Method and system for predicting tensile strength of polyimide assisted by large language model

Abstract

The application belongs to the technical field of artificial intelligence and new materials, and discloses a method and system for predicting the tensile strength of polyimide with the aid of a large language model. The method extracts the copolymer or homopolymer monomer of PI in the text by using a large language model LLM, converts the images of dianhydride and diamine into SMILES strings, and converts the LLM code into Morgan frequency-weighted-additive fingerprints, feature screening, model selection, parameter optimization and model interpretation, so as to quickly predict PI without TS value. The TS model for quickly predicting PI is built based on the large language model assisted machine learning whole process, and exhibits interdisciplinary innovation, and integrates data science, computational chemistry and artificial intelligence technology, and has the advantages of high efficiency, user friendliness and accuracy. The application provides intelligent technical support and professional knowledge guidance for researchers using AI tools, and can accelerate the in-depth development process and exploratory data analysis of polyimide.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and new materials technology, specifically to a method and system for predicting the tensile strength of polyimide using a large language model. The large language model-assisted machine learning process includes text extraction, image conversion, expert knowledge suggestions, molecular structure expression, model code writing, model analysis, and using the model to predict the tensile strength of polyimide. Background Technology

[0002] Polyimide (PI) is a high-performance engineering plastic with excellent strength and stiffness, outstanding heat resistance, and chemical stability. Its superior mechanical and thermal properties have led to its widespread application in the aerospace, automotive, and electronics industries. Designing experiments can be a challenging task, requiring a high level of expertise, and without extensive knowledge in the field, it is often difficult to develop optimal experimental designs. It is impossible for experimenters to exhaustively explore all possible two-component reactions of dianhydrides and diamines for material design and the discovery of novel polyimides. Therefore, researchers have attempted to use methods such as quantitative computation, machine learning, and molecular dynamics simulations to effectively predict the properties of novel polyimides.

[0003] While general artificial intelligence (AI) technology has developed rapidly in recent years, its integration with specialized technical fields lacks depth, and its generated technical solutions often cannot be practically applied. For example, in existing technology, Chinese invention document CN117668182A discloses a standardized intelligent question-answering method and system that integrates knowledge graphs and large language models. This method includes: constructing a professional knowledge graph of the target domain; fine-tuning the large language model based on the professional knowledge graph to generate a professional knowledge large language model; generating a structured answer to the target question based on the professional knowledge large language model and guiding prompts, and converting the structured answer into a triple answer; constructing an answer knowledge graph of the target question based on the triple answer, and rendering and displaying the answer knowledge graph. Through this approach, knowledge graph construction based on language models and fine-tuning of the large language model based on knowledge graphs can be achieved, enabling collaborative work between knowledge graphs and large language models. By mutually empowering knowledge graph technology and large language model technology, efficient knowledge management, professional knowledge intelligent question answering, and accurate question answering can be achieved, improving the quality and efficiency of solution decision-making. However, it only provides general methods and processes and cannot deeply integrate with specialized R&D needs.

[0004] Artificial intelligence (AI) has achieved unprecedented breakthroughs in screening, designing, and assisting in the synthesis and preparation of new materials, significantly shortening the R&D cycle and reducing R&D costs. Currently, in the field of polymer design, it has been proven feasible to use machine learning to identify the relationship between the microscopic chemical structure and various macroscopic properties of polymers. However, chemical scientists often lack knowledge of statistics and computer science, while many programmers are unfamiliar with the mechanisms and characterization techniques of materials. This limits the efficiency, accuracy, generalization ability, interpretability, practicality, and accessibility of material modeling in large-scale data environments. The emergence of Large Language Models (LLMs) has attracted widespread attention and has alleviated the challenges of interdisciplinary collaboration to some extent. ChatGPT, for example, is an AI-driven natural language processing tool that can understand and learn human language, engaging in dialogue and interaction based on context. GPT technology has ushered in a new era in the field of AI, finding widespread application in areas such as language recognition, machine translation, and speech generation, and driving the development of various ecosystems, including generative models, pre-trained models, and multimodal models (such as ProGen and BioGPT). While large language models have been studied for molecules and materials, they have not yet been used to learn representations of polymers. Most computational work has focused on predicting glass transition temperatures, with less attention paid to predicting tensile strength (TS). Furthermore, most machine learning research on polymers has been limited to homopolymer molecules, with research on copolymer molecules remaining largely unexplored. Therefore, utilizing large language models to assist the entire machine learning process—including image conversion, expert knowledge suggestions, molecular structure representation, model code writing, model analysis, and model prediction—is highly valuable for efficiently predicting the tensile strength of polymers (PIs). Summary of the Invention

[0005] To address the problems of existing technologies, this invention aims to overcome their shortcomings and provide a method and system for predicting the tensile strength of polyimide (PI) using a large language model (MLM). Based on the large language model, it assists in solving in-depth interdisciplinary research and development problems. Utilizing AI tools, it extracts copolymer or homopolymer monomers of PI from text using an LLM model, converts the images of dianhydrides and diamines into the string "SMILES," and uses LLM code to convert this into a Morgan frequency-weighted summation fingerprint, feature selection, model selection, parameter optimization, and model interpretation. The established model can quickly predict PI without a TS value. This method is low-cost, simple, efficient, provides complete and accurate data, requires no experiments, no complex calculations, and is pollution-free.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0007] A method for predicting the tensile strength of polyimide using a large language model includes the following steps:

[0008] S1. Using the Web of Science database, search for publicly available articles on synthesizing PIs, input LLM into a normalized language, and extract the names of dianhydrides and diamines and their corresponding TS values ​​from the text to form a dataset.

[0009] S2. Use LLM to convert the corresponding dianhydride and diamine chemical structure diagrams into SMILES strings, generate the structure in ChemOffice, modify it, and then convert it back into SMILES strings.

[0010] S3. Use LLM to write Python code to convert the SMILES string into a Morgan frequency fingerprint (MFF) and the corresponding substructure graph of the fingerprint;

[0011] S4. According to the LLM recommendation, the molar ratio of dianhydride and diamine is multiplied in MFF as a percentage, and the MFF containing two dianhydrides or diamines is summed according to expert knowledge.

[0012] S5. Using TS as the target variable and MFF as the variable, perform data preprocessing. For the initially selected dataset, randomly divide it into training and test sets for subsequent modeling.

[0013] S6. Use the RFE-nested XGBoost method to filter variables based on feature importance, retain the required number of features, and obtain the optimal subset of variables by comparing the five-fold cross-validation results of XGBoost modeling.

[0014] S7. Optimize the model parameters and use the XGBoost regression method to establish a fast prediction model for TS with co-aggregated PI.

[0015] S8. Based on the established rapid prediction model of TS for copolymerized and homopolymerized PI, predict the TS values ​​of PI with different ratios of dianhydride and diamine without TS values. Input these samples into the constructed machine learning model to quickly predict the TS values ​​of PI.

[0016] S9. Based on the built-in feature importance function of the XGBoost algorithm and the LIME method, obtain the positive and negative relationship between features and the target, and further explain the relationship between the variable and the target variable.

[0017] Preferably, in step S1, the required table examples are input into a customized LLM, and the published literature is input into the LLM to obtain a table of dianhydride and diamine names and tensile strength values, which are then organized and proofread.

[0018] The execution process of step S1, customizing the large language model for text mining, is as follows:

[0019] S11. Name the LLM file "Polyimide Pro", specifying its purpose as text content extraction. Importantly, provide examples in the backend.

[0020] S12. Upload the case study article (PDF) to guide the LLM to find the names of dianhydride and diamine and the tensile strength value in the article. At the same time, upload the table and image information in the article to reduce "illusions" in time.

[0021] S13. Create a standardized output table to output the correct answers from S12 into a table.

[0022] S14. On the LLM main page, select "Polyimide Pro" to jump to step S12 for testing until the correct table content is completed.

[0023] S15. Continue with the text conversion table for the next article.

[0024] Preferably, in step S2, the chemical structure diagrams of dianhydrides and diamines in the table are converted into SMILES strings using LLM, and the structures are generated, modified, and then converted back into SMILES strings in ChemOffice.

[0025] Preferably, in step S3, the LLM is required to write Python code to convert the table containing the string SMILES into an MFF table and output the substructure diagram corresponding to the fingerprint.

[0026] Preferably, in step S4, LLM polymer expert knowledge and reactivity ratio issues are input, and LLM provides suggestions to multiply the molar ratio of dianhydride and diamine in MFF as a percentage, and to sum the MFF values ​​of structures containing two dianhydrides or diamines.

[0027] Preferably, in steps S5-9, according to the data structure and the explicit requirements of the standardized language, Python code for the feature selection, modeling, parameter optimization, and model interpretation processes of the machine learning workflow is written using LLM.

[0028] Step S6: The standardized language for the machine learning process is written sequentially for the large language model as follows:

[0029] S61. Input: Table name (.csv), feature name, the first column is the target, and the second column onwards are variables;

[0030] S62. For the training and test sets, utilize the features to perform XGBoost modeling (including parameter names), and use the RMSE and R-squared results from five-fold cross-validation. 2 As an indicator;

[0031] S63. Output: Cross-validation results of the training set and results of the test set.

[0032] A system for predicting the tensile strength of polyimide using a large language model (LLM) for performing the aforementioned method is characterized by including a large language model (LLM) and a machine learning process:

[0033] Input module 1: Publicly available articles on the synthesis of PI and molecular diagrams of dianhydrides and diamines are used as input data;

[0034] Data Analysis Module 1: Executes LLM using the data obtained from the input module;

[0035] Output module 1: Table of dianhydride and diamine names and tensile strengths, and structural samples of dianhydrides and diamines;

[0036] Input Module 2: Dianhydride and diamine addition MFF and TS tables;

[0037] Data Analysis Module 2: Using the data obtained from the input module, execute the TS method based on the XGBoost model to predict the tensile strength of unlabeled polyimide for PIs with different dianhydrides and diamines.

[0038] Output module 2: Outputs the tensile strength of polyimide.

[0039] Compared with the prior art, the present invention has the following obvious and prominent substantive features and significant advantages:

[0040] 1. This invention discloses a method and system for predicting the tensile strength of polyimide (PI) using a large language model (LLM). This method accelerates the in-depth development and exploratory data analysis of PI by interdisciplinary researchers. Specifically, it utilizes a large language model (LLM) to extract copolymer or homopolymer monomers of PI from text, converts the images of dianhydrides and diamines into the string "SMILES," and uses LLM code to convert this into a Morgan frequency-weighted fingerprint, feature selection, model selection, parameter optimization, and model interpretation. The established model can quickly predict PI without a tensile strength (TS) value. This invention, based on the entire process of machine learning assisted by a large language model, demonstrates interdisciplinary innovation in its rapid PI TS prediction model, integrating data science, computational chemistry, and artificial intelligence technologies. It boasts advantages such as high efficiency, user-friendliness, and accuracy.

[0041] 2. The focus of this invention is the entire process of machine learning assisted by a large language model, including text extraction, image conversion, expert knowledge suggestions, molecular structure expression, model code writing, model analysis, and using the model to predict the tensile strength of polyimide. This provides technical support and professional knowledge guidance for interdisciplinary researchers to use AI tools to intelligently obtain the mechanical properties (tensile strength TS) of polyimide (PI).

[0042] 3. This invention, based on relevant literature published by the PI, uploads text to an LLM (Liquid Microsystem Manager) and inputs it into the LLM using a standardized language. Leveraging the advantages of the LLM, the names of dianhydrides and diamines, along with their corresponding TS (Structured Total Scale) values, are extracted from the text. This process is simple, fast, and efficient. Furthermore, the LLM is used to convert the corresponding dianhydride and diamine chemical structure diagrams into SMILES strings. The structures are then generated in ChemOffice, modified, and converted back into SMILES strings. The structural data and performance values ​​of the samples can be obtained from text and image information in just a few seconds, forming a unified dataset.

[0043] 4. Based on the existing expert knowledge and suggestions of LLM, this invention uses a weighted summation of the molar ratio of dianhydrides and diamines in percentage form to weight the molecular fractional flow (MFF). Taking into account the reactivity of dianhydrides and diamines in copolymerized PI polymers, the invention sums the MFFs of structures containing either two dianhydrides or diamines, providing professional polymer chemistry advice for interdisciplinary researchers. Python code written using LLM is used to convert the SMILES string into an MFF, providing support for interdisciplinary researchers in the field of computer science and enabling rapid expression of molecular structural information. Considering the poor interpretability of MFFs, LLM also includes code to obtain the corresponding substructure diagrams, providing direct interpretability and helping interdisciplinary researchers quickly solve problems.

[0044] 5. This invention utilizes LLM to write Python code to perform a series of machine learning processes, including feature selection, model selection, parameter optimization, and model interpretation. It provides professional assistance to interdisciplinary R&D personnel in computer science, enabling them to gain a deeper understanding of the relationship between molecular structure and TS, and thus helping modelers to deeply understand and apply the model.

[0045] 6. This invention helps researchers quickly obtain the TS value of PI in experimental samples. The process does not involve experiments or chemicals, does not generate environmental pollution, conforms to the concept of green environmental protection, and obtains the TS value of PI quickly, efficiently and economically. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the overall process of an embodiment of the present invention.

[0047] Figure 2 This embodiment of the invention utilizes LLM to extract text information graphs.

[0048] Figure 3 This embodiment of the invention utilizes LLM to convert molecular structures into SMILES string diagrams.

[0049] Figure 4 This is a Python code dialogue diagram illustrating how LLM outputs MFF and corresponding substructure graphs in an embodiment of the present invention.

[0050] Figure 5This is a dialogue diagram based on LLM suggestion weighted MFF in an embodiment of the present invention.

[0051] Figure 6 This is a diagram showing the REF nested XGBoost feature filtering results of an embodiment of the present invention.

[0052] Figure 7 The figure shows the results of leave-one-out cross-validation and independent test set for the XGBoost regression model predicting PI using TS in an embodiment of the present invention.

[0053] Figure 8 This is an example of the built-in feature importance graph of XGBoost and the positive / negative relationship graph between features and targets obtained by the LIME method in this embodiment of the invention.

[0054] Figure 9 The highlighted molecular structure diagram is used to explain the model in the embodiment of the present invention. Detailed Implementation

[0055] The above solution will be further described below with reference to specific embodiments. The preferred embodiments of the present invention are described in detail below:

[0056] Example 1:

[0057] See Figure 1-9 The method for predicting the tensile strength of polyimide using a large language model, as provided in this embodiment, includes the following steps:

[0058] S1. Input publicly available articles on the tensile strength of homopolymer or copolymer PI into the Large Language Model (LLM). The normalized language extracts the names of dianhydrides and diamines and their corresponding TS values ​​from the text to form a database.

[0059] S2. Use LLM to convert the corresponding dianhydride and diamine chemical structure diagrams into SMILES strings. Generate the structure in ChemOffice, modify it, and then convert it back into SMILES strings to form dianhydride and diamine strings and TS numerical datasets.

[0060] S3. Use LLM to write Python code to convert the SMILES string into a Morgan frequency fingerprint (MFF) and the corresponding substructure graph of the fingerprint;

[0061] S4. According to LLM recommendations, the molar ratio of dianhydride and diamine is weighted by MFF in percentage form, and the MFF of structures containing two dianhydrides or diamines is summed according to expert knowledge.

[0062] S5. Using TS as the target variable and MFF as the variable, perform data preprocessing. For the initially selected dataset, randomly divide it into training and test sets for subsequent modeling.

[0063] S6. Use LLM to output Python code, select the RFE nested XGBoost method to filter variables based on feature importance, retain the required number of features, and obtain the optimal subset of variables by comparing the five-fold cross-validation results of XGBoost modeling.

[0064] S7. Use LLM to output Python code, optimize the model parameters, and use the XGBoost regression method to build a fast prediction model for TS with co-aggregated PI.

[0065] S8. Based on the established rapid forecasting model, forecast experimental samples without TS values ​​that have different ratios of dianhydride and diamine PI.

[0066] S9. Based on the built-in feature importance function of the XGBoost algorithm and the LIME method, obtain the positive and negative relationship between features and the target, and further explain the relationship between variables and target variables;

[0067] The system for predicting the tensile strength of polyimide using a large language model (LLM) provided in this embodiment is used to execute the above method, and includes a large language model (LLM) and a machine learning process:

[0068] Input module 1: Publicly available articles on the synthesis of PI and molecular diagrams of dianhydrides and diamines are used as input data;

[0069] Data Analysis Module 1: Executes LLM using the data obtained from the input module;

[0070] Output module 1: Table of dianhydride and diamine names and tensile strengths, and structural samples of dianhydrides and diamines;

[0071] Input Module 2: Dianhydride and diamine addition MFF and TS tables;

[0072] Data Analysis Module 2: Using the data obtained from the input module, execute the TS method based on the XGBoost model to predict the tensile strength of unlabeled polyimide for PIs with different dianhydrides and diamines.

[0073] Output module 2: Outputs the tensile strength of polyimide.

[0074] This invention provides interdisciplinary researchers with professional polymer chemistry advice and computer programming support to quickly and accurately predict the total saturation (TS) of copolymerized polymers (PIs). Through an LLM-assisted machine learning workflow, it achieves the conversion from text to tables, from images to SMILES strings, MFF weighted summation calculations, model code writing, analysis of structure-property relationships, and prediction of PI TS using the model. This method leverages artificial intelligence to reduce the burden of manual labor and programming costs for interdisciplinary researchers, achieving advantages such as high cost-effectiveness, ease of operation, high efficiency, and high data integrity and accuracy. Furthermore, this method requires no actual chemical experiments or complex calculations and generates no environmental pollution, aligning with green and environmentally friendly research and development principles.

[0075] Example 2

[0076] This embodiment is an improvement on Embodiment 1, and further includes the following differences based on Embodiment 1:

[0077] In this embodiment, in step S1, the published article on PI synthesis and the table image of the formulation within the article are imported into LLM. The normalized language is input into LLM, requesting the output table with the following header: literature name|sample name|dianhydride 1|mollar ratio|dianhydride 2|mollar ratio|diamine 1|mollar ratio|diamine 2|mollar ratio|tensile strength (MPa)|T5% (°C). Figure 2 The names of dianhydrides and diamines and their corresponding TS values ​​were extracted from the text to form a dataset.

[0078] In this embodiment, in step S2, the corresponding dianhydride and diamine chemical structure diagrams are uploaded to LLM, requiring them to be converted into SMILES strings. The structures are then generated in ChemOffice, modified, and converted back into SMILES strings, as shown below. Figure 3 As shown, dianhydride and diamine strings and TS numerical datasets are formed.

[0079] This embodiment utilizes LLM to extract textual information and convert molecular structures to SMILES. These methods provide results within minutes, offering convenience and speed while reducing the cost of manual data entry and retrieval. Furthermore, LLM has storage and continuous learning capabilities, allowing for correction of "illusions" that may arise, thereby further improving accuracy. Based on reliable expert knowledge and interactive question-and-answer functionality from a large language model, this embodiment can understand polymer-specific contexts, generating responses based on the context and autonomously identifying the topic of the current dialogue, thus better meeting user needs.

[0080] Example 3

[0081] This embodiment is an improvement on Embodiments 1 and 2, and further includes the following differences based on Embodiments 1 and 2:

[0082] In this embodiment, step S3 involves using LLM to write Python code to convert the SMILES string into a Morgan Frequency Fingerprint (MFF) and its corresponding substructure graph. First, the process of inputting the Python code for generating the Extended Connection Fingerprint (ECFP) using LLM is described below. Figure 4 In the dialog box, enter 1. ① Output to an Excel spreadsheet, and enter the string "smiles": *C(=O)Nc1ccc(cc1)Oc1ccc(cc1)N1C(=O)c2c(C1=O)cc(cc2)*,

[0083] *C(=O)Nc1cccc(c1)C#Cc1cccc(c1)N1C(=O)c2c(C1=O)cc(cc2)* ② Convert the SMILES string to a molecule object using RDKit; ③ Use the AllChem.GetMorganFingerprintAsBitVect function to calculate the ECFP of the molecule with a specified radius (2 in this example). This ECFP will be represented as a binary string, with each bit indicating the presence of a specific substructure in the molecule; ④ Through

[0084] The `AllChem.GetMorganFingerprint` function generates an `info` dictionary, which can retrieve the substructure for each bit activation information, along with its corresponding atom index and radius. After running the code, the following results will be obtained: a binary string representing the ECFP fingerprint, such as '100001000110...'; an `info` dictionary where the keys are the activated bits in the binary string (i.e., bits with a value of '1'), and the values ​​are a list of tuples, each containing the atom index and the radius used to generate that bit. The output ECFP result effectively represents the molecular structure. Next, input an LLM, such as... Figure 4 Enter 2 in the dialog box, "Of these 1024 descriptors, only a few substructures correspond to them; explain the principle." The output shows that descriptors cannot always describe all the characteristics of each molecule in detail. Third, upload the literature containing the Morgan fingerprint (MFF), requiring a comparison between the MFF and ECFP. For example... Figure 4As shown in dialog box output 3, ECFP is a general fingerprint commonly used in various cheminformatics applications, including drug discovery, compound similarity search, and structure-activity relationship modeling. If your goal is to perform a broad analysis of molecular structures without focusing on specific properties, then ECFP may be more suitable. If the frequency of substructures in the dataset is important, especially if it relates to specific properties such as thermal conductivity (as described in the article), then MFF is more suitable. MFF captures not only the presence of substructures but also their frequency, providing insights into how these substructures contribute to certain properties. Next, we utilize the Python code for MFF output from LLM to help interdisciplinary researchers directly obtain descriptor values, such as... Figure 4 Dialog box input / output 4. Finally, considering the interpretability of fingerprints, the LLM is required to simultaneously obtain the substructure diagram corresponding to the MFF and highlight it, such as... Figure 4 Dialog box input / output 5.

[0085] In this embodiment, step 4, according to LLM recommendations, involves weighting the molar ratio of dianhydrides and diamines in MFFs as a percentage, and summing the MFFs containing two dianhydrides or diamines based on expert knowledge. Figure 5 As shown in the MLL dialogue diagram, following LLM recommendations, separate columns are created for the proportions of dianhydrides or diamines for each component. Furthermore, descriptors are weighted according to the proportions of the components. For example, if a formulation is a mixture of PMDA and ODA, its descriptor could be a weighted average of the descriptors for both.

[0086] This embodiment of the method provides professional code for molecular structure expression based on LLM, substructure diagrams to help interpret molecular fingerprint descriptors, and expert advice for two dianhydride or diamine cases of copolymer PI. LLM assistance offers low-cost, highly accurate, and disciplinaryly extensible interdisciplinary professional support.

[0087] Example 4

[0088] This embodiment is an improvement on embodiments one to three, and further includes the following differences based on embodiments one to three:

[0089] After preprocessing (removing descriptors with Pearson correlation coefficients > 0.90), the dataset was initially screened to obtain 77 descriptors. In this embodiment, step S6 utilizes LLM to output Python code, selects the RFE-nested XGBoost method to filter variables based on feature importance, obtains the optimal variable subset by comparing the five-fold cross-validation results of XGBoost modeling, and selects the optimal feature subset for modeling, reducing the original 77 features to 11 features. Figure 6 It can be seen that the 11th feature is the optimal one.

[0090] In this embodiment, step S7 utilizes LLM to output Python code, model, and optimize parameters. The optimal model parameters are as follows: learning_rate is 0.01, max_depth is 5, subsample is 0.7, colsample_bytree is 0.7, and n_estimators is 300. Based on the above model parameters, the five-fold cross-validation results of the established model are R... 2 The RMSE values ​​were 0.88 and 30.16, respectively. The results for the actual and predicted values ​​on the independent test set were R0 and R1, respectively. 2 The RMSE values ​​are 0.81 and 27.04 (e.g., ...). Figure 7 (As shown).

[0091] This embodiment utilizes an LLM (Limited Language Modeling) to automate Python code writing for feature engineering, modeling, and parameter optimization, establishing a model with good generalization ability and consistent performance. LLM offers expertise across various fields, particularly in language and text question answering, providing guidance and rapid code writing services for interdisciplinary researchers. It also evaluates the stability and reliability of data modeling methods, yielding good results and offering advantages such as low cost and environmental friendliness.

[0092] Example 5

[0093] This embodiment is an improvement on embodiments one to four, and further includes the following differences based on embodiments one to four:

[0094] In this embodiment, step S9 is described in the following example. Figure 8 According to the XGBoost algorithm, the importance ranking of the built-in features is as follows: A27979804 > A2041434490 > A1100037548 > A2246728737 > G3974328374 >

[0095] A4256002933>G3276511768>G1409751636>A4216335232>G2072597505>

[0096] A2698059862. The LIME method yielded positive and negative relationships between features and the target, indicating that A27979804, A2041434490, A1100037548, A4256002933, G3276511768, G1409751636, and A4216335232 are negatively correlated, while A2246728737, G3974328374, G2072597505, and A2698059862 are positively correlated.

[0097] In this embodiment, combined with Figure 8and Figure 9 This study explains the relationship between structural characteristics and target performance (TS) and establishes a QSPR model. In amines, the nitrogen atom pairs in the benzopyrrole ring and benzene ring are negatively correlated with the target performance, especially the ortho-position nitrogen atoms in the benzene ring, as evidenced by A27979804, A2041434490, and A1100037548. Only A2246728737 (methyl group) shows a strong positive correlation. A4256002933 and A4216335232 show a negative correlation with the ortho and meta carbon atoms of the amino group. A2698059862 is not prominent and therefore not discussed. In anhydrides, in the naphthoquinone structure of G3974328374, the more carbon atoms connected to the six-membered and five-membered rings, the more favorable the TS. In the naphthoquinone structure of G3276511768, the fewer oxygen atoms in the furan ring, the more favorable the TS. The carbon atoms of the carbon-oxygen double bond in the furan ring and the carbon atoms of the benzene ring connected to the S atom in the naphthoquinone structure show a weak negative correlation and a weak positive correlation, respectively, as shown in G1409751636 and G2072597505.

[0098] This embodiment uses an LLM (Learning Library) to automatically generate Python code to implement XGBoost's built-in feature importance function and the LIME method to obtain the positive and negative relationships between features and targets, providing tools for interpreting the relationship between structure and performance. By automatically generating code, LLM can help users save time in finding suitable code libraries and learning how to use them. LLM can enable non-data science experts to understand and perform data analysis tasks, making machine learning more democratized.

[0099] Example 6

[0100] This embodiment is an improvement on embodiments one through five, and further includes the following differences based on embodiments one through five:

[0101] In this embodiment, step S8 utilizes seven copolymer PIs with different ratios of dianhydride and diamine without TS values ​​collected from the literature, obtained in steps 1-2. These are then converted into MFF descriptors and a prediction test set is established in steps S3-4. These samples are fed into the rapid TS prediction model for PIs established in steps S5-7 to predict the TS values ​​of the copolymer PIs. The model is invoked using Python for prediction, and the prediction results are shown in Table 1 (Descriptor values ​​and prediction results for PIs without TS values).

[0102] Table 1

[0103]

[0104] This embodiment uses seven collected samples to build a pre-established model to predict their corresponding TS (transient thrombus). The process is quick and pollution-free, providing researchers with a rapid prediction tool and accelerating the research process.

[0105] Example 7

[0106] This embodiment is basically the same as all the embodiments described above, except that:

[0107] A system for predicting the tensile strength of polyimide using a large language model (LLM), performing the methods described in Examples 1-6, includes a base LLM and interconnected and communicating modules as follows:

[0108] Input module 1: Publicly available articles on the synthesis of PI and molecular diagrams of dianhydrides and diamines are used as input data;

[0109] Data Analysis Module 1: Executes LLM using the data obtained from the input module;

[0110] Output module 1: Table of dianhydride and diamine names and tensile strengths, and structural samples of dianhydrides and diamines;

[0111] Input Module 2: Dianhydride and diamine addition MFF and TS tables;

[0112] Data Analysis Module 2: Using the data obtained from the input module, execute the TS method based on the XGBoost model to predict the tensile strength of unlabeled polyimide for PIs with different dianhydrides and diamines.

[0113] Output module 2: Outputs the tensile strength of polyimide.

[0114] The system provided in this embodiment utilizes the expertise of LLM (Liquid Mover) across various fields to facilitate a comprehensive machine learning workflow for rapidly predicting the tensile strength (TS) values ​​of polyimides (PIs) from different dianhydrides and diamines. This includes text extraction, image conversion, expert knowledge suggestions, molecular structure expression, model code writing, model analysis, and using the model to predict the tensile strength of polyimides. In particular, LLM's professional coding capabilities and the integration of chemical knowledge with molecular structure expression and conversion to construct an XGBoost regression model for predicting TS values ​​help cross-disciplinary researchers efficiently discover novel PIs.

[0115] In summary, the method and system for predicting the tensile strength of polyimide using a large language model (LLM) provided in the above embodiments include the following core steps: S1, inputting publicly available articles on the tensile strength of homopolymer or copolymer PI into an LLM to extract the names of dianhydrides and diamines and their corresponding TS values ​​to form a database; S2, using an LLM to convert the corresponding dianhydride and diamine chemical structure diagrams into SMILES strings; S3, using an LLM to write Python code to convert the SMILES strings into Morgan frequency fingerprints (MFFs) and the corresponding substructure diagrams; S4, according to the LLM recommendations, expressing the molar ratio of dianhydrides and diamines as a percentage. S5. Perform data preprocessing; S6. Use LLM to output Python code to filter variables and model; S7. Use LLM to output Python code to optimize parameters; S8. Predict samples without TS values ​​based on the established fast prediction model; S9. Obtain the positive / negative relationship interpretation model between features and targets based on the built-in feature importance function of the XGBoost algorithm and the LIME method. The above embodiments are based on reliable literature real values, a fully intelligent large language model and a robust machine learning model. The process of establishing the PI monomer TS value prediction model has the advantages of intelligent acquisition, convenience and speed, and no chemical pollution.

[0116] The prediction methods described in the above embodiments focus on leveraging the expertise of LLM (Limited Learning Model) across various fields to assist the entire machine learning process. This includes memory storage and timely updates, as well as interaction with the LLM. The operation is simple, low-cost, and pollution-free, requiring only interdisciplinary researchers with intermediate-level domain knowledge. The methods in these embodiments involve uploading publicly available literature and standardized language, using LLM to extract text information and create tables, eliminating the need for mechanical copying and pasting, thus reducing labor costs and time. Based on LLM expert advice, the methods in these embodiments weight and sum structures containing two dianhydrides or diamines according to their reactivity ratios, creating innovative descriptor expressions. The methods in these embodiments utilize LLM to write Python code for a series of machine learning processes, including feature selection, model selection, parameter optimization, and model interpretation. This requires logical input and clearly defined output requirements, ensuring professionalism and directness. The methods in these embodiments establish a model for rapidly predicting the TS (Transmission Time) values ​​of Principal Injectors (PIs), quickly predicting the TS values ​​of candidate PIs, providing a reference for researchers developing new PIs, shortening development time, and reducing development costs. The methods described above accelerate the initial development process and exploratory data analysis for interdisciplinary researchers, reduce the difficulty for non-data science experts to understand and execute data, and reduce the process of repeated experiments and trial and error.

[0117] The method and system for predicting the tensile strength of polyimide using a large language model (LLM) provided in the above embodiments of this invention offer technical support and professional knowledge guidance for interdisciplinary researchers to intelligently obtain the mechanical properties (tensile strength TS) of polyimide (PI) using AI tools. This accelerates the in-depth development process and exploratory data analysis of homopolymer polyimide by interdisciplinary chemists. By using a large language model (LLM) to extract copolymers or homopolymer monomers of PI from text, and converting the images of dianhydrides and diamines into the string "SMILES", the LLM code is converted into Morgan frequency weighted sum fingerprints, feature selection, model selection, parameter optimization, and model interpretation. The established model can quickly predict PI without a TS value. This invention, based on the entire process of machine learning assisted by a large language model, demonstrates interdisciplinary innovation in its rapid PI TS prediction model, integrating data science, computational chemistry, and artificial intelligence technologies, and possesses advantages such as high efficiency, user-friendliness, and accuracy.

[0118] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made according to the purpose of the invention. Any changes, modifications, substitutions, combinations or simplifications made based on the spirit and principle of the technical solution of the present invention shall be equivalent substitutions. As long as they meet the purpose of the invention and do not deviate from the technical principle and inventive concept of the present invention, they shall fall within the protection scope of the present invention.

Claims

1. A method for predicting the tensile strength of polyimide using a large language model, characterized in that, Includes the following steps: S1. Using the Web of Science database, search for publicly available articles on synthesizing PIs, normalize the language input to a large language model (LLM), and extract the names of dianhydrides and diamines and their corresponding TS values ​​from the text to form a dataset. S2. Use LLM to convert the corresponding dianhydride and diamine chemical structure diagrams into SMILES strings, generate the structure in ChemOffice, modify it, and then convert it back into SMILES strings. S3. Use LLM to write Python code to convert the SMILES string into Morgan frequency fingerprint MFF and the corresponding substructure graph of the fingerprint. S4. According to the LLM recommendation, the molar ratio of dianhydride and diamine is multiplied in MFF as a percentage, and the MFF containing two dianhydrides or diamines is summed according to expert knowledge. S5. Using TS as the target variable and MFF as the variable, perform data preprocessing. For the initially selected dataset, randomly divide it into training and test sets for subsequent modeling. S6. Use the RFE-nested XGBoost method to filter variables based on feature importance, retain the required number of features, and obtain the optimal subset of variables by comparing the five-fold cross-validation results of XGBoost modeling. S7. Optimize the model parameters and use the XGBoost regression method to establish a fast prediction model for TS with co-aggregated PI. S8. Based on the established rapid prediction model of TS for copolymerized and homopolymerized PI, predict the TS values ​​of PI with different ratios of dianhydride and diamine without TS values. Input these samples into the constructed machine learning model to quickly predict the TS values ​​of PI. S9. Based on the built-in feature importance function of the XGBoost algorithm and the LIME method, obtain the positive and negative relationship between features and the target, and further explain the relationship between the variable and the target variable.

2. The method for predicting the tensile strength of polyimide using a large language model according to claim 1, characterized in that, Step S1 also includes inputting the required table examples into a customized LLM, inputting publicly available literature into the LLM, obtaining a table of dianhydride and diamine names and tensile strength values, and then organizing and proofreading it.

3. The method for predicting the tensile strength of polyimide using a large language model according to claim 2, characterized in that, The execution process of step S1, customizing the large language model for text mining, is as follows: S11. Name the LLM file "Polyimide Pro", specifying its purpose as text content extraction. Importantly, provide examples in the backend. S12. Upload case study articles to guide LLM to find the names of dianhydrides and diamines and tensile strength values ​​in the article, and upload the table and image information in the article to reduce "illusions" in time; S13. Create a standardized output table and output the correct answers from step S12 into a table. S14. On the LLM main page, select "Polyimide Pro" to jump to step S12 for testing until the correct table content is completed; S15. Continue with the text conversion table for the next article.

4. The method for predicting the tensile strength of polyimide using a large language model according to claim 1, characterized in that, In step S2, LLM is used to convert the chemical structure diagrams of dianhydrides and diamines in the table into SMILES strings. The structures are generated in ChemOffice, modified, and then converted back into SMILES strings.

5. The method for predicting the tensile strength of polyimide using a large language model according to claim 1, characterized in that, In step S3, LLM is required to write Python code to convert a table containing the string SMILES into a table in MFF and output the substructure diagram corresponding to the fingerprint.

6. The method for predicting the tensile strength of polyimide using a large language model according to claim 1, characterized in that, In step S4, LLM polymer expert knowledge and reactivity ratio are input, and LLM provides suggestions to multiply the molar ratio of dianhydride and diamine in MFF as a percentage, and to sum the MFF values ​​of structures containing two dianhydrides or diamines.

7. The method for predicting the tensile strength of polyimide using a large language model according to claim 1, characterized in that, In steps S5-S9, according to the data structure and the explicit requirements of the standardized language, Python code for the feature selection, modeling, and parameter optimization process of the machine learning workflow is written using LLM. Step S6: The standardized language for writing the machine learning workflow for each of the six major language models is as follows: S61. Input: table_name.csv, feature_name, the first column is the target, and the second column onwards are variables; S62. For the training and test sets, utilize features to perform XGBoost modeling, including parameter names, modeling, and RMSE and R-squared results using five-fold cross-validation. 2 As an indicator; S63. Output: Cross-validation results of the training set and results of the test set.

8. A system for predicting the tensile strength of polyimide using a large language model, characterized in that, The method for performing any one of claims 1-7 includes a large language model and a machine learning process: Input module 1: Publicly available articles on the synthesis of PI and molecular diagrams of dianhydrides and diamines are used as input data; Data Analysis Module 1: Executes LLM using the data obtained from the input module; Output module 1: Table of dianhydride and diamine names and tensile strengths, and structural samples of dianhydrides and diamines; Input Module 2: Dianhydride and diamine addition MFF and TS tables; Data Analysis Module 2: Using the data obtained from the input module, execute the TS method based on the XGBoost model to predict the tensile strength of unlabeled polyimide for PIs with different dianhydrides and diamines. Output module 2: Outputs the tensile strength of polyimide.

Images