method for constructing a hydrogen storage material database based on artificial intelligence technology, the overall workflow of which includes the automatic database expansion method is as follows: Figure 1 As shown, it includes the following 9 steps:

[0096] Step 1: Searching for and organizing papers;

[0097] Step 2: Downloading and archiving the paper;

[0098] Step 3: Text format conversion;

[0099] Step 4: Text preprocessing;

[0100] Step 5: Chemical Nominated Entity Recognition;

[0101] Step 6: Text classification;

[0102] Step 7: Text relation extraction;

[0103] Step 8: Tripartite archiving;

[0104] Step 9: Multi-terminal storage;

[0105] The specific implementation method of the above steps is as follows:

[0106] Step 1: Searching and organizing papers. Use academic search engines to obtain the names of papers related to hydrogen storage materials, and then organize them into the names of papers to be searched for use in Step 2. The search engines used in this embodiment include Baidu Scholar and Google Scholar.

[0107] Step 2, downloading and archiving papers: The papers that need further processing are obtained and archived through automated downloading based on the names of the papers to be searched compiled in Step 1.

[0108] The process of searching and archiving papers is as follows: Figure 2 As shown, the processing steps are as follows: inputting the paper title, web crawling, downloading the paper, and archiving the paper in various formats. The specific steps are as follows:

[0109] Step 2.1: After inputting the paper title, web crawling technology can be used to search for the corresponding paper title. However, when using academic search engines to obtain relevant information about papers, traditional manual operations cannot meet the requirements of big data machine learning. Therefore, this invention uses Python's Selenium automated script to simulate a browser in silent mode to replace traditional manual operations, thereby querying the results of the literature title to be searched and returning the paper information. Since the DOI number is the most useful information in the literature results, this invention selects the DOI number for subsequent processing.

[0110] Step 2.2, saving the above DOI number to the local Excel file, also involves a lot of repetitive operations. Manual operation cannot meet the requirements. Therefore, this invention uses Python's openpyxl to write to the Excel file. Then, the Requests command is used to call the applied Elsevier API to read the DOI number in the above Excel document and download the PDF document of the literature to be searched, so as to achieve simultaneous archiving and organization.

[0111] Taking the download and archiving process of the paper "An efficient mechanochemical synthesis of alpha-aluminumhydride…" as an example, the specific process is as follows: Figure 3 As shown, the specific steps are as follows:

[0112] First, the DOI number 10.1016 / j.jhazmat.2019.03.064 was obtained using Chromedriver's headless mode and archived in Excel.

[0113] Then, the paper is downloaded by calling the Elsevier API through Requests.

[0114] Finally, save the downloaded paper as a PDF to the corresponding folder, and the paper retrieval and archiving are complete;

[0115] Step 3, text format conversion: Since the paper downloaded and archived in Step 2 is in PDF format, and the text of the PDF paper cannot be directly processed in subsequent steps, it is necessary to perform text format conversion to convert the PDF paper document into Word and TXT format documents.

[0116] Meanwhile, this step also involves a large number of repetitive operations, which cannot be met manually. Therefore, this invention uses the Python Pdfminer_3k and Python-docx packages to convert PDF documents to Word documents, and then uses an external Win32 package to call WPS to convert the Word documents into TXT documents, which are named Raw_TXT. This achieves the organization and archiving of documents of the above formats. The text format conversion diagram is shown below. Figure 4 As shown, the converted text is archived locally by calling the system storage command;

[0117] Taking the text format conversion process of the paper "An efficient mechanochemical synthesis of alpha-aluminumhydride…" as an example, the specific process is as follows: Figure 5 As shown, its DOI number is 10.1016 / j.jhazmat.2019.03.064; the above paper is converted into Word format and then into TXT format and archived, thus completing the text format conversion of the paper;

[0118] Step 4, text preprocessing: Since there are easily confused words and a lot of irrelevant content in the text format conversion process, it is necessary to remove irrelevant content in the Raw_TXT after text format conversion in Step 3 by natural language processing methods.

[0119] Text preprocessing workflow as follows Figure 6 As shown, the specific steps are as follows:

[0120] Step 4.1: Perform common format conversions, replace easily misspelled words, truncate irrelevant content, remove non-English text, and strip merged words from the Raw_TXT file;

[0121] Step 4.2 requires segmenting the text after the above processing into sentences. Therefore, it is necessary to process the text for common stop words and then archive the segmented text in multiple formats, namely, archive it separately in Excel, Word and TXT formats.

[0122] Step 4.3: Perform word segmentation and content block extraction on the processed TXT document in sequence;

[0123] Step 4.4 involves word segmentation of the text. This is because the accuracy of chemical named entity recognition in subsequent steps depends on the accuracy of the pre-trained word vector model of the hydrogen storage material paper. Therefore, it is necessary to train word vectors on the segmented text to establish a word vector analysis model for subsequent database applications.

[0124] Step 4.5: The extracted content is then divided into sections such as introduction, keywords, abstract, main text, and citations, and these sections are archived for future use.

[0125] Taking the text preprocessing process of the paper "An efficient mechanochemical synthesis of alpha-aluminumhydride…" as an example, the specific process is as follows: Figure 7 As shown, the specific steps are as follows:

[0126] First, a list of commonly used conversion error words, an illustration of irrelevant content extraction, an illustration of non-English text removal, a list of commonly used words to be stripped and merged, and a list of commonly used stop words to be processed are provided.

[0127] Next, the process of word segmentation and sentence segmentation is demonstrated. A portion of the sentences is extracted from the Raw_TXT of the above paper and segmented into words and sentences. The content after word and sentence segmentation and the functions of the program used are listed in the figure.

[0128] Finally, the above paper is divided into blocks for extraction, and then the block-based abstract, keywords, introduction, and main text are archived separately, thus completing the text preprocessing;

[0129] Step 5, Chemical Named Entity Recognition: Since there is a lot of useless information in the preprocessed sentence text in Step 4, it will increase the amount of invalid work in subsequent relation extraction and ternary archiving. Relying on manual chemical named entity recognition will affect the accuracy of subsequent steps. Therefore, it is necessary to classify and filter the text by using a named entity recognition model adapted to hydrogen storage materials, and remove text that does not need to be extracted to reduce the invalid extraction in the future.

[0130] To address the aforementioned issues, the named entity recognition process in this patent incorporates methods from Chemdataextrator and SuperalloyDigger. The specific steps are as follows:

[0131] Step 5.1: Use regular expressions to expand and adapt the relevant performance values ​​and named entities of hydrogen storage materials to improve the accuracy of the chemical named entity recognition process for hydrogen storage materials.

[0132] Step 5.2: Use Word2vec and TF-IDF to train a word vector model to generate a high-frequency dictionary of chemical nominal entities for hydrogen storage materials, and apply this to subsequent steps;

[0133] Step 6, text classification: Since the text after text preprocessing in Step 4 contains invalid information, this patent selects sentence segments containing target information from the full text after sentence segmentation based on the named entity recognition results in Step 5, as the processing objects for subsequent relation extraction.

[0134] Step 7, text relation extraction. Although the text processed in Step 6 contains chemical named entities of hydrogen storage materials, it is still impossible to directly obtain the relationship between the various hydrogen storage material matrices, performance names, and performance values ​​in the sentences. Therefore, relation extraction is required. That is, the sentence relationship between various text parts in the text is extracted by natural language processing. The key to relation extraction is to search and match through part-of-speech relations.

[0135] Since traditional non-chemical text relation extraction is not effective in hydrogen storage materials, this patent proposes a text relation extraction method in this part. It mainly adopts the principle of limited relation extraction, and combines Dijkstra's shortest path algorithm, simple sequence matching algorithm and CoreNLP part-of-speech relation distance to extract relations.

[0136] The algorithm diagram of Dijkstra's shortest path algorithm is shown below. Figure 8 A schematic diagram of the simple sequence matching algorithm is shown below. Figure 9 A diagram illustrating the part-of-speech distance in CoreNLP is shown below. Figure 10 The specific steps for extracting text relations are as follows:

[0137] Step 7.1: Process the sentence segments containing target information after classification in Step 6. Since each word will have a sorting label to represent its position in the sentence after word segmentation, the position of the keyword to be extracted in the current sentence to be processed can be located, and the corresponding keyword label can be attached to it for subsequent processing.

[0138] Step 7.2: Use different search strategies for different keywords to extract to extract the corresponding data values;

[0139] Taking a portion of the text from the paper "An efficient mechanochemical synthesis of alpha-aluminumhydride…" as an example, bidirectional noun order matching was performed, and the part-of-speech relationships of the text are as follows: Figure 10 As shown in the diagram, its extraction schematic is as follows: Figure 11 As shown, the specific steps for processing the above text are as follows:

[0140] First, bidirectional sequential matching is performed based on keywords labeled with hydrogen storage matrix or additives. "MgH2" is a noun labeled with hydrogen storage matrix, and can be regarded as a keyword to be extracted.

[0141] Then, perform forward and backward sequential searches. The first word with a part-of-speech tag that is a noun and exists in the dictionary of similar words for data names can be regarded as the data name to be extracted, such as "temperature".

[0142] Next, the text with part-of-speech tags of numbers in the sentence is filtered, the location of the data value is obtained according to Dijkstra's shortest path, and the data value is extracted using the temperature and pressure regular expression after optimization and adaptation to hydrogen storage materials. The numerical value and unit of temperature and pressure are combined as a data value to be extracted.

[0143] Finally, the keywords, data names, and data values ​​are combined into a ternary list for output in subsequent steps.

[0144] Step 8, Ternary Archiving: Archive data in a ternary format of "keyword + data name + data value". A diagram illustrating this step is shown below. Figure 12 As shown, since database writing operations are required, the hydrogen storage matrix, additives, chemical formula, minimum hydrogen absorption and desorption temperature, minimum hydrogen absorption and desorption pressure, and their performance values ​​and units of the hydrogen storage material extracted in step 7 need to be structured and stored in the hydrogen storage material database in a ternary manner, and then organized and archived using the Python openpyxl package.

[0145] Step 9, Multi-terminal storage: Since the data extracted from the paper needs to be stored on multiple terminals for subsequent big data machine learning of hydrogen storage materials, this patent proposes a multi-terminal storage strategy of "local storage + MySQL + online hydrogen storage material database". The storage strategy is as follows: Figure 13 As shown; based on this strategy, the automatically extracted ternary data such as hydrogen storage matrix, additives, hydrogen absorption / desorption temperature and pressure are organized and archived with the classified texts of other papers at different stages.

[0146] Taking the papers numbered "010101", "010102", and "010103" in the database as examples, their multi-terminal storage results are as follows: Figure 14 As shown, the local storage portion uses server-side Excel for storage, the local area network storage portion uses MySQL for data storage, and the external network public storage portion uses an online hydrogen storage material database for data storage, with the URL http: / / www.hsmdatabase.cn / ;

[0147] At this point, the entire process is complete.

Claims

1. A method for constructing a hydrogen storage material database based on artificial intelligence technology, characterized in that... It includes the following 9 steps: Step 1: Searching and organizing papers. Use academic search engines to obtain the names of papers related to hydrogen storage materials, and then organize them into the names of papers to be searched for use in Step 2. The search engines used in this embodiment include Baidu Scholar and Google Scholar. Step 2, downloading and archiving papers: The papers that need further processing are obtained and archived through automated downloading based on the names of the papers to be searched compiled in Step 1. Step 3, text format conversion: Since the paper downloaded and archived in Step 2 is in PDF format, convert the PDF paper document to Word and TXT format documents. Step 4, text preprocessing: Remove irrelevant content from the Raw_TXT file after text format conversion in Step 3 using natural language processing methods; Step 5, Chemical Named Entity Recognition: The text is classified and filtered by a named entity recognition model adapted to hydrogen storage materials to remove text that does not need data extraction in order to reduce subsequent invalid extraction. Step 6, text classification: Based on the named entity recognition results in Step 5, select the sentence segments containing the target information from the full text after sentence segmentation, as the processing objects for subsequent relation extraction. Step 7, Text Relation Extraction: Using natural language processing methods, the sentence relationships between various text parts are extracted. The key to relation extraction is to search and match based on part-of-speech relationships. Step 8, Ternary Archiving: Archive the data in a ternary format of "keyword + data name + data value". Store the hydrogen storage matrix, additives, chemical formula, minimum hydrogen absorption / desorption temperature, minimum hydrogen absorption / desorption pressure and their performance values ​​and units extracted in Step 7 into the hydrogen storage material database in a structured and ternary format. Then, organize and archive the data using the Python openpyxl package. Step 9, multi-terminal storage: a multi-terminal storage strategy of "local storage + MySQL + online hydrogen storage material database"; based on this strategy, the automatically extracted ternary data of hydrogen storage matrix, additives, hydrogen absorption / desorption temperature and pressure are organized and archived with the classified texts of other papers at other stages. At this point, the entire process is complete.

2. The method for constructing a hydrogen storage material database based on artificial intelligence technology according to claim 1, characterized in that: The processing flow for querying and archiving papers in step 2 is as follows: paper title input, web crawling, paper download, and archiving of papers in various formats. The specific steps are as follows: Step 2.1: After inputting the paper title, web crawler technology can be used to search for the corresponding paper title. When using an academic search engine to obtain relevant information about the paper, the Selenium script in Python is used to simulate a browser in silent mode to replace traditional manual operation, thereby querying the results of the literature title to be searched and returning the paper information. Since the DOI number is the most useful information in the literature results, this invention selects the DOI number for subsequent processing. Step 2.2: When saving the above DOI number to the local Excel file, Python's openpyxl is used to write the Excel file. Then, the Requests command is used to call the requested Elsevier API to read the DOI number in the Excel file and download the PDF document of the literature to be searched, so as to simultaneously archive and organize the documents.

3. The method for constructing a hydrogen storage material database based on artificial intelligence technology according to claim 1, characterized in that: The specific method for text format conversion in step 3 is as follows: use the Python Pdfminer_3k and Python-docx packages to convert PDF documents into Word documents, then use the Win32 package to externally call WPS to convert the Word documents into TXT documents and name them Raw_TXT, thereby realizing the organization and archiving of the above-mentioned document formats, and using the system storage command to archive the format-converted text locally.

4. The method for constructing a hydrogen storage material database based on artificial intelligence technology according to claim 1, characterized in that: The specific steps of text preprocessing in step 4 are as follows: Step 4.1: Perform common format conversions, replace easily misspelled words, truncate irrelevant content, remove non-English text, and strip merged words from the Raw_TXT file; Step 4.2 requires segmenting the text after the above processing into sentences. Therefore, it is necessary to process the text for common stop words and then archive the segmented text in multiple formats, namely, archive it separately in Excel, Word and TXT formats. Step 4.3: Perform word segmentation and content block extraction on the processed TXT document in sequence; Step 4.4: Perform word segmentation on the text, train word vectors on the segmented text, and establish a word vector analysis model for the text, which will be used in subsequent database applications. Step 4.5: The extracted content is then divided into sections: introduction, keywords, abstract, main text, and citations. These sections are then archived for future use.

5. The method for constructing a hydrogen storage material database based on artificial intelligence technology according to claim 1, characterized in that: The specific steps of step 5, chemical nominate entity identification, are as follows: Step 5.1: Use regular expressions to expand and adapt the relevant performance values ​​and named entities of hydrogen storage materials to improve the accuracy of the chemical named entity recognition process for hydrogen storage materials. Step 5.2: Use Word2vec and TF-IDF to train a word vector model to generate a high-frequency dictionary of chemical nominal entities for hydrogen storage materials, and apply this to subsequent steps.

6. The method for constructing a hydrogen storage material database based on artificial intelligence technology according to claim 1, characterized in that: Step 7, text relation extraction, involves Dijkstra's shortest path algorithm, simple sequence matching algorithm, and CoreNLP's part-of-speech distance algorithm. The specific steps of text relation extraction are as follows: Step 7.1: Process the sentence segments containing target information after classification in Step 6. Since each word will have a sorting label to represent its position in the sentence after word segmentation, the position of the keyword to be extracted in the current sentence to be processed can be located, and the corresponding keyword label can be attached to it for subsequent processing. Step 7.2: Use different search strategies for different keywords to be extracted to extract the corresponding data values.

Images