A polymer thermal property prediction method based on gated multi-modal fusion and multi-task learning

Abstract

The application discloses a polymer thermal property prediction method based on a gating multimodal fusion and multi-task learning, comprising: obtaining a structure representation of a polymer repeating unit; extracting graph topology features from a molecular graph structure and substructure features from a molecular fingerprint in parallel through a double-path feature extraction module; generating a gating signal by using features of one mode through a cross gating fusion module, dynamically adjusting the information flow weight of features of another mode, and realizing adaptive deep fusion of the two features; inputting the fused multimodal features into a multi-task learning prediction module, combining task identification, and predicting various polymer thermal properties (such as T g , T m , T d , T c and SP). The application realizes the synergistic enhancement of multimodal information through a gating mechanism, processes the data sparsity problem through multi-task learning, improves the accuracy and generalization ability of the prediction model, and provides a powerful tool for the rapid screening and design of high-performance polymers.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and materials science, specifically to a method for predicting the thermal properties of polymers based on gated multimodal fusion and multi-task learning. Background Technology

[0002] Polymer materials are indispensable in modern industrial systems due to their structural diversity, tunable properties, and scalability. Among these, thermal properties, such as the glass transition temperature (T0), are crucial. g ), melting temperature (T) m ), decomposition temperature (T) d ), crystallization temperature (T) c The thermal properties (TPS) and softening point (SP) are key parameters that determine the polymer's processing window, upper limit of operating temperature, and final application performance. However, accurately obtaining these thermal properties presents significant challenges.

[0003] Traditional experimental characterization methods, such as differential scanning calorimetry (DSC) and thermogravimetric analysis (TGA), are reliable but are typically cumbersome, time-consuming, and costly, making them unsuitable for large-scale, high-throughput material screening. Computational simulation methods (such as molecular dynamics) can provide profound insights, but their enormous computational overhead makes them ill-suited for the vast chemical space of polymers.

[0004] In recent years, quantitative structure-property relationship (QSPR) models, especially machine learning-based models, have provided new avenues for accelerating materials discovery. Existing methods are mainly divided into two categories: one is based on traditional molecular descriptors or fingerprints, which rely on manually designed features and require extensive expert knowledge for feature engineering and screening, potentially overlooking crucial structural information; the other is based on graph neural networks (GNNs), which can automatically learn features from molecular graph structures, but typically requires a large amount of high-quality labeled data to achieve good performance and is prone to overfitting in the data-sparse polymer domain.

[0005] Furthermore, the various thermal properties of polymers are not independent; they all originate from the polymer's intrinsic chemical structure and intermolecular forces. However, most existing models employ a "single-task" learning model, modeling each property independently. This not only ignores the potential correlations between properties but also fails to address certain properties (such as T...). c When the amount of data (SP) is severely insufficient, model training becomes difficult and performance cannot be guaranteed. How to effectively integrate feature information from different sources and utilize the correlation between properties to overcome the challenge of data sparsity is a technical problem that urgently needs to be solved in the field of polymer informatics.

[0006] Therefore, to address the above problems, this invention provides a method for predicting the thermal properties of polymers based on gated multimodal fusion and multi-task learning, which combines graph neural networks, molecular fingerprinting, and multi-task learning strategies to efficiently and accurately predict the thermal properties of various polymers, a system, and a computer-readable storage medium. Summary of the Invention

[0007] The purpose of this invention is to provide a method, system, and storage medium for predicting the thermal properties of polymers based on gated multimodal fusion and multi-task learning. Through an innovative deep learning framework, it synergistically integrates topological information extracted from molecular graph structures and chemical substructure information extracted from molecular fingerprints, and utilizes a multi-task learning strategy to explore the intrinsic correlations between different thermal properties. This aims to solve the problems of limitations of single information sources, poor prediction of sparse data properties, and simple feature fusion mechanisms, thereby achieving high-precision and high-generalization prediction of the thermal properties of various polymers.

[0008] The objective of this invention is achieved through the following technical solution:

[0009] A method for predicting the thermal properties of polymers based on gated multimodal fusion and multi-task learning includes the following steps:

[0010] S1. Data Input and Preprocessing: Experimental data containing the structural representation of polymer repeating units and their corresponding thermal property labels are collected from public databases and literature as sample datasets for machine learning, and the structural representation of the polymer repeating units is normalized.

[0011] S2. Dual-path feature extraction: Establish a dual-path feature extraction model, which includes:

[0012] (a) A first feature extraction path for extracting graph topological features from the graph structure of polymer repeating units to obtain a first feature vector;

[0013] (b) A second feature extraction path for extracting substructure features from the molecular fingerprint of polymer repeating units to obtain a second feature vector;

[0014] S3, Cross-gated fusion: Construct a cross-gated fusion module, receive the first feature vector and the second feature vector, and generate a gating signal through the feature of one modality to adaptively adjust the information weight of the feature of the other modality, and output the fused multimodal feature vector;

[0015] S4. Multi-task learning and prediction: Construct a multi-task prediction model, receive the fused multimodal feature vector and task identifier, and simultaneously predict multiple thermal properties;

[0016] S5. Model Training and Optimization: Divide the dataset into training, validation, and test sets, train the model on the training set, and optimize the model hyperparameters using the hyperparameter search method.

[0017] S6. Model Performance Evaluation: The optimized model is evaluated using evaluation metrics, either through cross-validation or independent test sets.

[0018] S7. Model Interpretation and Analysis: Analyze the gating signal values ​​in the cross-gating fusion module to explain the model's dependence on different modal features.

[0019] Preferably, in step S1, the data preprocessing includes standardizing the representation of the polymer repeating unit structure to ensure that chemically equivalent structures have a unique representation.

[0020] Preferably, in step S2(a), the first feature extraction path is implemented using a graph isomorphic network (GINE), which aggregates neighbor node features and connection edge features simultaneously when updating node representations.

[0021] Preferably, in step S2(b), the molecular fingerprint is a hybrid molecular fingerprint combining the MACCS key fingerprint and the PubChem fingerprint.

[0022] Preferably, the graph topology feature vector is extracted through a graph isomorphic network (GINE); the substructure feature vector is obtained by calculating a hybrid molecular fingerprint (such as a combination of MACCS Keys and PubChem fingerprints) and passing it through a multilayer perceptron (MLP) network.

[0023] Preferably, in step S3, the cross-gating fusion module is implemented in the following way:

[0024] The second feature vector is transformed by a first nonlinear transformation to generate a first gating signal for element-wise multiplication with the first feature vector.

[0025] The first feature vector is transformed by a second nonlinear transformation to generate a second gating signal for element-wise multiplication with the second feature vector.

[0026] Preferably, in step S4, the multi-task prediction model is a unified deep neural network architecture, which can simultaneously predict the glass transition temperature (T) based on the task identifier. g ), melting temperature (T) m ), decomposition temperature (T) d ), crystallization temperature (T) c At least two of the following thermal properties: ) and softening point (SP).

[0027] Preferably, in step S4, the multi-task prediction model is a deep neural network architecture, and the task identifier vector is a one-hot encoded attribute selector vector.

[0028] Preferably, in step S5, the hyperparameter search method includes at least one of random search, Bayesian search, or grid search, and the hyperparameters include learning rate, number of network layers, number of hidden units, batch size, and optimizer type.

[0029] Preferably, in step S6, the evaluation index includes the coefficient of determination (R²). 2 ) and root mean square error (RMSE), where, by choosing, the coefficient of determination (R) 2 The optimal model is evaluated and determined by maximizing the combination of hyperparameters that minimizes the root mean square error (RMSE).

[0030] Preferably, both the first nonlinear transformation and the second nonlinear transformation are calculated using a multilayer perceptron (MLP) network and a sigmoid activation function, and are applied to the feature vector to be adjusted in an element-wise multiplication manner.

[0031] Preferably, in step S7, the model interpretation and analysis includes visualizing the gating signal values ​​to quantify the relative dependence of the model on graph topological features and substructure features during the prediction process.

[0032] Preferably, in step S1, the public database is the PolyInfo polymer database, and the literature is relevant literature from SCI-indexed journals.

[0033] Due to the application of the above technical solution, the present invention has the following beneficial effects compared with the prior art:

[0034] 1. This invention offers high precision and complementary information. Through an innovative gating fusion mechanism, it achieves adaptive deep fusion of global topological information and local chemical information of polymer molecules. Specifically, graph neural networks can effectively capture the global spatial structure and topological relationships of molecules, while molecular fingerprints excel at characterizing specific chemical substructure features. Through the intelligent gating mechanism, the model can dynamically weigh and integrate these two complementary information sources, fundamentally overcoming the limitations of single information representation, thereby significantly improving the accuracy and reliability of the prediction model.

[0035] 2. This invention effectively addresses the problem of data sparsity. The multi-task learning framework adopted in this invention effectively solves the industry challenge of sparse data on polymer thermal properties. In response to the large differences in the amount of data for different thermal properties, the multi-task learning mechanism realizes knowledge transfer and sharing from data-rich tasks to data-sparse tasks, enabling the model to learn more generalizable feature representations from limited data. This significantly improves the prediction accuracy for data-scarce properties. This breakthrough provides strong technical support for polymer materials research.

[0036] 3. This invention has strong generalization and wide application. By organically combining multimodal fusion and multi-task learning, the model not only performs well in various thermal property prediction tasks, but more importantly, it has good robustness and adaptability. This enables the invention to be used as a general computing tool for high-throughput screening and performance prediction of large-scale polymer virtual libraries, providing an efficient and reliable computing platform for new material research and development, and has important industrial application value.

[0037] 4. This invention possesses interpretability. It provides a unique analytical perspective. By deeply analyzing the gate signal values ​​generated by the cross-gated fusion module, researchers can intuitively understand the model's dependence on and preference for different feature modes when making predictions. This interpretability not only provides an important window into understanding the model's decision-making mechanism, significantly enhancing the model's credibility and practicality, but also provides valuable theoretical references and practical guidance for polymer molecular design and performance research. Attached Figure Description

[0038] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, some of the drawings in the following description are some embodiments of the present invention. For those skilled in the art, other drawings can be made based on these drawings without creative effort.

[0039] Figure 1 This is a scatter plot comparing the predicted and actual values ​​of Embodiment 1 of the present invention;

[0040] Figure 2 This is an analysis diagram of the cross-gating mechanism in Embodiment 1 of the present invention;

[0041] Figure 3 This is a performance comparison chart of multi-task learning and single-task learning in Embodiment 1 of the present invention. Detailed Implementation

[0042] To provide a clearer understanding of the technical features, objectives, and effects of this invention, specific implementation schemes are now described in detail.

[0043] The present invention will be further described below with reference to embodiments, but the present invention is not limited to the following embodiments. The implementation conditions used in the embodiments can be further adjusted according to different requirements of specific use, and the implementation conditions not specified are conventional conditions in the industry. The technical features involved in the various embodiments of the present invention can be combined with each other as long as they do not conflict with each other.

[0044] Example 1

[0045] See appendix Figure 1 ~Attached Figure 3 This embodiment provides a method for predicting the thermal properties of polymers based on gated multimodal fusion and multi-task learning, including the following steps:

[0046] S1. Data Input and Preprocessing:

[0047] First, collect data containing polymer structures and their corresponding thermal properties (T). g T m T d T c The polymer structure is represented by SMILES (Simplified Linear Input Canonicalization) strings of repeating units in the PolyInfo polymer database. To ensure the uniqueness and consistency of the model input, all SMILES strings are canonicalized to remove redundancy and chemically equivalent representations. The data used in this embodiment comes from the PolyInfo polymer database and has undergone rigorous screening and cleaning. A total of 16,420 valid data points were collected and screened in this embodiment. See Table 1 for an overview of the datasets collected in this embodiment.

[0048] Table 1

[0049] Glass transition temperature <![CDATA[T g ]]> K [167,739] 7262 Melting temperature <![CDATA[T m ]]> K [223,756] 2911 Thermal decomposition temperature <![CDATA[T d ]]> K [398,990.5] 5118 softening point SP K [239,723] 702 Crystallization temperature <![CDATA[T c ]]> K [153,650] 427

[0050] S2, Dual-path feature extraction:

[0051] For each preprocessed polymer SMILES, the model extracts features through parallel dual paths;

[0052] First Path (GNN Branch): This path aims to capture the topological structure information of molecules;

[0053] (1) Convert the SMILES string into a molecular graph object, where atoms are nodes and chemical bonds are edges;

[0054] (2) Initialize feature vectors for each node and each edge in the graph. Node features (80-dimensional) may include atom type, degree, hybridization mode, aromaticity, etc.; edge features (6-dimensional) may include bond type, conjugation, whether it is in a loop, etc.

[0055] (3) Input the graph with features into a graph neural network (GNN). In this embodiment, a graph isomorphism network (GINE) with edge features is preferably used, and its update rule is as follows:

[0056]

[0057] Among them, h v (k) It is the feature vector of node v in the k-th layer. E is the set of neighboring nodes of node v. uv The features are those of the edges connecting v and u. MLP is a multilayer perceptron. GINE updates the representation of the central node by iteratively aggregating the information of neighboring nodes and edges, which can effectively learn rich local chemical environments.

[0058] (4) After processing by a multi-layer GINE network, graph-level pooling operations (such as global summation pooling and global max pooling) are performed on the final feature vectors of all nodes to obtain the first feature vector h representing the topology of the entire molecular graph. GNN .

[0059] The second pathway (FP branch): This pathway aims to capture key chemical substructures present in the molecule;

[0060] (1) Convert the SMILES string into one or more molecular fingerprints. In this embodiment, a mixed molecular fingerprint (Mixed FP) is preferred. It is composed of 167-bit MACCS Keys and 881-bit PubChem Fingerprints, totaling 1048 dimensions. This mixed fingerprint covers a wide range of chemical information from basic functional groups to more complex molecular fragments.

[0061] (2) Input the obtained 1048-dimensional binary fingerprint vector into a multilayer perceptron (MLP) network to map it into a low-dimensional, dense real vector space, and obtain the second feature vector h. FP This step aims to learn the non-linear combination relationships between different fingerprint positions.

[0062] S3, Cross-Modal Gating and Fusion: This aims to achieve intelligent and adaptive fusion of information from two modalities.

[0063] (1) Fingerprint gating of the image: the second feature vector h FP The input is fed into a dedicated MLP network and passed through a Sigmoid activation function to generate a signal similar to h. GNN Gated vectors g of the same dimension GNN g GNN Each element in h has a value between 0 and 1, representing the importance weight of the corresponding dimension information; then, h GNN With g GNN Perform element-wise multiplication (Hadamard product) to obtain the graph feature vector h′ after weighting and filtering by fingerprint information. GNN =h GNN ⊙g GNN ;

[0064] (2) Image-based fingerprint gating: Symmetrically, the first feature vector h′ is... GNN The input is fed into another dedicated MLP and sigmoid function to generate a gating vector g used to modulate fingerprint features. FP h FP With g FP Perform element-wise multiplication to obtain h′ GNN =h GNN ⊙g GNN ;

[0065] (3) Final fusion: The two cross-adjusted feature vectors h′ are merged. GNN and h′ FP Concatenation is performed to form the final fused feature vector h. fused ;

[0066] In this way, the model can dynamically decide whether to rely more on the overall connectivity information provided by GNN or the specific functional group information provided by FP, based on the input molecules, achieving a deep fusion effect that is far superior to simple splicing.

[0067] S4, Multi-task learning and prediction: This aims to improve the overall performance of the model by leveraging the inherent correlation between thermal properties, especially for tasks with sparse data.

[0068] (1) Fuse the feature vector h fused As input, it is fed into the final estimator, which is usually also an MLP network;

[0069] (2) To achieve multi-task learning, an attribute selector vector is introduced; this is a one-hot encoded vector whose dimension is equal to the total number of properties to be predicted (5 in this embodiment); for example, predicting Tg When the vector is [1,0,0,0,0], predict T. m When [0, 1, 0, 0, 0], and so on; this vector and the fused feature vector h fused They are input together into the prediction header;

[0070] (3) The prediction head outputs a single scalar value, i.e., the predicted value of the corresponding property, based on the fusion features and task identifier.

[0071] (4) During model training, all five types of data are aggregated together, and each data point is accompanied by its corresponding attribute selector vector; the model is trained end-to-end on the entire aggregated dataset to optimize a common regression loss function (such as mean squared error loss).

[0072] S5. Model Training and Optimization: Divide the entire aggregated dataset (containing data of all five properties) into two sets: 80% as the training set and 20% as the test set. Use the training set to train and optimize the hyperparameters of the multi-task prediction model built in step S4.

[0073] (1) Model training: The Adam optimizer was used, with the initial learning rate set to 0.001 and the batch size set to 64. During training, an early stopping strategy was used to prevent overfitting.

[0074] (2) Hyperparameter optimization: Based on the model’s performance on the validation set (which can be separated from the training set), optuna is used to optimize the model’s hyperparameters. The optimized hyperparameters include the learning rate (e.g., searching within the range of 1e-4 to 1e-2), the number of layers in the MLP, and the number of hidden units. The optimization objective is to minimize the root mean square error (RMSE) on the validation set. After optimization, an optimal set of hyperparameters is obtained.

[0075] S6. Model performance evaluation: The optimal model optimized in step S5 is evaluated comprehensively and robustly using 5-fold cross-validation. The training set is randomly divided into 5 groups. Four groups are used for training and one group is used for testing each time, and this is repeated 5 times.

[0076] (1) Evaluation indicators: using the coefficient of determination (R²) 2 The root mean square error (RMSE) is used as an indicator to evaluate model performance.

[0077] (2) Evaluation Results: The performance evaluation results of different models are shown in Tables 2 and 3. As can be seen from the data in the tables, the multimodal fusion model (FP-GINE) proposed in this invention performs best in all property predictions. Figure 1 The scatter plot shows the predicted and true values ​​of the model on the test set. It can be seen that the predicted points are closely distributed around the diagonal, indicating that the model has high prediction accuracy and good fitting ability. The prediction performance (R²) of different machine learning models on the five properties is also shown. 2 See Table 2; see Table 3 for the prediction performance (RMSE) of different machine learning models for the five properties.

[0078] Table 2

[0079] ECFP+Random Forest 0.8853 0.8035 0.7695 0.6987 0.7193 RDKit + Random Forest 0.8913 0.8061 0.7438 0.7249 0.7556 GINE-Only 0.9065 0.8266 0.7173 0.7345 0.7704 FP-Only 0.9240 0.8456 0.8288 0.7431 0.7496 FP-GINE 0.9405 0.8710 0.8391 0.7660 0.8042

[0080] Table 3

[0081] ECFP+Random Forest 36.74 43.65 50.09 51.68 51.78 RDKit + Random Forest 35.77 43.36 52.81 49.38 48.31 GINE-Only 33.17 41.00 55.47 48.51 46.83 FP-Only 29.90 38.69 43.17 47.71 48.90 FP-GINE 26.47 35.37 41.84 45.54 43.25

[0082] S7. Model Explanation and Analysis: In order to understand the working mechanism of the model, the cross-gating module in step S3 and the multi-task learning effect in step S4 are analyzed.

[0083] (1) Analysis of gating mechanism: as shown in the appendix Figure 2 As shown, the average activation values ​​of FP gates and GNN gates were compared using radar charts; the results show that the model's performance in predicting T... g T m T d When the properties are equal, the weights of the FP gate are significantly higher than those of the GNN gate, indicating that the model relies more on the chemical substructure information provided by the fingerprint.

[0084] (b) Analysis of Multi-Task Learning Effectiveness: See attached Figure 3 As shown, the performance of the multi-task model (FP-GINE-MT) and the single-task model (FP-GINE-ST) was compared; the results show that for sparse data T, the performance of the multi-task model is better than that of the single-task model (FP-GINE-ST). c In addition to the SP property, the RMSE of the multi-task model is significantly lower than that of the single-task model, proving that multi-task learning effectively improves the predictive ability of the model through knowledge transfer.

[0085] This invention, through the above embodiments, fully demonstrates a method for predicting polymer thermal properties based on gated multimodal fusion and multi-task learning. By inputting the SMILES structure representation of any given polymer into the optimal prediction model trained by this invention, various key thermal properties of the polymer (such as T) can be obtained quickly and accurately. g T m T d T cThe prediction results of SP). The method of this invention, through innovative dual-path feature extraction, cross-gated fusion, and a multi-task learning framework, profoundly incorporates the influence of polymer graph topology and chemical substructure information on performance. The core prediction framework developed using this invention can not only be used for homopolymers in this embodiment, but its principle can also be extended to other complex polymer systems such as copolymers and blends, as well as the prediction of other properties such as mechanical properties, optical properties, and barrier properties, demonstrating good versatility and scalability.

[0086] Compared to traditional experimental trial-and-error methods or single machine learning models, the method of this invention has the following significant advantages: Higher accuracy: By using a gating mechanism to adaptively fuse multimodal features, the synergistic effect of different information sources is fully explored, resulting in prediction accuracy that is significantly higher than traditional methods and single-modal models; Higher efficiency and lower cost: This invention utilizes data-driven and deep learning to achieve end-to-end prediction from molecular structure to multiple properties, greatly shortening the evaluation cycle of material properties, avoiding a large number of expensive and time-consuming experiments, effectively reducing research costs and improving R&D efficiency; Enhanced designability: This invention can not only predict properties, but its interpretable gating mechanism can also provide clues for understanding structure-property relationships, thereby enabling more rational and targeted design of polymer structures, opening up a new and efficient path for developing novel polymers with specific thermal properties (such as high glass transition temperatures).

[0087] In summary, this invention offers high precision and complementary information. Through an innovative gating fusion mechanism, it achieves adaptive deep fusion of global topological information and local chemical information of polymer molecules. Specifically, graph neural networks effectively capture the global spatial structure and topological relationships of molecules, while molecular fingerprints excel at characterizing specific chemical substructure features. Through an intelligent gating mechanism, the model dynamically balances and integrates these two complementary information sources, fundamentally overcoming the limitations of single information representation and significantly improving the accuracy and reliability of the prediction model. Furthermore, this invention effectively addresses the data sparsity problem. The multi-task learning framework employed effectively solves the industry challenge of sparse polymer thermal property data. Addressing the significant differences in data volume across different thermal properties, the multi-task learning mechanism enables knowledge transfer and sharing from data-rich tasks to data-sparse tasks, allowing the model to learn more generalizable feature representations from limited data. This significantly improves the prediction accuracy for data-scarce properties. This groundbreaking advancement provides strong technical support for polymer materials research. The invention boasts high generalization and wide applicability. By organically combining multimodal fusion and multi-task learning, the model not only performs exceptionally well in various thermal property prediction tasks but, more importantly, exhibits excellent robustness and adaptability. This enables the invention to serve as a general-purpose computational tool, successfully applied to high-throughput screening and performance prediction of large-scale polymer virtual libraries, providing an efficient and reliable computational platform for new material development and possessing significant industrial application value. Furthermore, the invention offers interpretability. It provides a unique analytical perspective; through in-depth analysis of the gating signal values ​​generated by the cross-gating fusion module, researchers can intuitively understand the model's dependence on and preference for different feature modes when making predictions. This interpretability not only provides an important window into understanding the model's decision-making mechanism, significantly enhancing its credibility and practicality, but also provides valuable theoretical reference and practical guidance for polymer molecular design and performance research.

[0088] The above-described embodiments are merely one implementation of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention should be determined by the appended claims.

Claims

1. A method for predicting polymer thermal properties based on gated multimodal fusion and multi-task learning, characterized in that, Includes the following steps: S1. Data Input and Preprocessing: Provide a dataset containing the structural representation of polymer repeating units and their corresponding thermal property labels, and normalize the structural representation of the polymer repeating units; S2. Dual-path feature extraction: Establish a dual-path feature extraction model, which includes: (a) A first feature extraction path for extracting graph topological features from the graph structure of polymer repeating units to obtain a first feature vector; the first path includes molecular graph transformation, initialization of feature vector, and processing of a graph isomorphic network including a multilayer perceptron; the first feature vector is used to characterize the entire molecular graph topological structure. (b) A second feature extraction path for extracting substructure features from the molecular fingerprint of polymer repeating units to obtain a second feature vector; the second path includes molecular fingerprint conversion and low-dimensional vector space mapping processing based on multilayer perceptron; the second feature vector includes key chemical substructures; S3, Cross-gated fusion: Construct a cross-gated fusion module, receive the first feature vector and the second feature vector, and generate a gating signal through the feature of one modality to adaptively adjust the information weight of the feature of the other modality, and output the fused multimodal feature vector; S4. Multi-task learning and prediction: Construct a multi-task prediction model, receive the fused multimodal feature vector and task identifier, and simultaneously predict multiple thermal properties; S5. Model Training and Optimization: Divide the dataset into training, validation, and test sets, train the model on the training set, and optimize the model hyperparameters using the hyperparameter search method. S6. Model Performance Evaluation: The optimized model is evaluated using evaluation metrics, either through cross-validation or independent test sets. S7. Model Interpretation and Analysis: Analyze the gate signal values ​​in the cross-gated fusion module to explain the model's dependence on different modal features; In step S3, the cross-gating fusion module is implemented in the following way: The second feature vector is transformed by a first nonlinear transformation to generate a first gate signal for element-wise multiplication with the first feature vector; the first gate signal has the same dimension as the first feature vector, and each element value is between 0 and 1, which is used to characterize the importance weight of the corresponding dimension information; the feature information obtained by performing element-wise multiplication between the first gate signal and the first feature vector is determined as a graph feature vector. The first feature vector is transformed by a second nonlinear transformation to generate a second gate signal for element-wise multiplication with the second feature vector. The feature information obtained by performing element-wise multiplication between the second gate signal and the second feature vector is concatenated with the graph feature vector to obtain the fused multimodal feature vector; Both the first and second nonlinear transformations are implemented using a multilayer perceptron network and a sigmoid activation function.

2. The method for predicting polymer thermal properties based on gated multimodal fusion and multi-task learning according to claim 1, characterized in that, In step S1, the data preprocessing includes standardizing the representation of the polymer repeating unit structure to ensure that chemically equivalent structures have a unique representation.

3. The polymer thermal property prediction method based on gated multimodal fusion and multi-task learning according to claim 1, characterized in that, In step S2(a), the first feature extraction path is implemented using a graph isomorphic network, which simultaneously aggregates neighbor node features and connection edge features when updating node representations.

4. The polymer thermal property prediction method based on gated multimodal fusion and multi-task learning according to claim 1, characterized in that, In step S2(b), the molecular fingerprint is a hybrid molecular fingerprint combining the MACCS key fingerprint and the PubChem fingerprint.

5. The method for predicting polymer thermal properties based on gated multimodal fusion and multi-task learning according to claim 1, characterized in that, In step S4, the multi-task prediction model is a unified deep neural network architecture, which can simultaneously predict at least two of the following thermal properties: glass transition temperature, melting temperature, decomposition temperature, crystallization temperature, and softening point, based on the task identifier.

6. The method for predicting polymer thermal properties based on gated multimodal fusion and multi-task learning according to claim 1, characterized in that, In step S5, the hyperparameter search method includes at least one of random search, Bayesian search, or grid search, and the hyperparameters include learning rate, number of network layers, number of hidden units, batch size, and optimizer type.

7. The method for predicting polymer thermal properties based on gated multimodal fusion and multi-task learning according to claim 1, characterized in that, In step S6, the evaluation metrics include the coefficient of determination and the root mean square error, wherein the optimal model is evaluated and determined by selecting a combination of hyperparameters that maximizes the coefficient of determination and minimizes the root mean square error.

8. The method for predicting polymer thermal properties based on gated multimodal fusion and multi-task learning according to claim 1, characterized in that, In step S7, the model interpretation and analysis includes visualizing the gating signal values ​​to quantify the relative dependence of the model on graph topological features and substructure features during the prediction process.

Images