Analysis method, system and equipment of polymer performance prediction model and medium

Abstract

The invention discloses an analysis method, system and equipment of a polymer performance prediction model and a medium, and belongs to the technical field of artificial intelligence technology. The analysis method of the polymer performance prediction model comprises the following steps: constructing a data set of various polymers; determining molecular feature information of each molecular structural formula in the data set; constructing an algorithm warehouse integrated with a plurality of artificial intelligence algorithms; if the request information is received, determining a target parameter according to the request information; performing model training based on the target parameters to obtain a plurality of polymer performance prediction models; and evaluating the prediction effect of the polymer performance prediction model, and generating a corresponding analysis result. According to the method, the prediction effect of the polymer performance prediction model can be accurately evaluated.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an analysis method, system, device and medium for polymer performance prediction models. Background Technology

[0002] Polymer materials, as the cornerstone of modern industry, are widely used in key fields such as energy storage, biomedicine, and electronic information. However, the traditional trial-and-error R&D model is inefficient and cannot meet the demand for rapid design of high-performance materials. In recent years, breakthroughs in computer science and artificial intelligence technologies have provided a new paradigm for polymer material R&D: establishing "structure-property" relationships through machine learning modeling is expected to significantly shorten the R&D cycle and reduce costs.

[0003] Although machine learning-based modeling methods have shown significant advantages in efficiency and predictive ability, the sheer number of artificial intelligence algorithms makes it difficult to accurately assess the capabilities of polymer performance prediction models built by each algorithm, thus hindering effective decision-making in polymer research and development.

[0004] Therefore, how to accurately evaluate the predictive performance of polymer performance prediction models is a technical problem that needs to be solved by those skilled in the art. Summary of the Invention

[0005] The purpose of this application is to provide an analysis method, system, device, and medium for polymer performance prediction models, which can accurately evaluate the prediction effect of polymer performance prediction models.

[0006] To address the aforementioned technical problems, this application provides an analytical method for a polymer performance prediction model, comprising:

[0007] Construct datasets for various polymers; wherein the datasets include the molecular structural formulas of the corresponding polymers;

[0008] Determine the molecular feature information of each molecular structure in the dataset; wherein the types of molecular feature information include molecular descriptors, molecular fingerprints, and molecular graphs;

[0009] Build an algorithm repository that integrates multiple artificial intelligence algorithms;

[0010] If a request is received, the target parameters are determined based on the request; wherein, the target parameters include data selection information, polymer representation, and algorithm selection information, the data selection information is used to describe the dataset participating in model training, the polymer representation is used to describe the types of molecular feature information of the dataset participating in model training, and the algorithm selection information is used to describe the artificial intelligence algorithm participating in model training;

[0011] Based on the target parameters, multiple polymer performance prediction models are obtained through model training; wherein, in the training process of a single polymer performance prediction model, the number of datasets, polymer representations, and artificial intelligence algorithms used are all 1.

[0012] The predictive performance of the polymer performance prediction model is evaluated, and corresponding analysis results are generated.

[0013] Optionally, after obtaining multiple polymer performance prediction models, the following may also be included:

[0014] The polymer performance prediction model is interpreted to obtain the model interpretation result; wherein the model interpretation result includes the degree of contribution of each feature in the input data of the polymer performance prediction model to the model prediction result;

[0015] The parameters of the polymer performance prediction model are adjusted based on the model interpretation results.

[0016] Optionally, constructing datasets for various polymers includes:

[0017] The raw polymer data is read from the database, and the raw polymer data is formatted and quality evaluated. The raw polymer data with a quality evaluation score greater than a preset value is selected as the first intermediate data.

[0018] The first intermediate data is deduplicated using regular expressions, and the deduplicated data is then integrated to obtain the second intermediate data.

[0019] The second intermediate data is subjected to unit identification unification and outlier filtering operations to obtain the third intermediate data;

[0020] The third intermediate data is cleaned to obtain the fourth intermediate data, and the integrity of the fourth intermediate data is verified.

[0021] Set the fourth intermediate data that passed the data integrity verification as the alternative polymer data;

[0022] If the candidate polymer data includes property data, then a dataset containing the molecular structure formula and property data of the polymer is constructed based on the candidate polymer data; wherein, the property data includes any one or a combination of physical properties, chemical properties, thermodynamic properties and mechanical properties;

[0023] If the candidate polymer data does not contain property data, then a dataset containing the molecular structure formulas of the polymers is constructed based on the candidate polymer data.

[0024] Optionally, the predictive performance of the polymer performance prediction model is evaluated, including:

[0025] The first evaluation result is determined based on the model evaluation index of the polymer performance prediction model; wherein, the model evaluation index is a first type index or a second type index; the first type index includes the coefficient of determination, mean square error and mean absolute error, and the second type index includes the F1 score and the area under the ROC curve; the F1 score is used to describe the harmonic mean of precision and recall, and the ROC curve represents the receiver operating characteristic curve;

[0026] The second evaluation result is determined based on the training time and memory usage of the polymer performance prediction model;

[0027] Construct a scatter plot containing the actual values ​​and model predictions, and generate a third evaluation result based on the scatter plot.

[0028] Optionally, based on the target parameters, model training is performed to obtain multiple polymer performance prediction models, including:

[0029] The data selection information, polymer representation method, and algorithm selection information in the target parameters are combined to obtain multiple training configurations; wherein each training configuration includes one dataset, one polymer representation method, and one artificial intelligence algorithm;

[0030] Perform the corresponding model training operation according to each training configuration to obtain multiple polymer performance prediction models.

[0031] Optionally, performing the corresponding model training operation according to each training configuration includes:

[0032] Select the corresponding target dataset according to the training configuration;

[0033] The polymer representation in the training configuration is used to convert the molecular structural formulas in the target dataset into corresponding molecular feature information.

[0034] The artificial intelligence algorithm in the training configuration is used to perform the corresponding model training operation based on the target dataset.

[0035] Optionally, the predictive performance of the polymer performance prediction model is evaluated, and corresponding analysis results are generated, including:

[0036] The prediction performance of the polymer performance prediction models is evaluated to obtain the model performance evaluation value for each polymer performance prediction model.

[0037] The analysis results of the polymer performance prediction model are generated based on the target parameters and the model effect evaluation values, and the analysis results are displayed on the front-end page.

[0038] This application also provides an analysis system for polymer performance prediction models, the system comprising:

[0039] A dataset construction module is used to construct datasets for various polymers; wherein, the datasets include the molecular structural formulas of the corresponding polymers;

[0040] A polymer representation module is used to determine the molecular feature information of each molecular structure in the dataset; wherein the types of molecular feature information include molecular descriptors, molecular fingerprints, and molecular graphs;

[0041] The algorithm management module is used to build an algorithm repository that integrates multiple artificial intelligence algorithms;

[0042] A parameter determination module is used to determine target parameters based on a request message if a request message is received. The target parameters include data selection information, polymer representation, and algorithm selection information. The data selection information describes the dataset participating in model training, the polymer representation describes the types of molecular feature information of the dataset participating in model training, and the algorithm selection information describes the artificial intelligence algorithm participating in model training.

[0043] The model training module is used to train the model based on the target parameters to obtain multiple polymer performance prediction models; wherein, in the training process of a single polymer performance prediction model, the number of datasets, polymer representations and artificial intelligence algorithms used are all 1.

[0044] The evaluation module is used to evaluate the prediction effect of the polymer performance prediction model and generate corresponding analysis results.

[0045] This application also provides a storage medium storing a computer program thereon, which, when executed, implements the steps of the analysis method for the above-described polymer performance prediction model.

[0046] This application also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor, when calling the computer program in the memory, implements the steps of the analysis method for the polymer performance prediction model described above.

[0047] This application provides an analysis method for polymer performance prediction models. The method constructs a dataset containing polymer molecular structures, determines the molecular feature information of each molecular structure, and builds an algorithm repository integrating multiple artificial intelligence algorithms. The dataset includes the molecular structures of the corresponding polymers, providing basic data for model training. The types of molecular feature information include molecular descriptors, molecular fingerprints, and molecular graphs, providing rich feature inputs for model training. Upon receiving a request, corresponding target parameters can be determined based on the request information. These target parameters include data selection information, polymer representation methods, and algorithm selection information. Based on these target parameters, the datasets, types of molecular feature information, and artificial intelligence algorithms participating in model training can be determined. Based on these target parameters, this application performs model training operations to obtain multiple polymer performance prediction models. During the training of a single polymer performance prediction model, the number of datasets, polymer representation methods, and artificial intelligence algorithms used is always one. This independent model training method ensures the independence and consistency of each model, facilitating subsequent evaluation and comparison. This application evaluates the prediction performance of the polymer performance prediction models and generates corresponding analysis results, allowing users to fully understand the model's performance and thus achieve accurate evaluation of the polymer performance prediction models. This application also provides an analysis system for polymer performance prediction models, a storage medium, and a storage medium, which have the above-mentioned beneficial effects, and will not be elaborated here. Attached Figure Description

[0048] To more clearly illustrate the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 A flowchart illustrating an analysis method for a polymer performance prediction model provided in this application embodiment;

[0050] Figure 2 This application provides a platform overall architecture diagram as an embodiment.

[0051] Figure 3 This application provides a platform functional architecture diagram as an embodiment.

[0052] Figure 4 This is a diagram illustrating the data volume of a polymer dataset provided in an embodiment of this application.

[0053] Figure 5 This application provides a scatter plot comparing the actual and predicted values ​​of the glass transition temperature of a polymer using the Mordred descriptor and a random forest model, as shown in the embodiments of this application.

[0054] Figure 6 This is a schematic diagram of the learning curve for the visualization analysis portion of the platform evaluation layer provided in an embodiment of this application;

[0055] Figure 7 This is a schematic diagram of SHAP analysis, a visualization analysis component of the platform evaluation layer, provided in an embodiment of this application.

[0056] Figure 8 A data distribution histogram of a polymer glass transition temperature dataset provided in this application embodiment;

[0057] Figure 9 A three-dimensional view of a ranking list provided in an embodiment of this application;

[0058] Figure 10 This is a schematic diagram illustrating a platform polymer details module provided in an embodiment of this application;

[0059] Figure 11 This is a schematic diagram illustrating the effect of a platform dataset module provided in an embodiment of this application;

[0060] Figure 12 This is a schematic diagram of an optimal combination of display modules provided in an embodiment of this application;

[0061] Figure 13 This is a schematic diagram illustrating the effect of a lollipop image evaluation algorithm provided in an embodiment of this application;

[0062] Figure 14 This is a schematic diagram illustrating the effect of a lollipop diagram for evaluating polymer representations, provided in an embodiment of this application. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0064] Please see below. Figure 1 , Figure 1 This is a flowchart illustrating an analysis method for a polymer performance prediction model provided in an embodiment of this application.

[0065] Specific steps may include:

[0066] S101: Construct a dataset for various polymers.

[0067] This embodiment can be applied to electronic devices with data processing capabilities. This step can construct datasets of various polymers based on literature data, experimental data, and various publicly available polymer databases. The datasets include the molecular structural formulas of the corresponding polymers.

[0068] As a feasible implementation method, this embodiment can set up a corresponding dataset for each polymer. The dataset includes the molecular structural formula of the polymer, and a polymer can have at least one molecular structural formula.

[0069] S102: Determine the molecular feature information of each molecular structure in the dataset.

[0070] Based on the construction of the polymer dataset, this embodiment can determine the molecular feature information of each molecular structure in the polymer dataset. The types of molecular feature information include molecular descriptors (such as Mordred descriptors), molecular fingerprints, and molecular graphs.

[0071] S103: Build an algorithm repository that integrates multiple artificial intelligence algorithms;

[0072] To provide diverse model selection, this step involves building an algorithm repository integrating various artificial intelligence algorithms to meet the needs of different datasets and tasks. This algorithm repository can store, manage, and access different artificial intelligence algorithms, including machine learning algorithms and / or deep learning algorithms.

[0073] S104: If a request is received, the target parameters are determined based on the request.

[0074] The aforementioned request information can be transmitted from other terminals or input by the user. The target parameters are determined by parsing the request information. These target parameters include data selection information, polymer representation method, and algorithm selection information.

[0075] The data selection information describes the datasets used for model training (e.g., datasets corresponding to the glass transition temperatures of polymers); specifically, the data selection information can be the identifier or serial number of the dataset. The polymer representation describes the types of molecular feature information in the datasets used for model training, such as at least one of molecular descriptors, molecular fingerprints, and molecular diagrams. The algorithm selection information describes the artificial intelligence algorithms used for model training; specifically, the algorithm selection information can be the identifier or serial number of the artificial intelligence algorithm. Before model training, the molecular structures in the dataset need to be converted into corresponding molecular feature information so that the polymer performance prediction model can be trained using the artificial intelligence algorithm and the dataset containing the molecular feature information.

[0076] S105: Based on the target parameters, perform model training to obtain multiple polymer performance prediction models;

[0077] Based on the target parameters, this embodiment can train multiple polymer performance prediction models one by one according to the dataset, polymer representation and algorithm selection determined by the analysis request.

[0078] During the training of a single polymer performance prediction model, the number of datasets, polymer representations, and artificial intelligence algorithms used is always one. Specifically, each model uses only one dataset, one type of molecular feature information, and one artificial intelligence algorithm during training to ensure independence and consistency. Multiple polymer performance prediction models can be generated in this way. A polymer performance prediction model is a model used to predict the properties of polymers.

[0079] S106: Evaluate the prediction effect of the polymer performance prediction model and generate corresponding analysis results.

[0080] This step involves measuring the predictive ability of the polymer performance prediction model by calculating its performance indicators on the test data, thereby generating analysis results and achieving an accurate evaluation of the model.

[0081] Specifically, this embodiment can evaluate the prediction effect of the polymer performance prediction model to obtain the model effect evaluation value of each polymer performance prediction model; generate the analysis results of the polymer performance prediction model based on the target parameters and the model effect evaluation value, and display the analysis results on the front-end page.

[0082] The above analysis results may include the ranking of the prediction accuracy of various polymer performance prediction models, and may also include the influence of any one or more of the following factors on the prediction accuracy of polymer performance prediction models: dataset, molecular feature information, and artificial intelligence model.

[0083] This embodiment constructs a dataset containing polymer molecular structures, determines the molecular feature information of each molecular structure, and builds an algorithm repository integrating multiple artificial intelligence algorithms. The dataset includes the molecular structures of the corresponding polymers, providing basic data for model training; the types of molecular feature information include molecular descriptors, molecular fingerprints, and molecular graphs, providing rich feature inputs for model training. Upon receiving a request, the corresponding target parameters can be determined based on the request information. These target parameters include data selection information, polymer representation methods, and algorithm selection information; based on these target parameters, the datasets, types of molecular feature information, and artificial intelligence algorithms participating in model training can be determined. Based on the above target parameters, this embodiment performs model training operations to obtain multiple polymer performance prediction models. During the training of a single polymer performance prediction model, the number of datasets, polymer representation methods, and artificial intelligence algorithms used is always 1; this independent model training method ensures the independence and consistency of each model, facilitating subsequent evaluation and comparison. This embodiment evaluates the prediction performance of the polymer performance prediction models and generates corresponding analysis results, allowing users to fully understand the model's performance and thus achieve accurate evaluation of the polymer performance prediction models.

[0084] As for Figure 1 Further description of the corresponding embodiment: After obtaining multiple polymer performance prediction models, model interpretation can be performed on the polymer performance prediction models to obtain model interpretation results; the parameters of the polymer performance prediction models can then be adjusted based on the model interpretation results. The model interpretation results include the contribution degree of each feature in the input data of the polymer performance prediction model to the model prediction results. Specifically, this embodiment can perform model interpretation on the polymer performance prediction models using SHAP analysis, and set the obtained SHAP values ​​as the model interpretation results.

[0085] As for Figure 1In a further description of the corresponding embodiment, this embodiment can also construct datasets for various polymers in the following manner: reading raw polymer data from a database, formatting and quality evaluating the raw polymer data, and selecting raw polymer data with a quality evaluation score greater than a preset value as first intermediate data; performing deduplication on the first intermediate data based on regular expressions, and integrating the deduplicated data to obtain second intermediate data; performing unit identification unification and outlier filtering operations on the second intermediate data to obtain third intermediate data; performing data cleaning on the third intermediate data to obtain fourth intermediate data, and verifying the data integrity of the fourth intermediate data; setting the fourth intermediate data that passes the data integrity verification as candidate polymer data; if the candidate polymer data contains property data, then constructing a dataset containing the molecular structure formula and property data of the polymer based on the candidate polymer data; wherein, the property data includes any one or a combination of physical properties, chemical properties, thermodynamic properties, and mechanical properties; if the candidate polymer data does not contain property data, then constructing a dataset containing the molecular structure formula of the polymer based on the candidate polymer data.

[0086] As for Figure 1 In a further description of the corresponding embodiment, this embodiment can also evaluate the prediction effect of the polymer performance prediction model in the following ways: determining a first evaluation result based on the model evaluation index of the polymer performance prediction model; determining a second evaluation result based on the training time and memory usage of the polymer performance prediction model; constructing a scatter plot containing the true values ​​and model prediction values; and generating a third evaluation result based on the scatter plot. Based on this, this embodiment can also generate corresponding analysis results based on the above-mentioned first evaluation result, second evaluation result, and third evaluation result.

[0087] Furthermore, the aforementioned model evaluation metrics can be either first-class or second-class metrics; the first-class metrics include the coefficient of determination, mean squared error, and mean absolute error, while the second-class metrics include the F1 score and the area under the ROC curve (ROC-AUC); the F1 score is used to describe the harmonic mean of precision and recall, and the ROC curve represents the receiver operating characteristic curve.

[0088] The polymer performance prediction model described above can be used for either regression or classification tasks. If the polymer performance prediction model is used for regression, the model evaluation index can be a first-class index. If the polymer performance prediction model is used for classification, the model evaluation index can be a second-class index.

[0089] As for Figure 1As further described in the corresponding embodiment, the process of obtaining multiple polymer performance prediction models based on the target parameters in this embodiment includes: combining the data selection information, polymer representation method, and algorithm selection information in the target parameters to obtain multiple training configurations; and performing corresponding model training operations according to each training configuration to obtain multiple polymer performance prediction models. Each training configuration includes one dataset, one polymer representation method, and one artificial intelligence algorithm.

[0090] Accordingly, the process of performing the corresponding model training operation according to each training configuration includes: selecting the corresponding target dataset according to the training configuration; converting the molecular structure formula in the target dataset into the corresponding molecular feature information using the polymer representation in the training configuration; and performing the corresponding model training operation based on the target dataset using the artificial intelligence algorithm in the training configuration.

[0091] Machine learning-based modeling methods have shown significant advantages in efficiency and predictive power, but existing technologies still face several limitations in practical applications, hindering further development in this field. Current polymer performance prediction and design mainly rely on the following three methods:

[0092] (1) Traditional experimental method.

[0093] This method relies on repeated synthesis and characterization to obtain experimental data, but a single polymer property / performance test typically takes several weeks and is very costly, making it difficult to meet the needs of large-scale, high-throughput material screening. This huge time and cost investment creates a direct bottleneck, highlighting the urgent need for a more efficient method.

[0094] (2) Computer simulation.

[0095] This method employs molecular dynamics and density functional theory, which can replace experimental work to some extent. However, it is highly dependent on high-performance computing resources, making it difficult to quickly process and screen large sample data, thus limiting its scalability and wide applicability.

[0096] (3) Data-driven approach.

[0097] While machine learning-based data-driven methods offer significant advantages in data processing and prediction efficiency, they face the following bottlenecks:

[0098] Insufficient universality: Most studies focus only on a single property (such as glass transition temperature) or a specific polymer type (such as polyolefin), lacking universal models that span properties and systems;

[0099] Data quality and scale are limited: effective polymer data in public databases is scarce and scattered, and the labeling standards are inconsistent. There is an urgent need for a unified and standardized benchmark platform to support large-scale in-depth data mining.

[0100] There is a lack of unified standards for comparing model performance: different models have different evaluation standards and significant black box characteristics, making it difficult to analyze structure-property relationships and achieve structure-oriented optimization. Without a common metric, progress will be fragmented and collaboration will be hindered.

[0101] Furthermore, existing technologies suffer from shortcomings in tool integration, ease of operation, and interdisciplinary collaboration, further limiting the value of transforming technological achievements into engineering applications. This gap between research and application is a critical issue that urgently needs to be addressed.

[0102] To address the problems existing in the aforementioned related technologies, this embodiment provides a standardized benchmark platform and system for artificial intelligence algorithms, data, and representation methods for polymer research and development. This embodiment can provide researchers in polymer materials informatics with a comprehensive benchmark platform, covering data, algorithms, evaluation, and user interaction modules, constructing a complete process system from data acquisition, standardization processing, feature extraction to model training and performance evaluation. This solution is conducive to promoting innovation in polymer material design and development methods and can accelerate the transformation of academic research and engineering applications in this field.

[0103] The standardized benchmark platform for artificial intelligence algorithms, data, and representations for polymer research and development provided in this embodiment aims to provide standardized data, conduct standardized model training and evaluation, and integrate an open and easy-to-use web platform to provide a benchmark platform for comparing the performance of representations and algorithms in the polymer field, thereby achieving high efficiency and universality in the polymer performance prediction and design process.

[0104] This platform can adopt a distributed architecture with a front-end and back-end separation. The front-end is based on the React framework (a web development framework) to build a visual user interface, while the back-end uses FastAPI (a web framework) to build API (Application Programming Interface) services. Please see [link / reference]. Figure 2 , Figure 2This is a platform architecture diagram provided for an embodiment of this application. The platform architecture includes a user interface layer, a front-end application layer, a back-end service layer, and a data storage layer. The user interface layer includes a web browser or a mobile device browser; the front-end application layer includes the React framework, TypeScript (an open-source programming language), Ant Design (a user interface component library), Axios (a hypertext transfer protocol client), core front-end pages, and integrated components; the back-end service layer includes the FastAPI framework, the Python (a computer programming language) ecosystem, data processing models, machine / deep learning models, core back-end APIs, and data processing flows; the data storage layer can utilize a database to store CSV data files, etc.

[0105] The platform's backend functionality primarily includes the data layer, polymer representation, algorithm layer, evaluation layer, and model interpretation. Please refer to [link to relevant documentation]. Figure 3 , Figure 3 This is a platform functional architecture diagram provided in an embodiment of this application. The platform functions include a data layer, a polymer representation, an algorithm layer, an evaluation layer, and a model interpretation layer; the data layer includes a raw data pool and a normalized data pool; the polymer representation includes Mordred descriptors, molecular fingerprints, and graph neural networks; the algorithm layer includes supervised learning and unsupervised learning; the evaluation layer is used to implement basic evaluation parameters, computational efficiency indicators, and visualization analysis; and the model interpretation layer is used to implement SHAP analysis.

[0106] The data layer is used to collect various polymer property data and organize them into a standardized database; the polymer representation layer is used to extract molecular features using different types of representations; the algorithm layer is used to organize typical machine learning algorithms and build an algorithm repository; the evaluation layer is used to evaluate the model performance using multiple evaluation metrics and display it using visualization methods; and the model interpretation layer is used to provide basic interpretation of the model.

[0107] The data layer will be introduced below:

[0108] The actual success of any AI-driven materials discovery platform depends on the quality, comprehensiveness, and standardization of its underlying data. This platform directly addresses this fundamental requirement. The polymer data in the data layer primarily comes from original literature, experimental data, and various publicly available polymer databases, covering over 8,000 unique polymers and more than 200,000 polymer property data entries.

[0109] In polymer data, one type is the unlabeled polymer dataset, which only contains the molecular structure of polymers; the other type collects corresponding physical, chemical, thermodynamic, and mechanical properties for specific polymers. Please refer to [link to relevant documentation]. Figure 4 , Figure 4This is a data volume illustration of a polymer dataset provided in an embodiment of this application, showing the collected polymer properties and specific data entries. Figure 4The vertical axis, Number of Data, represents the amount of data, while the horizontal axis represents the properties of the polymer. Polymer properties specifically include: Glass transition temperature, Thermal decomposition temperature, Melting temperature, Bandgap chain (Egc), Intrinsic viscosity, Specific volume, Density, Power conversion efficiency (PCE_OPV_aug), Elongation at break, Tensile modulus, Tensile stress strength at break, Electrical conductivity, Liquid crystal phase transition temperature (Lc), Volume resistivity, Softening temperature, Gas permeability coefficient, Water absorption, Heat of fusion, Bandgap bulk (Egb), Crystallization temperature, Gas diffusion coefficient, Dynamic mechanical properties (storage modulus), Dielectric constant (EPS), Refractive index (Nc), and Ionization energy. Energy_Ei, Electron Affinity_Eea, Surface Tension, Solubility Parameter, Cohesive Energy Density, Dynamic Mechanical Properties Loss Tangent, Dielectric Loss Factor, Contact Angle, Tensile Stress Strength at Yield, Gas Solubility CoefficientDielectric loss tangent, isothermal weight loss time, heat of fusion molar conversion, specific heat capacity, and dynamic mechanical property loss modulus.

[0110] The raw polymer data is abundant but varies in format. Standardized databases process this raw data, ultimately organizing it into uniform data units and standard formats. The specific process includes the following steps A1 to A6:

[0111] Step A1: Raw data processing and quality assessment.

[0112] Raw data is typically stored in CSV (Comma-Separated Values) files, but often suffers from formatting issues, missing values, or record errors. This step should begin by checking the completeness of each field, such as whether the polymer SMILES (Simplified Molecular Input Line Entry System) structural formula is missing or contains illegal characters (such as spaces or special symbols). Numerical fields (such as glass transition temperature) should also be checked. The density (or other numerical values) must be verified to be a valid number, avoiding mixed text (such as "N / A not applicable" or unit symbols). For detected outliers or erroneous records, manual correction is required based on experimental records or literature data. For example, a record... The value is "250" The units of the data need to be labeled and prepared for subsequent conversion. This initial screening is crucial for data integrity. It represents degrees Fahrenheit.

[0113] During the data processing, this embodiment can perform quality assessment on data entries within the dataset. Data entries assessed as abnormal are directly deleted from the dataset, while data entries that cannot be determined are reported to manual processing.

[0114] Step A2: Deduplication and data integration based on regular expression matching.

[0115] Polymer data often suffers from duplication due to multiple data sources. Using the SMILES structure as the core identifier, regular expressions are employed to remove format differences (such as bracket placement or hydrogen atom visibility) to achieve standardized matching. For duplicate entries, the first occurrence is retained to ensure data traceability. If multiple data sources exist, different property descriptions of the same polymer need to be compared horizontally and integrated into a maximum feature set. For example, file A records the properties of a certain polymer... And density, B file adds its dielectric constant, and merges them to form a complete entry.

[0116] Step A3: Standardize the units of physical quantities.

[0117] Inconsistent units of measurement for polymer properties can severely impact data analysis. It is necessary to identify unit identifiers in the fields (e.g., ...). The column " "or" The conversion is standardized according to international standards. For example, converting Fahrenheit to Celsius: =( -32)×5 / 9, freezing point temperature 32 =0 For units not explicitly specified, common units should be used by default based on the field name (e.g., density defaults to [unit name]). (Molecular weight is expressed in g / mol). This standardization is crucial for accurate modeling.

[0118] Step A4: Outlier filtering.

[0119] Outliers may originate from measurement errors or data entry mistakes; the data distribution pattern should first be determined using a histogram. When the data approximately follows a normal distribution, according to... The principle is to identify and remove extreme outliers; when the data distribution is severely skewed, the interquartile range (IQR) method is used to filter the data. However, it should be noted that some "outliers" may be the true manifestation of special structures (such as abnormally high cross-linking caused by high cross-linking). (This can be determined by combining chemical knowledge.) It represents the standard deviation.

[0120] Step A5: Verification after data cleaning.

[0121] Check data integrity, compare records before and after cleaning to ensure that no valid data was deleted during cleaning; draw histograms or box plots to verify the effectiveness of outlier handling; and use knowledge of polymer science to determine the reasonableness of the data.

[0122] Step A6: Standardize storage and database construction.

[0123] For unlabeled basic polymer SMILES structured data, it is directly stored as a CSV file for easy reading and sharing; for data containing multi-dimensional polymer properties (such as basic attributes, physical / chemical properties, test conditions, etc.), it is stored in MySQL (a relational database management system). This structured approach ensures data accessibility, integrity, and scalability.

[0124] The data standardization process described in steps A1 to A6 is the cornerstone of this platform, providing a high-quality, reliable data foundation that is currently lacking in the field, and is crucial for building trustworthy artificial intelligence models.

[0125] The polymer notation is described below:

[0126] In the standardized dataset, the molecular structure of polymers is represented using the SMILES model. Three types of molecular features are extracted based on SMILES: Mordred descriptors (i.e., molecular descriptors), molecular fingerprints, and molecular graphs. These diverse representations are crucial for capturing different aspects of molecular information, providing flexibility for various modeling tasks.

[0127] Mordred (an open-source library designed for computational molecular descriptors) can calculate over 1800 molecular descriptors from molecular structures (SMILES), covering multidimensional features such as geometry, topology, electronic properties, and physicochemical properties. These descriptors quantify the structural characteristics of molecules using mathematical and statistical methods. The specific steps are: first, input the SMILES of repeating units or representative segments of the polymer; then, calculate the descriptors; finally, generate a CSV table containing all calculated descriptors, with each descriptor corresponding to a numerical value or category.

[0128] Molecular fingerprinting is a method of representing molecular structures using binary or counting vectors to capture key structural fragments or substructure information. This platform calculates and organizes a total of 13 different types / bits of molecular fingerprints, such as Morgan fingerprinting and MACCS (Molecular Access System Fingerprint), and finally stores them uniformly in CSV format.

[0129] Molecular graphs are a method of representing molecular structures as graph data, where atoms are nodes and chemical bonds are edges. This representation naturally captures the topological structure and chemical properties of molecules, such as atom types, bond types, and spatial arrangements. In a molecular graph, each node can contain rich feature information, such as atomic number, charge, and hybridization state, while edges can encode bond type, length, or angle. The specific steps for extracting a molecular graph are as follows: First, a graph neural network is used to transform the polymer repeating units into a graph structure, with atoms as nodes and bonds as edges. Node features include atom type and charge, while edge features include bond type and bond length. Second, graph embedding is implemented using node embedding and multi-layer aggregation. Third, global pooling is performed. The final node features are summed, averaged, or maximized to generate the representation vector of the entire molecular graph. The vector generated by the graph neural network is input into downstream task models and can be used for property prediction (such as regression models predicting glass transition temperature) and generative design (such as combining with generative adversarial networks (GANs) to generate target-oriented models). Tasks such as virtual polymer structures.

[0130] The provision of a variety of well-characterized representation methods in this embodiment is a key practical advantage, allowing researchers to select the most appropriate featureization technique for their specific problem, thus improving flexibility.

[0131] The algorithm layer is described below:

[0132] Currently, machine learning and deep learning can be mainly divided into several categories, such as supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning, each with its own unique characteristics.

[0133] Supervised learning is the most common machine learning paradigm. Its core idea is to train a model using labeled data (i.e., input and corresponding output) to predict new data. Supervised learning can be further divided into classification tasks and regression tasks. The goal of classification tasks is to divide data into different categories, such as polymer classification; regression tasks are used to predict continuous values, such as predicting the value of a certain property of a polymer. Typical algorithms include logistic regression, support vector machines (SVM), decision trees, and deep neural networks (such as convolutional neural networks (CNN) and transformers). The advantage of supervised learning is its high prediction accuracy, but its disadvantage is its strong dependence on labeled data, resulting in high data labeling costs.

[0134] Unlike supervised learning, unsupervised learning does not rely on labeled data but instead searches for hidden patterns or structures directly within the data. Common unsupervised learning methods include clustering (such as K-Means clustering and the density-based spatial clustering algorithm DBSCAN) and dimensionality reduction (such as Principal Component Analysis (PCA) and the t-distributed neighborhood embedding algorithm t-SNE). Clustering can be used for data analysis, while dimensionality reduction helps with data visualization or feature extraction. Furthermore, generative models (such as Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) also fall under the category of unsupervised learning, capable of generating new data samples that can be applied to data augmentation. The advantage of unsupervised learning is that it does not require labeled data, but the interpretability and evaluation of the models are often more challenging.

[0135] Semi-supervised learning lies between supervised and unsupervised learning, utilizing a small amount of labeled data and a large amount of unlabeled data for training. This approach is particularly suitable for scenarios where labeling is costly. Typical semi-supervised learning methods include self-training, co-training, and graph-based methods (such as graph convolutional networks, GCNs). Semi-supervised learning can effectively utilize unlabeled data to improve model performance, but its effectiveness depends on the quality and representativeness of the labeled data.

[0136] Reinforcement learning is a method that learns optimal policies through trial and error. An agent adjusts its behavior based on reward signals during interactions with the environment to maximize long-term cumulative rewards. Classic reinforcement learning algorithms include Q-Learning (a model-free reinforcement learning algorithm) and Policy Gradient, while deep reinforcement learning (such as Deep Q-Network (DQN) and Proximal Policy Optimization (PPO)) combines deep neural networks to handle more complex environments. The advantage of reinforcement learning lies in its ability to handle sequential decision-making problems, but the training process typically requires a large amount of interactive data, and the stability and interpretability of the algorithms remain research challenges.

[0137] Overall, machine learning and deep learning have diverse classification systems, each with its own advantages and disadvantages. Supervised learning relies on labeled data but provides accurate predictions, unsupervised learning is suitable for exploring the intrinsic structure of data, semi-supervised learning performs well when labeled data is limited, and reinforcement learning excels at sequential decision problems.

[0138] This platform has built a modular algorithm repository, integrating 25 cutting-edge algorithms from traditional machine learning and deep learning, forming a machine learning toolbox covering all scenarios related to polymer properties. The artificial intelligence algorithms in the repository include: supervised learning algorithms (9 types), tree models and ensemble methods (5 types), kernel methods and support vector regression (2 types), neural network models (3 types), probabilistic and Bayesian methods (2 types), and other regression methods (1 type); and unsupervised learning algorithms (Principal Component Analysis (PCA), t-distributed neighborhood embedding (t-SNE), and K-Means clustering.

[0139] Linear algorithms include: Linear Regression, Ridge Regression, Lasso, Elastic Net, Least Angle Regression (LARS), Orthogonal Matching Pursuit (OMP), Bayesian Regression, Automatic Association Decision (ARD), and Random Sample Consensus Algorithm (RANSAC).

[0140] Linear algorithms excel at handling linearly separable regression and classification problems in polymer research. These algorithms build predictive models through linear combinations of features, offering advantages such as high computational efficiency and strong interpretability, making them suitable for analyzing small to medium-sized datasets.

[0141] Tree models and ensemble algorithms include: Decision Tree, Random Forest (RF), Gradient Boosting (GBDT), Extreme Gradient Boosting (XGBoost), and CatBoost.

[0142] Tree models and ensemble methods perform well in handling nonlinear, high-dimensional data commonly found in polymer research. These methods improve model performance by constructing decision rule trees or combining multiple weak learners, and are particularly effective in modeling polymer structure-performance relationships. Decision trees are known for their intuitive interpretability, while random forests enhance model robustness through a bagging strategy. These algorithms can automatically capture nonlinear feature relationships and have a high tolerance for missing data, but careful parameter tuning is required to avoid overfitting.

[0143] Kernel methods and support vector regression algorithms include: Support Vector Machine (SVM) and Kernel Ridge Regression (KRR).

[0144] Kernel methods and support vector regression employ kernel functions to map low-dimensional data to a high-dimensional feature space, addressing nonlinear issues and making them suitable for small-sample, high-dimensional data analysis tasks in polymer research. These methods excel in small-sample scenarios and can handle complex data patterns through flexible kernel function selection; however, their computational complexity is high, making them unsuitable for large-scale datasets.

[0145] Neural network models include: Multilayer Perceptron (MLP), Recurrent Neural Network (RNN), and Graph Convolutional Network (GraphConv).

[0146] Neural network models excel at handling complex nonlinear relationships in polymer research. These methods possess powerful feature learning capabilities, but require large amounts of training data and significant computational resources.

[0147] Probabilistic and Bayesian methods include Gaussian Process Regression (GPR), Poisson Regression, and other regression methods such as K-Nearest Neighbors (KNN).

[0148] Probabilistic and Bayesian methods provide probabilistic predictions, suitable for uncertainty modeling. These methods offer interpretable probabilistic outputs and are well-suited for small datasets, but assumptions about feature independence can affect model accuracy.

[0149] The evaluation layer is described below:

[0150] The evaluation layer includes basic indicator evaluation, computational efficiency indicator evaluation, and visualization analysis functions. Specifically, in the basic indicator evaluation section, the regression task (i.e., the task of predicting the numerical properties of polymers) sets the coefficient of determination (COP). The mean squared error (MSE) and mean absolute error (MAE) were used as evaluation metrics; for classification tasks, accuracy, F1-Score, and ROC-AUC were used as evaluation metrics. Regarding computational efficiency, training time and memory usage were statistically analyzed.

[0151] In the visualization analysis section, scatter plots can be used to display the actual values ​​and model predictions, and learning curves can be used to dynamically demonstrate the model's fit. All metrics serve to measure the model's performance and are widely accepted; therefore, only a selection of metrics are introduced below.

[0152] Please see Figure 5 , Figure 5 This is a scatter plot comparing the actual and predicted values ​​of the glass transition temperature of a polymer using the Mordred descriptor and a random forest model, as provided in an embodiment of this application.

[0153] Coefficient of determination of the training set The mean squared error of the training set is 0.982. The mean absolute error of the training set is 223.978. The coefficient of determination for the test set is 10.188. The mean square error of the test set is 0.872. The mean absolute error of the test set is 1571.541. The value is 27.224, and the vertical axis is... This indicates the predicted glass transition temperature. This represents a glass transition temperature model combining the random forest algorithm and molecular descriptors, with the x-axis representing the glass transition temperature. This represents the actual glass transition temperature, ideal represents the target relationship between the predicted value and the standard value, Train represents the training set, and test represents the test set.

[0154] Please see Figure 6 , Figure 6 This is a schematic diagram of the learning curve for the visualization analysis portion of the platform evaluation layer provided in an embodiment of this application. The vertical axis represents the score. The x-axis represents the score (i.e., the coefficient of determination), the x-axis represents the training set size, the training score represents the training score, and the cross-validation score represents the cross-validation score.

[0155] Coefficient of determination ( The goodness of fit of a regression model (i.e., a model for predicting polymer properties) is an important metric in statistics used to evaluate the goodness of fit of the model and to measure the model’s ability to explain the variation in the dependent variable. The higher the value, the better the model fits the data. The core idea is to evaluate the explanatory power of the model by comparing its predicted values ​​with the actual values.

[0156] Coefficient of determination The calculation formula is as follows:

[0157] ;

[0158] Mean Squared Error (MSE) is an evaluation metric used in regression models to measure the average difference between model predictions and actual values. Its core idea is to quantify the model's predictive accuracy by calculating the squared mean of the prediction error.

[0159] The formula for calculating the mean squared error (MSE) is as follows:

[0160] ;

[0161] in, It is the first The actual value of each sample; It is the first Predicted values ​​for each sample; It is the sample size. This represents the mean of the actual values.

[0162] Mean Absolute Error (MAE) is a commonly used evaluation metric in regression models, measuring the average absolute difference between model predictions and actual values. Unlike Mean Squared Error (MSE), MAE directly takes the absolute value of the error, avoiding the influence of error direction and providing a more intuitive measure of error. The more accurate the model prediction, the larger the MAE value, indicating a greater prediction error.

[0163] The formula for calculating the Mean Absolute Error (MAE) is as follows:

[0164] ;

[0165] in, It is the first The actual value of each sample; It is the first Predicted values ​​for each sample; It refers to the number of samples.

[0166] The learning curve is an important tool for analyzing the relationship between model performance and the amount of training data or training time. By visualizing the trend of basic evaluation parameters as the sample size (or training epochs) increases, it helps diagnose whether the model is underfitting (high bias) or overfitting (high variance) and provides direction for model optimization.

[0167] The horizontal axis of the learning curve represents the training sample size (gradually increasing from small to large) or the number of training epochs (such as epochs in neural networks); the vertical axis represents error metrics such as accuracy metrics (such as the coefficient of determination). The two curves indicate the model's performance on the training set or on the validation set (or cross-validation set), respectively.

[0168] The explanation of the model interpretation layer is as follows:

[0169] The model interpretation section primarily employs SHAP (SHapley Additive exPlanations, a game theory-based model interpretation method). SHAP analysis quantifies the contribution of each feature to the model's prediction results. It provides global and local feature importance analysis by calculating the marginal contribution of features under different feature combinations.

[0170] The computational steps of SHAP analysis are as follows: Enumerate all possible subsets. Assuming there are M features, all... The prediction results of each subset are calculated; then the marginal contribution, i.e. the change in prediction before and after a feature is added to a subset, is calculated; finally, the marginal contributions of all subsets are averaged according to their weights (subset size) to obtain the SHAP value of the feature.

[0171] The formula for calculating the SHAP value is as follows:

[0172] ;

[0173] in, Let F represent the SHAP value, S be the feature set, and i represent the i-th feature. It is the predicted value of subset S. This represents the model's predicted value after adding feature i to the feature subset S. Indicates the size of subset S. This indicates the number of all features. See also: Figure 7 , Figure 7 This is a schematic diagram of SHAP analysis, a visualization analysis component of the platform evaluation layer, provided in an embodiment of this application.

[0174] Figure 7 The horizontal axis represents the SHAP value (the impact on the model output), and the vertical axis represents the number of rotational isomers (polymer nRot), average polarizability (polymer apol), number of nitrogen atoms in the polymer (polymer nN), average molecular weight (polymer AMW), bond polarizability (polymer bpol), number of hydrogen bond donors in the polymer (polymer nHBDn), second-order shape index (a parameter used to reflect the degree of branching or spatial complexity of the molecule) (polymer Kier2), number of hydrogen bond acceptors in the polymer (polymer nHBAcc), number of bases in the polymer (polymer nBase), and proportion of nonpolar surface area (polymer nP). High indicates a high eigenvalue, and Low indicates a low eigenvalue.

[0175] The platform's front end is displayed in a web format, primarily including a homepage, registration / login page, leaderboard page, data page, algorithm page, and guide page. This user-centric design is crucial for expanding the application of AI (artificial intelligence) tools in the field of polymer science.

[0176] The homepage provides a brief introduction to the platform's functions, polymer data, and usage methods, serving as the platform's interface. The leaderboard page summarizes and statistically analyzes the performance of various models, displaying the rankings in multiple ways. The data page shows the total amount of polymer data and the specific distribution of each property, as well as functions for exploring and displaying simple polymer properties. The algorithm page includes an introduction to the algorithm and the basic performance of the corresponding model. The guide page provides a user guide for the platform.

[0177] The platform's various pages contain the main functional modules of the platform's front end, such as: data distribution module, leaderboard module, detailed information module, dataset effect module, and optimal combination display module; the modules are described below:

[0178] The data distribution module displays the names and data volumes of all datasets on the platform, and uses histograms to show the specific distribution of each dataset. The histograms indicate the mean, median, maximum, and minimum values ​​for each dataset, allowing users to quickly understand the data distribution for this polymer property. Please refer to [link to relevant documentation]. Figure 8 , Figure 8 This is a data distribution histogram of a polymer glass transition temperature dataset provided in an embodiment of this application.

[0179] The horizontal axis represents the glass transition temperature of the polymer. The vertical axis represents frequency. The figure shows the distribution of polymer glass transition temperatures, with a minimum of -123.00, a maximum of 495.00, an average of 145.14, and a median of 138.00. The total number of data rows is 8055.

[0180] The leaderboard module summarizes the performance of all models on the platform. The leaderboard clearly displays the corresponding polymer property dataset, polymer representation, algorithm, and model performance evaluation value. Users can also impose constraints (such as specific algorithms or datasets) on the leaderboard to observe the model ranking under those constraints. The leaderboard is displayed not only in form but also as a 3D graph.

[0181] The limited database and polymer representation leaderboard provided in this embodiment may include a dataset name, a polymer representation method, an artificial intelligence algorithm model, a determination coefficient corresponding to the training set, a mean squared error corresponding to the training set, a mean absolute error corresponding to the training set, a determination coefficient corresponding to the test set, a mean squared error corresponding to the test set, and a mean absolute error corresponding to the test set.

[0182] Please see Figure 9 , Figure 9 A three-dimensional view of a ranking list provided in an embodiment of this application; Figure 9 The horizontal axis represents the algorithm, such as: Feedforward Neural Network (FNN), K Nearest Neighbor (KNN), Kernel Ridge Regression (KRR), Multilayer Perceptron (MLP), Random Forest (RF), Ridge Regression (Ridge), Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), and Graph Convolution (GraphConv). Figure 9The vertical axis (Z-axis) represents the polymer representation, while the coordinates in the vertical direction (Z-axis) represent the corresponding evaluation index values ​​(such as the coefficient of determination). In this embodiment, the corresponding bar chart can be colored according to the coefficient of determination. The intervals for the coefficient of determination can include: greater than or equal to 0.8, greater than or equal to 0.5 and less than 0.8, greater than or equal to 0 and less than 0.5, and less than 0.

[0183] In the detailed information module, users can select two different polymer properties as the x and y axes of a scatter plot to observe the correlation between polymer properties. The box on the right of the module displays the polymer structure formula and other detailed information for the selected data points. Please refer to [link to relevant documentation]. Figure 10 , Figure 10 This is a schematic diagram illustrating a platform polymer details module provided in an embodiment of this application. This module can display the files used (e.g., file 1 and file 2) and the corresponding columns for each file (e.g., column 1 shows the glass transition temperature). Listed as 2 thermal decomposition temperatures The vertical axis, Thermal decomposition temperature.csv, represents the thermal decomposition temperature. The horizontal axis, Glass transitiontemperature.csv, represents the glass transition temperature. The unit is Celsius. . Figure 10 The diagram shown may also include detailed molecular information, molecular properties (such as polymer name, simplified molecular linear input specification SMILES) and values, as well as different attributes.

[0184] The dataset performance module primarily displays the prediction performance of different models on this dataset. The bar chart of dataset performance is colored according to the coefficient of determination (e.g., beige for a coefficient of determination greater than 0.8, brown for values ​​between 0.5 and 0.8, etc.). The coloring effect can roughly indicate the difficulty of predicting on this polymer property dataset (e.g., datasets with more beige coloring are easier to predict because most models perform well on them). Please see [link to relevant documentation]. Figure 11 , Figure 11 This is a schematic diagram illustrating the effect of a platform dataset module provided in an embodiment of this application. The vertical axis (1-39) represents the dataset number, and the horizontal axis represents the number of samples. In this embodiment, the coefficient of determination is used. To distinguish the samples in the unified dataset, three ranges of determination coefficients were defined: greater than or equal to 0.8, greater than or equal to 0.5 but less than 0.8, and less than 0.5.

[0185] The optimal combination display module shows: training a model using different combinations of representations and algorithms, predicting a specific dataset, and recording the combination that performs best on that dataset; similarly, it processes datasets for all polymer properties, recording the best-performing combination for each dataset, and plotting it as a tree diagram. This tree diagram can be used to determine the universality of a representation or algorithm; the more frequently an algorithm / representation appears, the better its universality, as it performs best on most datasets. Please see [link to relevant documentation]. Figure 12 , Figure 12 This diagram illustrates an optimal combination of display modules provided in an embodiment of this application. The artificial intelligence algorithms shown include: HistGBDT (Hist Gradient Boosting Decision Tree), ExtraTrees, CatBoost (Gradient Boosting Framework), GBDT (Gradient Boosting Decision Tree), MLP (Multilayer Perceptron), XGBoost (Efficient Gradient Boosting Framework), Random Forest (RF), and RNN (Recurrent Neural Network); the polymer representation includes: Mordred (Molecular Descriptor), Morgan_2048 (2048-bit), and Morgan_1 (1024-bit). 024, RDKit_1024 (1024-bit topological fingerprint), Torsion_2048 (2048-bit dihedral conformation fingerprint), AtomPair_1024 (1024-bit atomic pair fingerprint), Morgan_512 (512-bit Morgan fingerprint), RDKit_512 ​​(512-bit topological fingerprint), AtomPair_2048 (2048-bit atomic pair fingerprint), RDKit_2048 (2048-bit topological fingerprint), and predefined functional fingerprints Maccs. Figure 12 The numbers in parentheses shown (such as 1, 2, 3, 4, 6, 9, 12, 18) are used to indicate the frequency of the algorithm and representation. The larger the number in parentheses, the higher the frequency.

[0186] The user-friendly interface, combined with these powerful visualization and analysis modules, makes the platform a highly useful tool for both novice and expert users, significantly lowering the barrier to leveraging artificial intelligence in polymer research and development.

[0187] This embodiment proposes a standardized benchmark platform for artificial intelligence algorithms, data, and representations in polymer research and development, which can solve problems such as low data utilization efficiency, poor algorithm adaptability, and lack of unified standards for performance comparison in existing technologies. The urgency of establishing such a platform stems from the current low research and development efficiency, which has become a major obstacle to innovation in this key field. The practicality of the proposed solution lies in its comprehensive and integrated approach. The main beneficial effects of this invention include:

[0188] Addressing data fragmentation: By constructing a standardized polymer structure-property database, the problem of data dispersion and fragmentation is fundamentally solved. This centralized and standardized resource is crucial for reliable and reproducible research.

[0189] Systematic and fair model comparison: By adopting a modular modeling framework, different algorithms can be systematically compared and verified on the same platform, which makes up for the shortcomings of insufficient cross-system modeling capabilities and inconsistent evaluation standards, thereby creating a more rigorous and transparent research environment.

[0190] Enhanced accessibility and ease of use: The platform also integrates polymer property prediction, model performance comparison, and visualization analysis functions, and lowers the technical barrier to use through an intuitive user interface, making advanced artificial intelligence tools available to a wider range of researchers and engineers.

[0191] Connecting research and application: This provides a comprehensive guiding tool for the intelligent development of polymer materials, combining scientific rigor with practical engineering needs. This facilitates the smooth transformation of research findings into industrial innovation.

[0192] Accelerating the innovation process: By simplifying data management, model development, and evaluation processes, the platform enables researchers to focus on discovery and design, thereby accelerating the pace of innovation in polymer science and engineering.

[0193] This platform acts as a much-needed catalyst for AI research in the polymer field, unleashing the full potential of AI in polymer R&D through standardized methodologies and the fostering of a collaborative ecosystem.

[0194] The process described in the above embodiments is illustrated below through examples in practical applications.

[0195] Example 1: Evaluation of the prediction effect of different algorithms on the glass transition temperature of polymers.

[0196] This embodiment compares the predictive performance of different algorithms through an interactive platform system, combining visualization and automated processes to provide researchers with intuitive decision support. The specific steps are as follows:

[0197] Step B1: Data loading and visualization.

[0198] Users import a dataset—polymer glass transition temperature—and can view the specific data distribution for this property on the front end. They can select the polymer representation using a selection box (Mordred is used as an example here); and the training and test sets are automatically split using a 7:3 ratio of five-fold cross-validation.

[0199] Step B2: Algorithm selection and model construction.

[0200] Users can select the following comparison algorithms from the algorithm repository via a drop-down menu: RF, SVM, XGBoost, Ridge, KNN, RNN, and KRR. The platform loads default hyperparameters based on the algorithm type, trains the model, and predicts the glass transition temperature of the polymer.

[0201] Step B3: Effectiveness evaluation.

[0202] The platform automatically evaluates the model's performance in predicting polymer glass transition temperatures, outputs a scatter plot comparing predicted and actual values, and provides evaluation metrics; the evaluation results are then uploaded to the leaderboard module.

[0203] Step B4: Compare the algorithm performance.

[0204] In the leaderboard interface, users specify the database "glass transition temperature" and the polymer notation "Mordred"; the platform displays the leaderboard under these conditions and visualizes it using a lollipop chart. Please see [link / reference]. Figure 13 , Figure 13 This is a schematic diagram illustrating the performance of a lollipop graph evaluation algorithm provided in an embodiment of this application. The vertical axis represents the coefficient of determination (Test) of the test set. The horizontal axis represents various algorithms, such as XGBoost, RF, Ridge, KNN, RNN, SVM, and KRR.

[0205] Example 2: Evaluation of the effectiveness of different representation methods in predicting the glass transition temperature of polymers.

[0206] This embodiment evaluates the performance differences of different polymer representation methods in predicting polymer glass transition temperatures, and clarifies the impact of different structures and different bit encoding methods on model accuracy. The specific steps are as follows:

[0207] Step C1: Data loading and visualization.

[0208] The data loading process is the same as step B1 in Example 1. Multiple polymer representations (such as Mordred, molecular fingerprint, etc.) are selected using a checkbox; the training and test sets are automatically divided using five-fold cross-validation at a 7:3 ratio.

[0209] Step C2: Algorithm selection and model construction.

[0210] Users can select an algorithm (RF for example) from the algorithm repository via a drop-down menu. The platform loads default hyperparameters based on the algorithm type, trains the model, and predicts the glass transition temperature of the polymer.

[0211] Step C3: Effectiveness evaluation.

[0212] Same as Example 1.

[0213] Step C4: Compare polymer representation methods.

[0214] Similarly, follow step B4 in Example 1. Please refer to... Figure 14 , Figure 14 This is a schematic diagram illustrating the effectiveness of a lollipop chart for evaluating polymer representation methods, provided in an embodiment of this application. The vertical axis represents the coefficient of determination (Test) of the test set. The horizontal axis represents various polymer representations, such as the molecular descriptor Mordred, the Morgan fingerprint of length 2048 (Morgan_2048), the Morgan fingerprint of length 1024 (Morgan_1024), the dihedral conformation fingerprint of length 2048 (Torsion_2048), the dihedral conformation fingerprint of length 1024 (Torsion_1024), the Morgan fingerprint of length 512 (Morgan_512), and the topological fingerprint of length 2048 (RDK). it_2048, Torsion_512 (dihedral fingerprint with a length of 512), Maccs (predefined functional fingerprint), RDKit_1024 (topological fingerprint with a length of 1024), AtomPair_2048 (atomic pair fingerprint with a length of 2048), AtomPair_1024 (atomic pair fingerprint with a length of 1024), RDKit_512 ​​(topological fingerprint with a length of 512), and AtomPair_512 (atomic pair fingerprint with a length of 512).

[0215] Example 3: Evaluating the predictive effect of the same representation and algorithm combination on the properties of different polymers.

[0216] This embodiment compares the generalization ability of the same representation and algorithm combination (e.g., a random forest model based on the Mordred descriptor) for multi-property prediction, covering electrical properties, mechanical properties, thermal properties, etc. The specific steps are as follows:

[0217] Step D1: Data loading and visualization;

[0218] Users import all polymer property datasets from the database. A polymer representation is selected using a checkbox (Mordred as an example); the training and test sets are automatically split using five-fold cross-validation at a 7:3 ratio.

[0219] Step D2: Algorithm selection and model construction;

[0220] Users can select an algorithm (RF, for example) from a dropdown menu in the algorithm repository. The platform loads default hyperparameters based on the algorithm type, trains various models, and predicts multiple polymer properties.

[0221] Step D3: Effectiveness Evaluation;

[0222] Similarly, step B3 of Example 1 is followed.

[0223] Step D4: Compare and predict the effects of different properties;

[0224] In the leaderboard interface, the representation "Mordred descriptor" and the algorithm "RF" are specified to generate the corresponding leaderboard and lollipop image (same as step B4 in Example 1).

[0225] Example 4: Applicability of various representations and algorithm combinations to datasets.

[0226] This embodiment evaluates the applicability of models constructed using various representations and algorithms in different types of databases (thermal performance, mechanical performance, etc.), that is, to realize the optimal combination display module function. The specific steps are as follows:

[0227] Step E1: Data loading and visualization.

[0228] This step is the same as step B1 in Example 1.

[0229] Step E2: Algorithm selection and model construction.

[0230] This step is the same as step B2 in Example 1. The core lies in the combination of multiple representations and algorithms.

[0231] Examples of combinations of representation and algorithm include: Mordred+XGBoost, MACCS+XGBoost, Mordred+RF, etc.

[0232] Step E3: Effectiveness evaluation.

[0233] This step is the same as step D3 in Example 3.

[0234] Step E4: Compare the generalization capabilities of different combinations.

[0235] The leaderboard section does not specify restrictions on datasets, polymer representations, or algorithms. For greater clarity, the platform generates a best-of-breed tree diagram. This tree diagram represents the best-performing model combination for each dataset. For example, in... The model makes predictions using all combinations on the dataset, retaining only the best-performing combination. Then, it changes to a different dataset, retaining only the best-performing combination on that dataset, and so on. Finally, it summarizes all the best-performing combinations, counts the frequency of each representation and algorithm, and creates a tree diagram. This is used to evaluate generalization ability.

[0236] This invention relates to the collection and systematic organization of multi-source data in the field of polymers, constructing a unified standard polymer property database. The creation of this high-quality, benchmark-ready data resource is a crucial and novel contribution, addressing the widespread problems of data fragmentation and inconsistency that severely hinder AI-driven polymer research and development.

[0237] The platform provided in this embodiment integrates various polymer representation methods (descriptors, fingerprints, graph-based methods) and a wide range of supervised / unsupervised learning algorithms, realizing a complete workflow design from polymer structure characterization, polymer molecular feature extraction, and model training and evaluation. This platform integrates end-to-end workflows for multiple methods.

[0238] This embodiment establishes a unified and transparent evaluation and benchmarking mechanism, and displays the evaluation results in the form of a leaderboard, enabling horizontal comparisons between different representation methods and algorithms, and providing a benchmark platform for comparing the performance of representation methods / algorithms. This ability to fairly benchmark and rank different modeling strategies is the core innovation, directly solving the current problem of a lack of standardized comparisons in this field.

[0239] Compared with existing technologies, the above-mentioned polymer artificial intelligence benchmark platform has the following technical effects: (1) By collecting and cleaning multi-source data, a unified and standardized polymer structure-property database is constructed, which solves the problems of data dispersion and inconsistent labeling, and provides a solid data foundation for cross-system modeling. This is a fundamental step towards more repeatable and credible artificial intelligence research in the field of polymers; (2) The platform integrates a variety of algorithms, including supervised learning and unsupervised learning, and realizes the whole process from structural representation to feature extraction to model training. The modular design allows for horizontal comparison between different algorithms, which helps to reveal the intrinsic relationship between structure and performance, improves the universality and prediction accuracy of the model, and meets the urgent need for standardized methodologies in this field; (3) The front-end and back-end separation architecture and interactive interface lower the technical threshold for platform use, enabling non-professionals to predict polymer properties through intuitive operation, which helps to realize the rapid transformation of scientific research results into engineering applications. This accessibility is crucial for promoting the rapid transformation of research results into practical engineering applications and promoting interdisciplinary cooperation; (4) Through multi-dimensional evaluation indicators and visualization methods, not only is the predictive performance of the model quantified, but also data support and decision-making basis are provided for experimental design and structural optimization through auxiliary means such as learning curves and cluster dimensionality reduction diagrams. These auxiliary tools provide valuable data support and decision-making basis for experimental design and structural optimization, guiding researchers to conduct more targeted and efficient research and development work.

[0240] It is evident that this platform has the potential to become an indispensable tool, addressing the urgent needs of the polymer R&D field for standardization, improved data quality, and access to artificial intelligence, thereby accelerating the discovery and design of next-generation polymer materials.

[0241] The present application provides an analysis system for a polymer performance prediction model, which may include:

[0242] A dataset construction module is used to construct datasets for various polymers; wherein, the datasets include the molecular structural formulas of the corresponding polymers;

[0243] A polymer representation module is used to determine the molecular feature information of each molecular structure in the dataset; wherein the types of molecular feature information include molecular descriptors, molecular fingerprints, and molecular graphs;

[0244] The algorithm management module is used to build an algorithm repository that integrates multiple artificial intelligence algorithms;

[0245] A parameter determination module is used to determine target parameters based on a request message if a request message is received. The target parameters include data selection information, polymer representation, and algorithm selection information. The data selection information describes the dataset participating in model training, the polymer representation describes the types of molecular feature information of the dataset participating in model training, and the algorithm selection information describes the artificial intelligence algorithm participating in model training.

[0246] The model training module is used to train the model based on the target parameters to obtain multiple polymer performance prediction models; wherein, in the training process of a single polymer performance prediction model, the number of datasets, polymer representations and artificial intelligence algorithms used are all 1.

[0247] The evaluation module is used to evaluate the prediction effect of the polymer performance prediction model and generate corresponding analysis results.

[0248] This embodiment constructs a dataset containing polymer molecular structures, determines the molecular feature information of each molecular structure, and builds an algorithm repository integrating multiple artificial intelligence algorithms. The dataset includes the molecular structures of the corresponding polymers, providing basic data for model training; the types of molecular feature information include molecular descriptors, molecular fingerprints, and molecular graphs, providing rich feature inputs for model training. Upon receiving a request, the corresponding target parameters can be determined based on the request information. These target parameters include data selection information, polymer representation methods, and algorithm selection information; based on these target parameters, the datasets, types of molecular feature information, and artificial intelligence algorithms participating in model training can be determined. Based on the above target parameters, this embodiment performs model training operations to obtain multiple polymer performance prediction models. During the training of a single polymer performance prediction model, the number of datasets, polymer representation methods, and artificial intelligence algorithms used is always 1; this independent model training method ensures the independence and consistency of each model, facilitating subsequent evaluation and comparison. This embodiment evaluates the prediction performance of the polymer performance prediction models and generates corresponding analysis results, allowing users to fully understand the model's performance and thus achieve accurate evaluation of the polymer performance prediction models.

[0249] Furthermore, it also includes:

[0250] The model interpretation module is used to interpret the polymer performance prediction models after obtaining multiple polymer performance prediction models, and obtain model interpretation results; wherein, the model interpretation results include the contribution degree of each feature in the input data of the polymer performance prediction model to the model prediction results; and is also used to adjust the parameters of the polymer performance prediction models according to the model interpretation results.

[0251] Furthermore, the process of constructing datasets for multiple polymers by the dataset construction module includes: reading raw polymer data from the database, formatting and quality assessing the raw polymer data, and selecting raw polymer data with a quality assessment score greater than a preset value as first intermediate data; performing deduplication on the first intermediate data based on regular expressions, and integrating the deduplicated data to obtain second intermediate data; performing unit identification unification and outlier filtering on the second intermediate data to obtain third intermediate data; performing data cleaning on the third intermediate data to obtain fourth intermediate data, and verifying the data integrity of the fourth intermediate data; setting the fourth intermediate data that passes the data integrity verification as candidate polymer data; if the candidate polymer data contains property data, then constructing a dataset containing the molecular structure formula and property data of the polymer based on the candidate polymer data; wherein, the property data includes any one or a combination of physical properties, chemical properties, thermodynamic properties, and mechanical properties; if the candidate polymer data does not contain property data, then constructing a dataset containing the molecular structure formula of the polymer based on the candidate polymer data.

[0252] Furthermore, the evaluation module evaluates the prediction performance of the polymer performance prediction model by: determining a first evaluation result based on the model evaluation metrics of the polymer performance prediction model; wherein the model evaluation metrics are either first-class or second-class metrics; the first-class metrics include the coefficient of determination, mean squared error, and mean absolute error, and the second-class metrics include the F1 score and the area under the ROC curve; the F1 score is used to describe the harmonic mean of precision and recall, and the ROC curve represents the receiver operating characteristic curve; determining a second evaluation result based on the training time and memory usage of the polymer performance prediction model; constructing a scatter plot containing the true values ​​and model prediction values, and generating a third evaluation result based on the scatter plot.

[0253] Furthermore, the process of the model training module training the model based on the target parameters to obtain multiple polymer performance prediction models includes: combining the data selection information, polymer representation method and algorithm selection information in the target parameters to obtain multiple training configurations; wherein, each training configuration includes one dataset, one polymer representation method and one artificial intelligence algorithm; and performing the corresponding model training operation according to each training configuration to obtain multiple polymer performance prediction models.

[0254] Furthermore, the process of the model training module performing the corresponding model training operation according to each training configuration includes: selecting the corresponding target dataset according to the training configuration; converting the molecular structure formula in the target dataset into the corresponding molecular feature information using the polymer representation in the training configuration; and performing the corresponding model training operation based on the target dataset using the artificial intelligence algorithm in the training configuration.

[0255] Furthermore, the process by which the evaluation module evaluates the prediction performance of the polymer performance prediction model and generates corresponding analysis results includes: evaluating the prediction performance of the polymer performance prediction model to obtain a model performance evaluation value for each polymer performance prediction model; generating analysis results of the polymer performance prediction model based on the target parameters and the model performance evaluation value; and displaying the analysis results on the front-end page.

[0256] Since the embodiments of the system part correspond to the embodiments of the method part, please refer to the description of the embodiments of the method part for the embodiments of the system part, and they will not be repeated here.

[0257] This application also provides a storage medium on which a computer program is stored, which, when executed, can perform the steps provided in the above embodiments. The storage medium may include various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0258] This application also provides an electronic device that may include a memory and a processor. The memory stores a computer program, and when the processor calls the computer program in the memory, it can implement the steps provided in the above embodiments. Of course, the electronic device may also include various network interfaces, power supplies, and other components.

[0259] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

[0260] It should also be noted that, in this specification, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. An analytical method for predicting polymer performance, characterized in that, include: Construct datasets for various polymers; wherein the datasets include the molecular structural formulas of the corresponding polymers; Determine the molecular feature information of each molecular structure in the dataset; wherein the types of molecular feature information include molecular descriptors, molecular fingerprints, and molecular graphs; Build an algorithm repository that integrates multiple artificial intelligence algorithms; If a request is received, the target parameters are determined based on the request; wherein, the target parameters include data selection information, polymer representation, and algorithm selection information, the data selection information is used to describe the dataset participating in model training, the polymer representation is used to describe the types of molecular feature information of the dataset participating in model training, and the algorithm selection information is used to describe the artificial intelligence algorithm participating in model training; Based on the target parameters, multiple polymer performance prediction models are obtained through model training; wherein, in the training process of a single polymer performance prediction model, the number of datasets, polymer representations, and artificial intelligence algorithms used are all 1. The predictive performance of the polymer performance prediction model is evaluated, and corresponding analysis results are generated.

2. The analytical method for the polymer performance prediction model according to claim 1, characterized in that, After obtaining multiple polymer performance prediction models, the following are also included: The polymer performance prediction model is interpreted to obtain the model interpretation result; wherein the model interpretation result includes the degree of contribution of each feature in the input data of the polymer performance prediction model to the model prediction result; The parameters of the polymer performance prediction model are adjusted based on the model interpretation results.

3. The analytical method for the polymer performance prediction model according to claim 1, characterized in that, The datasets for constructing various polymers include: The raw polymer data is read from the database, and the raw polymer data is formatted and quality evaluated. The raw polymer data with a quality evaluation score greater than a preset value is selected as the first intermediate data. The first intermediate data is deduplicated using regular expressions, and the deduplicated data is then integrated to obtain the second intermediate data. The second intermediate data is subjected to unit identification unification and outlier filtering operations to obtain the third intermediate data; The third intermediate data is cleaned to obtain the fourth intermediate data, and the integrity of the fourth intermediate data is verified. Set the fourth intermediate data that passed the data integrity verification as the alternative polymer data; If the candidate polymer data includes property data, then a dataset containing the molecular structure formula and property data of the polymer is constructed based on the candidate polymer data; wherein, the property data includes any one or a combination of physical properties, chemical properties, thermodynamic properties and mechanical properties; If the candidate polymer data does not contain property data, then a dataset containing the molecular structure formulas of the polymers is constructed based on the candidate polymer data.

4. The analytical method for the polymer performance prediction model according to claim 1, characterized in that, The prediction performance of the polymer performance prediction model is evaluated, including: The first evaluation result is determined based on the model evaluation index of the polymer performance prediction model; wherein, the model evaluation index is a first type index or a second type index; the first type index includes the coefficient of determination, mean square error and mean absolute error, and the second type index includes the F1 score and the area under the ROC curve; the F1 score is used to describe the harmonic mean of precision and recall, and the ROC curve represents the receiver operating characteristic curve; The second evaluation result is determined based on the training time and memory usage of the polymer performance prediction model; Construct a scatter plot containing the actual values ​​and model predictions, and generate a third evaluation result based on the scatter plot.

5. The analytical method for the polymer performance prediction model according to claim 1, characterized in that, Based on the target parameters, model training is performed to obtain multiple polymer performance prediction models, including: The data selection information, polymer representation method, and algorithm selection information in the target parameters are combined to obtain multiple training configurations; wherein each training configuration includes one dataset, one polymer representation method, and one artificial intelligence algorithm; Perform the corresponding model training operation according to each training configuration to obtain multiple polymer performance prediction models.

6. The analytical method for the polymer performance prediction model according to claim 5, characterized in that, Performing the corresponding model training operations according to each of the training configurations includes: Select the corresponding target dataset according to the training configuration; The polymer representation in the training configuration is used to convert the molecular structural formulas in the target dataset into corresponding molecular feature information. The artificial intelligence algorithm in the training configuration is used to perform the corresponding model training operation based on the target dataset.

7. The analytical method for the polymer performance prediction model according to claim 1, characterized in that, The predictive performance of the polymer performance prediction model is evaluated, and corresponding analysis results are generated, including: The prediction performance of the polymer performance prediction models is evaluated to obtain the model performance evaluation value for each polymer performance prediction model. The analysis results of the polymer performance prediction model are generated based on the target parameters and the model effect evaluation values, and the analysis results are displayed on the front-end page.

8. An analysis system for a polymer performance prediction model, characterized in that, include: A dataset construction module is used to construct datasets for various polymers; wherein, the datasets include the molecular structural formulas of the corresponding polymers; A polymer representation module is used to determine the molecular feature information of each molecular structure in the dataset; wherein the types of molecular feature information include molecular descriptors, molecular fingerprints, and molecular graphs; The algorithm management module is used to build an algorithm repository that integrates multiple artificial intelligence algorithms; A parameter determination module is used to determine target parameters based on a request message if a request message is received. The target parameters include data selection information, polymer representation, and algorithm selection information. The data selection information describes the dataset participating in model training, the polymer representation describes the types of molecular feature information of the dataset participating in model training, and the algorithm selection information describes the artificial intelligence algorithm participating in model training. The model training module is used to train the model based on the target parameters to obtain multiple polymer performance prediction models; wherein, in the training process of a single polymer performance prediction model, the number of datasets, polymer representations and artificial intelligence algorithms used are all 1. The evaluation module is used to evaluate the prediction effect of the polymer performance prediction model and generate corresponding analysis results.

9. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and the processor, when calling the computer program in the memory, implements the steps of the analysis method of the polymer performance prediction model as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores computer-executable instructions, which, when loaded and executed by a processor, implement the steps of the analysis method for the polymer performance prediction model as described in any one of claims 1 to 7.

Images