Intelligent extraction and processing method and system for literature information in coating field and readable storage medium

Abstract

The invention provides an intelligent extraction and processing method and system for literature information in the coating field and a readable storage medium. The intelligent extraction and processing method comprises the steps that S1, digital material science literatures containing texts, tables and images in the coating field are obtained; s2, extracting structured description information from an unstructured text of the digital material science literature; s3, extracting structured performance data from the semi-structured table of the digital material science literature; s4, extracting digital value pairs from the curve graph and the histogram of the digital material science literature; s5, associating the structured description information, the structured performance data and the digital value pair to the same material entity record based on the public identifier, and performing cleaning processing; and S6, outputting a material entity record. According to the method, an end-to-end and collaborative multi-modal processing assembly line is constructed by fusing a LangChain framework, searching, enhancing and generating, refining cue word engineering and a special automatic image data extraction technology.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of materials science and artificial intelligence. It is based on Large Language Model (LLM), Retrieval-Augmented Generation (RAG), Prompt Engineering, and automatic image data extraction technology. It is a method for integrated and automated extraction, alignment, cleaning, and knowledge fusion of text, tables, and images (including curves, line graphs, histograms, etc.) in coating literature. In particular, it relates to a method, system, and readable storage medium for intelligent extraction and processing of literature information in the field of coating. Background Technology

[0002] The research and performance optimization of coating materials play a crucial role in industries such as aerospace, shipbuilding, electronics, and building materials. Researchers and engineers need to efficiently and accurately obtain complex coating-related data from a vast amount of academic literature, which is distributed in multiple modalities within the literature. 1) Unstructured text: Describe the polymer matrix system (such as resin type, monomer, curing agent), synthesis method, experimental design and mechanism analysis in the main text.

[0003] 2) Semi-structured tables: These tables centrally display systematic experimental data and performance results, such as the mechanical properties, electrochemical parameters, and flame retardant ratings of different formulation samples. They represent the section with the highest data density.

[0004] 3) Structured images: These images present continuous and highly accurate performance change trends in graphs, line graphs, and histograms, such as electrochemical impedance spectroscopy (EIS), Tafel curves, and stress-strain curves. These images carry quantitative information that is difficult to describe in detail in text and tables.

[0005] Currently, literature mining technologies for the coating field mainly face the following multimodal and integrated challenges: 1. The inherent challenges of extracting information from text and tables: Incomplete table parsing: Traditional methods or general large language models have low recognition rates for complex tables (such as those containing merged cells and multi-level headers), resulting in missing or misaligned key performance data.

[0006] Alignment between text and tables is difficult: There is a close logical relationship between the text description and the table content (such as “Sample A” in the text corresponding to the complete set of data in a certain row of the table), but existing methods lack effective automatic alignment strategies, resulting in a break in the context information.

[0007] 2. Long-standing gaps and bottlenecks in image data extraction: A large amount of high-value information in the literature exists in the form of images, which traditional text mining methods cannot process at all, resulting in a huge gap in information collection. Manually extracting data points from images (such as reading impedance values ​​at a specific frequency from EIS curves) is extremely inefficient and prone to subjective errors, and cannot meet the needs of large-scale data mining.

[0008] 3. Lack of multimodal information fusion and collaboration: This is the core challenge currently facing technology. Data in text, tables, and images do not exist in isolation, but rather corroborate and complement each other. For example, the text may describe a method for preparing a new coating, the table may list the adhesion strength of the coating under different conditions, and the stress-strain curve in the image may show its complete mechanical behavior.

[0009] Existing methods lack a unified framework capable of automatically associating, aligning, and integrating these three types of information into a single data entity (such as "sample A"), forming a comprehensive and consistent data record. This fragmentation severely limits the ability to construct high-quality datasets from the literature that can be used for machine learning and deep analysis.

[0010] Therefore, it is necessary to study an intelligent extraction and processing method, system, and readable storage medium for literature information in the field of coatings to address the shortcomings of existing technologies and solve or mitigate one or more of the above-mentioned problems. Summary of the Invention

[0011] In view of this, the present invention provides a method, system and readable storage medium for intelligent extraction and processing of literature information in the field of coating. It is an intelligent mining and processing method for literature in the field of coating materials that integrates multimodal information such as text, tables and images. The method constructs an end-to-end, collaborative multimodal processing pipeline by integrating the LangChain framework, search enhancement generation, refined prompt word engineering and dedicated automatic image data extraction technology.

[0012] On the one hand, this invention provides a method for intelligent extraction and processing of literature information in the field of coatings, which includes the following steps: S1: Obtain digital materials science literature in the field of coatings that includes text, tables, and images; S2: Extract structured descriptive information from the unstructured text of the digitized materials science literature; S3: Extract structured performance data from the semi-structured tables of the digitized materials science literature; S4: Extract digitized numerical pairs from the graphs and histograms of the digitized materials science literature; S5: Based on public identifiers, associate the structured description information, structured performance data, and digital numerical values ​​with the same material entity record and perform cleaning processing; S6: Output material entity record.

[0013] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein S2 specifically includes: obtaining a customized prompt word template for the field of materials science for the polymer matrix, filler, curing process and performance description in the unstructured text; determining structured fields from the unstructured text based on the customized prompt word template using a large language model; and outputting the structured fields as the structured description information in JSON format.

[0014] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein S3 specifically includes: performing structural reconstruction processing on the semi-structured table to fill blank rows with inherited content; obtaining table extraction prompts based on the reconstructed semi-structured table; determining discrete performance values ​​from the semi-structured table based on the table extraction prompts using a large language model; and outputting the discrete performance values ​​as the structured performance data in JSON format.

[0015] In addition to the aspects and any possible implementations described above, an implementation is further provided in which S4 specifically includes: separating the curve and histogram from the digitized materials science literature; performing coordinate axis calibration and data point tracking on the curve and histogram to obtain the digitized numerical pairs of X and Y coordinate values; and determining sample identification and test conditions from the captions of the curve and histogram through natural language processing.

[0016] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein S5 specifically includes: matching associated content from the structured description information, the structured performance data and the digitized numerical pairs using the sample name as the public identifier; merging the matched associated content into the same material entity record; and performing unit conversion and naming standardization processing on the same material entity record.

[0017] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein the customized prompt word template in S2 specifically includes: obtaining prompt words containing polymer matrix type, monomer and curing agent for coating type; obtaining prompt words containing type, composition and preparation method for filler and modifier; obtaining prompt words for corresponding performance parameters for mechanical properties, corrosion resistance and flame retardant properties respectively; and outputting the structured fields in JSON format from the large language model by constraining the few-shot example.

[0018] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein the structure reconstruction process in S3 specifically includes: converting the semi-structured table in XML format to HTML format; for blank rows with the same content in the first column of multiple consecutive rows, filling the blank rows with inherited content by inheriting the content of the first row; and applying the table extraction prompt words to the filled table to determine the discrete performance value.

[0019] 8. The intelligent extraction and processing method for literature information in the field of coatings according to claim 5, characterized in that, the cleaning process for the same material entity record in step S5 specifically includes: if there is a conflict between the structured performance data, the digitized numerical pair and the structured description information, the final value is determined according to the rule that table data takes precedence over image data over text data; multiple expressions of polymer types are unified into a standard name through a large language model; and unified standardization processing is performed on the time format and coating thickness.

[0020] In accordance with the aspects described above and any possible implementation, a system for intelligent extraction and processing of literature information in the field of coatings is further provided, the system comprising: The document acquisition module retrieves digital materials science literature in the field of coatings, including text, tables, and images. The text information extraction module extracts structured descriptive information from the unstructured text of the digitized materials science literature; The table information extraction module extracts structured performance data from the semi-structured tables of the digital materials science literature; The image data extraction module extracts digitized numerical pairs from the curves and histograms of the digitized materials science literature; Multimodal information alignment and cleaning module: Based on public identifiers, the structured description information, structured performance data and digital numerical values ​​are associated with the same material entity record and a cleaning process is performed; The information output module outputs material entity records.

[0021] In accordance with the aspects and any possible implementations described above, a computer-readable storage medium is further provided, the computer-readable storage medium including a program for intelligent extraction and processing of literature information in the field of coatings, wherein when the program for intelligent extraction and processing of literature information in the field of coatings is executed by a processor, the steps of the method for intelligent extraction and processing of literature information in the field of coatings are implemented.

[0022] Compared with the prior art, the present invention can achieve the following technical effects: Multimodal information fusion connects the data chains of text, tables, and images. Under a unified framework, this invention simultaneously incorporates key information from text, semi-structured tables, and curve-like images into its processing scope. It not only extracts traditional text descriptions but also automatically parses table fields and image data, enabling the complete acquisition of multi-dimensional information on "formulation-process-performance" in coating literature. Customized prompts and extraction templates for the coating field significantly improve accuracy. By designing fields for coating systems and engineering prompt words, and combining the output format constraints of few-shot examples, this invention can stably output standardized JSON results in highly complex material contexts. The accuracy in identifying fields such as coating material main category, monomer, curing agent, and filler is close to 100%, which is significantly better than general text extraction methods. High-quality data cleaning and standardization significantly reduce manual processing costs. By combining rule-based scripts with large language models to assist in cleaning, this invention can automatically complete tasks such as unit unification, naming standardization, and format standardization, greatly reducing the workload of manual inspection and correction, and improving the consistency and usability of the final dataset. Good scalability Although this invention has customized the prompt words and field system for the field of coatings, the overall process and system architecture are also applicable to relevant literature on other material systems (such as corrosion, alloy design, battery materials, etc.). It can be migrated and applied simply by adjusting the domain vocabulary and field definitions, and has good expansion space.

[0023] Of course, any product implementing this invention does not necessarily need to achieve all of the technical effects described above at the same time. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 is an example diagram of text information storage in Excel provided by an embodiment of the present invention; Figure 2 is an example diagram of table parsing results provided in an embodiment of the present invention; Figure 3 is an image information extraction diagram provided by an embodiment of the present invention; Figure 4 is an example diagram of corpus construction provided by an embodiment of the present invention; Figure 5 is a flowchart of the large model extraction module provided in an embodiment of the present invention. Detailed Implementation

[0026] To better understand the technical solution of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0027] It should be understood that the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0028] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0029] This invention provides an intelligent method for extracting and processing literature information in the field of coatings. This method targets academic literature commonly found in materials science, which includes text, tables, and images. It can accurately extract key information from three different modalities: unstructured, semi-structured, and image-based documents, and unify multi-source information to the same material entity, ultimately forming a structured material entity record that can be directly used for database construction or downstream analysis.

[0030] Figure 1 A schematic diagram of the overall method flow provided by an embodiment of the present invention is shown. Figure 1 As shown, the method mainly includes the following steps.

[0031] Step S1: Obtain digital materials science literature in the field of coatings that contains text, tables, and images.

[0032] Specifically, digital materials science literature is primarily in PDF format, but also includes some open-access papers in Word or HTML formats. The process begins by batch downloading PDF files of papers from target journals or conferences using a literature management system or web crawler. In one embodiment, targeted acquisition can be performed on journals focusing on polymer composites, anti-corrosion coatings, and flame-retardant materials, such as Progress in Polymer Science, Corrosion Science, and Polymer Degradation and Stability. The acquired documents typically include a cover, abstract, main text, references, and numerous tables, graphs, and other images. It should be noted that the acquired documents can be processed individually or in batches. In batch processing, the PDF files can be standardized in filename format to facilitate subsequent source tracing.

[0033] In one possible implementation, the acquisition step also includes format checking of the PDF file. If the file is encrypted or a scanned image, the scanned version is first converted into a searchable text PDF using an optical character recognition tool to ensure that the subsequent text extraction module can read the content normally.

[0034] Step S2: Extract structured descriptive information from the unstructured text of the digitized materials science literature using the text information extraction module.

[0035] The main body of digital materials science literature typically contains a large amount of narrative description, such as polymer matrix types, filler types, curing processes, and performance characteristics. This information is scattered throughout paragraphs in natural language and is a typical example of unstructured text. The core of the text information extraction module lies in using large language model prompt word templates deeply customized for the field of materials science to achieve high-precision field extraction.

[0036] Step S21: For the polymer matrix, filler, curing process and performance description in the unstructured text, obtain a customized prompt word template for the field of materials science.

[0037] In one embodiment, the customized prompt template is divided into multiple sub-templates, each corresponding to a different material component. For example, the prompt template for the polymer matrix includes key fields requiring the model to identify the main chain type (e.g., polyurethane, epoxy resin, polyurea), monomer name (e.g., isophorone diisocyanate IPDI, hexamethylenediamine), and curing agent type (e.g., aliphatic amine, aromatic amine). For the filler component, the prompt template requires the model to identify the filler's chemical composition (e.g., graphene oxide, carbon nanotubes, silicon dioxide), surface modification method (e.g., silane coupling agent treatment, plasma modification), and added mass fraction.

[0038] Specifically, a typical polymer matrix extraction prompt template could be designed as follows: You are a professional materials science information extraction expert. Please extract polymer matrix-related information for the coating system from the following paragraphs, strictly adhering to the JSON format. Fields should include: matrix type, main chain chemical name, main monomer 1, main monomer 2, curing agent type, specific name of curing agent, curing temperature, and curing time. If a field is missing, fill in null. Output only JSON, without any explanation. A few-shot example is then provided, such as three annotated paragraphs and their correct JSON output, to constrain the model's output format and accuracy.

[0039] Step S22: Determine structured fields from the unstructured text using a large language model based on the customized prompt word template.

[0040] In practice, the main text of the document is first segmented, typically inputting each segment sequentially into the large language model. For longer paragraphs, they can be further divided into fragments of 500-800 tokens each to avoid exceeding the context length limit. Each fragment is paired with a corresponding custom prompt word template for invocation. Currently available large language models include, but are not limited to, open-source or closed-source models with a context window of 128k or more.

[0041] For example, when processing a text describing "the formation of a polyurea coating by reacting isophorone diisocyanate (IPDI) with polytetrahydrofuran ether diol (PTMEG) to generate a prepolymer, followed by the addition of 3,5-dimethylthiotoluene diamine (DMTDA) as a curing agent, and curing at 80°C for 7 days," the model will output the following under cue word constraints: { "Matrix type": "Polyurea" Main chain chemical name: "Polyurea" "Main monomer 1": "Isophorone diisocyanate", "Main monomer 2": "Polytetrahydrofuran ether diol", "Curing agent type": "Aromatic amine" "Specific name of curing agent": "3,5-Dimethylthiotoluenediamine" "Curing temperature": "80℃" "Curing time": "7 days" } Step S23: Output the structured fields as the structured description information in JSON format.

[0042] The extracted results from all paragraphs are merged. If the same field appears multiple times, the previous text takes precedence or the value with the highest frequency is used. Finally, a complete structured description of the document in JSON format is formed. This JSON includes, but is not limited to, multiple major categories of fields such as polymer matrix, fillers and modifiers, curing process parameters, anti-corrosion performance description, mechanical performance description, and flame retardant performance description.

[0043] It should be noted that, in order to improve extraction accuracy, in a preferred embodiment, a few-shot example chain can be added to the prompt words, providing up to 5-7 high-quality annotated examples, covering academic writing with different expression styles, such as mixed Chinese and English, coexistence of abbreviations and full names, and different unit expressions, thereby significantly reducing the risk of illusion in large language models.

[0044] Step S3: Extract structured performance data from the semi-structured tables of the digital materials science literature using the table information extraction module.

[0045] In digital materials science literature, the most common form of performance data presentation is tables. These tables typically have complex characteristics such as multi-level headers, spanning rows and columns, individual rows located in the middle or at the bottom, and blank rows inheriting content from the preceding text, making them typical semi-structured data. Directly using general table recognition tools often leads to content misalignment or loss, thus requiring a dedicated table information extraction module.

[0046] Step S31: Perform a structure reconstruction process on the semi-structured table to fill the blank rows with inherited content.

[0047] First, export the tables from the PDF as XML or HTML. Most mainstream PDF parsing tools can preserve table structure tags. Then, reconstruct the structure to address the common phenomenon in literature tables of "multiple rows of performance data with the same content in the first column".

[0048] For example, a common table format has "Sample Number S1, S2, S3" in the first column, while subsequent rows have empty first columns. In reality, the sample numbers should be inherited from the first column. If this is not reconstructed, directly feeding the data into a large language model will cause the model to be unable to determine which sample the empty rows belong to.

[0049] In one embodiment, structural reconstruction specifically includes the following sub-steps: Step S71: Convert the semi-structured table in XML format into HTML format for easier subsequent parsing.

[0050] Step S72: For empty rows with multiple consecutive rows having the same content in the first column, fill the empty rows with inherited content by inheriting the content of the first row. Specifically, it traverses each row from top to bottom. If the content of the first column cell in the current row is empty and the cell above it is not empty, then fill the current row with the content of the nearest non-empty first column above it, until a new non-empty first column content is encountered. For cases where multi-level headers are merged, the upward search method is also used to fill the empty rows.

[0051] For example, in the original table, sample S1 corresponds to 5 rows of performance data. The first column of the first 4 rows is empty. After reconstruction, the first column of these 4 rows is filled with S1, thus forming a complete rectangular structure table.

[0052] Step S73: Apply the table extraction prompts to the filled table to determine the discrete performance values.

[0053] After reconstruction, the complete HTML table, along with specially designed table extraction prompts, is input into the large language model. Table extraction prompts typically require the model to output a JSON array on a sample-by-sample basis, with each sample containing performance values ​​such as tensile strength, elongation at break, adhesion, salt spray test time, LOI value, and corresponding units and test standards.

[0054] A typical prompt for table extraction is: "You are a professional materials performance data extraction expert." Below is a complete HTML table after structural reconstruction. Please extract the performance data for each sample number, outputting a JSON array format. Each object should contain the sample name and all performance parameters. If cells contain ±standard deviation, please retain them; if there are test conditions such as "ASTM D638," please place them in the corresponding `test_standard` field. Output only JSON.

[0055] By combining the above-mentioned structural reconstruction with prompt words, the accuracy rate can be increased from 60-70% with direct extraction to over 95%, especially when dealing with complex multi-span tables.

[0056] Step S32: Obtain table extraction prompts based on the reconstructed semi-structured table.

[0057] The table extraction prompts are not fixed, but dynamically select different sub-templates at runtime based on the actual number of columns in the table and the keywords in the table header. For example, if the table header contains terms related to "Salt spray test", the extraction template for corrosion resistance is automatically loaded; if it contains terms like "Tensile strength" or "Elongation at break", the extraction template for mechanical properties is loaded, thus achieving higher precision field mapping.

[0058] Step S33: Using a large language model, extract prompt words from the table to determine discrete performance values ​​from the semi-structured table.

[0059] After receiving the reconstructed HTML table and corresponding prompts, the large language model parses the content line by line, automatically identifies numerical values, units, test standards, and possible subscripts and superscripts, and finally outputs structured performance data in JSON format.

[0060] For example, for a table of mechanical properties containing 6 samples, the model output is a JSON array containing 6 objects, each object having fields such as: { Sample Name: EP-Zn-5 Tensile strength (MPa): 68.5 Tensile strength_standard deviation: 3.2, "Elongation at break (%)": 4.8 Test Standard: GB / T 1040.2-2006 } Step S34: Output the discrete performance values ​​as structured performance data in JSON format.

[0061] The extraction results from all tables are summarized to form the complete structured performance data JSON of the document, which maintains the same top-level key name as the structured description information JSON obtained in step S2, so as to facilitate subsequent alignment and merging.

[0062] Step S4: Extract digital numerical pairs from the curves and histograms of the digitized materials science literature using the image data extraction module.

[0063] A large amount of experimental data in digital materials science literature is presented in the form of images, such as tensile stress-strain curves, salt spray corrosion rate histograms, and thermogravimetric analysis curves. These images contain important digital information, but due to their unstructured nature, they cannot be directly parsed into numerical data by computers. The image data extraction module aims to extract key numerical pairs from these curves and histograms, and combine this with caption information to determine sample identification and test conditions.

[0064] Step S41: Separate the curves and histograms from the digitized materials science literature.

[0065] In one embodiment, the image content and text content in the document are first separated using a PDF parsing tool. Embedded images in PDF files can be directly extracted as independent image files; for scanned documents, page segmentation is required to distinguish image areas from text areas. After separation, the extracted images are classified to identify portions belonging to line graphs and histograms. Typically, line graphs have continuous line characteristics, while histograms have a distinct columnar structure; automatic classification can be achieved through edge detection and shape analysis techniques in image processing.

[0066] For example, when processing a paper on the mechanical properties of polymer composites, images containing tensile stress-strain curves and histograms showing the comparison of elongation at break of different samples can be extracted from the text. These images will be stored separately and labeled with the original paper page number and figure number for later reference.

[0067] Step S42: Perform coordinate axis calibration and data point tracking on the curve and histogram to obtain the digitized numerical pairs of X and Y coordinate values.

[0068] After separating the target image, the coordinate axes in the image need to be calibrated to determine the mapping relationship between the image pixel positions and actual values. Specifically, the X-axis and Y-axis positions in the image are first identified, usually by detecting straight line features and scale markings. Then, the scale values ​​and units on the coordinate axes are read, for example, "strain / %" for the X-axis and "stress / MPa" for the Y-axis, and the pixel distance corresponding to the scale intervals is recorded. In this way, a linear or non-linear mapping relationship from pixels to actual values ​​is established.

[0069] In one possible implementation, for tracking data points on a curve, contour tracking technology from image processing is used to extract a series of pixel positions along the curve path, and then the pixels are converted into actual numerical pairs according to the coordinate axis mapping relationship. For example, for a tensile stress-strain curve, multiple key points from the starting point to the fracture point can be extracted to form numerical pairs such as "strain 0.5, stress 10.2" and "strain 1.0, stress 20.5".

[0070] For histogram data extraction, the value represented by each bar is determined by identifying the bar height and the corresponding X-axis label. For example, in a histogram showing the tensile strength of different samples, data such as "Sample A, tensile strength 55.3" and "Sample B, tensile strength 62.7" can be extracted.

[0071] It should be noted that image resolution and scanning quality may affect extraction accuracy. Therefore, in a preferred embodiment, low-resolution images can be preprocessed, such as sharpening or noise removal, to improve the accuracy of data point tracking.

[0072] Step S43: Determine the sample identifier and test conditions from the captions of the curves and histograms using natural language processing.

[0073] The image itself does not contain contextual information such as sample name or test conditions; this information is usually found in the caption or text description below the image. In this step, key fields, such as sample number, test standard, and test environment, are extracted from the caption using natural language processing techniques.

[0074] For example, regarding the caption " Figure 3 The tensile stress-strain curve of sample EP-5%GO at room temperature, tested according to GB / T 1040.2-2006, can be extracted. The sample identifier is "EP-5%GO", the test condition is "room temperature", and the test standard is "GB / T 1040.2-2006". This information will be correlated with the digitized values ​​extracted from the image to form a complete record.

[0075] In one embodiment, if the figure caption information is incomplete, a further search can be conducted in the text for descriptive paragraphs related to the figure number to supplement the missing fields. This method, which combines figure caption and text search, can significantly improve the completeness of information extraction.

[0076] Step S5: The multimodal information alignment and cleaning module associates the structured description information, the structured performance data, and the digitized numerical values ​​with the same material entity record based on the public identifier and performs cleaning processing.

[0077] Through the aforementioned steps, structured information has been extracted from text, tables, and images, respectively. However, these pieces of information are not yet linked, making them difficult to use directly for materials database construction or performance analysis. The core of the multimodal information alignment and cleaning module lies in using common identifiers to unify information from different sources into the same material entity, and resolving data conflicts and format inconsistencies through cleaning processes.

[0078] Step S51: Using the sample name as the public identifier, match associated content from the structured description information, the structured performance data, and the digitized numerical pairs.

[0079] In digital materials science literature, sample names are typically unique identifiers used to distinguish different materials or experimental groups. For example, a sample name might be "EP-Zn-5" or "PU-GO-3%", appearing in text descriptions, tabular data, and image captions. In this step, the information extracted from the three modalities is matched using the sample name as the key.

[0080] Specifically, the process begins by extracting all sample names from the structured description information to form a sample list. Then, iterates through the structured performance data and digitized numerical pairs, searching for records with names matching the sample names in the list, and merging the matched entries into the same group. If there are inconsistencies in the spelling of sample names, such as "EP-Zn-5" and "EP-Zn5", corrections are made through string similarity matching or semantic understanding using a large language model.

[0081] For example, for the sample "PU-GO-3%", its polymer matrix is ​​"polyurethane" and its filler is "graphene oxide, 3%". Its tensile strength is "58.6 MPa" and its stress-strain curve key points are "strain 1.2, stress 30.5" extracted from the text. This information will be associated with the same record.

[0082] Step S52: Merge the matched associated content into the same material entity record.

[0083] After matching is complete, the information from different modalities is merged into a single complete material entity record according to a predefined field structure. This record is stored in JSON format and contains multiple field groups, including sample name, material composition information, preparation process parameters, performance data, and test conditions.

[0084] In one embodiment, duplicate or missing fields need to be handled during the merging process. For example, if the filler addition amount described in the text is "3.0%" while it is "3.2%" in the table, the final value needs to be determined through subsequent cleaning steps. The merged records provide a unified data view for subsequent data analysis.

[0085] Step S53: Perform unit conversion and naming standardization processing on the same material entity record.

[0086] Because the units and naming conventions of data in the literature may be inconsistent—for example, the unit of tensile strength may be "MPa" or "GPa", and the polymer name may be "polyurethane" or "PU"—standardization is required to ensure data consistency.

[0087] In one possible implementation, unit conversion is achieved through predefined unit conversion rules. For example, all tensile strength units are standardized to "MPa", and coating thickness units are standardized to "μm". For the standardization of polymer names and chemical compositions, the semantic understanding capabilities of a large language model are used to unify different expressions into standard names, such as unifying "polyurea" and "Polyurea" into "polyurea".

[0088] Step S81: If there is a conflict between the structured performance data, the digitized numerical pairs, and the structured description information, the final value shall be determined according to the rule that tabular data takes precedence over image data over text data.

[0089] During the multimodal information merging process, the same field may yield different values ​​from different sources. For example, the text describes the tensile strength of sample "EP-Zn-5" as "60 MPa," the table shows "58.6 MPa," and the image curve extracts a value of "59.2 MPa." To resolve conflicts, a priority rule is adopted: tabular data has the highest priority due to its highest degree of structure, followed by image data, and finally text data. Therefore, in the above example, the final value is "58.6 MPa."

[0090] Step S82: Unify the various expressions of polymer types into a standard name through a large language model.

[0091] Polymer names may have multiple expressions in the literature, such as "epoxy resin," "Epoxy," and "EP," and these differences can affect data retrieval and analysis. By leveraging the semantic understanding capabilities of a large language model, all relevant expressions are mapped to a standard name. For example, given the input "Epoxy coating with 5% ZnO," the model outputs the standard name "epoxy resin," while retaining additive information.

[0092] Step S83: Perform uniform standardization processing on the time format and coating thickness.

[0093] Time formats in the literature may be expressed as "24 h", "24 hours", "1 day", etc., which should be uniformly converted to the unit of "hours", such as "24 hours". Coating thickness may be expressed as "100 μm", "0.1 mm", etc., which should be uniformly converted to the unit of "μm", such as "100 μm". This standardization process can significantly improve the consistency and comparability of data, laying the foundation for subsequent material performance analysis.

[0094] Step S6: Output the material entity record.

[0095] After completing the extraction, alignment, and cleaning of multimodal information, a complete material entity record is finally output. This record is stored in JSON format and contains all key information extracted from the literature, such as sample name, material composition, preparation process, performance data, and testing conditions. The output record can be directly imported into a materials database to support materials design, performance prediction, or literature reviews.

[0096] In one embodiment, the output records can also be customized according to the needs of downstream applications. For example, if used to build a polymer coating database, the anti-corrosion and mechanical properties fields can be highlighted; if used for flame retardant material analysis, fields such as thermogravimetric analysis data and limiting oxygen index can be highlighted.

[0097] In one possible implementation, the output record also includes metadata information, such as the source of the document, DOI number, and extraction time, to facilitate data traceability and version management. This approach can effectively support the batch processing and long-term storage of large-scale literature data.

[0098] The following provides several specific embodiments for steps S4 and S5 to cover different scenarios and application needs in digital materials science literature processing.

[0099] In one embodiment, for the image data extraction module in step S4, the stress-strain curves in literature on the mechanical properties of polymer composite materials are considered for processing. First, all images are separated from the literature using a PDF parsing tool, and the curves are classified by shape features. Then, the curves are calibrated on the coordinate axes, identifying the X-axis as "strain / %" and the Y-axis as "stress / MPa," and determining the pixel distance corresponding to the scale intervals. Next, contour tracking technology is used to extract key points on the curves, forming a series of numerical pairs, such as "strain 0.2, stress 5.6," "strain 0.8, stress 18.9," etc. Finally, from the figure captions… Figure 4 The sample is identified as "PU-3%GO" from the tensile property curve of the sample PU-3%GO. Through the above steps, the complete digital information in the curve can be obtained.

[0100] In another embodiment, the image data extraction module in step S4 processes the histogram of salt spray test results from anti-corrosion coating literature. First, the histogram image is extracted, with the X-axis representing the "sample number" and the Y-axis representing the "corrosion rate / mm·year⁻¹". Then, the value corresponding to each bar height is extracted, such as "Sample EP-1, corrosion rate 0.12", "Sample EP-2, corrosion rate 0.08". Finally, the test conditions are extracted from the caption as "salt spray test time 1000 hours, test standard ISO9227". This method can accurately obtain performance comparison data from the histogram.

[0101] In one embodiment, for the multimodal information alignment and cleaning module in step S5, a document about polyurethane-based anti-corrosion coatings is processed. The text extracts the matrix of sample "PU-Zn-5" as "polyurethane," the filler as "zinc oxide, 5% addition," and the curing conditions as "80℃, 6 hours." The table extracts the tensile strength as "52.3 MPa" and the salt spray test time as "1200 hours." The image extracts the key points of the stress-strain curve as "strain 1.5, stress 35.2." Using "PU-Zn-5" as the common identifier, the above information is merged into a complete record. During the cleaning process, it is found that the salt spray test time described in the text is "1100 hours," which conflicts with the table data. According to the priority rule, the table data "1200 hours" is taken as the final value.

[0102] In another embodiment, for the multimodal information alignment and cleaning module in step S5, a document on epoxy resin-based flame-retardant coatings is processed. The matrix of sample "EP-FR-10" is extracted as "Epoxy," and the filler is "aluminum hydroxide, 10% addition," from the text; the limiting oxygen index is extracted as "28.5%" from the table; and the key points of the thermogravimetric analysis curve are extracted as "temperature 300℃, mass loss 5%." During the merging process, "Epoxy" is standardized to "epoxy resin" using a large language model. Simultaneously, the temperature unit in the thermogravimetric analysis is standardized to "℃," and the mass loss unit is standardized to "%." This standardization process ensures data consistency.

[0103] In one possible implementation, the multimodal information alignment and cleaning module in step S5 addresses inconsistencies in sample names within the literature. For example, the sample name might be "PU-GO-3wt%" in the text, "PU-GO-3%" in the table, and "PU / GO-3%" in the image caption. Through string similarity matching and semantic understanding, these three expressions are unified to "PU-GO-3%". Subsequently, the curing temperature "75℃" in the text, the tensile strength "48.9 MPa" in the table, and the elongation at break "5.2%" in the image are merged into the same record. This method effectively solves the problem of inconsistent sample name expressions in the literature.

[0104] In one embodiment, the output material entity records from step S6 are designed for applications involving the construction of a polymer coating database. The output JSON records highlight corrosion resistance fields, such as salt spray test time, corrosion rate, and coating thickness, while retaining mechanical property fields such as tensile strength and elongation at break. Each record includes a literature DOI number and extraction date for data traceability. This customized output method directly supports downstream database construction and performance screening.

[0105] In another embodiment, the output material entity records from step S6 are considered for application scenarios involving flame retardant material performance analysis. The output JSON records highlight fields such as thermogravimetric analysis data, limiting oxygen index, and combustion rate, while retaining material composition information such as polymer matrix type and flame retardant dosage. The records are exported in batches as CSV files for easy import into analysis software. This approach can meet the needs of researchers in specific fields.

[0106] In one possible implementation, the image data extraction module in step S4 processes the curves in low-quality scanned documents. Since scanned images may contain blur or noise, they are first preprocessed, including grayscale conversion, noise removal, and edge enhancement. Subsequently, enhanced edge detection technology is used to identify the coordinate axis positions, and manual correction is used to assist in determining the scale values. For example, when processing an older document on polyurea coatings, the image resolution is low, but after preprocessing and correction, key points of the stress-strain curves, such as "strain 0.3, stress 8.5" and "strain 1.1, stress 25.6," can still be extracted. This approach effectively addresses the challenges of low-quality images.

[0107] In one embodiment, the multimodal information alignment and cleaning module in step S5 handles the batch extraction of multiple documents. For 100 sample records extracted from 10 documents, initial matching is first performed using the sample name as the key. Then, for samples with possible duplicate names (e.g., "EP-1" in different documents may refer to different materials), the document DOI number is introduced as an auxiliary identifier to ensure the uniqueness of the match. During the cleaning process, all tensile strength units are standardized to "MPa", coating thickness units are standardized to "μm", and polymer names are standardized using a large language model. This batch processing method can significantly improve the efficiency of large-scale document data processing.

[0108] In one possible implementation, the image data extraction module in step S4 processes complex images containing multiple curves. For example, in a paper on the effect of different filler contents on the performance of polyurethane coatings, a single graph contains stress-strain curves for three different samples. First, different curves are distinguished using color segmentation technology. Then, key points of each curve are extracted, and corresponding sample identifiers, such as "PU-1%GO", "PU-3%GO", and "PU-5%GO", are extracted from the graph caption. The extraction result forms an independent list of numerical pairs for each curve. This method can handle the need for multi-object extraction in complex images.

[0109] In one embodiment, the multimodal information alignment and cleaning module in step S5 addresses inconsistencies in test conditions. For example, for sample "EP-Zn-10," the text describes the salt spray test conditions as "5% NaCl solution, 35℃," the table shows "5% salt solution, 35℃," and the image caption reads "Salt spray test, 35℃." Through semantic understanding using a large language model, these three descriptions are unified to "5% NaCl solution, 35℃," and recorded as the final test condition. This approach effectively solves the problem of inconsistent test condition descriptions.

[0110] In one possible implementation, the output material entity record from step S6 is designed for applications involving integration with existing material databases. The field structure of the output record is aligned with the field definitions of the target database; for example, the tensile strength field is named "tensile_strength_MPa," and the coating thickness field is named "coating_thickness_um." Simultaneously, the output record is directly transmitted to the database server via an API interface, avoiding the cumbersome manual import process. This approach allows for seamless integration into existing data processing workflows.

[0111] In one embodiment, the image data extraction module in step S4 handles cases where the histogram contains error bars. For example, in a literature on the adhesion of epoxy resin coatings, the histogram displays the adhesion values ​​of different samples and includes standard deviation error bars. First, the height of each bar is extracted as the average value, and then the upper and lower limits of the error bars are extracted as the standard deviation range, forming records such as "Sample EP-1, adhesion 5.2, standard deviation 0.3" and "Sample EP-2, adhesion 4.8, standard deviation 0.4". This method can completely preserve the statistical information in the image.

[0112] In one possible implementation, the multimodal information alignment and cleaning module in step S5 addresses the issue of missing data in the literature. For example, for sample "PU-FR-5," the text does not provide the curing temperature, the table does not provide the elongation at break, and the image does not provide the test standard. When merging records, missing fields are marked as "not provided," and the possible curing temperature range, such as "70-80℃," is inferred from the literature context during subsequent cleaning as a reference value. This approach maximizes the preservation of information integrity.

[0113] In one embodiment, the output material entity record from step S6 is considered for data visualization analysis applications. The numerical data contained in the output record can be directly used to generate stress-strain curves or performance comparison histograms, helping researchers intuitively understand differences in material properties. Simultaneously, the test condition field retained in the record can be used to filter performance data under specific conditions. This approach effectively supports the intuitive analysis and comparison of material properties.

[0114] In one possible implementation, the image data extraction module in step S4 processes complex graphs containing multiple axes. For example, in a paper on the thermal properties of polymer coatings, a single graph contains both thermogravimetric analysis (TGA) and derivative TGA curves, corresponding to two Y-axis respectively. First, the scale values ​​and units of the two Y-axis are identified. Then, key points of the two curves are extracted to form two sets of numerical pairs. Finally, sample identification and test conditions are extracted from the graph captions. This approach can handle the complex extraction requirements of multi-axis images.

[0115] In one embodiment, the multimodal information alignment and cleaning module in step S5 handles sample matching across different documents. For example, when processing a series of documents on polyurea coatings, it was found that multiple documents contained the sample "PU-5%SiO2," but the performance data and test conditions differed. When merging records, a document source field was introduced to distinguish the test results from different documents and to indicate the differences in test conditions, such as "Document A, 1000 hours of salt spray test" and "Document B, 1500 hours of salt spray test." This approach supports cross-document material performance comparison analysis.

[0116] Example 1: Text Information Extraction like Figure 4 and Figure 5 As shown, this module is based on a large language model and RAG architecture, specifically optimized for the coating field. The process is organized using the LangChain framework, combining semantic vector retrieval with prompt-guided generation to accurately extract structured information from the literature text. Specialized prompts are refined prompt templates designed specifically for the characteristics of coating materials, explicitly requiring the model to extract key fields such as polymer matrix system, fillers and modifiers, coating substrate, and curing process. Data priority rules are built into the prompts, guiding the model to prioritize higher-precision data sources (e.g., tables > images > text) when conflicts exist, laying the foundation for subsequent multimodal alignment.

[0117] Each task has a corresponding prompt template, some examples of which are shown in Table 1. These prompts explicitly instruct the LLM to extract specific information from the input article fragment. The generated long text is then passed to the corresponding LLMChain instance to perform the specific information extraction task. The LLMChain combines the predefined prompt template with the knowledge extraction and uses few shots to ensure that the large model outputs data in JSON format.

[0118] Task Module Prompt words Coating type Please extract details about the coating system including polymer matrix type, monomer / pre-polymer, curing agent, and inhibitor loading method. Fillers and Modifiers Please extract details about fillers and modifiers including their type, composition, and preparation method. Coating substrate Please extract details about the coating substrate including its type, grade,elemental composition, and surface pre-treatment. Curing process Please extract details about the coating curing process including curingmechanism, schedule, post-curing conditions, and total curing time. Sample preparation Please extract details about sample preparation including dry coatingthickness and application method. Mechanical properties Please extract details about mechanical properties including tensile strength,adhesion strength, hardness, and self-healing efficiency. Corrosion resistance Please extract details about anticorrosion performance including corrosivemedium, test duration, EIS parameters, and potentiodynamic polarizationresults. antifouling performance 请提取有关防污性能的详细信息,包括测试的污损生物、抗菌率和防污机制。 Flame retardancy 请提取有关阻燃性的详细信息,包括极限氧指数(LOI)、热释放速率峰值(PHRR)、总热释放量(THR)和点火时间。 Surface properties 请提取有关表面性能的详细信息,包括水接触角和表面能特性。 Other functional properties 请提取有关其他功能特性的详细信息,如超疏水性、疏冰性或抗静电性能。 Table 1 The generated long text is passed to the corresponding LLMChain instance to perform specific information extraction tasks. LLMChain combines predefined prompt word templates for knowledge extraction and uses few-shot examples to ensure that the output of the large model is in a stable JSON format.

[0119] This invention designs an automated batch task workflow for question-and-answer structured extraction based on the ZhipuAI platform. Each text content is combined with a prompt word and placed into a request, constructed in a standardized JSON request format, and processed uniformly through the batch API interface provided by ZhipuAI. Each batch task can upload up to 100 requests, with an average processing time of 30 minutes, and multiple tasks can be processed simultaneously. The text information is extracted in batches, and an example of storing the text information in an Excel file is shown below. Figure 1 As shown.

[0120] Extracting information from tables This invention designs an independent table information extraction process for accurately extracting corrosion efficiency data from structured tables in literature. The overall process includes a table parsing stage and a large language model structure extraction stage.

[0121] First, the XML file is converted to a simplified HTML format so that tools like BeautifulSoup can be used to extract the table row by row. To address potential structural issues in the table, such as "inherited blank rows"—that is, when multiple consecutive rows have the same content in the first column, only the name is displayed in the first row, with the remaining rows left blank—the specific situation is as follows: Figure 2 As shown, a dedicated table reconstruction algorithm was designed to ensure the integrity and accuracy of the data.

[0122] We designed a dedicated table to extract prompts, and also used a few-shot technique to ensure that the output of the large model was in our specified JSON format. Then, we submitted a batch task, downloaded the JSON results, and used a script to extract the data into a unified Excel spreadsheet.

[0123] Image information extraction like Figure 3 As shown, this invention integrates the Auto_fig_extractor project to build an automated image data extraction process, specifically designed for extracting digital data from graphs, line charts, and histograms in literature.

[0124] Image separation: First, locate and separate the chart images from the literature.

[0125] Data digitization: Using the algorithms provided by Auto_fig_extractor, the image is preprocessed, the coordinate axes are calibrated, and the data points are tracked, and finally the original (X, Y) value pairs of the data points are output.

[0126] Caption Analysis: Simultaneously, NLP technology is used to analyze the captions, extracting information such as sample identification, test conditions, and performance parameter names from the images, providing context for subsequent data alignment.

[0127] Multimodal information alignment and cleaning Information alignment: Based on common identifiers (such as sample name and DOI number), information extracted from text, tables, and images is associated and matched. For example, the formulation of "sample A" described in the text, the hardness value of "sample A" in the table, and the stress-strain curve data of "sample A" in the image are automatically integrated into the same data record.

[0128] Intelligent data cleaning: To address the diversity of coating data, a strategy combining rules and models is adopted.

[0129] Regularized data cleaning (Python script): For data with clear patterns, such as unit conversion (e.g., unifying "μm" to "mm"), and time format standardization, Python scripts are written for automated processing.

[0130] Model-assisted cleaning (large language model): For irregular or complex data requiring semantic understanding, special prompt words are designed, and a large language model is invoked for cleaning and standardization. For example, for multiple expressions of polymer types (such as "EP", "Epoxy resin", "epoxy resin"), the prompt words guide the model to unify them into the standard name "Epoxy".

[0131] This invention uses these four modules in common. The accuracy of this invention will be described below.

[0132] To verify the effectiveness of the proposed information extraction strategy, this study constructed a manually annotated validation set using 100 samples from real-world literature. The extraction accuracy of the text-based RAG extraction path and the table-based question-answering extraction path were evaluated. The evaluation used field-level matching as the basic standard, along with statistical metrics such as precision, recall, and F1 score.

[0133] (1) Precision: The proportion of fields that the model predicts correctly out of the total number of predictions. It is used to measure the "precision" of the extracted results.

[0134] (2) Recall: The proportion of correct fields that the model successfully predicted out of all fields that should be extracted. It is used to measure the "completeness" of the extraction results.

[0135] (3) F1 score: the harmonic mean of precision and recall, which is a comprehensive evaluation of the overall performance of the model.

[0136] In this context, TP represents correctly extracted fields (True Positive), FP represents incorrectly extracted fields (False Positive), and FN represents fields that were not extracted (False Negative).

[0137] To ensure the comprehensiveness of the evaluation, this invention performs field-level comparisons on each record in the structured output, verifying whether the field names, values, and units are consistent with the manual annotations.

[0138] The evaluation results show that the accuracy of each module is approximately 0.9019, the recall is approximately 0.9665, and the F1 score is approximately 0.9325, indicating that the model's recognition performance is good. The accuracy of identifying the main material categories, curing agents, monomers, and fillers is very high, almost 100%. However, when identifying specific proportions, there may be instances where prompts do not cover areas where identification errors occur. Performance and process identification is generally accurate. Some issues exist, such as antifouling performance; the test methods are not mentioned in the main text, and quantitative data on antifouling is much less than that on anticorrosion, making it difficult to extract numerical data; only test methods and specific effects can be extracted, without comparative analysis.

[0139] This invention addresses the challenges of extracting, aligning, and fusing multimodal data in coating literature by constructing an end-to-end processing pipeline encompassing text extraction, table extraction, image data extraction, multimodal information alignment, and cleaning. By integrating the LangChain framework, Retrieval Enhancement Generation (RAG), refined prompt engineering, and dedicated automatic image data extraction technology, it achieves integrated and automated extraction, alignment, cleaning, and knowledge fusion of multimodal information (text, tables, and images) from coating literature. Verification shows that this method achieves an accuracy of 0.9019, a recall of 0.9665, and an F1 score of 0.9325. The accuracy in identifying core fields such as coating material categories approaches 100%, significantly improving the completeness, accuracy, and efficiency of scientific data mining in the coating field, and exhibiting good cross-material system scalability.

[0140] Overall, the model performs well in text mining tasks, but there is still room for improvement in performance recognition.

[0141] The above provides a detailed description of a method, system, and readable storage medium for intelligent extraction and processing of literature information in the field of coatings, as well as embodiments of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas; furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

[0142] Certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that hardware manufacturers may use different names to refer to the same component. This specification and claims do not distinguish components based on differences in name, but rather on differences in function. The terms "comprising" and "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising / including but not limited to". "Approximately" means that within an acceptable margin of error, those skilled in the art can solve the technical problem and substantially achieve the technical effect within a certain margin of error. The following descriptions in the specification are preferred embodiments for carrying out this application; however, these descriptions are for the purpose of illustrating the general principles of this application and are not intended to limit the scope of this application. The scope of protection of this application shall be determined by the appended claims.

[0143] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a product or system comprising a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a product or system. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the product or system that includes said element.

[0144] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0145] The foregoing description illustrates and describes several preferred embodiments of this application. However, as previously stated, it should be understood that this application is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the application concept described herein through the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of this application should be within the protection scope of the appended claims.

Claims

1. A method for intelligent extraction and processing of literature information in the field of coatings, characterized in that, The intelligent extraction and processing method for literature information in the field of coatings includes the following steps: S1: Obtain digital materials science literature in the field of coatings that includes text, tables, and images; S2: Extract structured descriptive information from the unstructured text of the digitized materials science literature; S3: Extract structured performance data from the semi-structured tables of the digitized materials science literature; S4: Extract digitized numerical pairs from the graphs and histograms of the digitized materials science literature; S5: Based on public identifiers, associate the structured description information, structured performance data, and digital numerical values ​​with the same material entity record and perform cleaning processing; S6: Output material entity record.

2. The intelligent extraction and processing method for literature information in the field of coatings according to claim 1, characterized in that, S2 specifically includes: obtaining a customized prompt word template for the field of materials science for the polymer matrix, filler, curing process and performance description in the unstructured text; determining structured fields from the unstructured text based on the customized prompt word template using a large language model; and outputting the structured fields as the structured description information in JSON format.

3. The intelligent extraction and processing method for literature information in the field of coatings according to claim 1, characterized in that, S3 specifically includes: performing structural reconstruction processing on the semi-structured table to fill blank rows with inherited content; obtaining table extraction prompts based on the reconstructed semi-structured table; determining discrete performance values ​​from the semi-structured table using a large language model based on the table extraction prompts; and outputting the discrete performance values ​​as structured performance data in JSON format.

4. The intelligent extraction and processing method for literature information in the field of coatings according to claim 1, characterized in that, S4 specifically includes: separating curves and histograms from the digitized materials science literature; performing coordinate axis calibration and data point tracking on the curves and histograms to obtain the digitized numerical pairs of X and Y coordinate values; and determining sample identification and test conditions from the captions of the curves and histograms through natural language processing.

5. The intelligent extraction and processing method for literature information in the field of coatings according to claim 1, characterized in that, S5 specifically includes: matching related content from the structured description information, the structured performance data, and the digitized numerical pairs using the sample name as the public identifier; merging the matched related content into the same material entity record; performing unit conversion and naming standardization processing on the same material entity record; and cleaning the same material entity record after performing unit conversion and naming standardization processing.

6. The intelligent extraction and processing method for literature information in the field of coatings according to claim 2, characterized in that, The customized prompt word template in S2 specifically includes: obtaining prompt words containing polymer matrix type, monomer and curing agent for coating type; obtaining prompt words containing type, composition and preparation method for filler and modifier; obtaining prompt words for corresponding performance parameters for mechanical properties, corrosion resistance and flame retardancy; and constraining the large language model to output the structured fields in JSON format through few-shot examples.

7. The intelligent extraction and processing method for literature information in the field of coatings according to claim 3, characterized in that, The structural reconstruction process in S3 specifically includes: converting the semi-structured table in XML format into HTML format; filling empty rows with the same content in the first column by inheriting the content of the first row; and applying the table extraction prompt words to the filled table to determine the discrete performance value.

8. The intelligent extraction and processing method for literature information in the field of coatings according to claim 5, characterized in that, The cleaning process performed on the same material entity record in S5 specifically includes: if there is a conflict between the structured performance data, the digitized numerical pairs, and the structured description information, the final value is determined according to the rule that tabular data takes precedence over image data over text data; multiple expressions of polymer types are unified into a standard name through a large language model; and unified standardization processing is performed on the time format and coating thickness.

9. A system for intelligent extraction and processing of literature information in the field of coatings, based on the method for intelligent extraction and processing of literature information in the field of coatings as described in any one of claims 1-8, characterized in that, The intelligent extraction and processing system for literature information in the coating field includes: The document acquisition module retrieves digital materials science literature in the field of coatings, including text, tables, and images. The text information extraction module extracts structured descriptive information from the unstructured text of the digitized materials science literature; The table information extraction module extracts structured performance data from the semi-structured tables of the digital materials science literature; The image data extraction module extracts digitized numerical pairs from the curves and histograms of the digitized materials science literature; Multimodal information alignment and cleaning module: Based on public identifiers, the structured description information, structured performance data and digital numerical values ​​are associated with the same material entity record and a cleaning process is performed; The information output module outputs material entity records.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a program for intelligent extraction and processing of literature information in the field of coatings. When the program is executed by a processor, it implements the steps of the method for intelligent extraction and processing of literature information in the field of coatings as described in any one of claims 1 to 8.

Images