Perovskite literature data intelligent extraction method and system based on OCR and LLM

Abstract

The present application relates to the technical field of document data processing and intelligent information extraction, and proposes a perovskite document data intelligent extraction method and system based on OCR and large language model. The method first accurately screens documents through a double-layer scoring mechanism of title pre-verification and PDF text secondary verification; then separates the graphic and text areas through OCR layout analysis, guides the multi-modal visual language model to deeply analyze scientific images and integrate them into text through structured prompts, and generates integrated graphic and text documents; then, based on the large language model, the complete chemical composition is extracted through the double strategy of direct matching of chemical formula and indirect inference of molar ratio, the information loss is reduced through key field fingerprint deduplication and relaxed retention strategy, and finally the structured data table is output. The present application combines multi-modal visual understanding and large model semantic reasoning, realizes the full-process automatic extraction of multi-source heterogeneous information, and has higher extraction accuracy and robustness in complex scientific research literature scenarios.

Description

Technical Field

[0001] This invention relates to the field of literature data processing technology, specifically to a method and system for extracting perovskite literature data based on multi-model collaboration. Background Technology

[0002] In recent years, third-generation photovoltaic technology, represented by perovskite solar cells, has made significant progress in the fields of materials science and new energy. Along with the continued surge in research interest, the number of related academic papers has exploded. These papers contain a wealth of crucial information, including material chemical composition, device structure design, fabrication process parameters, and photoelectric conversion efficiency, which is of great value in advancing the field. However, how to achieve efficient and accurate data extraction from the massive amount of perovskite literature remains a critical problem that urgently needs to be solved.

[0003] While existing technologies have made some progress in document processing and information extraction, they still have significant shortcomings when dealing with complex, multimodal perovskite document scenarios. In the document acquisition stage, traditional web crawling methods lack effective semantic filtering and content verification mechanisms, making it difficult to guarantee the domain relevance of the acquired documents, thus leading to a waste of storage and computing resources. In the document parsing stage, existing PDF parsing and OCR technologies have limited capabilities when handling complex layouts, multi-column structures, and embedded charts, often resulting in incomplete or even missing key information extraction.

[0004] Furthermore, perovskite literature contains a wealth of image data with high information density, such as scanning electron microscope images, XRD patterns, and device performance curves. These images carry important structural and performance information. However, existing methods mostly focus on textual information processing, with insufficient analysis and utilization of image data, limiting the completeness and depth of overall information extraction. Finally, in the data extraction stage, traditional methods often rely on rule templates or keyword matching strategies, lacking a deep understanding of contextual semantics. This makes it difficult to adapt to the diversity of expression and structure in perovskite literature, resulting in the need to improve the accuracy and robustness of the extraction results.

[0005] Patent CN121860795A discloses a method for multi-source data knowledge extraction in the entire audit process. It uses a semantic model and process dependency graph in the audit domain to extract knowledge triples and aggregate cross-process associations. Primarily targeting financial audit business scenarios, it addresses issues such as fragmented knowledge and insufficient consistency within the process. However, this solution focuses on text-based business data and time-series process processing, and does not address complex layout parsing of scientific literature PDFs, multimodal image and text information fusion, or structured parameter extraction in materials science. This differs significantly from the technical requirements for perovskite scientific literature.

[0006] Patent CN121882037A discloses a method for constructing a large-scale knowledge extraction model in the military field. It optimizes model performance through phased training, dynamic data matching, and reinforcement learning. The core lies in the training and iteration of the large-scale model in the field. It is only for processing pure text data in the military field and does not have the capabilities of automated document collection and screening, PDF multimodal parsing, chemical formula intelligent extraction, and structured output of scientific research data. It cannot adapt to the characteristics of perovskite documents, which are characterized by mixed text and images, complex structure, and high information density.

[0007] Compared to existing technologies, current large-model-based text information extraction methods primarily target single text data, lacking the ability to process complex layout structures and multimodal information such as images and charts. Furthermore, some multimodal document processing methods rely heavily on rule-driven approaches or module splicing, resulting in limited semantic understanding and cross-modal fusion capabilities. This invention integrates OCR parsing, multimodal visual language models, and large language models to construct a unified semantic modeling and segmented reasoning mechanism. This enables collaborative parsing and structured extraction of textual and graphical information from perovskite documents. Multi-stage verification and data retention strategies enhance data integrity and system robustness, thereby achieving higher extraction accuracy and applicability in complex scientific literature scenarios.

[0008] Therefore, there is an urgent need to develop an efficient and intelligent information extraction method for perovskite literature to improve the comprehensive processing capability of multi-source heterogeneous data, thereby better supporting knowledge mining and scientific research in this field. Summary of the Invention

[0009] This invention aims to overcome the problem of insufficient efficiency in document processing and information extraction in existing technologies, and proposes an intelligent extraction method and system for perovskite document data based on OCR and Large Language Model (LLM).

[0010] This invention addresses the shortcomings of existing technologies in document crawling, PDF parsing, multimodal information utilization, and data extraction accuracy by constructing a multi-model collaborative processing framework that integrates OCR technology and large language model capabilities. By introducing a semantic filtering mechanism to improve document retrieval efficiency, and combining high-precision document parsing with multimodal information understanding capabilities, it achieves unified modeling and deep mining of text and image data, thereby significantly improving the completeness and accuracy of data extraction with minimal human intervention.

[0011] The first aspect of this invention relates to an intelligent extraction method for perovskite literature data based on OCR and LLM, comprising the following steps: S1. Based on academic database API, literature retrieval and intelligent screening are performed. By constructing a two-layer verification mechanism that integrates title pre-verification and PDF content secondary verification, and combining multi-dimensional scoring of perovskite keywords and exclusion keywords, the system automatically filters and downloads PDF files of academic literature in the field of perovskite solar cells. Based on the verification results, the literature is stored in the valid directory and the rejection directory. Key parameter information of perovskite solar cells is collected, including chemical composition ratio, cell architecture, material information and performance parameters. S2. Perform OCR layout analysis on the collected PDF documents, separate text and image regions, extract Markdown-formatted text content, and download image resources to the local machine; load a multimodal visual language model, and use structured analysis prompts designed for scientific paper images to guide the model to perform in-depth semantic analysis on each image from multiple dimensions, including image type recognition, key visual feature extraction, data measurement value analysis, annotation information recognition, and scientific conclusion summary, generating image semantic description text; insert the image semantic description text into the Markdown document in a structured block format below the corresponding image position, forming a structured document that integrates text and image semantics; S3. Based on a large language model, segmented reasoning and information extraction are performed on the integrated text and image structured document: The long text Markdown document is divided into overlapping continuous text segments, and the large language model is called for semantic parsing for each segment; a multi-level data field system covering composition ratios, fixed attributes, and performance parameters is constructed; a dual extraction strategy combining direct matching and indirect inference of chemical formulas is adopted to obtain the complete chemical composition. When the large language model fails to extract a chemical formula, the result of the automatic chemical formula extraction module is automatically called to supplement it and the element composition ratio is synchronously backfilled; a unique fingerprint identifier is constructed by combining key fields to achieve cross-segment data deduplication, and a lenient data retention strategy is adopted, which retains the record as long as the chemical formula or any material process information exists, without requiring the completeness of performance data, and finally outputs a structured data table.

[0012] More preferably, step S1 specifically includes: Step 1.1: Configure the search keyword set, which includes core terms in the perovskite field, including perovskite, solar cell, photovoltaic, MAPbI, FAPbI, CsPbI and their combinations. At the same time, configure the exclusion keyword set to filter literature in non-target fields. Step 1.2: Call the academic database API to obtain the literature metadata, which includes DOI, title, author, journal, publication year, and abstract information. Perform the first layer of verification on the obtained literature—keyword pre-verification at the title and abstract level. If the title or abstract contains excluded keywords, it will be directly removed. If the perovskite keywords are found, it will be retained as a candidate literature. Step 1.3: Construct a content validator to perform a second layer of verification on the downloaded PDF file—extract the text content of the first few pages through document parsing, count the total number of hits of perovskite keywords and the total number of hits of excluded keywords, and calculate the verification score. When the score of perovskite keywords reaches the set threshold and the score of excluded keywords is lower than the set threshold, it is judged as a valid perovskite document. Step 1.4: Save the verified PDF files to the specified directory, record the literature metadata to an Excel spreadsheet, and move the PDF files that fail verification to the rejection directory to ensure that subsequent processing only applies to valid perovskite literature; Step 1.5: Collect key parameter information for perovskite solar cells from relevant literature, including chemical composition ratios, cell architecture, material information, and performance parameters. More preferably, step S2 specifically includes: Step 2.1: Call the OCR service to perform layout analysis and text recognition on the PDF file, obtain the Markdown format text content of each page, and at the same time obtain the image resource links and layout detection images to achieve automatic separation of text areas and image areas; Step 2.2: Download the identified image resources to the local output directory and establish a mapping record between image paths and corresponding pages; Step 2.3: Load the multimodal visual language model and use structured analysis prompts designed for the characteristics of images in scientific papers. The prompts guide the model to perform deep semantic analysis on the image from the following dimensions: (a) Image type recognition, distinguishing between charts, microscopic images, schematic diagrams, and molecular structure types; (b) Key visual element extraction, identifying the main components and contrast relationships in the image; (c) Data and measurement value analysis, extracting numerical values, scales, coordinate axis labels, and data trends; (d) Label information recognition, extracting text labels, legends, color codes, and highlighted areas; (e) Scientific conclusion summary, summarizing the scientific findings and key conclusions supported by the image. Step 2.4: Perform multimodal analysis on each image sequentially to generate semantic description text, save the analysis results to a JSON file, and perform a memory cleanup operation after each analysis. Step 2.5: Scan the image reference marks in the Markdown document line by line, and insert the corresponding image semantic description text into the original image position in a reference block format to form a structured document that integrates text semantics and image semantics, and save it.

[0013] More preferably, step S3 specifically includes: Step 3.1: Define the data extraction field system for perovskite solar cells, including composition ratio fields, fixed attribute fields, and performance parameter fields. The composition ratio fields cover FA ratio, MA ratio, Cs ratio, Pb ratio, Sn ratio, I ratio, Br ratio, and Cl ratio. The fixed attribute fields cover cell architecture, complete chemical formula, electron transport layer material, hole transport layer material, antisolvent type, buried interface passivator, perovskite layer additives, interface passivator, processing method, annealing temperature, and annealing time. The performance parameter fields cover photoelectric conversion efficiency, open-circuit voltage, short-circuit current density, fill factor, effective area, and data source. Step 3.2: Construct an automatic chemical formula extraction module and adopt a dual extraction strategy to obtain the complete chemical composition: The direct matching strategy uses regular expressions to sequentially match chemical formulas in LaTeX format, standard format, and nested bracket format, and cleans and removes duplicates from the extraction results; The indirect inference strategy, when the precursor solution formulation only provides the mass ratio of each component, converts the mass ratio to the molar ratio based on the standard molecular weight of each component, and then infers the cation ratio and constructs the complete chemical formula according to the perovskite chemical formula specification. Step 3.3: The integrated text and image long text Markdown document generated in S2 is divided into continuous text segments with a preset overlap length. The large language model is called for data extraction for each segment. The system prompt words guide the model to extract data according to priority. Chemical formulas and material process parameters are mandatory priority extraction items, and performance parameters are conditional extraction items. Step 3.4: Post-process the JSON results output by the large language model, including format parsing, data deduplication, and field validation; construct a unique fingerprint identifier by combining chemical formulas, battery architecture, material information, and data source to deduplicate duplicate data across segments; when the large language model fails to extract a chemical formula, automatically call the results of the automatic chemical formula extraction module described in Step 3.2 to supplement it, and simultaneously calculate the corresponding elemental composition ratio and fill it into the corresponding fields; Step 3.5: Employ a lenient data retention strategy. For each extracted record, determine whether it contains valid information. The criteria are that at least one of the following: chemical formula, battery architecture, electron transport layer material, or hole transport layer material, is not null. If the criteria are met, the record is retained, even if some or all of its performance parameters are missing. Records are discarded only if both the chemical formula and all material processing information are null. This minimizes the loss of valid information from the literature. Finally, organize the extraction results according to the field system and output them to an Excel spreadsheet.

[0014] A second aspect of the present invention relates to a system for implementing the intelligent extraction method for perovskite literature data based on OCR and LLM, characterized in that it comprises: Perovskite literature acquisition module, multimodal literature parsing module, and intelligent data extraction module; The perovskite literature acquisition module is based on academic database API and open access API for literature retrieval and download. It has a built-in two-layer validator that integrates title pre-verification and PDF content secondary verification, combined with a multi-dimensional scoring mechanism of perovskite keywords and exclusion keywords, to automatically acquire verified and valid perovskite literature PDF files, and automatically separate files that fail verification to the rejection directory. The multimodal document parsing module integrates OCR layout analysis service and multimodal visual language model to perform layout structure parsing on PDF documents to separate text and image regions. It uses a structured analysis prompting mechanism for images in scientific papers to perform multi-dimensional deep semantic analysis on images and structurally integrates the image semantic analysis results into the corresponding text positions in the format of citation blocks to generate a structured Markdown document that integrates text and image semantics. The intelligent data extraction module performs segmented reasoning on the integrated text and image structured document based on a large language model. It divides long documents into overlapping continuous text segments and extracts key parameters segment by segment. It integrates a dual extraction strategy of direct matching and indirect inference of chemical formulas, uses combined key fields to construct a unique fingerprint identifier to achieve cross-segment deduplication, and adopts a lenient data retention strategy to retain incomplete records that contain valid material information. Finally, it outputs a structured data table of key parameters of perovskite solar cells.

[0015] This invention constructs a multi-model collaborative processing system integrating OCR and a large language model to achieve efficient parsing and structured extraction of multi-source heterogeneous information from perovskite literature, thereby improving data acquisition efficiency and extraction accuracy in complex literature scenarios. The main innovations of this invention are as follows: (1) This invention addresses the problem of insufficient relevance screening accuracy in traditional document crawling by constructing a multi-dimensional content verification mechanism. Traditional methods rely solely on keyword matching for document screening, making it difficult to effectively distinguish between perovskite-related and adjacent disciplines, resulting in a large amount of irrelevant documents occupying storage and computing resources. This invention designs a two-layer screening system combining title pre-verification and PDF content secondary verification: in the title pre-verification stage, preliminary screening is conducted by comprehensively utilizing domain keyword hit rates and keyword exclusion filtering strategies; in the content verification stage, PDF metadata and text content are extracted using document parsing technology, and a multi-dimensional scoring mechanism is constructed based on keyword statistics to quantitatively evaluate document relevance. Through this method, not only are documents from non-target domains effectively excluded, but the accuracy of document acquisition and overall processing efficiency are also significantly improved.

[0016] (2) This invention achieves deep semantic understanding of perovskite literature by integrating OCR layout analysis with a multimodal visual language model. Perovskite literature contains a large number of high-information-density images, such as microscopic structure diagrams, diffraction patterns, and performance curves. Traditional methods mostly focus on text processing and cannot fully utilize image information. This invention first performs structural analysis on PDF documents using layout parsing technology to separate text and image regions; then, it introduces a visual language model and designs an analysis prompt mechanism for scientific images, analyzing image content from multiple dimensions such as image type recognition, key feature extraction, data parsing, and semantic summarization; finally, it integrates the analysis results with text information and inserts them into the corresponding positions in the document in a structured form, thereby forming an integrated text-image representation. This method fully explores the value of image data in the literature and achieves unified modeling of multimodal information.

[0017] (3) This invention addresses the problem of insufficient adaptability of traditional rule-based matching by using a large language model for intelligent extraction and an automatic chemical formula inference method. Perovskite literature is expressed in diverse forms, and traditional rule-based or keyword-based methods are difficult to adapt to complex semantic scenarios. This invention constructs a multi-level data field system covering material composition, device structure, and performance parameters, and uses a large language model for semantic parsing and information extraction. Simultaneously, in terms of chemical formula extraction, a dual strategy combining direct matching and indirect inference is designed: pattern matching extraction is performed when the chemical formula is explicitly given in the literature; when only the precursor formulation is provided, the molar ratio is calculated based on the relationship between mass ratio and molecular weight, thereby constructing the target chemical formula. This method effectively improves the accuracy and adaptability of data extraction.

[0018] (4) This invention improves the completeness and robustness of data extraction through segmented reasoning and data retention strategies. Addressing the issues of long texts exceeding the model's processing limit and single documents containing multiple samples, this invention divides documents into overlapping text segments and extracts information from each segment using a large language model; simultaneously, it constructs unique identifiers through key fields to achieve data deduplication. Furthermore, a retention strategy is employed during data processing, retaining records containing valid material information even if complete performance parameters cannot be extracted, thereby minimizing information loss. This method effectively adapts to the problem of incomplete document data and significantly improves the overall data extraction effect.

[0019] The advantages of this invention are: by constructing a multi-model collaborative processing system that integrates content verification, OCR layout analysis, multimodal visual understanding, and large language model extraction, an automated process from document acquisition to structured data generation is achieved; the accuracy of document screening is improved through a multi-dimensional content verification mechanism; information utilization is enhanced through image-text fusion methods; data processing capabilities are strengthened through intelligent extraction and inference mechanisms; and system robustness and data integrity are improved through segmented reasoning and data retention strategies. This invention has high practical value in perovskite document data processing and knowledge base construction, and can effectively support data analysis and scientific research in related fields. Attached Figure Description

[0020] The accompanying drawings are provided to give a visual representation of the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0021] Figure 1 This is a flowchart illustrating the present invention; Figure 2 This is a diagram of the extraction parameters structure in Embodiment 1. Detailed Implementation

[0022] The embodiments are provided to illustrate the purpose, technical methods, and advantages of the present invention in more detail. The present invention will be further described below with reference to the embodiments. The specific embodiments described herein are for illustrative purposes only and do not constitute a limitation thereof.

[0023] Example 1

[0024] This embodiment relates to the intelligent extraction method for perovskite literature data based on OCR and Large Language Model (LLM) of the present invention. The flowchart is shown below. Figure 1 As shown, the steps include: Step S1: Conduct literature searches based on academic database APIs, filter and download PDF files of perovskite solar cell-related literature through keyword matching and content verification mechanisms. The specific steps are as follows: Step 1.1: In the perovskite literature collection scenario, configure a set of literature search keywords, including core terms such as perovskite, solar cell, photovoltaic, MAPbI, FAPbI, and CsPbI. At the same time, configure an exclusion keyword set to filter out literature in non-target fields, thereby ensuring the domain relevance of the search results.

[0025] Step 1.2: Call the academic database API to obtain literature metadata, including DOI, title, author, journal, publication year and abstract information, and perform keyword pre-validation based on the title and abstract to initially screen candidate literature.

[0026] Step 1.3: Construct a content validator to perform secondary verification on candidate documents. This involves extracting the text content from the first few pages of the PDF, counting the number of occurrences of perovskite keywords and excluded keywords, and determining whether the document belongs to the target domain based on a preset threshold.

[0027] Step 1.4: Obtain the PDF download link and download the full text of the document through the open access API. Perform integrity checks on the downloaded file and remove documents that fail the verification.

[0028] Step 1.5: Save the verified PDF file to the specified directory, and record the literature metadata in the data table to provide a data foundation for subsequent processing.

[0029] Step S2: Perform OCR parsing and multimodal semantic analysis on the collected PDF documents to generate structured documents. The specific steps are as follows: Step 2.1: Call the OCR service to perform layout analysis and text recognition on the PDF document, obtain the text content and layout structure information of the page, and recognize the image and chart areas in the document.

[0030] Step 2.2: Extract image resources from the document and establish a mapping relationship between image paths and corresponding pages to achieve the association management of text and image content.

[0031] Step 2.3: Load the multimodal visual language model and construct an analysis prompting mechanism for images in scientific papers to guide the model to analyze images from multiple dimensions such as image type recognition, key visual feature extraction, data parsing, and semantic summarization.

[0032] Step 2.4: Perform semantic analysis on each image to generate structured description information, and save the analysis results as an intermediate data file.

[0033] Step 2.5: Insert the image analysis results into the corresponding text positions to form a Markdown structured document that integrates text and images, thereby achieving a unified representation of text and image information.

[0034] Step S3: Perform segmented reasoning and information extraction on structured documents based on a large language model. The specific steps are as follows: Step 3.1: Construct a data field system for perovskite solar cells, including composition ratio fields, fixed attribute fields, and performance parameter fields, to standardize data extraction objectives. Specific extraction parameters are as follows: Figure 2 As shown.

[0035] Step 3.2: Construct an automatic chemical formula extraction module, which uses regular expressions to match chemical formulas in different formats and supports molar ratio calculation based on precursor formulations, thereby generating complete chemical composition information.

[0036] Step 3.3: Divide the long text document after image and text fusion into multiple overlapping text segments, and call the large language model to extract information for each segment to ensure the integrity and contextual continuity of the long document processing.

[0037] Step 3.4: Parse and post-process the output of the large language model, including data format standardization, duplicate data removal, and field completion. If no chemical formula is extracted, call the automatic chemical formula extraction module to complete it.

[0038] Step 3.5: Adopt a data retention strategy to retain data records that contain key material information but lack some performance parameters, so as to make the most of the effective information in the literature, and finally output the sorted data as a structured table.

[0039] To verify the effectiveness of the method of the present invention, 20 academic papers in the field of perovskite solar cells were used as test samples. Based on the results of manual annotation, the extraction accuracy of the model on four types of core parameters was evaluated. The extraction accuracy results are shown in Table 1.

[0040]

[0041] Example 2

[0042] This embodiment provides a system for implementing the intelligent extraction method of perovskite literature data based on OCR and LLM in Embodiment 1, including: Perovskite literature acquisition module, multimodal literature parsing module, and intelligent data extraction module; The perovskite literature acquisition module performs literature retrieval and download based on academic database API and open access API. It uses a content validator to perform keyword matching and relevance verification on the literature, and is used to automatically obtain valid PDF files of perovskite literature. The multimodal document parsing module integrates OCR services and a multimodal visual language model to perform layout analysis and text recognition on PDF documents, and semantic analysis and fusion of image resources to generate structured Markdown documents. The intelligent data extraction module performs segmented reasoning on structured documents based on a large language model, and outputs a structured data table of key parameters of perovskite solar cells through automatic extraction of chemical formulas and systematic extraction of data fields.

[0043] The above description is merely an embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the concept and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent extraction of perovskite literature data based on OCR and LLM, characterized in that, Includes the following steps: S1. Based on academic database API, literature retrieval and intelligent screening are performed. By constructing a two-layer verification mechanism that integrates title pre-verification and PDF content secondary verification, and combining multi-dimensional scoring of perovskite keywords and exclusion keywords, the system automatically filters and downloads PDF files of academic literature in the field of perovskite solar cells. Based on the verification results, the literature is stored in the valid directory and the rejection directory. Key parameter information of perovskite solar cells is collected, including chemical composition ratio, cell architecture, material information and performance parameters. S2. Perform OCR layout analysis on the collected PDF documents, separate text and image regions, extract Markdown-formatted text content, and download image resources to the local machine; load a multimodal visual language model, and use structured analysis prompts designed for scientific paper images to guide the model to perform in-depth semantic analysis on each image from multiple dimensions, including image type recognition, key visual feature extraction, data measurement value analysis, annotation information recognition, and scientific conclusion summary, generating image semantic description text; insert the image semantic description text into the Markdown document in a structured block format below the corresponding image position, forming a structured document that integrates text and image semantics; S3. Based on a large language model, segmented reasoning and information extraction are performed on the integrated text and image structured document: The long text Markdown document is divided into overlapping continuous text segments, and the large language model is called for semantic parsing for each segment; a multi-level data field system covering composition ratios, fixed attributes, and performance parameters is constructed; a dual extraction strategy combining direct matching and indirect inference of chemical formulas is adopted to obtain the complete chemical composition. When the large language model fails to extract a chemical formula, the result of the automatic chemical formula extraction module is automatically called to supplement it and the element composition ratio is synchronously backfilled; a unique fingerprint identifier is constructed by combining key fields to achieve cross-segment data deduplication, and a lenient data retention strategy is adopted, which retains the record as long as the chemical formula or any material process information exists, without requiring the completeness of performance data, and finally outputs a structured data table.

2. The intelligent extraction method for perovskite literature data based on OCR and LLM as described in claim 1, characterized in that, Step S1 includes the following steps: Step 1.1: Configure the search keyword set, which includes core terms in the perovskite field, including perovskite, solar cell, photovoltaic, MAPbI, FAPbI, CsPbI and their combinations. At the same time, configure the exclusion keyword set to filter literature in non-target fields. Step 1.2: Call the academic database API to obtain the literature metadata, which includes DOI, title, author, journal, publication year, and abstract information. Perform the first layer of verification on the obtained literature—keyword pre-verification at the title and abstract level. If the title or abstract contains excluded keywords, it will be directly removed. If the perovskite keywords are found, it will be retained as a candidate literature. Step 1.3: Construct a content validator to perform a second layer of verification on the downloaded PDF file—extract the text content of the first few pages through document parsing, count the total number of hits of perovskite keywords and the total number of hits of excluded keywords, and calculate the verification score. When the score of perovskite keywords reaches the set threshold and the score of excluded keywords is lower than the set threshold, it is judged as a valid perovskite document. Step 1.4: Save the verified PDF files to the specified directory, record the literature metadata to an Excel spreadsheet, and move the PDF files that fail verification to the rejection directory to ensure that subsequent processing only applies to valid perovskite literature; Step 1.5: Collect key parameter information of perovskite solar cells from valid literature, including chemical composition ratio, cell structure, material information and performance parameters.

3. The intelligent extraction method for perovskite literature data based on OCR and LLM as described in claim 1, characterized in that, Step S2 includes the following steps: Step 2.1: Call the OCR service to perform layout analysis and text recognition on the PDF file, obtain the Markdown format text content of each page, and at the same time obtain the image resource links and layout detection images to achieve automatic separation of text areas and image areas; Step 2.2: Download the identified image resources to the local output directory and establish a mapping record between image paths and corresponding pages; Step 2.3: Load the multimodal visual language model and use structured analysis prompts designed for the characteristics of images in scientific papers. The prompts guide the model to perform deep semantic analysis on the image from the following dimensions: (a) Image type recognition, distinguishing between charts, microscopic images, schematic diagrams, and molecular structure types; (b) Key visual element extraction, identifying the main components and contrast relationships in the image; (c) Data and measurement value analysis, extracting numerical values, scales, coordinate axis labels, and data trends; (d) Label information recognition, extracting text labels, legends, color codes, and highlighted areas; (e) Scientific conclusion summary, summarizing the scientific findings and key conclusions supported by the image. Step 2.4: Perform multimodal analysis on each image sequentially to generate semantic description text, save the analysis results to a JSON file, and perform a memory cleanup operation after each analysis. Step 2.5: Scan the image reference marks in the Markdown document line by line, and insert the corresponding image semantic description text into the original image position in a reference block format to form a structured document that integrates text semantics and image semantics, and save it.

4. The intelligent extraction method for perovskite literature data based on OCR and LLM as described in claim 1, characterized in that, Step S3 includes the following steps: Step 3.1: Define the data extraction field system for perovskite solar cells, including composition ratio fields, fixed attribute fields, and performance parameter fields. The composition ratio fields cover FA ratio, MA ratio, Cs ratio, Pb ratio, Sn ratio, I ratio, Br ratio, and Cl ratio. The fixed attribute fields cover cell architecture, complete chemical formula, electron transport layer material, hole transport layer material, antisolvent type, buried interface passivator, perovskite layer additives, interface passivator, processing method, annealing temperature, and annealing time. The performance parameter fields cover photoelectric conversion efficiency, open-circuit voltage, short-circuit current density, fill factor, effective area, and data source. Step 3.2: Construct an automatic chemical formula extraction module and adopt a dual extraction strategy to obtain the complete chemical composition: The direct matching strategy uses regular expressions to sequentially match chemical formulas in LaTeX format, standard format, and nested bracket format, and cleans and removes duplicates from the extraction results; The indirect inference strategy, when the precursor solution formulation only provides the mass ratio of each component, converts the mass ratio to the molar ratio based on the standard molecular weight of each component, and then infers the cation ratio and constructs the complete chemical formula according to the perovskite chemical formula specification. Step 3.3: The integrated text and image long text Markdown document generated in S2 is divided into continuous text segments with a preset overlap length. The large language model is called for data extraction for each segment. The system prompt words guide the model to extract data according to priority. Chemical formulas and material process parameters are mandatory priority extraction items, and performance parameters are conditional extraction items. Step 3.4: Post-process the JSON results output by the large language model, including format parsing, data deduplication, and field validation; construct a unique fingerprint identifier by combining chemical formulas, battery architecture, material information, and data source to deduplicate duplicate data across segments; when the large language model fails to extract a chemical formula, automatically call the results of the automatic chemical formula extraction module described in Step 3.2 to supplement it, and simultaneously calculate the corresponding elemental composition ratio and fill it into the corresponding fields; Step 3.5: Employ a lenient data retention strategy. For each extracted record, determine whether it contains valid information. The criteria are that at least one of the following: chemical formula, battery architecture, electron transport layer material, or hole transport layer material, is not null. If the criteria are met, the record is retained, even if some or all of its performance parameters are missing. Records are discarded only if both the chemical formula and all material processing information are null. This minimizes the loss of valid information from the literature. Finally, organize the extraction results according to the field system and output them to an Excel spreadsheet.

5. A system for implementing the intelligent extraction method for perovskite document data based on OCR and LLM as described in any one of claims 1-4, characterized in that, include: The modules include: perovskite literature acquisition module, multimodal literature parsing module, and intelligent data extraction module. The perovskite literature acquisition module is based on academic database API and open access API for literature retrieval and download. It has a built-in two-layer validator that integrates title pre-verification and PDF content secondary verification, combined with a multi-dimensional scoring mechanism of perovskite keywords and exclusion keywords, to automatically acquire verified and valid perovskite literature PDF files, and automatically separate files that fail verification to the rejection directory. The multimodal document parsing module integrates OCR layout analysis service and multimodal visual language model to perform layout structure parsing on PDF documents to separate text and image regions. It uses a structured analysis prompting mechanism for images in scientific papers to perform multi-dimensional deep semantic analysis on images and structurally integrates the image semantic analysis results into the corresponding text positions in the format of citation blocks to generate a structured Markdown document that integrates text and image semantics. The intelligent data extraction module performs segmented reasoning on the integrated text and image structured document based on a large language model. It divides long documents into overlapping continuous text segments and extracts key parameters segment by segment. It integrates a dual extraction strategy of direct matching and indirect inference of chemical formulas, uses combined key fields to construct a unique fingerprint identifier to achieve cross-segment deduplication, and adopts a lenient data retention strategy to retain incomplete records that contain valid material information. Finally, it outputs a structured data table of key parameters of perovskite solar cells.

Images