Polymer compound conductive network structure screening method, system and device
By predicting the electrical properties of conductive nanoparticle networks using an incremental graph neural network model, this approach solves the problem of capturing multi-level structural information in existing technologies, achieving efficient and accurate screening of conductive network structures and providing theoretical support for novel conductive polymer nanocomposites.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTHWEST MEDICAL UNIV
- Filing Date
- 2026-01-28
- Publication Date
- 2026-05-12
AI Technical Summary
Existing technologies struggle to capture complex multi-level structural information, making it difficult to predict the polymer properties of conductive polymer nanocomposites. Furthermore, experimental methods are time-consuming, costly, and limited to molecular-scale observations.
An incremental graph neural network model is used to learn the physical property laws from three-dimensional structural data. The electrical performance of conductive nanoparticle networks is predicted by graph attention network. By combining the incremental learning strategy and graph structure dataset, a method for screening conductive network structures is constructed.
Accurate prediction of the conductivity and connectivity of conductive network structures reduces experimental costs, improves the accuracy and efficiency of prediction, and provides a theoretical basis for conductive polymer nanocomposites.
Smart Images
Figure CN122024927A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of conductive polymer nanocomposite materials technology, specifically to a method, system, and apparatus for screening conductive network structures of polymer composites. Background Technology
[0002] With the continuous advancement of science and technology, the research and development of novel materials has become a key force driving the development of modern science and technology. Especially in fields such as electronics, energy, and medicine, the demand for high-performance materials is increasing daily. Conductive polymer nanocomposites (CPNs) have attracted widespread attention due to their unique physical, chemical, and electrical properties. In recent years, scientists have conducted numerous studies on the self-assembly, phase structure, and electrical properties of CPN composite systems using experimental methods. However, most of these studies are based on experimental methods, adjusting the electrical properties of CPNs through trial and error. However, the vast parameter space of the composite system structure of CPNs means that determining the optimal synthesis conditions requires significant economic and time costs; furthermore, limited by the technical means of detection instruments, experimental methods also struggle to observe the local conformation and dynamic evolution of polymer chains at the molecular scale in real time.
[0003] A feature extraction method based on SMILES strings and monomer ratios is proposed. The monomers are formed into homopolymers through a weighted chain. The GCN method is used to predict Morgan fingerprints and Modradar fingerprints by aggregating information from adjacent nodes and edges in the polymer chain, and then predicts the glass transition temperature (Tg) and thermal decomposition temperature (Td).
[0004] However, due to the complex polydisperse, multiscale, and multi-level structures of polymers, this method cannot capture complex multi-level structural information and is difficult to predict the properties of polymers affected by this. Summary of the Invention
[0005] To address the shortcomings of existing technologies in capturing complex multi-level structural information and predicting the properties of polymers affected by it, this invention proposes a method, system, and device for screening conductive network structures of polymeric composites. By employing an incremental graph neural network model, it directly learns the physical property laws from three-dimensional structural data, studies the influence mechanism of conductive nanoparticle networks of different concentrations on the electrical properties of CPNs under the same polymer matrix (homogeneous polymer system), and further clarifies the structure-property relationship between the microstructure and electrical properties of CPNs, thereby solving the problems existing in the prior art.
[0006] A method for screening the conductive network structure of polymer complexes includes the following steps: Molecular dynamics simulation trajectory data of polymer composite systems with different concentrations of carbon nanotubes were obtained, and the three-dimensional coordinates of particles in each frame of the molecular dynamics simulation trajectory data were normalized. Based on the normalized 3D coordinates of the particles, the particle nodes are sequentially encoded and one-hot encoded according to the particle type and the molecule ID to which they belong, generating a node attribute matrix. At the same time, an adjacency matrix is generated according to the bonding relationship between particles to construct the corresponding graph structure dataset. Using the node attribute matrix and adjacency matrix as inputs and the overall conductivity corresponding to each frame of graph structure as output, a prediction model based on graph attention network is constructed. An incremental learning strategy is adopted to input the graph structure dataset into the prediction model in order of increasing concentration for training. Specifically, the model parameters obtained from the previous concentration training are used as the initial parameters for the next higher concentration data training, and the process is iterated until the training of all concentration data is completed. Using a trained prediction model, the conductivity of the conductive network structure of the polymer complex to be screened is predicted, and the attention score matrix of the network structure is extracted. Based on the attention score matrix, the connectivity of the conductive network structure of the polymer complex to be screened is quantitatively analyzed. Based on the predicted conductivity and connectivity analysis results, the optimal conductive network structure of the polymer complex is screened.
[0007] Furthermore, the overall conductivity corresponding to the structure of each frame is calculated using the macroscopic resistance method.
[0008] Furthermore, the calculation process for the carbon nanotube (CNT) concentration is as follows: CNT concentration = .
[0009] Furthermore, the prediction model based on graph attention network includes a residual connection module, a multi-head attention mechanism, and a global pooling strategy that integrates differentiable pooling.
[0010] Furthermore, the quantitative analysis of the connectivity of the polymer complex conductive network structure to be screened based on the attention score matrix specifically includes: reconstructing the network using the top N edges with the highest attention scores as connection thresholds, calculating the topological indices of the reconstructed network, and analyzing its connectivity based on the topological indices; the topological indices include: average degree, connected components, global efficiency, clustering coefficient, network density, and shortest path length.
[0011] The present invention also includes a screening system for conductive network structures of polymer complexes, comprising: The acquisition module is used to acquire molecular dynamics simulation trajectory data of polymer composite systems with different concentrations of carbon nanotubes, and to normalize the three-dimensional coordinates of particles in each frame of the molecular dynamics simulation trajectory data. The dataset construction module is used to sequentially encode particle nodes based on normalized particle 3D coordinates, perform one-hot encoding based on particle type and its molecule ID, generate a node attribute matrix, and generate an adjacency matrix based on the bonding relationship between particles to construct the corresponding graph structure dataset. The model training module is used to construct a prediction model based on a graph attention network, taking the node attribute matrix and adjacency matrix as inputs and the overall conductivity corresponding to each frame of graph structure as output. An incremental learning strategy is adopted to input the graph structure dataset into the prediction model in order of increasing concentration for training. Specifically, the model parameters obtained from the previous concentration training are used as the initial parameters for the next higher concentration data training, and the process is iterated until the training of all concentration data is completed. The screening module is used to predict the conductivity of the polymer complex conductive network structure to be screened using a trained prediction model, and extract the attention score matrix of the network structure; quantitatively analyze the connectivity of the polymer complex conductive network structure to be screened based on the attention score matrix; and screen out the optimal polymer complex conductive network structure based on the predicted conductivity and connectivity analysis results.
[0012] The present invention also includes a computer device for screening conductive network structures of polymeric complexes, comprising: a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method for screening conductive network structures of polymeric complexes.
[0013] The present invention also includes a readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, are used to perform the steps of the polymer composite conductive network structure screening method.
[0014] This invention provides a method for screening the conductive network structure of polymer complexes, which has the following advantages: This invention uses graph structure encoding to accurately convert the three-dimensional spatial positions and connections of atoms / molecules into computer-processable graph data, fully preserving the topological information of the system. Graph attention networks can deeply mine the complex, nonlinear interactions between nodes (particles) in the graph, thereby accurately establishing the mapping relationship between the microscopic network structure and macroscopic conductivity. At the same time, incremental learning is used to transfer knowledge learned at low concentrations to high concentrations, simulating the natural process of the gradual formation and evolution of conductive networks as the concentration increases. This allows the model to understand the systematic influence of concentration, a key parameter, on the network structure, thus enabling more accurate prediction of performance at unknown concentrations. Based on attention scores, the relationship between network connectivity efficiency and conductivity at different CNT concentrations is explained, further demonstrating that the interpretability of this incremental graph attention network can help understand the conductivity mechanism of composite materials at different CNT concentrations. This lays an important theoretical foundation for capturing complex multi-level structural information, better predicting the properties of polymers affected by this structure, and further exploring novel conductive polymer nanocomposites. Attached Figure Description
[0015] Figure 1 This is a flowchart of the polymer complex conductive network structure screening method in an embodiment of the present invention. Detailed Implementation
[0016] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0017] This invention proposes a graph network method based on an attention mechanism for screening the optimal conductive network structure of polymer complexes, such as... Figure 1 As shown, the specific steps include: S1. Data preprocessing, the specific process is as follows: Figure 1 As shown, the specific steps include: A training and validation dataset featuring different CNT concentrations with an aspect ratio of 20.2 was constructed using hybrid particle field molecular dynamics simulations of the dynamic trajectories of CNTs / homogeneities. Ten CNT concentrations (total CNT particles / (total CNT particles + total homopolymer matrix particles)) ranging from 1% to 10% were used. The 3D coordinates of all particles at each concentration were normalized to remove the influence of periodic boundary conditions. Particle node IDs and bonding edge types were sequentially encoded, and node attributes (e.g., particle type, molecule ID) were one-hot encoded. The overall conductivity of each frame was calculated, resulting in the node attribute matrix, adjacency matrix, and overall conductivity calculated using the macroscopic resistance method for each frame. The conductivity of each frame was then segmented and converted into classification labels. For example, the conductivity was divided into five intervals: 0, 0-0.01, 0.01-0.02, 0.02-0.03, and 0.03-1, corresponding to class labels 0, 1, 2, 3, and 4. Then we obtained a graph input dataset that could actually be used for training.
[0018] S2. Model Construction: Using the node attribute matrix and adjacency matrix as inputs and the overall conductivity corresponding to the graph structure of each frame as output, construct a prediction model based on graph attention network.
[0019] We improve upon standard GCN (Graph Convolutional Network) and GAT (Graph Attention Network) models by introducing residual modules and integrating a global pooling strategy based on Differentiable Pooling (DiffPool). We also employ a 5-head attention mechanism, a dropout rate of 0.2, batch normalization, a learning rate of 0.001, ReLU as the activation function, and Adam as the optimizer. Both GCN and GAT networks are set to three layers.
[0020] S3. Training the model: An incremental learning strategy is used to input the graph structure dataset into the prediction model in order of increasing concentration for training. Specifically, the model parameters obtained from training at the previous concentration are used as the initial parameters for training the next higher concentration data. This process is iterated until all concentration data are trained.
[0021] Specifically, after preprocessing the molecular dynamics simulation data, it is converted into the standard input form of a graph network, including node attribute information and edge connection information. The batch dataset is then input into the model constructed in S2, and the internal parameters of the model are adjusted through backpropagation and gradient descent to continuously reduce the loss until the preset target threshold or number of iterations is reached.
[0022] S4. Using the trained prediction model, predict the conductivity of the polymer complex conductive network structure to be screened, and extract the attention score matrix of the network structure; quantitatively analyze the connectivity of the polymer complex conductive network structure to be screened based on the attention score matrix; and select the optimal polymer complex conductive network structure based on the predicted conductivity and connectivity analysis results.
[0023] This invention employs an incremental graph neural network method to explore a broader combinatorial parameter space, visually studying the relationship between the structure and conductivity of CPN conductive nanoparticles. Based on the interpretation of the relationship between network connectivity efficiency and conductivity at different CNT concentrations using attention scores, and the nonlinear transformation of the network feature space distribution by the GAT model, this further demonstrates that the interpretability of this incremental graph attention network can contribute to understanding the conductivity mechanism of composite materials at different CNT concentrations. This provides guidance for further material synthesis and conductivity regulation mechanisms, laying an important theoretical foundation for the further exploration of novel conductive polymer nanocomposites.
[0024] Experiments have shown that: (1) Model GCN and GAT network modules: Due to the limitations of datasets at different concentrations, this invention aims to learn the common relationship between structure and conductivity at different concentrations by using an incremental learning method to train the model. Training starts with data at a 1% concentration, and then the parameters of this model are used as the starting point for training on data at a 2% concentration, and this process continues. This allows for the sharing of data at different concentrations and provides a finely tuned prediction model for each concentration. To adapt to this training method, the standard GCN (Graph Convolutional Network) and GAT (Graph Attention Network) models are improved by introducing a residual module and integrating a global pooling strategy of Differentiable Pooling (DiffPool). A 5-head attention mechanism is also adopted, with a dropout rate of 0.2, batch normalization, a learning rate of 0.001, ReLU as the activation function, and Adam as the optimizer. The number of network layers in both GCN and GAT is set to three. Both GCN and GAT provide regression and classification prediction versions. In regression prediction, mean squared error (MSE) is used as the loss function; while in classification prediction, cross entropy is used as the loss function.
[0025] Network topology indices are of great significance in modern computational chemistry and complex network modeling, especially in conveying the connectivity and efficiency of network structures, capturing complex connectivity and topological permutations within networks. This invention employs six important network topology indices to represent key features of a network structure frame, including Average Degree (AD), Clustering Coefficient (CC), Network Density, Shortest Path (SP), Global Efficiency (GE), and Connected Component (CC).
[0026] (2) Conductivity regression and classification prediction module; GCN and GAT were used to conduct experiments on the conductivity regression and classification prediction module respectively, and appropriate models were selected, as shown in Table 1: Table 1. Experimental results using GCN and GAT For attribute classification prediction, the above indicators were obtained by averaging the results of three experiments. All indicators are better when they are closer to 1. As can be seen in Table 1, the GCN model performs well on all indicators, while the GAT model performs poorly. Therefore, it can be concluded that the GCN model should be preferred for attribute classification tasks.
[0027] Table 2 Prediction errors of GCN and GAT As shown in Table 2, all prediction errors of the GAT model are between 0 and 0.00001, indicating that the GAT model achieves a high level of accuracy in predicting conductivity. Since the GAT model outperforms the GCN model in predicting all concentrations, it can be considered that the GAT model should be preferred for attribute regression prediction tasks in this invention.
[0028] (3) Structural feature analysis of training data: To further analyze the relationship between the conductive network structure and concentration of polymer composites, this invention constructs a complex network using the traditional resistance network method for the training set, utilizing six core network analysis indicators. These indicators can reflect the network connectivity to a certain extent. Specifically, firstly, the CNT particle network in each frame is reconstructed into a new network according to the traditional resistance method. There are 1000 frames at each concentration, that is, 1000 networks. Then, one network is taken every 100 frames in each concentration, so 10 networks can be taken at each concentration. Then, the six traditional indicators of these 10 networks are calculated and averaged to obtain the six traditional indicators for each concentration shown in Table 3. The self-organizing emergence ability and dynamic structural adaptability of conductive polymer nanocomposites at different concentrations are analyzed through the changes in these indicators.
[0029] Table 3 Comparison of key indicators of complex networks under different CNT concentrations To further explain the impact of different CNT concentrations on the structure formed by the conductive network, and thus the resulting changes in the overall conductivity of the polymer composite, this invention uses the GAT model to perform regression predictions on conductivity from CP concentrations ranging from 1% to 8%. Through incremental training, training models from 1% to 8% and their corresponding attention score matrices were obtained, while data from 9% and 10% concentrations were used for the test set. This invention extracts the attention score matrix of CNTs (i.e., the pairs of C particles) from the attention score matrix and calculates the variance and coefficient of variation of this matrix. Furthermore, to visualize the variation patterns in the attention score matrix, the attention score of the same carbon nanotube at each concentration was extracted, revealing the changing patterns of network structure and the importance of node pair connections under different concentrations.
[0030] To compare the spatial distribution of the training and validation sets, all data were first divided into three parts: 80% of the data with concentrations of 1%-8% was used as the training set, the remaining 20% as the validation set, and all data with concentrations of 9%-10% were used as the test set, which was not used for model training but directly for prediction. All prediction results were then compared with the results calculated by the corresponding macroscopic resistance method to calculate the corresponding evaluation index, thereby evaluating the model's performance and generalization ability on different concentration datasets.
[0031] The distribution of the CNT network structure feature space before and after training with the GAT model was compared. The initial distribution of 1000 training frames of CP1-CP8 was compared with the spatial distribution after nonlinear transformation and mapping by the GAT model. First, the GAT model was trained with CP1 data. Then, the maximum length of the edges with the top 200 attention scores of the second layer CNT network in the model (other values such as 100 and 300 were calculated and found to be the most suitable) was used as the threshold to redefine whether any two nodes are connected. After reconstructing the network, six traditional indicators of the network structure (average degree, connected components, global efficiency, clustering coefficient, network density, and shortest path length of the largest connected component) were calculated. Then, TSNE was applied to reduce it to 2 dimensions and perform maximum and minimum value normalization.
[0032] This invention first uses hybrid particle field molecular dynamics to simulate the dynamic trajectory of CNT / homopolymers with an aspect ratio of 20.2 at concentrations ranging from 1% to 10%. The conductivity of each frame is calculated using the macroscopic resistance method. Next, 1000 frames after equilibrium simulation at each concentration (1%-8%) are used as training data, with 80% serving as the training set and 20% as the validation set. All simulation data at concentrations of 9%-10% are used as the test set. Then, through incremental training of GCN and GAT models, it is found that GAT performs best in conductivity regression prediction. Subsequently, the attention score matrix is output based on the model results. By analyzing the coefficient of variation of the attention score matrix of the CNT network, as well as the conductivity at different concentrations and six structural characteristic indicators of the network constructed using the traditional resistance method, the process of the conductive network unfolding gradually at low concentrations, then undergoing internal structural reorganization and optimization at medium concentrations, and finally reaching a stage of excessively dense or overly connected network at high concentrations is derived. Furthermore, it was found that the conductive network structure with a CP7 concentration has the best connectivity efficiency. The research value of conductive network structures with CP8 and higher concentrations gradually decreases due to cost-effectiveness issues. Finally, the network was reconstructed using the top 200 of the second-layer attention scores of the GAT model as a threshold. By comparing it with the feature space of the input data, it was proved that the attention scores of the GAT model are reliable for interpretability analysis of conductive networks. This interpretable graph neural network can be used for property prediction and network structure analysis of polymer composite materials.
[0033] Based on the above inventive concept, the present invention also proposes a polymer composite conductive network structure screening system, comprising: The acquisition module is used to acquire molecular dynamics simulation trajectory data of polymer composite systems with different concentrations of carbon nanotubes, and to normalize the three-dimensional coordinates of particles in each frame of the molecular dynamics simulation trajectory data.
[0034] The dataset construction module is used to sequentially encode particle nodes based on normalized coordinate data, perform one-hot encoding according to particle type and its molecule ID, generate a node attribute matrix, and generate an adjacency matrix based on the bonding relationship between particles to construct the corresponding graph structure dataset.
[0035] The model training module is used to construct a prediction model based on a graph attention network, taking the node attribute matrix and adjacency matrix as inputs and the overall conductivity corresponding to each frame of graph structure as output. An incremental learning strategy is adopted to input the graph structure dataset into the prediction model in order of increasing concentration for training. Specifically, the model parameters obtained from the previous concentration training are used as the initial parameters for the next higher concentration data training, and the process is iterated until the training of all concentration data is completed.
[0036] The screening module is used to predict the conductivity of the polymer complex conductive network structure to be screened using a trained prediction model and extract its attention score matrix; quantitatively analyze the connectivity of the polymer complex conductive network structure based on the attention score matrix; and screen out the optimal polymer complex conductive network structure based on the predicted conductivity and connectivity analysis results.
[0037] The present invention also proposes a computer device for screening conductive network structures of polymer complexes, comprising: a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method for screening conductive network structures of polymer complexes.
[0038] The present invention also proposes a readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, are used to perform the steps of the polymer complex conductive network structure screening method.
[0039] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for screening the conductive network structure of polymer composites, characterized in that, Includes the following steps: Molecular dynamics simulation trajectory data of polymer composite systems with different concentrations of carbon nanotubes were obtained, and the three-dimensional coordinates of particles in each frame of the molecular dynamics simulation trajectory data were normalized. Based on the normalized 3D coordinates of the particles, the particle nodes are sequentially encoded and one-hot encoded according to the particle type and the molecule ID to which they belong, generating a node attribute matrix. At the same time, an adjacency matrix is generated according to the bonding relationship between particles to construct the corresponding graph structure dataset. Using the node attribute matrix and adjacency matrix as inputs and the overall conductivity corresponding to the graph structure of each frame as output, a prediction model based on graph attention network is constructed. An incremental learning strategy is adopted to input the graph structure dataset into the prediction model in order of increasing concentration for training. Specifically, the model parameters obtained from training at the previous concentration are used as the initial parameters for training the next higher concentration data. This process is iterated until training of all concentration data is completed. Using a trained prediction model, the conductivity of the conductive network structure of the polymer complex to be screened is predicted, and the attention score matrix of the network structure is extracted. Based on the attention score matrix, the connectivity of the conductive network structure of the polymer complex to be screened is quantitatively analyzed. Based on the predicted conductivity and connectivity analysis results, the optimal conductive network structure of the polymer complex is screened.
2. The method for screening conductive network structures of polymer complexes according to claim 1, characterized in that, The overall conductivity corresponding to the structure of each frame was calculated using the macroscopic resistance method.
3. The method for screening optimal conductive network structures of polymer composites according to claim 1, characterized in that, The calculation process for the carbon nanotube (CNT) concentration is as follows: CNT concentration = .
4. The method for screening conductive network structures of polymer composites according to claim 1, characterized in that, The prediction model based on graph attention network includes a residual connection module, a multi-head attention mechanism, and a global pooling strategy that integrates differentiable pooling.
5. The method for screening conductive network structures of polymer composites according to claim 1, characterized in that, The quantitative analysis of the connectivity of the polymer complex conductive network structure to be screened based on the attention score matrix specifically includes: reconstructing the network using the top N edges with the highest attention scores as connection thresholds, calculating the topological indices of the reconstructed network, and analyzing its connectivity based on the topological indices; the topological indices include: average degree, connected components, global efficiency, clustering coefficient, network density, and shortest path length.
6. A screening system for conductive network structures of polymeric complexes, characterized in that, include: The acquisition module is used to acquire molecular dynamics simulation trajectory data of polymer composite systems with different concentrations of carbon nanotubes, and to normalize the three-dimensional coordinates of particles in each frame of the molecular dynamics simulation trajectory data. The dataset construction module is used to sequentially encode particle nodes based on normalized particle 3D coordinates, perform one-hot encoding based on particle type and its molecule ID, generate a node attribute matrix, and generate an adjacency matrix based on the bonding relationship between particles to construct the corresponding graph structure dataset. The model training module is used to construct a prediction model based on graph attention network by taking the node attribute matrix and adjacency matrix as input and the overall conductivity corresponding to the graph structure of each frame as output. An incremental learning strategy is adopted to input the graph structure dataset into the prediction model in order of increasing concentration for training. Specifically, the model parameters obtained from training at the previous concentration are used as the initial parameters for training the next higher concentration data. This process is iterated until training of all concentration data is completed. The screening module is used to predict the conductivity of the polymer complex conductive network structure to be screened using a trained prediction model, and extract the attention score matrix of the network structure; quantitatively analyze the connectivity of the polymer complex conductive network structure to be screened based on the attention score matrix; and screen out the optimal polymer complex conductive network structure based on the predicted conductivity and connectivity analysis results.
7. A computer device for screening the conductive network structure of polymer complexes, characterized in that, include: A memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the polymer composite conductive network structure screening method according to any one of claims 1-5.
8. A readable storage medium, characterized in that, The readable storage medium stores a computer program, which includes program instructions that, when executed by a processor, perform the steps of the polymer composite conductive network structure screening method according to any one of claims 1-5.