Composite material preparation process parameter extraction method based on large model assistance

Abstract

The invention discloses a composite material preparation process parameter extraction method based on large model assistance. The specific process comprises the steps of text conversion and division, processing of a tagged literature set by using a large model, manual cleaning and standard data set construction, basic knowledge base construction, literature set process paragraph recognition, retrieval enhancement and cue word engineering, parameter extraction and output and dynamic knowledge base updating. The extraction method can adapt to different types of composite materials and different process description modes by constructing a field dynamic knowledge base and combining a vector retrieval technology, solves the problem that the performance of a traditional method is suddenly reduced in a vertical scene, provides structured process parameter data for researchers, reduces the time cost of literature reading and information extraction, and improves the extraction efficiency. And formula design, process optimization and performance prediction of the novel composite material are accelerated, and innovative development of the field of material science is promoted.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a method for extracting process parameters for composite material preparation based on a large model. Background Technology

[0002] Composite materials are widely used in aerospace, automotive manufacturing, and new energy fields due to their excellent mechanical properties and chemical stability. Meanwhile, the continuous accumulation of a wealth of research literature, patents, and technical reports on composite material preparation, containing key process parameters such as resin type, curing temperature, and fiber ratio, serves as an important basis for developing new composite materials and optimizing preparation processes.

[0003] Researchers mainly obtain these parameters by manually reading literature, which not only relies on the researchers' professional experience, but also suffers from problems such as low efficiency and limited coverage, making it easy to miss key information and restricting the iterative speed of composite material research and development.

[0004] Existing parameter extraction methods based on rules, machine learning, or natural language processing (NLP) suffer from limitations. Due to the diverse types of composite materials (such as resin-based, ceramic-based, and metal-based) and the complex preparation processes (such as curing, sintering, and impregnation), rule design is cumbersome and lacks adaptability, making it difficult to cover different scenarios. Furthermore, the scarcity of labeled data in the field of composite materials limits the model's ability to understand contextual semantics and makes it difficult to handle complex linguistic phenomena such as ambiguous expressions and variations of technical terms, resulting in low parameter extraction accuracy.

[0005] In recent years, large language models (such as GPT and Deepseek) have demonstrated outstanding performance in natural language understanding and generation. Their powerful contextual semantic understanding capabilities have provided a new technical approach to solving the problem of extracting parameters from composite material literature, and have become a key support for improving the efficiency of materials research and development.

[0006] However, directly using large models for parameter identification presents numerous challenges. For instance, while large models possess vast amounts of training data in general domains, their capabilities are limited when facing niche scenarios or vertical fields, making it difficult to adapt to high demands. In complex contexts, due to diverse variations in expression, large models lack the ability to accurately identify and correlate parameters, leading to parameter matching errors and reduced reliability of identification results. Large models cannot handle recurring, unfinishable tasks. Furthermore, the default output of large models lacks unified parameter category definitions and structured format specifications, often resulting in ambiguous expressions that make the extracted results difficult to apply directly, significantly reducing their usability. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method for extracting process parameters for composite material preparation based on a large model. This extraction method, by constructing a dynamic domain knowledge base and combining it with vector retrieval technology, can adapt to different types of composite materials and different process description methods. It solves the problem of drastic performance degradation of traditional methods in vertical scenarios, provides researchers with structured process parameter data, reduces the time cost of literature reading and information extraction, accelerates the formulation design, process optimization, and performance prediction of new composite materials, and promotes innovative development in the field of materials science.

[0008] A method for extracting process parameters for composite material preparation based on a large model is described below: Step 1, Text Conversion and Segmentation: Convert the texts related to composite material preparation into the target format in batches, and extract a portion of the literature as a set of annotated documents; Step 2: Process the annotated literature set using a large model: Identify the categories of synthetic parameter data to be extracted, use a large model to determine whether each paragraph in the literature contains preparation process content, and then use the large model again to extract parameters according to the above data categories for paragraphs containing preparation process to obtain preliminary annotation results. Step 3, manual cleaning and standard dataset construction: The initial annotation results of the large model are manually checked and cleaned to correct parameter identification errors, supplement missing parameters, and standardize the expression format, and finally form a standard paragraph extraction dataset, which includes the original text of the preparation process paragraph and its corresponding structured parameters. Step 4, Basic Knowledge Base Construction: Extract the original text of paragraphs from the standard paragraph dataset and combine them with structured parameters to form paragraph-parameter question-answer pairs. Then, vectorize the question-answer pairs to build a vertical domain knowledge base for the large model, which will be used for subsequent similar case retrieval. Step 5, Identification of process paragraphs in the literature collection: After excluding irrelevant paragraphs using a large model, all paragraphs in the remaining literature collection are judged to identify and extract paragraphs containing composite material preparation processes; Step Six, Retrieval Enhancement and Prompt Word Engineering: For the preparation process paragraphs extracted in Step Five, vector retrieval is first performed in the knowledge base to obtain the paragraph-parameter question-answer pairs with the highest similarity. The retrieved question-answer pairs are then used as examples and written into the large model prompt words along with the paragraphs to be extracted, data category requirements, and output format specifications to form a prompt word structure of example guidance + task instructions. Step 7, Parameter Extraction and Output: Input the constructed prompt words into the large model. The large model extracts parameters from the preparation process paragraphs of the extracted literature set according to the example format. The large model is required to output only the recognized parameters and not mention the parameters that cannot be determined, so as to avoid generating hallucinatory information. Finally, the structured parameter extraction results are output. Step 8, Dynamic Knowledge Base Update: For the results of subsequent extraction errors, manual extraction is performed, and the original text of the incorrectly extracted paragraphs is combined with the manually revised structured parameters to form new paragraph-parameter question-answer pairs. Then, the question-answer pairs are vectorized to dynamically update the vertical domain knowledge base of the large model, which is used to improve the ability to retrieve similar high-difficulty cases in the future.

[0009] As a preferred embodiment of the above technical solution, in step one, the target format can be any one of Markdown, TXT, HTML, DOCX, PDF, XLSX, ODS, CSV, TSV, JSON, XML, STEP, IGES, MAT, CDB, INP, ODT, RTF, WPS, JPG, or PNG.

[0010] As a preferred embodiment of the above technical solution, in step one, the number of documents extracted shall not be less than 20.

[0011] As a preferred embodiment of the above technical solution, in step two, the data categories to be extracted include: raw material parameters: resin type, curing agent type, curing agent ratio, coupling agent type and dosage, additive type and ratio, fiber type, fiber morphology, fiber specifications (diameter / length), fiber volume fraction, number of fibers, and fiber surface treatment method; mixing / impregnation parameters: prepreg resin content, mixing speed, impregnation temperature, and impregnation time; molding process parameters: molding process type, layup sequence, number of layup layers, layup angle deviation, mold temperature, and mold pressure holding method; and curing process parameters: curing temperature, curing pressure, curing time, heating rate, cooling rate, parameters for each heat preservation / pressure holding stage, and curing time. Environmental vacuum level, inert gas protection type; Post-treatment parameters: post-treatment temperature, post-treatment time, surface treatment process; Mechanical properties: tensile strength, tensile modulus, flexural strength, flexural modulus, compressive strength, shear strength, impact strength, interlaminar shear strength, fatigue strength, fracture toughness; Thermal properties: glass transition temperature (Tg), heat distortion temperature, thermal decomposition temperature, coefficient of linear expansion, thermal conductivity; Chemical properties: corrosion resistance (acid resistance, alkali resistance, salt spray resistance), aging resistance (thermal oxidative aging, UV aging), water absorption; Electrical properties: volume resistivity, dielectric constant, electromagnetic shielding effectiveness; Special properties: porosity, fiber dispersion uniformity, flame retardancy rating, biocompatibility.

[0012] As a preferred embodiment of the above technical solution, in step two, the extracted data categories can be supplemented, including prepreg preparation parameters, molding auxiliary parameters, and post-processing parameters. The above supplementary data categories can be dynamically added or deleted according to the specific application scenario of the composite material, and user-defined new parameter categories are also supported.

[0013] As a preferred embodiment of the above technical solution, in step five, the irrelevant paragraphs excluded include the introduction, references, and author information.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By leveraging the strong semantic understanding capabilities of large models and the design of prompt words guided by a small number of labeled examples, the shortcomings of traditional methods in handling complex contexts and terminology variations are resolved.

[0015] 2. By leveraging the vector retrieval mechanism of a dynamic knowledge base, the method can quickly adapt to the material type (e.g., from resin-based to ceramic-based) and process description method in vertical fields without retraining the model, thus solving the problem of limited scenarios in traditional machine learning methods.

[0016] 3. By manually cleaning and constructing a dynamic standard pattern knowledge base, the criterion rate of the extracted results can be ensured. This knowledge base can be directly used for data analysis, process optimization, and other scenarios in composite material research and development, reducing subsequent data processing costs. Attached Figure Description

[0017] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The present invention will now be described in further detail with reference to the accompanying drawings: like Figure 1 The method for extracting process parameters for composite material preparation based on a large model is shown below. Step 1, Text Conversion and Segmentation: Convert the texts related to composite material preparation into the target format in batches, and extract a portion of the literature as a set of annotated documents; Step 2: Process the annotated literature set using a large model: Identify the categories of synthetic parameter data to be extracted, use a large model to determine whether each paragraph in the literature contains preparation process content, and then use the large model again to extract parameters according to the above data categories for paragraphs containing preparation process to obtain preliminary annotation results; among them, the Deepseek large model is preferred.

[0020] Step 3, manual cleaning and standard dataset construction: The initial annotation results of the large model are manually checked and cleaned to correct parameter identification errors, supplement missing parameters, and standardize the expression format, and finally form a standard paragraph extraction dataset, which includes the original text of the preparation process paragraph and its corresponding structured parameters. Step 4, Basic Knowledge Base Construction: Extract the original text of paragraphs from the standard paragraph dataset and combine them with structured parameters to form paragraph-parameter question-answer pairs. Then, vectorize the question-answer pairs to build a vertical domain knowledge base for the large model, which will be used for subsequent similar case retrieval. Step 5, Identification of process paragraphs in the literature collection: After excluding irrelevant paragraphs using a large model, all paragraphs in the remaining literature collection are judged to identify and extract paragraphs containing composite material preparation processes; Step Six, Retrieval Enhancement and Prompt Word Engineering: For the preparation process paragraphs extracted in Step Five, vector retrieval is first performed in the knowledge base to obtain the paragraph-parameter question-answer pairs with the highest similarity. The retrieved question-answer pairs are then used as examples and written into the large model prompt words along with the paragraphs to be extracted, data category requirements, and output format specifications to form a prompt word structure of example guidance + task instructions. Step 7, Parameter Extraction and Output: Input the constructed prompt words into the large model. The large model extracts parameters from the preparation process paragraphs of the extracted literature set according to the example format. The large model is required to output only the recognized parameters and not mention the parameters that cannot be determined, so as to avoid generating hallucinatory information. Finally, the structured parameter extraction results are output. Step 8, Dynamic Knowledge Base Update: For the results of subsequent extraction errors, manual extraction is performed, and the original text of the incorrectly extracted paragraphs is combined with the manually revised structured parameters to form new paragraph-parameter question-answer pairs. Then, the question-answer pairs are vectorized to dynamically update the vertical domain knowledge base of the large model, which is used to improve the ability to retrieve similar high-difficulty cases in the future.

[0021] In this embodiment, in step one, the target format can be any one of Markdown, TXT, HTML, DOCX, PDF, XLSX, ODS, CSV, TSV, JSON, XML, STEP, IGES, MAT, CDB, INP, ODT, RTF, WPS, JPG, and PNG.

[0022] In this embodiment, in step one, the number of documents extracted is no less than 20.

[0023] In this embodiment, in step two, the data categories to be extracted include: raw material parameters: resin type, curing agent type, curing agent ratio, coupling agent type and dosage, additive type and ratio, fiber type, fiber morphology, fiber specifications (diameter / length), fiber volume fraction, number of fibers, and fiber surface treatment method; mixing / impregnation parameters: prepreg resin content, mixing speed, impregnation temperature, and impregnation time; molding process parameters: molding process type, layup sequence, number of layup layers, layup angle deviation, mold temperature, and mold pressure holding method; and curing process parameters: curing temperature, curing pressure, curing time, heating rate, cooling rate, parameters for each heat preservation / pressure holding stage, and curing environment. Vacuum degree, inert gas protection type; Post-processing parameters: post-processing temperature, post-processing time, surface treatment process; Mechanical properties: tensile strength, tensile modulus, flexural strength, flexural modulus, compressive strength, shear strength, impact strength, interlaminar shear strength, fatigue strength, fracture toughness; Thermal properties: glass transition temperature (Tg), heat distortion temperature, thermal decomposition temperature, coefficient of linear expansion, thermal conductivity; Chemical properties: corrosion resistance (acid resistance, alkali resistance, salt spray resistance), aging resistance (thermal oxidative aging, UV aging), water absorption; Electrical properties: volume resistivity, dielectric constant, electromagnetic shielding effectiveness; Special properties: porosity, fiber dispersion uniformity, flame retardancy rating, biocompatibility.

[0024] In this embodiment, the extracted data categories in step two can be supplemented, including prepreg preparation parameters, molding auxiliary parameters, and post-processing parameters. The above-mentioned supplementary data categories can be dynamically added or deleted according to the specific application scenario of the composite material, and user-defined new parameter categories are also supported.

[0025] Specifically, the prepreg preparation parameters include impregnation temperature, impregnation time, and resin content control range; the molding auxiliary parameters include release agent type, pressurization rate, and holding time; the post-treatment parameters include post-curing temperature, post-curing time, cooling rate, and annealing process parameters; the composite material application scenarios include high-strength composite materials for aerospace and low-cost composite materials for automotive lightweighting; and the custom-added parameter categories include flame retardant type and addition ratio, and antioxidant type and addition amount.

[0026] In this embodiment, in step five, the irrelevant paragraphs excluded include the introduction, references, and author information.

[0027] The following description, in conjunction with specific embodiments, will illustrate the process. Example 1: Extraction of Curing Process Parameters for Resin-Based Carbon Fiber Composite Materials (1) Download 500 papers with titles containing “Carbon Fiber Reinforced Polymer” and “composite” from Elsevier, parse the PDFs into Markdown, and randomly select 20 papers as the annotated literature set.

[0028] (2) Select the resin type, curing agent type, curing agent ratio, fiber type, fiber volume fraction, curing temperature, curing pressure, and curing time. Call Deepseek, temperature=0.1, and the prompt template is as follows: "You are a composite materials process expert. Please determine whether the following paragraph contains composite materials preparation processes. Answer only 'yes' or 'no': {paragraph}."

[0029] For paragraphs that return "yes", call the large model again: "Output in JSON format: resin type, curing agent type, curing agent ratio, fiber type, fiber volume fraction, curing temperature, curing pressure, curing time parameters. If not mentioned, write 'null': {paragraph}".

[0030] (3) Manual parallel proofreading, marking and extracting errors and omissions to form a standard paragraph extraction dataset S.

[0031] (4) Organize each record in dataset S into question-answer pairs: Q: "original paragraph text", A: "structured JSON". Use the bge-large-zh-v1.5 model to vectorize them and convert them into a knowledge base.

[0032] (5) Call the large model for the remaining literature set, call Deepseek, temperature=0.1, the prompt template is as follows: "You are a composite materials process expert. Please determine whether the following paragraph contains composite materials preparation processes. Answer only 'yes' or 'no': {paragraph}."

[0033] For paragraphs that return "yes", first use the knowledge base from step 4 to perform vector retrieval, take the top-3 similar question-answer pairs, and concatenate them into prompt words: Example 1: ... Example 2: ... Example 3: ... Task: Output the following parameters in JSON format: resin type, curing agent type, curing agent ratio, fiber type, fiber volume fraction, curing temperature, curing pressure, and curing time. Write 'null' if not mentioned: {paragraph}.

[0034] Ten documents were randomly reviewed manually. Any incorrectly extracted paragraphs were checked, and after correct manual annotation, the results were appended to the standard paragraph extraction dataset S. Steps 4-6 were then repeated.

[0035] Example 2: Extraction of preparation parameters for silicon carbide ceramic matrix composites (1) Download 500 papers with titles containing “Silicon carbide ceramic matrix” and “composites” from Google Scholar, parse the HTML into TXT, and randomly select 20 papers as the annotated literature set.

[0036] (2) Select the fiber type, fiber morphology, fiber volume fraction, impregnation temperature, impregnation time, annealing temperature, and annealing time. Call GPT, temperature=0.1, and the prompt template is as follows: "You are a composite materials process expert. Please determine whether the following paragraph contains composite materials preparation processes. Answer only 'yes' or 'no': {paragraph}."

[0037] For paragraphs that return "yes", call the large model again: "Output in JSON format: fiber type, fiber morphology, fiber volume fraction, impregnation temperature, impregnation time, annealing temperature, annealing time, write 'null' if not mentioned: {paragraph}".

[0038] (3) Manual parallel proofreading, marking and extracting errors and omissions to form a standard paragraph extraction dataset S.

[0039] (4) Organize each record in S into question-answer pairs: Q: "original paragraph text", A: "structured JSON". Use the Qwen3-Embedding model to vectorize them and convert them into a knowledge base.

[0040] (5) Call the large model for the remaining literature set, call GPT, temperature=0.1, the prompt template is as follows: "You are a composite materials process expert. Please determine whether the following paragraph contains composite materials preparation processes. Answer only 'yes' or 'no': {paragraph}."

[0041] For paragraphs that return "yes", first use the knowledge base from step 4 to perform vector retrieval, take the top-3 similar question-answer pairs, and concatenate them into prompt words: "Example 1: ... Example 2: ... Example 3: .... Task: Output the following in JSON format: tree fiber type, fiber morphology, fiber volume fraction, impregnation temperature, impregnation time, annealing temperature, annealing time parameters, write 'null' if not mentioned: {paragraph}".

[0042] (6) Manually review 10 documents randomly. Look for any incorrectly extracted paragraphs. After correct manual annotation, append the results to the standard paragraph extraction dataset S, and repeat steps 4-6 again.

[0043] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A composite material preparation process parameter extraction method based on large model assistance, characterized by: The specific process is as follows: Step one, text conversion and division: batch convert the text related to composite material preparation into the target format, and extract part of the literature as the annotated literature set; Step two, use large model to process the annotated literature set: determine the type of synthetic parameter data to be extracted, use the large model to determine whether each paragraph in the literature contains the preparation process content, and then use the large model again to extract parameters according to the above data categories for the paragraphs containing the preparation process, to obtain the preliminary annotation results; Step three, manual cleaning and standard data set construction: manually correct and clean the initial annotation results of the large model, correct parameter recognition errors, supplement missing parameters, and standardize the expression format, to finally form a standard paragraph extraction data set, which includes the original text of the preparation process paragraph and its corresponding structured parameters; Step four, basic knowledge base construction: form paragraph-parameter question and answer pairs by combining the original text of the paragraph in the standard paragraph extraction data set with the structured parameters, and then perform vectorization processing on the question and answer pairs to construct a vertical domain knowledge base for the large model, which is used for subsequent similar case retrieval; Step five, process paragraph identification in literature set: after excluding irrelevant paragraphs using the large model, all paragraphs in the remaining literature set are judged to identify and extract paragraphs containing composite material preparation processes; Step six, retrieval enhancement and prompt word engineering: for the preparation process paragraphs extracted in step five, first perform vector retrieval in the knowledge base to obtain the most similar paragraph-parameter question and answer pairs, and use the retrieved question and answer pairs as examples together with the to-be-extracted paragraph, data category requirements, and output format specifications to write prompt words into the large model, forming a prompt word structure of example guidance + task instruction; Step seven, parameter extraction and output: input the constructed prompt words into the large model, which extracts parameters from the preparation process paragraphs in the extracted literature set according to the example format, and requires the large model to only output the recognized parameters, and labels the parameters that cannot be determined as not mentioned to avoid generating illusory information, and finally outputs the structured parameter extraction results; Step eight, dynamic knowledge base update: for subsequent extracted error results, perform manual extraction, and form new paragraph-parameter question and answer pairs by combining the extracted paragraph original text with the manually revised structured parameters, then perform vectorization processing on the question and answer pairs, and dynamically update the vertical domain knowledge base of the large model to improve the subsequent similar high-difficulty case retrieval capability.

2. The method of claim 1, wherein the method is based on a large model assisted composite material preparation process parameter extraction method. In step one, the target format can be any of Markdown, TXT, HTML, DOCX, PDF, XLSX, ODS, CSV, TSV, JSON, XML, STEP, IGES, MAT, CDB, INP, ODT, RTF, WPS, JPG, and PNG.

3. The method of claim 1, wherein the method is based on a large model assisted composite material preparation process parameter extraction method. In step one, the number of extracted literature is not less than 20.

4. The composite material preparation process parameter extraction method based on large model assistance according to claim 1, characterized in that: In step two, the data categories that need to be extracted include raw material parameters: resin model, curing agent model, curing agent proportion, coupling agent model and dosage, additive type and proportion, fiber model, fiber form, fiber specification, fiber volume fraction, fiber surface treatment method; mixing / impregnation parameters: prepreg resin content, mixing speed, impregnation temperature, impregnation time; molding process parameters: molding process type, layering sequence, layering number, layering angle deviation, mold temperature, mold pressure retention method; curing process parameters: curing temperature, curing pressure, curing time, heating rate, cooling rate, parameters of each holding / pressure holding stage, curing environment vacuum degree, inert gas protection type; post-processing parameters: post-processing temperature, post-processing time, surface treatment process; mechanical properties: tensile strength, tensile modulus, bending strength, bending modulus, compression strength, shear strength, impact strength, interlaminar shear strength, fatigue strength, fracture toughness; thermal properties: glass transition temperature, heat distortion temperature, thermal decomposition temperature, linear expansion coefficient, thermal conductivity; chemical properties: corrosion resistance, aging resistance, water absorption; electrical properties: volume resistivity, dielectric constant, electromagnetic shielding effectiveness; special properties: porosity, fiber dispersion uniformity, flame retardant grade, biocompatibility. In step two, the extracted data categories can also be supplemented, including prepreg preparation parameters, molding auxiliary parameters, post-processing parameters, and the above-mentioned supplementary data categories can be dynamically added or deleted according to the application scenario of the specific composite material, and user-defined new parameter categories are also supported. In step five, the irrelevant passages that are excluded include the introduction, references, and author information.

5. The method of claim 4, wherein the method is based on a large model assisted composite material preparation process parameter extraction method. ​ 6. The method of claim 1, wherein: ​

Images