Polymer glass transition temperature prediction method and device based on adaptive optimization algorithm and graph neural network, and medium

Abstract

The invention relates to a polymer glass transition temperature prediction method and device based on an adaptive optimization algorithm and a graph neural network, and a medium. The prediction method comprises the following steps: converting polymer molecular structure data into graph structure data containing node features and edge features; acquiring training parameter configuration information; constructing a graph neural network model comprising a graph convolution layer, a graph pooling layer and a full connection layer based on the training parameter configuration information, and training by adopting an improved RAdam optimization algorithm; after multiple batches of gradients are accumulated through a gradient accumulation strategy, model parameter updating is executed, the model performance is dynamically evaluated based on the mean absolute error and the mean square error of the verification set, and optimal model parameters are stored; and predicting the Tg of the polymer by using the optimized model. Compared with the prior art, the polymer Tg prediction precision and the training efficiency are synchronously improved, and the low-flux bottleneck of traditional experimental measurement and the precision-scale contradiction of computational simulation are broken through.

Description

Technical Field

[0001] The present invention relates to the intersection of materials science and machine learning, and in particular to a method, device, and medium for predicting the glass transition temperature of polymers based on an adaptive optimization algorithm and a graph neural network. Background Art

[0002] Accurate prediction of the glass transition temperature (Tg) of polymers is crucial for materials science and industrial applications. Existing prediction methods are mainly divided into two categories: experimental measurement and computational simulation. However, these traditional methods all have significant limitations. For example, experimental methods such as differential scanning calorimetry (DSC) and dynamic mechanical analysis (DMA), although considered the gold standard, are inefficient. A single experiment takes hours or even days, and high-quality samples need to be prepared, which makes it difficult to meet the needs of high-throughput material research and development. Computational chemistry methods such as molecular dynamics (MD) simulations and density functional theory (DFT) calculations have attempted to break through the experimental bottleneck, but MD relies on a preset force field and is difficult to accurately describe the flexible conformational changes of polymer chains and the dipole effects of polar groups, resulting in high prediction errors; the computational complexity of DFT increases exponentially with the number of atoms, and can only handle short-chain molecules, and cannot be extended to long-chain polymer systems in practical applications.

[0003] Recently, the application of artificial intelligence technology in the field of materials science has provided a new approach for polymer Tg prediction. CN119446329A proposed a prediction method based on molecular descriptors and deep learning models, which optimized the database and constructed a deep learning model through feature engineering. However, this method did not fully utilize the graph characteristics of the polymer molecular structure, and the model was not interpretable enough. CN119317967A emphasized the effect of molecular weight on Tg, and made predictions by graphically representing the polymer chemical structure and combining it with an artificial neural network. However, its model training efficiency was low, and it did not mention dynamic optimization strategies to adapt to the complexity of different polymer systems. CN115062181A focused on convolutional neural networks, encoding the SMILES string of polymer monomers into a binary image for prediction. However, the SMILES string has problems of information loss and fuzzy structural analysis when representing complex polymer structures, and it did not fully utilize the advantages of graph neural networks in processing molecular graph data.

[0004] In summary, existing technologies for predicting polymer glass transition temperatures suffer from low efficiency, insufficient precision, limited applicability, and poor model interpretability. Developing a method that can efficiently and accurately predict polymer Tg has become a pressing technical challenge. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned existing technologies and provide a polymer glass transition temperature prediction method, equipment, and medium based on an adaptive optimization algorithm and a graph neural network. By combining an improved stochastic gradient optimization algorithm (RAdam) with multi-scale molecular feature modeling, the polymer Tg prediction accuracy and training efficiency are simultaneously improved, breaking through the low-throughput bottleneck of traditional experimental measurements and the accuracy-scale contradiction of computational simulation.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] A first aspect of the present invention provides a method for predicting the glass transition temperature of a polymer based on an adaptive optimization algorithm and a graph neural network, comprising the following steps:

[0008] S1: Construct a polymer glass transition temperature dataset and convert the polymer molecular structure data into graph structure data containing node features and edge features;

[0009] S2: Get training parameter configuration information;

[0010] S3: Based on the training parameter configuration information in S2, a graph neural network model is constructed, including graph convolutional layers, graph pooling layers, and fully connected layers. The improved RAdam optimization algorithm is used for training. The learning rate is dynamically adjusted by calculating the effective sliding average length. According to the flag in the training parameter configuration information, phased training is performed, including pre-training and fine-tuning phases.

[0011] S4: After accumulating multiple batches of gradients through the gradient accumulation strategy, the model parameters are updated. The model performance is dynamically evaluated based on the mean absolute error and mean square error of the validation set, and the optimal model parameters are saved.

[0012] S5: Use the optimized model in S4 to predict the Tg of polymer.

[0013] Furthermore, S1 specifically includes the following processes:

[0014] Collecting raw data of polymer molecular structure and preprocessing the collected raw data;

[0015] For the pre-processed polymer molecular structure data, the type and hybridization state of each atom in the molecule are used as node features, and the type and bond length of each chemical bond in the molecule are used as edge features;

[0016] The processed node and edge features are integrated to construct complete graph structure data, forming a polymer glass transition temperature dataset containing node features and edge features.

[0017] Furthermore, the raw data of the polymer molecular structure is collected, and the specific process of preprocessing the collected raw data includes:

[0018] Obtain structural information and glass transition temperature data of polymer molecules;

[0019] Clean the collected raw data to remove duplicate, erroneous, and incomplete records;

[0020] Standardize data and unify data formats and units;

[0021] Analyze the molecular structure and extract the type and hybridization state of each atom in the molecule, and the type and bond length of each chemical bond.

[0022] Furthermore, for the pre-processed polymer molecular structure data, the type and hybridization state of each atom in the molecule are used as node features, and the type and bond length of each chemical bond in the molecule are used as edge features. The specific process includes:

[0023] Based on the preprocessed polymer molecular structure data, a node record is created for each atom, and the atom type and hybridization state are encoded as a node feature vector;

[0024] Create edge records for each pair of bonded atoms, encoding the bond type and bond length as edge feature vectors;

[0025] Use graph data structures to integrate nodes and edges with features to construct a complete molecular graph;

[0026] All molecular maps and their corresponding glass transition temperature data were combined to form a complete data set.

[0027] Furthermore, in S2, the training parameter configuration information includes: model structure, data path, initial learning rate, training rounds, batch size, pre-training stage flag, and fine-tuning stage flag. These flags are used to indicate different stages of model training.

[0028] The model structure includes the settings of graph convolution layer, graph pooling layer, and fully connected layer;

[0029] The data path includes storage locations for training and validation data.

[0030] Furthermore, S3 specifically includes the following steps:

[0031] According to the training parameter configuration information in S2, the parameters of each layer of the graph neural network model are determined, and the graph convolution layer, graph pooling layer, and fully connected layer are constructed in sequence;

[0032] Initialize model parameters to prepare for training using the improved RAdam optimization algorithm;

[0033] During training, the improved RAdam optimization algorithm is used to calculate the first-order moment and second-order moment exponential moving average of the gradient, and the effective sliding average length ρ is calculated according to the current training step number. t , when ρ t When ρ > 5, a variance correction term is introduced to adjust the learning rate, t When ≤5, the traditional momentum update rule is used;

[0034] Perform phased training based on the flag in the training parameter configuration information: If the flag indicates the pre-training phase, the model parameters are initialized with the global learning rate and trained with a large dataset. When switching to the fine-tuning phase, the pre-training parameters are inherited and the local learning rate is adjusted according to the specific dataset.

[0035] Furthermore, the improved RAdam optimization algorithm in the present invention is a stochastic gradient optimization method for deep learning model training. The traditional Adam optimization algorithm is improved by introducing a variance correction mechanism, thereby improving the convergence speed and training stability of the model. First, the exponential moving average of the first-order moment and the second-order moment of the gradient of the model parameters is calculated, and then the effective sliding average length is dynamically calculated according to the current number of training steps. If the effective sliding average length is greater than 5, the algorithm will introduce a variance correction term to adjust the learning rate to suppress the fluctuation of the gradient variance; if the effective sliding average length is less than or equal to 5, the traditional momentum update rule is adopted.

[0036] Furthermore, S4 specifically includes the following steps:

[0037] Initialize the variables required for gradient accumulation and the validation set evaluation indicators;

[0038] Use forward propagation to calculate the loss, then perform backpropagation to calculate the gradient and accumulate it to a specified batch, perform parameter updates, and dynamically evaluate the model performance based on the mean absolute error and mean square error of the validation set;

[0039] Save the optimal model parameters to optimize the model performance.

[0040] Furthermore, the improved RAdam optimization algorithm incorporates a gradient accumulation strategy, allowing gradients to be accumulated over multiple training batches before performing a single parameter update. This feature not only improves model training efficiency on large datasets but is also particularly suitable for scenarios with limited graphics memory.

[0041] Furthermore, S5 specifically includes the following steps:

[0042] Convert the polymer molecular structure that needs to be predicted into graph structure data;

[0043] The graph structure data is input into the optimized model. The model extracts local features through the graph convolution layer, reduces the feature dimension through the graph pooling layer, and performs feature fusion in the fully connected layer, and finally outputs the predicted value of the polymer glass transition temperature Tg.

[0044] A second aspect of the present invention provides an electronic device comprising a memory and a processor, wherein the processor is configured to execute a program in the memory, thereby implementing the above-mentioned polymer glass transition temperature prediction method based on the adaptive optimization algorithm and graph neural network.

[0045] A third aspect of the present invention provides a storage medium comprising computer-executable instructions, which, when executed by a computer processor, is used to execute the above-mentioned polymer glass transition temperature prediction method based on the adaptive optimization algorithm and graph neural network.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1) The improved RAdam optimization algorithm effectively solves the problem of gradient oscillation in the traditional Adam algorithm at the beginning of training, significantly improving the convergence speed and training stability of the model. The dynamic variance correction mechanism makes the learning rate adjustment more accurate, especially the ability to automatically adapt to the optimal learning rate at different training stages, greatly shortening the training time and improving training efficiency. Secondly, the use of a graph neural network architecture can fully utilize the graph structure characteristics of polymer molecules, deeply explore the interactions between atoms and long-range dependencies, and thus achieve high-precision prediction of the polymer glass transition temperature (Tg). The collaborative work of the graph convolution layer, graph pooling layer, and fully connected layer not only improves the model's ability to capture the features of complex molecular structures, but also effectively reduces the computational complexity, enabling the model to process large-scale polymer datasets.

[0048] 2) The modular training process design of the present invention enhances the flexibility and reusability of the system. By dynamically loading external configuration files and parsing command line parameters, it can quickly adapt to different training tasks and data sets, reducing manual intervention and repetitive work. The seamless switching and parameter update mechanism between the pre-training and fine-tuning stages enables the model to maintain good performance at different stages, further improving the generalization ability and adaptability of the model. Finally, the addition of the breakpoint resume function effectively avoids training failures caused by accidental interruptions, ensures the reliability and integrity of long-term training tasks, and provides users with a more stable and efficient model training experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a schematic flow chart of the polymer glass transition temperature prediction method based on the adaptive optimization algorithm and graph neural network in the present invention.

[0050] Figure 2 Schematic diagram of the process of the adaptive optimization method based on dynamic variance correction in the present invention. DETAILED DESCRIPTION

[0051] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. Component models, material names, connection structures, circuit structures, control methods, algorithms, and other features not explicitly described in this technical solution are considered common technical features disclosed in the prior art.

[0052] Example 1

[0053] This embodiment provides a polymer glass transition temperature prediction method based on an adaptive optimization algorithm and a graph neural network. Figure 1 , including the following steps:

[0054] S1: Construct a polymer glass transition temperature dataset and convert the polymer molecular structure data into graph structure data containing node features and edge features;

[0055] In the specific implementation, S1 includes the following processes:

[0056] Collecting raw data of polymer molecular structure and preprocessing the collected raw data;

[0057] For the pre-processed polymer molecular structure data, the type and hybridization state of each atom in the molecule are used as node features, and the type and bond length of each chemical bond in the molecule are used as edge features;

[0058] The processed node and edge features are integrated to construct complete graph structure data, forming a polymer glass transition temperature dataset containing node features and edge features.

[0059] In specific implementation, the raw data of polymer molecular structure is collected, and the specific process of preprocessing the collected raw data includes:

[0060] Obtain structural information and glass transition temperature data of polymer molecules;

[0061] Clean the collected raw data to remove duplicate, erroneous, and incomplete records;

[0062] Standardize data and unify data formats and units;

[0063] Analyze the molecular structure and extract the type and hybridization state of each atom in the molecule, and the type and bond length of each chemical bond.

[0064] In specific implementation, for the pre-processed polymer molecular structure data, the type and hybridization state of each atom in the molecule are used as node features, and the type and bond length of each chemical bond in the molecule are used as edge features. The specific process includes:

[0065] Based on the preprocessed polymer molecular structure data, a node record is created for each atom, and the atom type and hybridization state are encoded as a node feature vector;

[0066] Create edge records for each pair of bonded atoms, encoding the bond type and bond length as edge feature vectors;

[0067] Use graph data structures to integrate nodes and edges with features to construct a complete molecular graph;

[0068] All molecular maps and their corresponding glass transition temperature data were combined to form a complete data set.

[0069] The technical principle of step S1 is mainly based on converting polymer molecular structure data into a graph structure data format that can be processed by graph neural networks. Its core technical principles include the following aspects:

[0070] Raw data collection and preprocessing: A preliminary data set is formed by collecting structural information and glass transition temperature data of polymer molecules. The data is then cleaned to remove duplicate, erroneous, and incomplete records to ensure accuracy and reliability. The data is then standardized to unify the data format and units, eliminating inconsistencies caused by different data sources and laying the foundation for subsequent feature extraction and model training. Finally, the molecular structure is analyzed to extract the types and hybridization states of each atom in the molecule, as well as the types and bond lengths of each chemical bond. This is a key prerequisite for converting the molecular structure into graph-structured data.

[0071] Graph structure data construction: Based on the pre-processed polymer molecular structure data, a node record is created for each atom, and the atom type and hybridization state are encoded as node feature vectors. The atom type is directly related to the atomic number of the element, reflecting the type of atom, while the hybridization state reflects the orbital hybridization form of the atom when forming a molecule, such as sp, sp 2 、sp 3Etc., which is of great significance for understanding the bonding characteristics and electron distribution of atoms. At the same time, edge records are created for each pair of bonding atoms, and the type of chemical bond (such as single bond, double bond, triple bond, etc.) and bond length are encoded as edge feature vectors. The type of chemical bond reflects the connection mode and bond energy between atoms, while the bond length reflects the spatial distance between atoms. This information together determines the geometric configuration and stability of the molecule. The graph data structure is used to integrate the nodes and edges with features to construct a complete molecular graph, which achieves an accurate representation of the molecular structure of the polymer. Finally, all molecular graphs and their corresponding glass transition temperature data are combined to form a complete data set, which provides a rich data foundation for the subsequent graph neural network model training, enabling the model to learn the intrinsic relationship between molecular structure and glass transition temperature.

[0072] Through step S1, the complex polymer molecular structure data is converted into graph structure data suitable for graph neural network processing, which realizes the effective extraction and encoding of molecular structure information and provides high-quality data support for subsequent model training and prediction.

[0073] S2: Get training parameter configuration information;

[0074] In specific implementation, in S2, the training parameter configuration information includes: model structure, data path, initial learning rate, training rounds, batch size, pre-training stage flag, and fine-tuning stage flag. These flags are used to indicate different stages of model training;

[0075] The model structure includes the settings of graph convolution layer, graph pooling layer, and fully connected layer;

[0076] The data path includes storage locations for training and validation data.

[0077] The technical principle of step S2 is mainly based on the efficient management and flexible configuration of the deep learning model training process. Its core technical principles are as follows:

[0078] By combining loading external configuration files with parsing command-line parameters, comprehensive parameter information required for model training is obtained to ensure the flexibility and configurability of model training. The external configuration file provides default values ​​for key parameters such as the model structure (including the specific settings of the graph convolution layer, graph pooling layer, and fully connected layer), data path (storage location of training and validation data), initial learning rate, training rounds, batch size, etc., providing a basic framework for model training. The command-line parameter parsing unit uses the argparse library to parse the command-line parameters entered by the user. These parameters can dynamically overwrite or supplement the default parameters in the configuration file, such as the device index (used to specify the computing device) and the breakpoint resume flag (used to control whether to continue training from the historical training state), thereby meeting the user's personalized needs in different training environments and tasks.

[0079] The training parameter configuration information also includes pre-training and fine-tuning flags, which indicate different stages of model training. The pre-training flag guides the model's initial training on a large-scale dataset. At this stage, the model parameters are initialized in a more general way, and hyperparameters such as the learning rate are set relatively stably, allowing the model to learn the basic characteristics and patterns of polymer molecular structures. The fine-tuning flag instructs the model to perform further optimization on a specific dataset. At this stage, the model inherits the parameters from the pre-training phase and adjusts hyperparameters such as the local learning rate based on the characteristics of the specific dataset to improve the model's performance on specific tasks.

[0080] Through step S2, the model training process can be efficiently and flexibly managed and executed based on the acquired training parameter configuration information, ensuring that the model can achieve the best training effect at different training stages and providing reliable performance guarantee for subsequent model predictions.

[0081] S3: Based on the training parameter configuration information in S2, a graph neural network model is constructed, including graph convolutional layers, graph pooling layers, and fully connected layers. The improved RAdam optimization algorithm is used for training. The learning rate is dynamically adjusted by calculating the effective sliding average length. According to the flag in the training parameter configuration information, phased training is performed, including pre-training and fine-tuning phases.

[0082] In the specific implementation, S3 specifically includes the following steps:

[0083] According to the training parameter configuration information in S2, the parameters of each layer of the graph neural network model are determined, and the graph convolution layer, graph pooling layer, and fully connected layer are constructed in sequence;

[0084] Initialize model parameters to prepare for training using the improved RAdam optimization algorithm;

[0085] During training, the improved RAdam optimization algorithm is used to calculate the first-order moment and second-order moment exponential moving average of the gradient, and the effective sliding average length ρ is calculated according to the current training step number. t , when ρ t When ρ > 5, a variance correction term is introduced to adjust the learning rate, t When ≤5, the traditional momentum update rule is used;

[0086] Perform phased training based on the flag in the training parameter configuration information: If the flag indicates the pre-training phase, the model parameters are initialized with the global learning rate and trained with a large dataset. When switching to the fine-tuning phase, the pre-training parameters are inherited and the local learning rate is adjusted according to the specific dataset.

[0087] In specific implementation, the improved RAdam optimization algorithm in the present invention is a stochastic gradient optimization method for deep learning model training. The traditional Adam optimization algorithm is improved by introducing a variance correction mechanism, thereby improving the convergence speed and training stability of the model. First, the exponential moving average of the first-order moment and the second-order moment of the gradient of the model parameters is calculated, and then the effective sliding average length is dynamically calculated according to the current number of training steps. If the effective sliding average length is greater than 5, the algorithm will introduce a variance correction term to adjust the learning rate to suppress the fluctuation of the gradient variance; if the effective sliding average length is less than or equal to 5, the traditional momentum update rule is adopted.

[0088] The technical principle of step S3 is mainly based on the combination of graph neural network and improved RAdam optimization algorithm to achieve efficient prediction of polymer glass transition temperature. It is specifically reflected in the following aspects:

[0089] Based on the training parameter configuration information in S2, the parameters of each layer of the graph neural network model are determined, and the graph convolution layer, graph pooling layer, and fully connected layer are constructed in sequence. The graph convolution layer is used to extract local features of the molecular graph by sliding the convolution kernel on the molecular graph to extract features and capture the interactions between atoms. The graph pooling layer is used to reduce the dimensionality of the graph data, while retaining key information and reducing the amount of computation. The fully connected layer is used to fuse the features extracted by the graph convolution layer and the graph pooling layer to output the final polymer Tg prediction result.

[0090] Initialize the model parameters and prepare for training using the improved RAdam optimization algorithm. During training, the improved RAdam optimization algorithm is used to calculate the first-order moment and second-order moment exponential moving average of the gradient, and the effective sliding average length ρ is calculated according to the current training step number. t , when ρ t When ρ > 5, a variance correction term is introduced to adjust the learning rate, t When ≤ 5, the traditional momentum update rule is used. The improved RAdam optimization algorithm improves the traditional Adam optimization algorithm by introducing a variance correction mechanism, thereby improving the model's convergence speed and training stability.

[0091] Phased training strategy: Training is performed in phases, depending on the flag in the training parameter configuration. If the flag indicates the pre-training phase, the model parameters are initialized using a global learning rate and trained on a large dataset, enabling the model to learn the fundamental characteristics and patterns of polymer molecular structures. During the fine-tuning phase, the pre-training parameters are inherited, and the local learning rate is adjusted based on the specific dataset to further optimize the model and improve its performance on specific tasks.

[0092] Through step S3, the construction of the graph neural network model is combined with the improved RAdam optimization algorithm to achieve efficient prediction of the glass transition temperature of the polymer. At the same time, the generalization ability and adaptability of the model are improved through a phased training strategy.

[0093] S4: After accumulating multiple batches of gradients through the gradient accumulation strategy, the model parameters are updated. The model performance is dynamically evaluated based on the mean absolute error and mean square error of the validation set, and the optimal model parameters are saved.

[0094] In the specific implementation, S4 specifically includes the following steps:

[0095] Initialize the variables required for gradient accumulation and the validation set evaluation indicators;

[0096] Use forward propagation to calculate the loss, then perform backpropagation to calculate the gradient and accumulate it to a specified batch, perform parameter updates, and dynamically evaluate the model performance based on the mean absolute error and mean square error of the validation set;

[0097] Save the optimal model parameters to optimize the model performance.

[0098] In practice, the improved RAdam optimization algorithm incorporates a gradient accumulation strategy, allowing gradients to be accumulated over multiple training batches before performing a single parameter update. This feature not only improves model training efficiency on large datasets but is also particularly suitable for scenarios with limited graphics memory.

[0099] The technical principle of step S4 is mainly based on the gradient accumulation and dynamic evaluation mechanism to achieve efficient and stable model training and save the optimal parameters. The details are as follows:

[0100] Using a gradient accumulation strategy, parameter updates are performed after accumulating gradients over multiple mini-batches. The gradient for a single mini-batch is first calculated and accumulated. Once the set batch size is reached, the parameters are updated, and the accumulated gradients are reset to continue training. This strategy simulates the effects of large-batch training, improving model performance on large-scale datasets, reducing memory usage, and adapting to scenarios with limited graphics memory. During training, the model is regularly evaluated on the validation set, and error metrics (mean absolute error and mean squared error) between the predicted and true values ​​are calculated to monitor model performance. Based on the validation set evaluation results, when the model performance reaches optimal, the current model parameters are automatically saved to ensure the final model has optimal predictive performance and avoid overfitting and performance degradation.

[0101] Through step S4, combined with the gradient accumulation and dynamic evaluation mechanism, the model parameters are effectively updated and the optimal parameters are saved, which improves the model training efficiency and performance and ensures the prediction accuracy of the model.

[0102] S5: Use the optimized model in S4 to predict the Tg of polymer.

[0103] In the specific implementation, S5 specifically includes the following steps:

[0104] Convert the polymer molecular structure that needs to be predicted into graph structure data;

[0105] The graph structure data is input into the optimized model. The model extracts local features through the graph convolution layer, reduces the feature dimension through the graph pooling layer, and performs feature fusion in the fully connected layer, and finally outputs the predicted value of the polymer glass transition temperature Tg.

[0106] The technical principle of step S5 is mainly based on the feature extraction and fusion of polymer molecular structure data by the graph neural network model, thereby achieving accurate prediction of the glass transition temperature (Tg). The details are as follows:

[0107] Preprocess the polymer molecular structure to be predicted and convert it into graph-structured data. This step is a critical pre-processing operation that ensures the model can recognize and process the input data. Specifically, it involves extracting the type and hybridization state of each atom in the molecule as node features, extracting the type and length of each chemical bond as edge features, and integrating these features into graph-structured data.

[0108] The graph structure data is input into the optimized graph neural network model in S4. The model uses graph convolution layers to extract local features from the molecular graph data, capturing information about atomic interactions and the local molecular structure. The graph pooling layer performs dimensionality reduction on the extracted features, preserving key features while reducing computational complexity. Finally, the fully connected layer globally fuses the reduced features, integrating all local and global information to output a predicted value for the polymer's glass transition temperature (Tg).

[0109] Example 2

[0110] Based on the solution of Comparative Example 1, this embodiment provides an adaptive optimization method based on dynamic variance correction, which can be used in the corresponding steps of the embodiment, see Figure 2 The core of this approach is to achieve efficient updates of model parameters through an improved RAdam optimizer, and to improve polymer Tg prediction performance by combining molecular graph structural feature encoding. It also uses a specific graph neural network architecture, such as the KGNNModel, which includes graph convolutional layers, graph pooling layers, and fully connected layers.

[0111] The graph convolution layer slides the convolution kernel across the polymer molecular graph to extract local features such as atomic interactions, providing a foundation for subsequent analysis. The graph pooling layer performs dimensionality reduction on the graph data, preserving key information while reducing computational effort and improving model training efficiency. The fully connected layer fuses the previously extracted features to output the final polymer Tg prediction result. The collaborative work of these three layers effectively addresses the cross-scale information fragmentation problem in traditional methods and improves the model's ability to capture polymer molecular structural features.

[0112] The adaptive optimization method of dynamic variance correction is: introducing the exponential moving average strategy to track the first-order moment (mean) and second-order moment (variance) of the gradient respectively, and the calculation formula is: m t =β1m t-1 +(1-β1)g t ,v t =β2v t-1 +(1-β2)g t 2 ) Among them, g t is the current batch gradient, m t and v t are the exponential moving averages of the mean and variance of the gradient, respectively. β1 and β2 are the decay rates of the exponential moving averages used to compute the first moment (mean) and second moment (variance) of the gradient.

[0113] In specific implementation, bias correction and adaptive learning rate adjustment are performed by calculating the bias correction factor 1-β1 t and 1-β2 t The mean and variance are estimated unbiasedly to obtain the corrected mean. The theoretical maximum sliding average length ρ is introduced ∞ =2 / (1-β2)-1, dynamically calculate the current effective sliding average length When ρ t When ρ > 5, the learning rate is adjusted through the variance correction term to suppress the gradient variance fluctuation and avoid the early gradient oscillation problem of the traditional Adam algorithm; when ρ t When <5, the conventional momentum update strategy is adopted to ensure stability in the initial stage of training.

[0114] Specifically, when ρ t When >5, a variance correction term is introduced Suppress early gradient oscillation; when ρ t When ≤5, traditional momentum update is used to ensure stability in the initial stage of training; weight decay is achieved by introducing L2 regularization to avoid overfitting.

[0115] In specific implementation, the graph neural network GCNN is used to model the graph structure of polymer molecules, convert molecular fragments (such as side chains and cross-linking groups) into node features, and capture the interactions between atoms and long-range dependencies through graph convolution operations to solve the problem of cross-scale information fragmentation in traditional methods.

[0116] During implementation, the number of GCNN convolutional layers, attention mechanism parameters, and optimizer hyperparameters (such as learning rate and batch size) are dynamically adjusted according to the structural complexity of the training data. Adaptive strategies are used to replace hard-coded configurations to improve the model's generalization ability for different polymer systems.

[0117] Example 3

[0118] This embodiment provides an electronic device with the following specific configuration: In terms of hardware, the electronic device is equipped with high-speed, large-capacity memory and a high-performance processor. The memory can be DDR4 or DDR5 type, with a capacity of 32GB or more to meet the storage and call requirements of large-scale deep learning models; the processor is selected to have multiple cores, high main frequency, and powerful floating-point computing capabilities, such as Intel Core i9 or Xeon series, or AMD Ryzen Threadripper series, to ensure efficient operation of the model training and prediction process. The memory pre-stores the program code corresponding to the polymer glass transition temperature prediction method based on the adaptive optimization algorithm and graph neural network of the present invention. The program code is written in Python and fully utilizes the rich library functions and optimization tools of deep learning frameworks such as PyTorch or TensorFlow.

[0119] In practical applications, researchers or corporate R&D departments input polymer molecular structure data into the electronic device. This data comes from an internal experimental database or a publicly available polymer material dataset. According to a preset program, the processor first preprocesses the input molecular structure data, converting it into graph structure data. It then sequentially constructs a graph neural network model consisting of a graph convolutional layer, a graph pooling layer, and a fully connected layer. The model's hyperparameters are initialized by loading training parameter configuration information, such as a learning rate of 0.001, 100 training rounds, and a batch size of 32. Next, an improved RAdam optimization algorithm is used to initiate the model training process. During training, a gradient accumulation strategy is used to perform a parameter update after accumulating four batches of gradients. After each round of training, the mean absolute error and mean square error are calculated based on the validation set to dynamically evaluate model performance. When the model's performance on the validation set reaches optimality, the current model parameters are automatically saved to a specified location in memory. Finally, when the glass transition temperature of a new polymer material needs to be predicted, it is only necessary to input the molecular structure data of the polymer to be predicted into the electronic device. The processor calls the optimized model and its parameters. After forward propagation calculation, the predicted glass transition temperature value of the polymer can be output. The whole process is efficient and accurate, providing strong technical support for the design and development of polymer materials.

[0120] Example 4

[0121] This embodiment provides a storage medium containing computer executable instructions. Specifically, the storage medium can be a USB flash drive, a mobile hard drive, an optical disk, a solid-state drive, or a network storage device, etc., which stores program codes that can be directly executed by a computer processor after compilation. These program codes are written in Python and developed based on the PyTorch deep learning framework. They integrate core components such as the improved RAdam optimization algorithm and can realize all the functions of the polymer glass transition temperature prediction method based on the adaptive optimization algorithm and the graph neural network. When researchers or relevant technicians connect the storage medium to a computer device and start the corresponding execution program, the computer processor will read and execute the instructions in the storage medium, thereby completing the entire process from preprocessing of polymer molecular structure data, construction and training of the graph neural network model, to final glass transition temperature prediction. For example, in a materials science laboratory, researchers can insert a USB flash drive storing the program code into an experimental workstation, run the program, and input the collected polymer molecular structure data. Driven by the improved RAdam optimization algorithm, the workstation quickly and accurately outputs the prediction results of the polymer glass transition temperature according to the operating procedures defined by the program code, providing key data support for the research and development of new polymer materials.

[0122] The above description of the embodiments is intended to facilitate understanding and use of the invention by those skilled in the art. It will be apparent that those skilled in the art can readily make various modifications to these embodiments and apply the general principles described herein to other embodiments without requiring inventive effort. Therefore, the present invention is not limited to the above-described embodiments. Improvements and modifications made by those skilled in the art based on the disclosure of the present invention, without departing from the scope of the present invention, should be within the scope of protection of the present invention.

Claims

1. A polymer glass transition temperature prediction method based on adaptive optimization algorithm and graph neural network, characterized in that: The following steps are involved: S1: Construct a polymer glass transition temperature dataset and convert the polymer molecular structure data into graph structure data containing node features and edge features; S2: Get training parameter configuration information; S3: Based on the training parameter configuration information in S2, a graph neural network model is constructed, including graph convolutional layers, graph pooling layers, and fully connected layers. The improved RAdam optimization algorithm is used for training. The learning rate is dynamically adjusted by calculating the effective sliding average length. According to the flag in the training parameter configuration information, phased training is performed, including pre-training and fine-tuning phases. S4: After accumulating multiple batches of gradients through the gradient accumulation strategy, the model parameters are updated. The model performance is dynamically evaluated based on the mean absolute error and mean square error of the validation set, and the optimal model parameters are saved. S5: Use the optimized model in S4 to predict the Tg of polymer.

2. The method for predicting polymer glass transition temperature based on adaptive optimization algorithm and graph neural network according to claim 1, characterized in that: S1 specifically includes the following processes: Collecting raw data of polymer molecular structure and preprocessing the collected raw data; For the pre-processed polymer molecular structure data, the type and hybridization state of each atom in the molecule are used as node features, and the type and bond length of each chemical bond in the molecule are used as edge features; The processed node and edge features are integrated to construct complete graph structure data, forming a polymer glass transition temperature dataset containing node features and edge features.

3. The method for predicting polymer glass transition temperature based on adaptive optimization algorithm and graph neural network according to claim 2, characterized in that: The specific process of collecting the raw data of the polymer molecular structure and preprocessing the collected raw data includes: Obtain structural information and glass transition temperature data of polymer molecules; Clean the collected raw data to remove duplicate, erroneous, and incomplete records; Standardize data and unify data formats and units; Analyze the molecular structure and extract the type and hybridization state of each atom in the molecule, and the type and bond length of each chemical bond.

4. The method for predicting polymer glass transition temperature based on adaptive optimization algorithm and graph neural network according to claim 2, characterized in that: For the preprocessed polymer molecular structure data, the specific process of taking the type and hybridization state of each atom in the molecule as node features and the type and bond length of each chemical bond in the molecule as edge features includes: Based on the preprocessed polymer molecular structure data, a node record is created for each atom, and the atom type and hybridization state are encoded as a node feature vector; Create edge records for each pair of bonded atoms, encoding the bond type and bond length as edge feature vectors; Use graph data structures to integrate nodes and edges with features to construct a complete molecular graph; All molecular maps and their corresponding glass transition temperature data were combined to form a complete data set.

5. The method for predicting polymer glass transition temperature based on adaptive optimization algorithm and graph neural network according to claim 1, characterized in that: In S2, the training parameter configuration information includes: model structure, data path, initial learning rate, training rounds, batch size, pre-training stage flag, and fine-tuning stage flag. These flags are used to indicate different stages of model training; The model structure includes the settings of graph convolution layer, graph pooling layer, and fully connected layer; The data path includes storage locations for training and validation data.

6. The method for predicting polymer glass transition temperature based on adaptive optimization algorithm and graph neural network according to claim 1, characterized in that: In S3, the following steps are specifically included: According to the training parameter configuration information in S2, the parameters of each layer of the graph neural network model are determined, and the graph convolution layer, graph pooling layer, and fully connected layer are constructed in sequence; Initialize model parameters to prepare for training using the improved RAdam optimization algorithm; During training, the improved RAdam optimization algorithm is used to calculate the first-order moment and second-order moment exponential moving average of the gradient, and the effective sliding average length ρ is calculated according to the current training step number. t , when ρ t When ρ > 5, a variance correction term is introduced to adjust the learning rate, t When ≤5, the traditional momentum update rule is used; Perform phased training based on the flag in the training parameter configuration information: If the flag indicates the pre-training phase, the model parameters are initialized with the global learning rate and trained with a large dataset. When switching to the fine-tuning phase, the pre-training parameters are inherited and the local learning rate is adjusted according to the specific dataset.

7. The method for predicting polymer glass transition temperature based on adaptive optimization algorithm and graph neural network according to claim 1, characterized in that: S4 specifically includes the following steps: Initialize the variables required for gradient accumulation and the validation set evaluation indicators; Use forward propagation to calculate the loss, then perform backpropagation to calculate the gradient and accumulate it to a specified batch, perform parameter updates, and dynamically evaluate the model performance based on the mean absolute error and mean square error of the validation set; Save the optimal model parameters to optimize the model performance.

8. The method for predicting polymer glass transition temperature based on adaptive optimization algorithm and graph neural network according to claim 1, characterized in that: S5 specifically includes the following steps: Convert the polymer molecular structure that needs to be predicted into graph structure data; The graph structure data is input into the optimized model. The model extracts local features through the graph convolution layer, reduces the feature dimension through the graph pooling layer, and performs feature fusion in the fully connected layer, and finally outputs the predicted value of the polymer glass transition temperature Tg.

9. An electronic device comprising a memory and a processor, characterized in that: The processor is used to execute the program in the memory to implement the polymer glass transition temperature prediction method based on the adaptive optimization algorithm and graph neural network as described in any one of claims 1 to 8.

10. A storage medium containing computer-executable instructions, characterized in that: The storage medium of the computer executable instructions is used to execute the polymer glass transition temperature prediction method based on the adaptive optimization algorithm and graph neural network according to any one of claims 1 to 8 when executed by a computer processor.

Images