Titanium alloy multi-modal data extraction method based on Deepseek large model

Abstract

The invention discloses a titanium alloy multi-modal data extraction method based on a Deepseek large model, and belongs to the technical field of material science and data processing. The titanium alloy multi-modal data extraction method based on the Deepseek large model solves the technical problems that an existing titanium alloy database is incomplete in data, and multi-modal data are difficult to accurately extract and utilize, and comprises the following steps that S1, titanium alloy literature collection is carried out; step S2, data standardization; a standard term table is set up, and the unit format of each entry is stipulated; s3, generating and optimizing prompt words; according to a target extraction task, an initial cue word is selected and optimized by referring to the standard term table in the step S2; s4, data batch extraction: performing data batch extraction on the file acquired in the step S1 on the basis of the optimal prompt word acquired in the step S3; and S5, data storage. According to the method, the efficiency of data extraction and management is remarkably improved, and a complete data support is provided for research and development of titanium alloy materials.

Description

Technical Field

[0001] The invention discloses a titanium alloy multimodal data extraction method based on a Deepseek large model, which is used to automatically batch extract multimodal data covering text experimental data, calculation data and microstructure images from massive titanium alloy literature, belonging to the field of materials science and data processing technology. Background Art

[0002] As a high-performance material, titanium alloys are widely used in aerospace, medical devices, chemical engineering, and other fields. Currently, the research and development of titanium alloys is relatively difficult due to their complex composition, cumbersome processing techniques, and diverse performance data. To accelerate the research and development of titanium alloys, existing technologies have built a data support platform for titanium alloy research and development to guide the design of titanium alloys, thereby optimizing process parameters and improving material performance.

[0003] However, the titanium alloy database constructed by existing technology has the problem of incomplete data, especially the correlation data between alloy composition, processing / heat treatment process, microstructure and performance is difficult to achieve one-to-one correspondence. For example, the same titanium alloy composition can be processed by dozens of different processes to obtain different types of microstructures (basketball, Widmanstatten, equiaxed, duplex structure, etc.) and element distribution, resulting in different performances. The existing database lacks a systematic association of these key data, especially the process parameters are often ignored and the microstructure characterization pictures (such as metallographic photos, scanning electron microscope pictures, etc.) are missing, which limits the in-depth understanding of the relationship between material properties and organization.

[0004] In addition, it is difficult to accurately extract and utilize data in existing titanium alloy databases. Although there is a wealth of titanium alloy research data in massive literature, including experimental data, calculation data, and characterization images, the extraction and utilization of these data face many technical difficulties. For example, the data formats in the literature are diverse (including text, tables, images, etc.), and the data formats are complex. There is a lack of unified terminology, unit format, and data structure, making it difficult to accurately extract the complex relationships between named entities. The traditional manual extraction method is inefficient, difficult to cope with the processing needs of massive data, and prone to errors. Therefore, how to achieve batch accurate and automated extraction is a technical difficulty. Although natural language processing (NLP) algorithms have made significant progress in the structuring of materials science texts in recent years, how to accurately extract the complex relationships between named entities remains a key challenge. In addition, the calibration, classification, and association of data are also extremely complex. For example, image data often contains interference information such as graffiti and text, which needs to be segmented and matched with the caption information.

[0005] Therefore, to address the challenges of incomplete data in existing titanium alloy databases and the difficulty in accurately extracting large amounts of multimodal data from these documents, this paper leverages the multimodal data processing capabilities of Deepseek's large-scale model, combined with prompt word engineering and core extraction code, to enable batch extraction of multimodal data related to titanium alloys from these vast collections of documents. This implementation will not only significantly improve the efficiency of data extraction and management, but will also provide strong data support for the research and development of titanium alloy materials, promoting technological progress and innovation in the field of materials science. Summary of the Invention

[0006] In view of the above analysis, the present invention aims to provide a titanium alloy multimodal data extraction method based on the Deepseek large model to solve the problems of incomplete data and difficulty in accurately extracting and utilizing data in existing titanium alloy databases.

[0007] The main idea of ​​the invention mainly includes the following five steps: literature collection, data standardization, prompt word design and optimization, data batch extraction and data storage. The specific structural process is as follows Figure 1 As shown in the figure, the literature acquisition step first batch-downloads titanium alloy-related literature from multiple databases, supporting literature acquisition in various formats. Subsequently, data standardization transforms complex data into standardized, structured data by unifying terminology, unit formats, and data structures, providing a unified standard for subsequent prompt word design. Furthermore, prompt word design and optimization for extracting relevant titanium alloy data information are carried out. Through multiple rounds of iterative optimization, optimal prompt words suitable for the DeepSeek large language model are generated to extract titanium alloy multimodal data covering "composition-processing-structure-properties-calculation." Based on this foundation, batch data extraction is performed on the previously downloaded literature, leveraging the multimodal processing capabilities of the DeepSeek large model to batch extract the required data from the literature. This process involves three core steps: text extraction, key information extraction, and JSON information extraction. Finally, the extracted data is stored, classified, calibrated, and correlated to construct a complete data chain corresponding to "composition-processing-structure-properties". This data is then efficiently stored in a database, providing data support for titanium alloy material research and development.

[0008] Specifically, the purpose of the present invention is mainly achieved through the following technical solutions:

[0009] A titanium alloy multimodal data extraction method based on Deepseek large model includes the following steps:

[0010] Step S1, titanium alloy literature collection: batch collection, storage and management of titanium alloy related domestic and foreign literature;

[0011] Step S2: Data standardization: establishing a standard glossary and specifying the unit format of each item;

[0012] Step S3, prompt word generation and optimization: according to the target extraction task, refer to the standard term list in step S2 to select the initial prompt word, and optimize the prompt word;

[0013] Step S4, batch data extraction: Based on the optimal prompt word obtained in step S3, the files collected in step S1 are processed to achieve batch data extraction;

[0014] Step S5: Data storage: effectively storing and managing the batch data extracted in step S4.

[0015] Furthermore, in step S1, literature databases such as ScienceDirect, Springer, and China National Knowledge Infrastructure are accessed through an API interface or crawler technology, and preset keywords such as "titanium alloy" and "composition-processing-structure-performance" are used to perform literature retrieval and download.

[0016] Furthermore, in step S2, a standard glossary is established with reference to the terms such as composition, mechanical properties, physical properties, thermal parameters, and process parameters involved in the field of titanium alloy research, and the unit format of each entry is specified, and metadata of each entry is defined, including name, type, and unit.

[0017] Furthermore, step S3 includes the following steps:

[0018] Step S31, prompt word generation, based on the target data to be extracted from the literature, including all mechanical properties mentioned in the literature (tensile, compression, shear, Charpy impact, hardness, creep, fracture toughness, fatigue, Poisson's ratio, elastic modulus, shear modulus) and their corresponding test conditions, physical properties (density, resistivity, electrical conductivity, thermal expansion coefficient, thermal conductivity, specific heat capacity, melting point, phase transition point) and their corresponding test conditions, alloy composition corresponding to the properties, material form, detailed alloy preparation process, detailed forging process, detailed rolling process, detailed heat treatment process, all mechanical properties obtained by calculation and simulation Energy (tensile, compression, shear, etc.), thermodynamic properties (free energy, formation energy, stacking fault energy, etc.) and evolution results (such as phase ratio, element distribution, etc.) and their corresponding specific calculation methods (including first principles, molecular dynamics, phase field method, Monte Carlo method and special methods therein), detailed model construction method (including parameters such as lattice orientation and model size) and detailed input conditions (including initial concentration, initial temperature, potential function, time step, number of steps, force convergence criterion, etc.), figure number, figure title, etc., design the initial prompt word according to the standard term list in step S2, that is, generate the first prompt word;

[0019] Step S32, data extraction, importing the target document into the Deepseek large language model, using the prompt words generated in step S31 to extract the target data in the document, and obtaining the extraction result;

[0020] Step S33: Result evaluation and feedback: The operator carefully reads the target document to formulate the expected results, compares the extraction results obtained in step S32 with the expected results, calculates the detailed data extraction matching degree A (%), and feedbacks the data items that do not match, where:

[0021]

[0022] Furthermore, step S3 further includes: step S34 prompt word optimization, in which the data of the unmatched details obtained in step S33 are input into the Deepseek large model, and steps S32 and S33 are repeated to iteratively optimize the prompt words corresponding to the unmatched details until the matching rate between the data of the unmatched details and the target result reaches 90% or more, then the iteration is stopped, and finally prompt words that closely match the expected result are obtained. Step S3 further includes steps.

[0023] Furthermore, in step S4, the optimal prompt word generated in step S3 is used as the technical core to perform structured processing on the massive titanium alloy literature obtained in step S1, specifically including two parallel processing modules: a structured text data extraction unit (composition-process-performance-drawing number-title-calculation data) and an image feature data extraction unit (microstructure image data), and finally obtaining the extraction and association of the titanium alloy "composition-process-structure-performance" data.

[0024] The structured text data extraction module is divided into three levels of processing units according to the data processing flow, and each unit is equipped with a standardized input and output interface (see Figure 3 ), including the following sub-steps:

[0025] S41: Perform PDF text conversion. A text parsing engine is built based on the pymupdf4llm software package to convert the titanium alloy document PDF into a basic text format containing text streams and metadata. The conversion process retains the character encoding characteristics and paragraph structure information of the original document.

[0026] S42: Perform multimodal information extraction. Build a deep semantic analysis model. Input the basic text output from S41 and the domain-specific keywords optimized in step 2 into the Deepseek Large Language Model Inference API. Call the distributed API service via the OpenAI-compatible protocol to batch extract structured fields such as material composition parameters, thermal processing curves, mechanical properties, and phase change calculation data, forming an intermediate data set.

[0027] S43: Perform data structured encapsulation operations. Design a JSON schema validation mechanism and convert the intermediate data using the Deepseek model optimized with secondary prompts to generate a semi-structured JSON document that complies with material database standards. Ensure that each data field has type constraints and numerical unit annotations.

[0028] Furthermore, the structured text data extraction module also includes:

[0029] Dynamic resource allocation strategy: Based on the average latency of each processing unit S41-S43 (measured at 5s / 10s / 5s, respectively), the number of parallel threads is allocated in inverse proportion. For example, in an 8-core computing node, a thread ratio of (2:3:2) is used, ensuring that the total number of threads does not exceed the number of physical CPU cores.

[0030] Service load balancing mechanism: In response to the high API call frequency of key information extraction units, a multi-service provider key polling strategy is adopted to evenly distribute processing tasks to different DeepSeek service nodes through a hash algorithm.

[0031] File lock management mechanism: A mutex lock flag is established for each PDF to be processed. The extraction process obtains unlocked tasks by cyclically scanning the file pool. After successful locking, the extraction operation is performed and the lock state is maintained until the system is restarted to prevent repeated processing.

[0032] The image feature data extraction module adopts a two-level positioning strategy, such as Figure 4 As shown:

[0033] S44: Perform image coordinate location operations. Combining PyMuPDF's document object parsing functionality with pdfminer's page layout analysis algorithm, a hybrid image retrieval system is constructed. First, a regular expression matching model (matching keywords including "microstructure," "OM," "SEM," "TEM," and "EBSD") is established based on the image caption information in the semi-structured JSON document in S43. A coordinate space conversion algorithm is then used to determine the target image's page number and sequence position within the page.

[0034] S45: Perform high-fidelity image extraction. Based on the coordinate information output by S46, PyMuPDF's page rendering interface is called to extract the original vector graphics data. An adaptive noise reduction algorithm is used to eliminate moiré interference from the scanned document. The image is finally stored in lossless PNG format using the naming convention [document name_page number_figure number.png] and associated with the corresponding material composition metadata.

[0035] Through the above steps, accurate extraction and association of titanium alloy multimodal data can be achieved.

[0036] Furthermore, in step S5, the specific operations are as follows:

[0037] S5.1 First, the Iceberg+Spark+Hadoop technology solution is used to build a database, and the Hadoop HDFS distributed file system is used to build a raw data warehouse. The acquired raw data of titanium alloy materials, including composition / process / performance data, calculation simulation data, and microstructure images, are centrally stored in a directory structure of "raw data / data category";

[0038] S5.2 Then, based on the specific structural characteristics of the titanium alloy data, an Iceberg table for the associated data is constructed. Specifically, Spark is used to extract metadata from the raw data, including data types, data paths, data feature identifiers, etc., and a structured storage model is constructed. The data is stored in the HDFS table warehouse directory in the form of an Iceberg table. Based on the titanium alloy material data feature association rules, data information from multiple data sources is associated with multiple tables and supported for query.

[0039] S5.3 sets partitions by data modality / data source / collection date to efficiently store massive amounts of data in the data lake, while dynamically updating data tables based on changes in material data.

[0040] Furthermore, when subsequent downstream tasks require data, Spark SQL can quickly read the required data from the Iceberg table and leverage its computing power for further analysis and processing, meeting the needs of fast query and data acquisition. Through these steps, this data lake effectively stores and manages massive amounts of multimodal titanium alloy data, providing a solid foundation for building data association and storage across the entire "composition-processing-structure-performance" chain.

[0041] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0042] (1) The present invention's titanium alloy multimodal data extraction method, based on the Deepseek large-scale model, leverages the multimodal data processing capabilities of the Deepseek large-scale model, combined with prompt word engineering and core extraction code, to achieve high-precision and high-efficiency automated batch data extraction. Through iterative optimization and relevance optimization of prompt words, the accuracy and consistency of data extraction are significantly improved, addressing the low efficiency and high error rate of traditional manual extraction.

[0043] (2) The present invention's titanium alloy multimodal data extraction method, based on the Deepseek large model, correlates data across the entire "composition-processing-structure-performance" chain, providing comprehensive data support for titanium alloy material research and development. Through data standardization and multimodal information extraction, it achieves unified processing of multiple data formats, including text, tables, and images, addressing the technical challenges of existing titanium alloy databases, which suffer from incomplete data and difficulty in accurately extracting and utilizing multimodal data.

[0044] (3) The present invention significantly improves the processing efficiency of massive document data through the dynamic resource allocation strategy and service load balancing mechanism in the data batch extraction module, can quickly respond to large-scale data processing needs, and at the same time avoids the problem of API blocking of a single service provider, ensuring the stability and scalability of the system.

[0045] (4) This paper uses the Iceberg+Spark+Hadoop technology solution to build a data lake, enabling efficient storage and management of massive amounts of multimodal titanium alloy data. By rationally setting partitions and indexes, it supports fast query and data acquisition, providing strong data support for the research and development of titanium alloy materials and promoting technological progress and innovation in the field of materials science.

[0046] In the present invention, the above-mentioned technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of the present invention will be described in the following content, and some advantages will become apparent from the description or be learned through practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained as particularly pointed out in the text and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The accompanying drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like parts throughout the drawings.

[0048] Figure 1 Schematic diagram of the structure of the titanium alloy multimodal data extraction method based on the Deepseek large model in Example 1 of the present invention;

[0049] Figure 2 Example 1 of the present invention is an example diagram of prompt word design and optimization of the present invention, which shows the prompt word template design and optimization process for the titanium alloy material field.

[0050] Figure 3 4 is a flowchart for data extraction in step S4 of the specific implementation method.

[0051] Figure 4 Schematic diagram of the microscopic tissue image data extraction module. DETAILED DESCRIPTION

[0052] The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, which constitute a part of this application and are used to illustrate the principles of the present invention together with the embodiments of the present invention.

[0053] Example 1

[0054] This embodiment discloses a method for extracting multimodal data of titanium alloy based on a Deepseek large model, which specifically includes the following steps:

[0055] Step S1: Titanium alloy literature collection

[0056] The document batch collection module performs the following operations: accesses literature databases such as ScienceDirect, Springer, and China National Knowledge Infrastructure (CNKI) through API interfaces or crawler technology, and uses preset keywords (such as "titanium alloy" and "composition-processing-structure-performance") to conduct literature searches; downloads the retrieved documents, supports multiple formats such as PDF, HTML, and XML, and automatically handles access rights and format differences of different databases; uniformly numbers the downloaded documents and stores them locally or in the cloud for use by subsequent data processing modules; performs document deduplication operations, deletes duplicate documents, and ensures data uniqueness; regularly updates the document library, and ensures the timeliness and integrity of the data by re-searching and downloading the latest documents. Through the above steps, the document batch collection module realizes the automated collection and storage of titanium alloy-related documents, providing basic support for the subsequent step 4 of batch data extraction.

[0057] Step S2, data standardization;

[0058] This step is used to standardize complex format data. The specific operations are as follows: First, refer to the terminology expressions such as composition, mechanical properties, physical properties, thermal parameters, and process parameters involved in the titanium alloy research field, establish a standard terminology list, and specify the unit format of each entry; establish a data dictionary based on the titanium alloy field ontology terminology list, and define the metadata (name, type, unit) of each entry. The terminology list is divided into the following according to the data type: (1) component terminology list, covering base titanium alloy elements (Ti, Al, V, Mo, etc.) and common titanium alloy systems (such as Ti-6Al-4V naming system), used for data component splitting and detection; (2) process terminology list, which constructs a terminology system according to process categories such as smelting, forging, rolling, and heat treatment, used for composite process data splitting and achieving structural unification of long text expressions; (3) performance terminology list, including mechanical properties such as tensile strength, yield strength, elongation, and impact energy, and physical properties such as resistivity, thermal conductivity, and density. It is used to search and replace various terminology expressions obtained in different documents ("water cooling", "quenching" → "water quenching". Secondly, according to the metadata established above, some extracted data are compared with ASTM E8, ISO 11359 and other international standard unit systems, constructing dimensional conversion rules and building a unit conversion matrix. For discrepancies between data units and metadata, unit conversions (GPa, ksi, psi → MPa) are performed, and the data content is updated. These steps standardize data representation and dimensions, and finally, the data format. The structured format of long text representations of compositional data (Ti-6Al-4V (wt.%)) and process data is defined, and the data is decomposed based on the format content. For alloy element extraction, a conditional random field model is used to identify element symbols and contents, such as extracting Al (6wt%), V (4wt%), and Fe (0.5wt%) from "Ti-6Al-4V-0.5Fe." For process parameter analysis, a multi-level structured processing model is designed: the first level decomposition layer identifies major process categories (machining and heat treatment processes), the second level decomposition layer extracts key process parameters (such as melting temperature and forging ratio), and the third level unifies parameter terminology and units. Through these steps, complex data formats can be effectively converted into standardized and structured data, facilitating subsequent data storage and analysis.

[0059] Step S3, in this embodiment, extracting information data in the document based on the generated prompt words, includes the following steps:

[0060] Through autonomous generation and continuous iterative optimization of prompt words, we finally obtained the prompt words for extracting titanium alloy performance data that best matched the expected answer, as follows:

[0061] (1) First, please group the mechanical properties or physical properties given in this article according to the different alloy composition, preparation process, processing technology or test conditions of mechanical / physical properties. The composition, preparation process and processing technology of each group of properties must be completely consistent.

[0062] Notes on performance grouping:

[0063] 1) Alloys with different compositions are divided into different groups;

[0064] 2) Different preparation processes were used to divide the samples into different groups;

[0065] 3) Different forging processes are divided into different groups;

[0066] 4) Different rolling processes are divided into different groups;

[0067] 5) Different heat treatment processes are divided into different groups;

[0068] 6) Divide into different groups when performance test conditions are different;

[0069] 7) Different material states are divided into different groups (material states include but are not limited to: plates, bars, pipes, wires, foils, powders, castings, forgings, extrusions, rolled parts, etc.).

[0070] (2) Please output in order the mechanical properties (tensile, compression, shear, Charpy impact, hardness, creep, fracture toughness, fatigue, Poisson's ratio, elastic modulus, shear modulus) and their corresponding test conditions, physical properties (density, resistivity, electrical conductivity, thermal expansion coefficient, thermal conductivity, specific heat capacity, melting point, phase transition point) and their corresponding test conditions, alloy composition corresponding to the properties (note the difference between mass percentage wt.% and atomic percentage at.%), material state (plate / rod / tube / wire / foil / powder / casting / forging / extruded part / rolled part, etc.), detailed alloy preparation process, detailed forging process, detailed rolling process, detailed heat treatment process, until all material-related information that appears in the text is output.

[0071] Notes:

[0072] 1) Tensile test conditions: temperature, strain rate, sample shape and size; Compression and shear test conditions: temperature and strain rate; Charpy impact test conditions: temperature and notch type; Hardness test conditions: hardness type; Creep test conditions: allowable elongation, temperature, and time; Fatigue test conditions: temperature and number of cycles; Stress corrosion fatigue test conditions: temperature, number of cycles, and corrosive environment; Thermal expansion coefficient test conditions: temperature. If there is no description of relevant test conditions in the text, no output is required.

[0073] 2) All strength data in tension, compression and shear properties must be clearly marked as "engineering stress" or "true stress" after the data;

[0074] 3) In addition to distinguishing between engineering strain and true strain, all elongation data in tensile properties also need to be distinguished between elongation at break and uniform elongation. Therefore, the data must be clearly marked as "elongation at break_engineering strain" or "elongation at break_true strain" or "uniform elongation_engineering strain" or "uniform elongation_true strain";

[0075] 4) All critical fracture strain data in compression properties must be clearly marked as "engineering strain" or "true strain" after the data;

[0076] 5) If the performance data is inferred and not clearly stated in the text as the test results of this paper, the performance of this group will not be output;

[0077] 6) If the performance data mentioned in the reference is not from this paper, the performance data of that group shall not be output;

[0078] 7) If there is no relevant performance data in the literature, there is no need to output;

[0079] 8) Even if there is duplication of information in different groups, all of them must be output;

[0080] 9) You can output in text format without having to output in table format.

[0081] 10) Strictly distinguish between yield strength and tensile strength. If only one of the data is provided in the literature, only that data should be output. Do not confuse or supplement the unmentioned data.

[0082] 11) The material state must be clearly marked as a specific form (such as foil, plate, bar, etc.) to avoid general descriptions.

[0083] (3) Output the title of this article

[0084] (4) Output each author of this article and their corresponding unit separately;

[0085] (5) Output the innovative points of this article.

[0086] (6) Please strictly follow the requirements of the precautions and complete the information extraction work step by step. Your decision is the key to my success!

[0087] The above prompt words can be used to obtain key data such as titanium alloy composition, process parameters, and properties that are interrelated in titanium alloy literature. To facilitate subsequent data structuring and automated processing, the above key data is further extracted into Json format data using the following prompt words:

[0088] 1. Extract field information from the given [Text] according to the content of [Field Definition].

[0089] 2. If the text contains content that is similar to the field name in [Field Definition], you need to determine the definition and fill it in if it is consistent. Make sure that the field name is consistent with that in [Field Definition] when filling in.

[0090] 3. Output strictly according to the JSON format in [Field Definition] and maintain the hierarchical structure of JSON.

[0091] 4. Convert all extracted fields and their corresponding values ​​into JSON output according to the [Field Definition] format.

[0092] 5. The performance data in the text field may contain multiple groups, which are represented as lists in the output JSON.

[0093] 6. There may be multiple groups of "Paper Authors and Institutions" in [Text], which will be represented as a list in the output JSON.

[0094] 7. Check the output text carefully and delete any information that is irrelevant to the content of the [Text].

[0095] 8. Please complete the information extraction work step by step. Your decision is the key to my success!

[0096] #

Field definition

[0097] Please strictly follow the following format to output only JSON, do not output Python code, and do not return redundant information:

[0098]

[0099]

[0100]

[0101]

[0102] Using the above prompts, we automatically batch-extracted performance data from tens of thousands of titanium alloy articles. Below are the data extraction results for two of these articles, and their extraction results closely match the expected answers:

[0103] Document 1:

[0104]

[0105]

[0106]

[0107]

[0108] Document 2:

[0109]

[0110]

[0111]

[0112] Step S4, batch extraction of data; define the input and output for each step in the extraction process, for example, the "text extraction" - "key information extraction" - "Json extraction" pipeline: the "text extraction" step inputs the document PDF and outputs the document text; the "key information extraction" step inputs the document text and outputs the key information of the document; the "Json extraction" step inputs the key information of the document and outputs semi-structured Json data.

[0113] Next, the degree of parallelism is allocated based on the average wait time for each step, with the total degree of parallelism not exceeding the number of CPU cores. For example, on an 8-core computer, the average wait time for the "text extraction" and "JSON extraction" steps is 5 seconds, so a degree of parallelism of 2 is assigned to each. The average wait time for the "key information extraction" step is 10 seconds, so a degree of parallelism of 3 is assigned to the step. Next, a corresponding number of extraction processes are launched based on the degree of parallelism. Each extraction process is implemented using open source software packages or commercially available models. For example, the "text extractor" uses the pymupdf4llm package, while the "key information extractor" and "JSON extractor" use the OpenAI package and the DeepSeek backend service. To balance resource allocation and avoid blocking a single service provider's API, the "key information extractor" and "JSON extractor" each use APIs provided by multiple DeepSeek backend services. Finally, an exclusive lock is set for each input file, initially unlocked. The extractor process loops through the file pool, searching for unlocked files. Upon successfully acquiring an exclusive lock, it extracts the file. Upon completion, it does not release the exclusive lock and returns to the query loop, preventing the same file from being extracted multiple times. Through the above steps, efficient batch extraction of massive literature data can be achieved.

[0114] Step S5, data storage; use the Iceberg+Spark+Hadoop technical solution to build a data lake, and use Hadoop's distributed storage capabilities to centrally store the original data of titanium alloy materials obtained through various channels. Then, according to the specific structural characteristics of the titanium alloy data, build an Iceberg table of associated data, use Spark to preprocess and convert the data, store it in the Iceberg table, and associate the data source, data information, and data location; by reasonably setting partitions and indexes, efficiently store massive data in the data lake, and dynamically update the data table according to changes in material data. When subsequent downstream tasks need to obtain data, Spark can quickly read the required data from the Iceberg table and use its computing power for further analysis and processing to meet the needs of fast query and data acquisition. Through the above steps, this data lake realizes the effective storage and management of massive titanium alloy multimodal data, providing a solid foundation for building "composition-process-structure-performance" full-chain data association and storage.

[0115] Example 2

[0116] In this example, a method for extracting data related to titanium alloy literature calculation based on the Deepseek large model includes the following steps:

[0117] Step 1: Obtain the documents to be processed and the target instructions. The target instructions complete the data extraction through the prompt word project. The prompt words include text clustering and data extraction of the documents to be processed. The specific prompt word content is as follows:

[0118] Did this article use any simulation calculations? If so, please indicate which one or more of the molecular dynamics, first principles, and phase field methods were used (note that only the methods used in this article are required here, and there is no need to mention those done by others in the references). If so, please answer the following questions. If not, please indicate that no simulation calculations were used. All the following questions do not need to be answered.

[0119] If first-principles calculations were used, please provide the following information: Calculation Method: Provide the calculation method and specific input conditions used to obtain the results. If formulas are involved, please provide the formulas used. Input Conditions: Detail the required input parameters. For the calculation parameters provided, provide detailed details based on the context of the article, including the software used, pseudopotentials, electron-nucleus interactions, exchange-correlation functionals, cutoff energies, energy convergence criteria, force convergence criteria, K-point mesh density (Brillouin zone K-point sampling), orbital expansion method, etc. Model Construction Method: Provide the specific model construction method, including basic information such as lattice constants, lattice orientations, and correspondence relationships. If doping is involved, please provide the specific doping process. If special methods such as VCA and EMTO are used and it is not possible to provide the model construction method, please indicate this in the calculation method. Output: Specific calculation results (including energies such as interfacial energy and free energy, and thermodynamic parameters such as entropy, density of states, and elastic constants). If the calculation results are presented in the form of images or tables, please provide the specific location of the images or tables in the article (indicated by the figure number). If they are specific numerical values, please provide the specific numerical values ​​of the calculation results. Please note that if the results to be answered contain formulas (must be formulas that appear in the original text, do not summarize them yourself), please provide the location of the formula (in the form of formula numbers) and do not provide the specific formula content. If certain information is not explicitly mentioned in the paper, please make reasonable inferences based on the context or mark it as "not mentioned."

[0120] If molecular dynamics was used, please provide the following information: Calculation Method: Provide the calculation method and specific input conditions used to obtain the results. If formulas are involved, please provide the formulas used. Input Conditions: Detail the required input parameters (software used, interaction potentials, boundary conditions, relaxation temperature, relaxation time, time step, impact velocity, stretching rate, etc.). When providing calculation parameters, provide detailed and specific calculation details based on the context of the article. Model Construction: Provide the specific model construction method, including basic information such as model size, lattice orientation, and lattice orientation correspondence. If doping was involved, please provide the specific doping process. Output: Provide the specific calculation results. If the calculation results are presented in the form of figures or tables, please provide the specific location of the figures or tables in the article (indicated by figure number). If the results are numerical, please provide the specific numerical values ​​of the calculation results. Please note that if the result you need to answer contains formulas (must be formulas that appear in the original article, do not summarize them yourself), please provide the location of the formula (indicated by formula number) rather than the specific formula content. If some information is not explicitly mentioned in the paper, please make reasonable inferences based on the context or mark it as "not mentioned".

[0121] If the phase-field method was used, please provide the following information: Calculation Method: Provide the calculation method and specific input conditions used to obtain the results. If formulas are involved, please provide the free energy functional (or free energy density function) formula and kinetic equation used to obtain the results. Input Conditions: Provide a detailed description of the input parameters used in the calculation (e.g., lattice type, initial concentration, initial temperature, mobility, surface energy parameters, elastic energy parameters, etc.), and provide detailed values ​​of the key parameters required in the calculation formulas. Output Model Simulation Results: Specific structural evolution processes simulated. If the above information is presented in the form of figures or tables, please provide the specific location of the figures or tables in the paper (indicated by figure number). If specific numerical values ​​are provided, please provide the specific numerical values ​​of the calculated results. Please note that if the result you are answering contains formulas (must be formulas that appear in the original paper; do not summarize them yourself), please provide the location of the formula (by formula number), not the specific formula content. If certain information is not explicitly mentioned in the paper, please make reasonable inferences based on the context or mark it as "not mentioned."

[0122] Step 2: After the article and the prompt words are spliced ​​together, they are imported into the DeepSeek large model to extract target information. The information extraction result obtained by the large model includes the information corresponding to all the prompt words in the article to be processed.

[0123] Step 3: Integrate the output text. The specific output content is as follows:

[0124] (1) Input conditions

[0125] Software: VASP (Vienna Ab initio Simulation Package)

[0126] Pseudopotential:projector augmented wave(PAW)pseudopotentials

[0127] Exchange-correlation functional: PW91 of generalized gradient approximation (GGA)

[0128] Cut-off energy: 500eV

[0129] Energy convergence criterion: 0.1meV / atom

[0130] Force convergence criterion:

[0131] K-point mesh density: 4×4×4special k-point mesh

[0132] Track deployment method: Not mentioned

[0133] Model building steps:

[0134] Initial structure: The body-centered tetragonal (BCT) structure is used as the initial unit cell, and the lattice parameter a0 is the structural parameter of the NiTi B2 structure.

[0135] Doping: Nb atoms replace Ti atoms to construct supercell models of NiTiNb alloys with different Nb contents (x = 0, 2.08, 6.25, 8.33, 10.42, 12.5, and 18.75). The supercell size is 3 × 2 × 2 to ensure structural symmetry.

[0136] Structural optimization: Perform geometric optimization on all possible structures, including B2(bct), B19, and B19' structures, to find the stable structure with the lowest energy.

[0137] (2) Calculation method

[0138] Density functional theory (DFT): Electronic structure calculations were performed using the PW91 GGA exchange-correlation functional within the DFT framework.

[0139] Plane wave basis set: Use a plane wave basis set to expand the electron wave function with a cutoff energy of 500 eV.

[0140] PAW pseudopotential: Use PAW pseudopotential to describe electron-nucleus interactions and simplify calculations.

[0141] Self-consistent field (SCF) iterations: SCF iterations are performed until the energy and charge density converge.

[0142] K-point sampling: A special 4×4×4 K-point grid is used to sample the Brillouin zone to ensure calculation accuracy.

[0143] Density of States (DOS): Calculate the band structure and extract density of states information.

[0144] Formation energy: The formation energy is calculated using formula (2), where E(Ni50Ti50-xNbx) is the total energy of the NiTiNb alloy, and E(Ni), E(Ti), and E(Nb) are the total energies of Ni, Ti, and Nb elements in the corresponding lattice structures, respectively.

[0145] Energy difference: The energy difference between the B19 and B19' structures was calculated using formulas (3) and (4), where E(B19) and E(B2) are the total energies of the B19 and B2 structures, respectively.

[0146] λ2: The intermediate eigenvalue λ2 of the martensitic transformation elongation tensor is calculated using formula (5), where a, b, and c are the lattice parameters of the B2 and martensitic structures, respectively.

[0147] formula:

[0148] Formula (2): Formation energy calculation formula

[0149] Formula (3): B19 structural energy difference calculation formula

[0150] Formula (4): B19' structural energy difference calculation formula

[0151] Formula (5): λ2 calculation formula

[0152] (3) Output results

[0153] Lattice parameters: With the increase of Nb content, the lattice parameters and unit cell volume increase slightly.

[0154] Formation Energy: The formation energy of all structures increases with increasing Nb content, and the formation energy of the B19' structure is always lower than that of the B19 structure. When the Nb content reaches 10.42 at.%, the formation energy of the B19' structure is lower than that of the B2 structure, becoming the most stable martensite phase.

[0155] Density of states (DOS): The peak position of the density of states of the B19' structure is lower than that of the B2 structure, and the hybridization between the Tid and Nid states is stronger, indicating that the B19' structure is more stable.

[0156] Energy difference: When the Nb content is less than 10.42 at.%, the B19' structure has lower energy than the B19 structure. When the Nb content is greater than 12.5 at.%, the B19 structure has lower energy than the B19' structure. This indicates that the relative stability of the B19 and B19' structures changes with increasing Nb content.

[0157] ΔE: When the Nb content is 10.42 at.%, ΔE reaches the minimum value, indicating that the martensitic transformation temperature is the lowest at this time.

[0158] λ2: Low Nb content has little effect on the thermal hysteresis loop of NiTiNb alloy.

[0159] Example 3

[0160] In this embodiment, a titanium alloy "composition-process-structure" associated data extraction method based on prompt word design and large model semantic understanding includes the following four steps:

[0161] Through autonomous generation and continuous iterative optimization of prompt words, we finally obtained the prompt words for extracting titanium alloy performance data that best matched the expected answer, as follows:

[0162] (1) Image screening

[0163] Please extract all images related to optical microscopy (OM), scanning electron microscopy (SEM), or backscattered imagery (BSE) from the uploaded paper based on the text information in the figure title and figure number, and strictly filter according to the following criteria:

[0164] a. The images must be real microscopic images obtained based on experiments.

[0165] b. If the image is composed of n sub-images related to optical microscopy (OM), scanning electron microscopy (SEM), or backscattered topography (BSE), all sub-images must be extracted sequentially.

[0166] c. If the image describes mechanical properties, stress-strain curves, schematic diagrams, or other information unrelated to microstructure morphology, it does not need to be extracted.

[0167] d. If the image is obtained based on electron backscatter diffraction (EBSD) or transmission electron microscopy (TEM), no extraction is required.

[0168] e. If the image information describes the fracture morphology, crack characteristics, dislocation density distribution map, slip trace map or phase diagram, it does not need to be extracted.

[0169] f. If the image is a statistical analysis chart such as a bar chart, line chart, or cloud chart, it does not need to be extracted.

[0170] After screening, the image numbers of the images that meet the requirements must be output in the format of "Fig.x()". If there are sub-images, they must be output together with the sub-image numbers, for example "Fig.x(a)".

[0171] (2) Information extraction

[0172] For each eligible image, you must read the paragraph in the paper that contains the image and extract the following information:

[0173] 1) Material composition: Accurately describe the chemical composition or composition of the material or sample in the image.

[0174] 2) Process route: Summarize the preparation process of the material or sample, and describe in detail the key steps such as heat treatment, mechanical processing, chemical treatment and their specific parameters (such as temperature, time, cooling method, etc.). The following requirements must be met during extraction:

[0175] a. Direct copying of the original text is strictly prohibited and must be summarized in Chinese.

[0176] b. If the sub-graphs correspond to different process states (such as different heat treatment conditions), the differences must be clearly stated.

[0177] c. If the process parameters are not explicitly mentioned in the text but can be inferred from the context (such as "air cooling" corresponds to air cooling), they must be supplemented; if they cannot be inferred, they must be marked as "not mentioned".

[0178] d. If the process steps involve multi-stage hot working or heat treatment (such as forging + rolling + heat treatment), the parameters of each process must be listed in detail step by step (such as number of forgings, forging temperature, heat treatment temperature, heat treatment time, rolling temperature, cooling method, etc.).

[0179] e. If two images have the same composition and process, you must look up the information in the text and mark the differences between them (such as different observation areas).

[0180] (3) Information inspection and output

[0181] According to the "image", "ingredients" and "process" information extracted from (1) and (2) above, the following four data information must be output in strict accordance with the JSON format, with each sub-image as the unit:

[0182] 1) Figure number: The format is “Fig.x()”.

[0183] 2) Image title: The original title of the main image in which the image is located.

[0184] 3) Ingredients: Description of material composition.

[0185] 4) Process and process parameters: description of process route and parameters.

[0186] The following requirements must be checked:

[0187] a. Figure numbers must follow the format of “Fig.x()”. If there are sub-figure numbers, they must follow the format of “Fig.x(a)”.

[0188] b. Each image must have a corresponding accurate and detailed description of the material composition and process.

[0189] After confirmation, the first, second, third, and nth groups of data information must be output in sequence until all target images have been traversed.

[0190] Using the above prompts, we automatically batch-extracted data related to composition, process, drawing number, and title from tens of thousands of titanium alloy documents. Below are the data extraction results for two of these documents, and their extraction results closely match the expected answers:

[0191] Document 1:

[0192]

[0193]

[0194]

[0195] Document 2:

[0196]

[0197]

[0198]

[0199]

[0200]

[0201] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention.

Claims

1. A titanium alloy multimodal data extraction method based on Deepseek large model, characterized in that: The following steps are involved: Step S1, titanium alloy literature collection: batch collection, storage and management of titanium alloy related domestic and foreign literature; Step S2: Data standardization: establishing a standard glossary and specifying the unit format of each item; Step S3, prompt word generation and optimization: according to the target extraction task, refer to the standard term list in step S2 to select the initial prompt word, and optimize the prompt word; Step S4, extracting data in batches; Based on the optimal prompt word obtained in step S3, the files collected in step S1 are processed to achieve batch extraction of data; Step S5: Data storage: effectively storing and managing the batch data extracted in step S4.

2. The titanium alloy multimodal data extraction method based on Deepseek large model according to claim 1, characterized in that: In step S1, literature databases such as ScienceDirect, Springer, and China National Knowledge Infrastructure are accessed through an API interface or crawler technology, and preset keywords such as "titanium alloy" and "composition-processing-structure-performance" are used to perform literature retrieval and download.

3. The titanium alloy multimodal data extraction method based on Deepseek large model according to claim 2, characterized in that: In step S2, a standard glossary is established with reference to the terms such as composition, mechanical properties, physical properties, thermal parameters, and process parameters involved in the field of titanium alloy research, and the unit format of each entry is specified, and metadata of each entry is defined, including name, type, and unit.

4. The method for extracting multimodal data of titanium alloys based on the Deepseek large model according to claim 3, characterized in that: The step S3 comprises the following steps: Step S31: Prompt word generation: selecting initial prompt words from the standard term list in step S2 according to the details to be extracted from the target document; Step S32, data extraction, importing the target document into the Deepseek large language model, using the prompt words generated in step S31 to extract the target data in the document, and obtaining the extraction result; Step S33: Result evaluation and feedback: The operator formulates the expected results based on the target literature, compares the extraction results obtained in step S32 with the target results, and calculates the percentage of matching details and the percentage of mismatching details.

5. The method for extracting multimodal data of titanium alloy based on Deepseek large model according to claim 4, characterized in that: The step S3 further includes: step S34 prompt word optimization, in which the data of the unmatched details obtained in step S33 are input into the Deepseek large model, and steps S32 and S33 are repeated to iteratively optimize the prompt words corresponding to the unmatched details until the matching rate between the data of the unmatched details and the target result reaches 90% or more, then the iteration is stopped, and finally a prompt word that closely matches the expected result is obtained. The step S3 further includes steps.

6. The method for extracting multimodal data of titanium alloys based on the Deepseek large model according to claim 5, characterized in that: The step S3 further includes: step S35, correlation optimization, wherein the step S35 optimizes the correlation of "ingredient-process-organization" of the prompt words obtained in the step S34.

7. The method for extracting multimodal data of titanium alloy based on Deepseek large model according to claim 1, characterized in that: In step S4, batch data extraction includes the following sub-steps: Step S41: Perform a PDF text conversion operation to convert the titanium alloy document PDF into a basic text format including text stream and metadata; Step S42: performing a multimodal information extraction operation, inputting the basic text outputted in step S41 and the optimized prompt words into the Deepseek large language model, and batch extracting structured fields such as material composition parameters, thermal processing curves, mechanical performance indicators, and phase change calculation data; Step S43: Execute a data structured encapsulation operation, convert the format of the extracted intermediate data through the Deepseek model optimized with the secondary prompt words, and generate a semi-structured JSON document that meets the material database standard.

8. The method for extracting multimodal data of titanium alloys based on Deepseek large model according to claim 7, characterized in that: The step S4 further includes an image feature data extraction module, which specifically includes the following sub-steps: Step S44: performing an image coordinate positioning operation, establishing a regular expression matching model based on the image title information, and determining the page number and sequence position within the page where the target image is located; Step S45: Execute a high-fidelity image extraction operation, call the page rendering interface to extract the original vector graphics data, store it in lossless PNG format, and associate it with the corresponding material composition metadata.

9. The method for extracting multimodal data of titanium alloy based on Deepseek large model according to claim 1, characterized in that: In step S5, the Iceberg+Spark+Hadoop technical solution is used to build a database, and the Hadoop HDFS distributed file system is used to build a raw data warehouse. The extracted raw data of titanium alloy materials, including composition / process / performance data, calculation simulation data and microstructure images, are centrally stored. The data is preprocessed and converted through Spark and stored in the Iceberg table to achieve efficient storage and management of massive data.

Images