A method for predicting the pyrolysis performance of polymer materials based on process parameters and structural information

Abstract

This application relates to the field of computational materials science and technology, and in particular to a method for predicting the pyrolysis performance of polymer materials based on process parameters and structural information. The method includes the following steps: First, acquiring a dataset; the dataset includes material structural information and material processing information; then, processing the dataset and dividing the processed dataset into a training set, a validation set, and a test set; next, constructing a deep learning prediction model coupling polymer material processing information and structural information; finally, training the deep learning prediction model using the training set, evaluating the model training effect using the validation set, and inputting the test set into the trained deep learning prediction model to predict the pyrolysis performance of the polymer material. This application proposes a deep learning prediction model that uses polymer material processing information to assist in optimizing the material structure, accurately describing the impact of the processing process on material properties.

Description

Technical Field

[0001] This application relates to the field of computational materials science and technology, and in particular to a method for predicting the pyrolysis performance of polymer materials based on process parameters and structural information. Background Technology

[0002] Pyrolysis-resistant materials are those that maintain stability in high-temperature, high-speed airflow, and corrosive environments. Under extreme conditions, their stability and durability make them indispensable materials in many industrial sectors, including aerospace, chemical, and energy fields. Resin-based pyrolysis-resistant materials, as a current major research focus, have attracted widespread attention from researchers.

[0003] With the rapid development of polymer materials, the research and development of novel polymer materials is urgently needed. However, due to the large candidate space and complex reaction mechanisms in polymer materials, developing novel polymer materials solely through experiments is inefficient, time-consuming, and costly. Although researchers have demonstrated that the structure-property relationship of polymer materials can be successfully established using machine learning algorithms, there is currently no model capable of predicting the degradation performance of polymer materials. Summary of the Invention

[0004] This application provides a method for predicting the pyrolysis performance of polymer materials based on process parameters and structural information. By constructing a deep learning prediction model that assists in optimizing the material structure based on the processing information of polymer materials, the method accurately describes the impact of the processing process on the material properties.

[0005] This application provides a method for predicting the pyrolysis performance of polymer materials based on process parameters and structural information, comprising the following steps: First, acquiring a dataset; the dataset includes material structural information and material processing information; then, processing the dataset and dividing the processed dataset into a training set, a validation set, and a test set; next, constructing a deep learning prediction model that couples polymer material processing information with structural information; finally, training the deep learning prediction model using the training set, evaluating the model training effect using the validation set, and inputting the test set into the trained deep learning prediction model to predict the pyrolysis performance of polymer materials.

[0006] In some exemplary embodiments, acquiring a dataset includes: acquiring material structure information and multiple different types of material processing information; wherein, the material structure information includes the structure and pyrolysis performance data of polymer materials; and the types of material processing information include curing atmosphere, curing temperature, curing time, curing agent, curing agent ratio, performance test atmosphere, and heating rate.

[0007] In some exemplary embodiments, processing the dataset includes: processing the material structure information and material processing information in the dataset respectively; converting the material structure information into a material structure data file that can provide input parameters for machine learning, and converting the material processing information into a processing information file that can provide processing information support for model training.

[0008] In some exemplary embodiments, during the processing of material processing information, the material processing information is transformed into a processing information file that can provide data support for model training, and the processing information file is transformed into a one-dimensional vector processing information file.

[0009] In some exemplary embodiments, the deep learning prediction model uses two fully connected layers to obtain a special processing matrix of material structure and processing information. Through matrix fusion, a total material information matrix is ​​obtained. Then, the matrix is ​​mapped to the material fracture performance through three convolutional layers and one fully connected layer.

[0010] In some exemplary embodiments, constructing a deep learning prediction model that couples polymer material processing information with structural information includes: constructing a graph convolutional neural network and a fully connected neural network respectively; using the fully connected neural network to process the material processing information in the dataset to obtain a structural processing matrix; fusing the material's structural matrix and structural processing matrix to represent the material's structure and processing information; and using the graph convolutional neural network to realize the mapping from the matrix to the material's pyrolysis performance.

[0011] In some exemplary embodiments, the graph convolutional neural network includes convolutional layers and fully connected layers, wherein the convolutional layers are used to expand the edge feature matrix using a Gaussian kernel function to describe the material structure information and obtain the material structure matrix.

[0012] In some exemplary embodiments, the fully connected neural network consists of multiple fully connected layers with linear transformations and nonlinear activation functions, used to extract high-order semantic representations of processing conditions, and to combine the processing information of polymer materials with an initialized structure matrix, using two fully connected layers to transform it into an information matrix of the same size as the graph convolutional neural network.

[0013] In some exemplary embodiments, a fully connected neural network is used to process the material processing information in the dataset to obtain a structural processing matrix. This includes: after the material processing information is input, using one-heat encoding to partition the curing temperature and curing time of the material to obtain multiple temperature ranges, and then combining the curing time with each temperature range; matching the one-dimensional vector of material processing information with different nodes in the structural matrix to obtain two matrices with the same dimension in the node direction; and obtaining a structural processing matrix with fused material structure processing information through matrix fusion.

[0014] In some exemplary embodiments, fusing the material's structure matrix and structural processing matrix to represent the material's structure and processing information includes: weighting the material's structure matrix and structural processing matrix with equal weights to obtain the final material representation, as shown in the following formula:

[0015] h total =0.5*h structh +0.5*h proch

[0016] Among them, h total For the final material representation; h structh The structure matrix representing the material; h proch Represents the structural processing matrix.

[0017] The technical solution provided in this application has at least the following advantages:

[0018] This application provides a method for predicting the pyrolysis performance of polymer materials based on process parameters and structural information, comprising the following steps: First, acquiring a dataset; the dataset includes material structural information and material processing information; then, processing the dataset and dividing the processed dataset into a training set, a validation set, and a test set; next, constructing a deep learning prediction model coupling polymer material processing information and structural information; finally, training the deep learning prediction model using the training set, evaluating the model training effect using the validation set, and inputting the test set into the trained deep learning prediction model to predict the pyrolysis performance of the polymer material. On one hand, this application solves the problem of performance differences in the same polymer material due to different curing temperatures. By combining the structural information of the polymer material with processing information, the influence of processing information such as curing temperature and curing time on material performance is clarified while describing the polymer material structure. On the other hand, the method provided by this application can accelerate experimental design, predict the influence of different curing temperatures and times on material performance through the model, guide experimental design, optimize processing technology, improve synthesis efficiency, and reduce experimental costs. Attached Figure Description

[0019] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations do not constitute a limitation on the embodiments, and unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0020] Figure 1 This is a flowchart illustrating a method for predicting the pyrolysis performance of polymer materials based on process parameters and structural information, as provided in an embodiment of this application.

[0021] Figure 2 This is a schematic diagram illustrating the input of material structure information provided in an embodiment of this application.

[0022] Figure 3 This is a schematic diagram illustrating the transformation of the processing information matrix provided in an embodiment of this application.

[0023] Figure 4 This is a schematic diagram of the screening results of the feature Pearson correlation coefficient provided in the embodiments of this application.

[0024] Figure 5 This is a schematic diagram of the model framework provided for an embodiment of this application.

[0025] Figure 6 This is a schematic diagram of Gaussian kernel function dimensionality increase provided in an embodiment of this application.

[0026] Figure 7 The image shows the prediction results of the model training set provided in the embodiments of this application.

[0027] Figure 8 The image shows the prediction results of the model test set provided in the embodiments of this application. Detailed Implementation

[0028] As can be seen from the background technology, there is currently no model that can predict the pyrolysis performance of polymer materials.

[0029] To address the aforementioned technical problems, this application provides a method for predicting the pyrolysis performance of polymer materials based on process parameters and structural information, comprising the following steps: First, acquiring a dataset; the dataset includes material structural information and material processing information; then, processing the dataset and dividing the processed dataset into a training set, a validation set, and a test set; next, constructing a deep learning prediction model coupling polymer material processing information and structural information; finally, training the deep learning prediction model using the training set, evaluating the model training effect using the validation set, and inputting the test set into the trained deep learning prediction model to predict the pyrolysis performance of the polymer material. This application proposes a deep learning prediction model that uses polymer material processing information to assist in optimizing the material structure, accurately describing the impact of the processing process on material properties.

[0030] The embodiments of this application will now be described in detail with reference to the accompanying drawings, so that those skilled in the art can better understand and implement this application.

[0031] The following are specific examples of the implementation process of the technical solution to be protected in this application. However, this application may also adopt other implementation methods different from those described herein. Those skilled in the art can adopt different technical solutions under the guidance of the ideas in this application to realize the scope of this application. Therefore, this application is not strictly limited to the specific implementation examples below.

[0032] See Figure 1 This application provides a method for predicting the pyrolysis performance of polymer materials based on process parameters and structural information, comprising the following steps:

[0033] Step S1: Obtain the dataset; the dataset includes material structure information and material processing information.

[0034] Step S2: Process the dataset and divide the processed dataset into training set, validation set and test set.

[0035] Step S3: Construct a deep learning prediction model that couples polymer material processing information with structural information.

[0036] Step S4: Train the deep learning prediction model using the training set, evaluate the model training effect using the validation set, and input the test set into the trained deep learning prediction model to predict the pyrolysis performance of polymer materials.

[0037] This application aims to construct a performance prediction model to predict the pyrolysis performance of polymer materials. The main index of pyrolysis performance is Tp. 5％ (5% thermal decomposition temperature), T 10％ (10% thermal decomposition temperature), M 残重 (Final residual weight percentage after pyrolysis). Simultaneously, the material undergoes a curing process during experimental testing. This processing method has a significant impact on the material's pyrolysis performance, especially the 5% thermal decomposition temperature. Currently, researchers have only studied the material's structure without combining processing information with the material's structure, leading to inaccurate material descriptions. For example, the Tg of the same polymer material cured at 200℃ differs from that cured at 400℃. 5％ There are certain performance gaps. In order to enable the prediction model to accurately capture the impact of the processing on the material, this application proposes a deep learning prediction model for polymer material processing information to assist in optimizing the material structure. It constructs a multimodal material information description, that is, it uses material structure combined with processing information to accurately describe the impact of the processing on the material performance.

[0038] The method comprises the following steps: First, literature data on polymer materials is collected, including their structural, processing, and performance information. Then, semantic transformation is used to convert this data into machine learning input parameters. Simultaneously, feature engineering is performed to enrich the group information within the material structure, and the processing information is converted into a one-dimensional vector for subsequent operations. Next, a deep learning prediction model coupling the processing and structural information of polymer materials is constructed to predict their pyrolysis performance. This model mainly includes processing the material structure information matrix, processing the material structure and processing information matrix, and the main body of the prediction model. Finally, the collected dataset and the constructed deep learning prediction model are used for training to establish a prediction model that jointly uses material structure and processing information to predict the pyrolysis performance of polymer materials.

[0039] Specifically, the method for predicting the pyrolysis performance of polymer materials based on process parameters and structural information provided in this application is generally divided into three parts: material structure information processing, material processing information processing, model construction, and model training.

[0040] In some embodiments, obtaining the dataset in step S1 includes: obtaining material structure information and multiple different types of material processing information respectively; wherein, the material structure information includes the structure and pyrolysis performance data of polymer materials; the types of material processing information include curing atmosphere, curing temperature, curing time, curing agent, curing agent ratio, performance test atmosphere, and heating rate.

[0041] In some embodiments, step S2 involves processing the dataset, including: processing the material structure information and material processing information in the dataset respectively; converting the material structure information into a material structure data file that can provide input parameters for machine learning, and converting the material processing information into a processing information file that can provide data support for model training.

[0042] In step S1, the structure and pyrolysis properties of the polymer material are collected and stored in a CSV table. For processing information, this application collects seven types of processing information: curing atmosphere, curing temperature, curing time, curing agent, curing agent ratio, performance testing atmosphere, and heating rate. In step S2, a coding method based on the topological structure of polymer materials (CN118675665A) is used to process the material structure information in the dataset, encoding each group and its bonding relationship in the polymer material, thereby representing the polymer structure information in the form of a topological graph, that is, converting the material into a data file that can provide input parameters for machine learning. For the pyrolysis performance file, it mainly includes the material's T... 5％ T 10％ M 残重 .

[0043] In some embodiments, during the material processing information processing in step S2, the material processing information is transformed into a processing information file that can provide data support for model training, and the processing information file is transformed into a one-dimensional vector processing information file. That is, the seven types of processing information are transformed into a one-dimensional vector processing information file, providing data support for subsequent model training.

[0044] In some embodiments, the deep learning prediction model in step S3 uses two fully connected layers to obtain a special processing matrix of material structure and processing information. Through matrix fusion, a material overall information matrix is ​​obtained, and then the matrix is ​​mapped to the material fracture performance through three convolutional layers and one fully connected layer.

[0045] To integrate the material's structure and processing information and map it onto the material's pyrolysis performance, this application establishes two neural network models in step S3. A fully connected neural network (multilayer perceptron) is used to process the material's processing information, and a graph convolutional neural network is used to process the complex overall information structure of the polymer material.

[0046] In some embodiments, step S3, which involves constructing a deep learning prediction model that couples polymer material processing information with structural information, includes:

[0047] Step S301: Construct a fully connected neural network and a graph convolutional neural network respectively.

[0048] Step S302: Use a fully connected neural network to process the material processing information in the dataset to obtain the structural processing matrix.

[0049] Step S303: Matrix fusion is performed between the initial structure matrix and the structure processing matrix to obtain the overall information matrix of the material; the mapping from the overall material information to the pyrolysis performance is realized through a graph convolutional neural network.

[0050] In some embodiments, the fully connected neural network consists of multiple fully connected layers with linear transformations and nonlinear activation functions, used to extract higher-order semantic representations of processing conditions, and to combine the processing information of polymer materials with an initialized structure matrix, using two fully connected layers to transform it into an information matrix of the same size as the graph convolutional neural network.

[0051] In some embodiments, the graph convolutional neural network includes convolutional layers and fully connected layers, wherein the convolutional layers are used to expand the edge feature matrix using a Gaussian kernel function to describe the material structure information and obtain the material structure matrix.

[0052] In some embodiments, step S302 uses a fully connected neural network to process the material processing information in the dataset to obtain a structural processing matrix, including: after the material processing information is input, using one-heat encoding to partition the curing temperature and curing time of the material to obtain multiple temperature ranges, and then combining the curing time with each temperature range; matching the one-dimensional vector of material processing information with different nodes in the structural matrix to obtain two matrices with the same dimension in the node direction; and obtaining a structural processing matrix of material structure fusion processing information through matrix fusion.

[0053] In some embodiments, step S303 involves matrix fusion of the initial structure matrix and the structural processing matrix to obtain the overall material information matrix. This includes weighting the material's structure matrix and structural processing matrix with equal weights to obtain the final material representation, as shown in the following formula:

[0054] h total =0.5*h structh +0.5*h proch

[0055] Among them, h total For the final material representation; h structh The structure matrix representing the material; h proch Represents the structural processing matrix.

[0056] After completing the three steps of material structure information processing, material processing information processing, and model building, this application inputs the prepared dataset into the deep learning prediction model and divides the dataset into three parts: training set, validation set, and test set.

[0057] The deep learning prediction model is trained using a training set. By training on the training set, structural features that affect performance optimization are discovered. The network parameter weights are retained for further iteration after validation on the validation set. Finally, the network model selected from the validation set is used for final validation on the test set to determine the model's prediction performance on polymer materials.

[0058] The following is a detailed description of the method for predicting the pyrolysis performance of polymer materials based on process parameters and structural information provided in this application, through a specific embodiment.

[0059] A method for predicting the pyrolysis performance of polymer materials based on process parameters and structural information includes the following steps:

[0060] Step 1: Obtaining the Dataset. First, collect the material's structural information, processing information, and pyrolysis performance. Specifically, first, consult relevant literature on polymer materials to find the material's structural diagrams and TGA (Thermogravimetric Analysis) curves. Use data extraction tools to extract the material's pyrolysis performance; in this application, WebPlotDigitizer is used. Simultaneously, find the material's curing information in the material preparation and characterization sections, including curing atmosphere, curing temperature, curing time, type of curing agent, the mass percentage of the curing agent in the material, and the testing atmosphere and heating rate, as the material's processing information.

[0061] Step Two: Processing the Dataset. Based on the existing patent "A Coding Method and System Based on the Topological Structure of Polymer Materials," this application enriches the information representation of materials. This mainly includes enriching the representation content of each node in the material structure and processing the material's curing information into a one-dimensional vector matrix. During data collection, the curing process of most materials is carried out under three atmospheres: air, vacuum, and nitrogen. To convert this into a data format, this application uses three numerical expressions—0.79, 0.00, and 1.00—for mathematical representation to distinguish between them. Figure 3 As shown in the figure. Here, 0.79 represents the nitrogen content in the air atmosphere, 250 and 400 represent the curing temperatures of 250℃ and 400℃ respectively, 2 and 1 represent the curing time at the corresponding temperatures, the subsequent 0 indicates the presence and proportion of curing agent in the material, 1.00 indicates that the material performance test was conducted under a nitrogen atmosphere, and 10 indicates a heating rate of 10℃ / min.

[0062] The structural matrix of the material is then transformed using the method described above. Figure 2The results were presented. The node feature matrix was enriched through feature engineering. Specifically, the physical information of the nodes was extracted using the RDKit toolkit. RDKit is an open-source toolkit for chemical information processing, written in C++ and providing a Python interface, enabling easy processing of information such as molecular structures, chemical reactions, and chemical properties in a Python environment. RDKit offers rich chemical information processing capabilities, including molecular representation, similarity comparison, drug design, and chemical data analysis. Based on 2D and 3D molecular manipulation of compounds, machine learning methods were used for compound descriptor generation, fingerprint generation, compound structural similarity calculation, and 2D and 3D molecular visualization.In this method, the following physicochemical information of the nodes is mainly calculated using RDKit: MolWt (molecular weight) is used to calculate the molecular mass of the molecule; TPSA (topological polar surface area) is used to calculate the area of ​​the polar region on the surface of the molecule; NumHDonors (number of hydrogen bond donors) is used to calculate the number of hydrogen bond donors (such as -OH and -NH groups) in the molecule; NumHAcceptors (number of hydrogen bond acceptors) is used to calculate the number of hydrogen bond acceptors (such as oxygen and nitrogen atoms) in the molecule; NumAromaticRings (number of aromatic rings) is used to calculate the number of aromatic rings in the molecule; C hi0 (zero-order nitrogen-hydrogen connectivity) calculates the zero-order connectivity of a molecule; Kappa1 (Kappa1 exponent) represents the flexibility or symmetry of the molecular skeleton; BertzCT (Bertz complexity) calculates the Bertz coefficient of molecular complexity; FrictionCSP3 (Csp3 carbon ratio) is the proportion of sp3 mixed carbons in a molecule; NumAliphaticRings (number of aliphatic rings) calculates the number of aliphatic rings in a molecule; NumSaturatedRings (number of saturated rings) calculates the number of fully saturated rings in a molecule. NumRotatableBonds (number of rotatable bonds) calculates the number of non-cyclic single bonds in a molecule, representing the number of rotatable bonds; NumHeteroatoms (number of heteroatoms) calculates the number of non-carbon and non-hydrogen atoms (such as oxygen, nitrogen, sulfur, etc.) in a molecule; RingCount (number of rings) calculates the total number of rings in a molecule (including aromatic and non-aromatic); HeavyAtomCount (number of heavy atoms) calculates the number of atoms in a molecule excluding hydrogen; Applications: used to estimate the complexity and relative size of molecules; NHOHCount (number of NH and OH groups) calculates the number of NH and OH groups in a molecule; NOCount (number of N and O) calculates the number of nitrogen and oxygen atoms in a molecule; ExactMolWt (exact molecular weight) calculates the exact molecular mass of a molecule (considering isotopic distribution); MolMR (molar refractive index) calculates the molar refractive index of a molecule (related to volume and polarizability); HallKierAlpha (Hall-Kier α parameter) indicates the degree of branching of the molecular skeleton; BalabanJ (BalabanJ index) is a topological index of a molecule used to describe the compactness of the molecular skeleton. Considering that there may be some highly correlated features among the features, this application uses the Pearson correlation coefficient to calculate the correlation between the features and filters more than 90% of the parameters. The final result is as follows. Figure 4 As shown.

[0063] Step 3: Construct a deep learning prediction model that couples polymer material processing information with structural information.

[0064] After step two, the data input parameters for machine learning in this application have been processed. To incorporate the processed information into the neural network model, this application employs a multimodal neural network design. The framework diagram is as follows: Figure 5 As shown.

[0065] First, this application divides the data into the material structure matrix and the material structure information matrix. In order to clearly represent the influence of processing information on the material structure matrix, this application establishes a fully connected neural network for information integration.

[0066] This fully connected neural network consists of multiple layers of linear transformations and nonlinear activation functions, used to extract high-order semantic representations of processing conditions. First, for the input processing vector, considering that most polymer materials are processed using a stepped curing process, to clearly describe the effects of different temperatures and times on the material, this application uses one-hot encoding to process the curing temperature and time of the material after the vector input. Specifically, the curing temperature is divided into several intervals: 100-150°C (inclusive), 150-200°C (inclusive), 200-250°C (inclusive), 250-300°C (inclusive), 300-350°C (inclusive), and 350-400°C (inclusive), and the time interval is combined with each interval. Then, a linear layer maps the result to a higher-dimensional feature space (64 dimensions). To clearly describe the influence of processing information on each node, this application matches a one-dimensional vector with the structure matrix, resulting in two matrices with the same dimension in the node direction. Through matrix fusion, a matrix combining the material structure and processing information is obtained. This matrix is ​​then enhanced with a nonlinear transformation, and finally, an embedded representation with the same dimension as the structure matrix is ​​output through another linear layer. This embedding can be intuitively viewed as a global vector representation of the influence of processing conditions on the overall material properties, ultimately forming the structural processing matrix mentioned in this method.

[0067] The embedded representation of this processing vector will then be fused with the features of the initial input (e.g., by splicing or weighted averaging) to jointly participate in the final prediction of material properties. By introducing this module, the model can simultaneously consider the effects of both structural and processing levels, thereby achieving material property modeling that more closely reflects real-world manufacturing processes.

[0068] Subsequently, this application merges the structural processing matrix with the structural matrix.

[0069] Structural information matrix h structh With structural processing information matrix h proch Fusion is performed in a high-dimensional feature space to generate the final material representation h. totalThis method assumes that the two matrices are equally important; therefore, this application performs a weighted average of the two matrices with equal weights.

[0070] Subsequently, regarding the structure of the graph convolutional neural network model, the main function includes data and model loading, model training, optimization, and evaluation. During model training, this application uses MAEloss, suitable for regression tasks, employs the Adam momentum optimizer to accelerate convergence, and adjusts the learning rate using the `adjust_learning_rate` function to ensure model stability. To reduce the impact of outliers on model training, this application uses Class AverageMeter to track the model's training data and defines a method for calculating the average metric, facilitating the calculation and storage of relevant data during model training and enabling subsequent data analysis.

[0071] For convolutional layers, since the matrix processing in this application is not a traditional normalized adjacency matrix, the application cannot use traditional convolution kernels to operate on the matrix. Therefore, this application uses a Gaussian kernel function, which has a similar processing method to convolution kernels, to process the feature matrix.

[0072] The Gaussian kernel function represents features by mapping distance information to a high-dimensional space. It captures the distances between atoms and represents these distances as Gaussian functions. This is particularly important for atomic relationships in molecular structures, as the distances and interactions between atoms are crucial determinants of molecular properties.

[0073] The formula for the Gaussian kernel function is:

[0074]

[0075] The edge feature matrix is ​​processed using a Gaussian kernel function. The two-dimensional tensor of the edge feature matrix is ​​increased in dimension by the Gaussian kernel function, transforming it into a three-dimensional tensor. Here, this application defines the minimum y-value of the Gaussian kernel function as 0, the maximum value as 20, and the step size (parameter of the kernel function) as 2. This achieves the transformation of the edge feature matrix from a two-dimensional tensor to a three-dimensional tensor, with a length of 11 in the z-axis direction. The results are shown in [link to results]. Figure 6 .

[0076] For the node feature matrix, this application uses indexing to concatenate the features of the node itself and the features of its neighbors. Thus, this application constructs both the edge feature matrix and the node feature matrix into a three-dimensional tensor.

[0077] Subsequently, this application performs a summation operation on the matrix by adding the node feature matrix, edge feature matrix, and node feature matrix. At the same time, using a fully connected layer, this synthesized matrix is ​​transformed into a node feature matrix plus node feature matrix operation, thus fusing the edge information with the node information.

[0078] The resulting matrix is ​​divided into two parts: the `nbr_filter` matrix and the `nbr_core` matrix. The `nbr_filter` matrix is ​​activated by the sigmoid activation function, while the `nbr_core` matrix is ​​activated by the softplus activation function. The sigmoid activation function has a large gradient when the input is close to 0, and a small gradient when the input is very large or very small. This is used to normalize or filter neighbor features, ensuring the output value is between 0 and 1, which can be considered a probability or weight. The softplus activation function has a gradient close to the input value when the input is large, and a gradient close to 0 when the input is small. The `nbr_core` matrix is ​​activated by the softplus function to perform a smooth non-linear transformation on neighbor features to maintain feature continuity and non-negativity, avoiding gradient vanishing or exploding. By combining the sigmoid and softplus activation functions, different feature transformation and processing logics can be implemented. The sigmoid activation function can filter and select features, making certain feature values ​​close to 0 or 1, thereby highlighting important features or suppressing unimportant features.

[0079] The softplus activation function performs a smooth, non-linear transformation on features, preserving their continuity and non-negativity, making it suitable for scenarios requiring further processing. This design helps the network automatically adjust the scale and importance of features during the learning process, thereby improving the model's performance and generalization ability.

[0080] Subsequently, this application expands the three-dimensional tensor in the z-axis direction into a two-dimensional tensor and performs a summation operation to achieve a new expression of the overall nodal information of the material.

[0081] At this point, the convolutional layers are complete. The overall model employs a special processing matrix consisting of three convolutional layers and one fully connected layer to handle material structure and processing information. The final fully connected layer maps the matrix to the material's degradation performance. This completes the framework of the overall prediction model.

[0082] Step 4: Train the deep learning prediction model using the training set. The training of the prediction model is mainly based on the dataset collected above. The dataset is divided into three parts: a training set, a validation set, and a test set, which are used for model training, parameter tuning, and result verification, respectively.

[0083] The training, validation, and test sets are configured in a 6:2:2 ratio. The model's optimization algorithm is stochastic gradient descent with a learning rate of 0.001 and 500 iterations. At each iteration, the model's predictive performance is validated using the validation set. The primary validation metric is the coefficient of determination (R²). 2 ), where R 2 The closer R is to 1, the better the model's prediction performance on the overall dataset. 2 If the value is the current maximum, then the model parameters are saved to provide model support for subsequent predictions.

[0084] After 500 iterations of training, the model with the best determination coefficient is selected to test its prediction performance on the test set, and the results of the test set are used as the evaluation of the model's prediction performance on the generalized dataset.

[0085] Based on existing polymer material structure and neural network models, this application collects processing information and constructs a fully connected neural network model to process structural and processing information, accurately expressing the influence of processing information on the structure. At the same time, by allocating equal weights, the material's structure matrix and structural information matrix are combined to realize a deep learning prediction model of material structure-processing-performance.

[0086] To verify the method provided in this application for predicting the pyrolysis performance of polymer materials based on process parameters and structural information, 1200 pieces of polymer structural and processing information were collected, and the performance was set to T. 5％ The following training results were obtained through model training, see below. Figure 7 and Figure 8 .

[0087] Compared with existing technologies, the method for predicting the pyrolysis performance of polymer materials based on process parameters and structural information provided in this application has the following advantages:

[0088] (1) In terms of data processing, this application expands the traditional structural data of polymer materials to structural data plus processing information. It uses three numerical expressions, 0.79, 0.00 and 1.00, to mathematically represent the curing atmosphere for differentiation. At the same time, it uses a one-hot format to compile the curing temperature and curing time of the material, which are divided into several intervals: 100-150 (inclusive), 150-200 (inclusive), 200-250 (inclusive), 250-300 (inclusive), 300-350 (inclusive), and 350-400 (inclusive). The time is then combined with each interval.

[0089] (2) Regarding the processing information in fully connected neural networks, in order to clearly describe the impact of processing information on structural data, this application does not simply sum the processing information and structural data, but uses matrix operations to process the one-dimensional vector processing information with different nodes in the structural matrix, and uses fully connected layers to obtain a mathematical matrix of the same size as the structural matrix, namely the processing structure matrix. This application uses the combination of the structural matrix and the processing structure matrix to describe the material structure and processing information, and ultimately achieves accurate prediction of the pyrolysis performance of polymer materials.

[0090] To verify the model's predictive performance for different curing processes of the same material, this application extracted one of the materials and used the trained model to perform T... 5％ Performance predictions are shown in Table 1:

[0091] Table 1 Model for T 5％ Performance prediction results

[0092] PI-PyF10-200 200 494 477.42 474.36 PI-PyF10-350 350 513.81 477.42 483.50 PI-PyF20-200 200 481 475.65 472.68 PI-PyF20-350 350 505.49 475.65 491.73

[0093] As shown in Table 1, under the same conditions, the material's T value increases as the final curing temperature changes from 200℃ to 350℃. 5％ There has been an increase, but the models without processing technology do not predict the same material properties because the material structure is expressed the same way. However, with the processing information added, the effect of curing temperature on material properties can be correctly expressed.

[0094] Based on the above technical solutions, this application provides a method for predicting the pyrolysis performance of polymer materials based on process parameters and structural information, comprising the following steps: First, acquiring a dataset; the dataset includes material structural information and material processing information; then, processing the dataset and dividing the processed dataset into a training set, a validation set, and a test set; next, constructing a deep learning prediction model coupling polymer material processing information and structural information; finally, training the deep learning prediction model using the training set, evaluating the model training effect using the validation set, and inputting the test set into the trained deep learning prediction model to predict the pyrolysis performance of the polymer material. On the one hand, this application solves the problem of performance differences in the same polymer material due to different curing temperatures. By combining the structural information of the polymer material with processing information, the influence of processing information such as curing temperature and curing time on material performance is clarified while describing the structure of the polymer material. On the other hand, the method provided by this application can accelerate experimental design, predict the influence of different curing temperatures and times on material performance through the model, guide experimental design, optimize processing technology, improve synthesis efficiency, and reduce experimental costs.

[0095] Those skilled in the art will understand that the above-described embodiments are specific examples of implementing this application, and in practical applications, various changes in form and detail may be made without departing from the spirit and scope of this application. Any person skilled in the art can make their own modifications and alterations without departing from the spirit and scope of this application; therefore, the scope of protection of this application should be determined by the scope defined in the claims.

Claims

1. A method for predicting the pyrolysis performance of polymer materials based on process parameters and structural information, characterized in that, Includes the following steps: Obtain the dataset; the dataset includes material structure information and material processing information; The dataset is processed, and the processed dataset is divided into a training set, a validation set, and a test set. Construct a deep learning prediction model that couples polymer material processing information with material structure information; The deep learning prediction model is trained using a training set and evaluated using a validation set. The test set is then input into the trained deep learning prediction model to predict the pyrolysis performance of polymer materials. Obtain the dataset, including: Material structure information and processing information for multiple different types of materials are acquired separately; among them, The material structure information includes data on the structure and pyrolysis performance of polymer materials; The types of material processing information include curing atmosphere, curing temperature, curing time, curing agent, curing agent ratio, performance testing atmosphere, and heating rate. Constructing a deep learning prediction model that couples polymer material processing information with structural information, including: Construct fully connected neural networks and graph convolutional neural networks respectively; A fully connected neural network is used to process the material processing information in the dataset to obtain the structural processing matrix; The structural processing matrix and the material structure matrix are fused to represent the material structure and processing information; a graph convolutional neural network is used to map the matrix to the material's fracture properties. The graph convolutional neural network includes convolutional layers and fully connected layers. The convolutional layers are used to expand the edge feature matrix using a Gaussian kernel function to describe the material structure information and obtain the material structure matrix. The fully connected neural network consists of multiple fully connected layers with linear transformations and nonlinear activation functions. It is used to extract high-order semantic representations of processing conditions and combine the processing information of polymer materials with the initialized structure matrix. The two fully connected layers are used to transform the information matrix into an information matrix of the same size as the structure matrix input to the graph convolutional neural network. The material processing information in the dataset is processed using a fully connected neural network to obtain a structural processing matrix, including: After the material processing information is input, the curing temperature and curing time of the material are divided into multiple temperature ranges using a unique thermal coding method. Then, the curing time is combined with each temperature range. The material processing information of a one-dimensional vector is matched with different nodes in the structural matrix to obtain two matrices with the same dimension in the node direction; the structural processing matrix with fused material and structural processing information is obtained by matrix fusion.

2. The method for predicting the pyrolysis performance of polymer materials based on process parameters and structural information according to claim 1, characterized in that, Processing the dataset includes: The material structure information and material processing information in the dataset are processed respectively; The material structure information is transformed into a material structure data file that can provide input parameters for machine learning, and the material processing information is transformed into a processing information file that can provide data support for model training.

3. The method for predicting the pyrolysis performance of polymer materials based on process parameters and structural information according to claim 2, characterized in that, During the processing of the material processing information, the material processing information is transformed into a processing information file that can provide data support for model training, and the processing information file is transformed into a one-dimensional vector processing information file.

4. The method for predicting the pyrolysis performance of polymer materials based on process parameters and structural information according to claim 1, characterized in that, The deep learning prediction model uses two fully connected layers to obtain a special processing matrix of material structure and processing information. Through matrix fusion, a total material information matrix is ​​obtained. Then, the matrix is ​​mapped to the material fracture performance through three convolutional layers and one fully connected layer.

5. The method for predicting the pyrolysis performance of polymer materials based on process parameters and structural information according to claim 1, characterized in that, The material's structure matrix and structural processing matrix are fused to represent the material's structure and processing information, including: The final material representation is obtained by weighting the material's structure matrix and structural processing matrix with equal weights, as shown in the following formula: h total =0.5*h structh +0.5*h proch Among them, h total For the final material representation; h structh The structure matrix representing the material; h proch Represents the structural processing matrix.

Images