Composite material performance prediction and process optimization method based on neural network

Abstract

The invention provides a composite material performance prediction and process optimization method based on a neural network, and the method comprises the steps: firstly collecting multi-source data in the preparation and test process of a composite material, carrying out the preprocessing of the data, screening key feature variables as input variables, constructing a feedforward artificial neural network model, and predicting and outputting the performance indexes of the composite material. And training the model, performing iterative optimization on model parameters, and optimizing composite material process parameters by using the optimized model based on a reverse optimization strategy of a genetic algorithm to obtain an optimal process parameter combination. The invention provides a scientific, efficient and reliable tool for design and optimization of composite materials, and particularly has wide application prospects in high-requirement industries such as aerospace and the like.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of composite material performance prediction and process optimization, and relates to a composite material performance prediction and process optimization method based on a neural network. BACKGROUND

[0002] Composite materials have been widely used in aerospace, automotive, wind energy and other fields due to their excellent mechanical properties and lightweight characteristics, and have become an important choice for structural materials. However, the design and manufacturing process of composite materials is very complex, involving the optimization of multiple process parameters and the accurate prediction of material properties. In order to efficiently optimize the performance of composite materials under different process conditions, traditional design methods rely heavily on experimental data and theoretical calculations, which not only consumes time and resources, but also is difficult to comprehensively cover all combinations of process parameters.

[0003] At present, although some researches have adopted machine learning methods to predict the performance of composite materials, due to the highly nonlinear, complex and multi-scale characteristics of the production process and material properties of composite materials, most existing prediction models are difficult to simultaneously process a large number of input features and give accurate prediction results. In addition, existing methods usually do not provide a detailed physical explanation when considering material damage behavior and failure modes, resulting in certain limitations in complex engineering applications. SUMMARY

[0004] In view of the problems of insufficient prediction accuracy of composite material mechanical properties, dependence on experience for process design, and low optimization efficiency in the prior art, the application proposes a composite material performance prediction and process optimization method based on a neural network, which combines data-driven models and intelligent optimization algorithms, and realizes accurate evaluation of material properties and efficient optimization of process parameters by fully utilizing process parameters, material properties and experimental data, thereby providing systematic support for the design, preparation and engineering application of composite materials.

[0005] The technical scheme adopted by the application is as follows:

[0006] A composite material performance prediction and process optimization method based on a neural network includes the following steps:

[0007] S1: data collection and preprocessing;

[0008] The multi-source data in the composite material preparation and testing process is collected, mainly including four categories of process parameters, material properties, external working conditions and mechanical performance indicators, wherein the process parameters, material properties and external working conditions are input variables, and the mechanical performance indicators are output variables. The process parameters are, for example, fiber volume fraction, hot pressing temperature, holding time, hot pressing pressure and cooling rate; the material properties are, for example, resin type, fiber type, interface modification method and resin content; the external working condition parameters are, for example, loading rate, loading temperature, humid heat aging time and ultraviolet irradiation time; and the performance indicators include tensile strength (MPa), elastic modulus (GPa) and main damage modes (such as interface debonding, resin crack and fiber fracture).

[0009] The collected raw data is preprocessed to improve the data quality, and the processed data is used to construct a data set sample, the total number of samples is denoted as N, and a training data set is constructed, wherein each sample is composed of a group of input variables and corresponding output variables.

[0010] For continuous variables in the raw data, Z-score standardization method is used for standardization processing, and Min-Max normalization method is used for normalization processing.

[0011] The category variables in the raw data are processed by one-hot encoding, the missing values are filled by mean imputation, and the outliers are cleaned by box plot and 3σ criterion.

[0012] S2: feature extraction and selection;

[0013] The preprocessed input variables are extracted and screened, the key feature variables affecting the performance of the composite material are determined as input features, and the non-key feature variables are deleted. Preferably, the redundant features are preliminarily removed by Pearson correlation coefficient analysis method, and the key influence variables are identified by gradient boosting tree-based feature importance sorting method. The final data sample set formed after screening is used for modeling and training, comprehensively covering the process, material and environmental working condition characteristics of the composite material, and the data sample set is divided into training set and test set according to the ratio of 8:2.

[0014] S3: model construction and training;

[0015] A feedforward artificial neural network model is constructed for predicting the performance of composite materials. The model includes an input layer, hidden layers, and an output layer. The input layer receives input features and contains several nodes, the number of which corresponds to the number of selected key feature variables. The hidden layers are configured with a multi-layer structure, each containing several neurons and an activation function. Each input feature is linearly combined with its corresponding weight parameters in each hidden layer, processed by the activation function, and then output to the next layer of neurons, achieving nonlinear feature extraction and mapping. The output layer outputs target performance indicators, including tensile strength, elastic modulus, and damage type. Initial model parameters are determined, including the number of hidden layers, the number of neurons in each layer, and the types of activation functions.

[0016] The role of neurons in the hidden layers is to receive multiple input signals from the previous layer, generate intermediate features through weighted summation and bias adjustment, and introduce nonlinear transformations using activation functions. This enables the model to handle complex coupling relationships, prevents the network from degenerating into a linear mapping, and enhances the model's expressiveness and convergence efficiency. This training and forward propagation process is executed layer by layer in the hidden layers, ensuring that the input variables undergo progressive abstraction and high-order feature reconstruction within the model.

[0017] During training, mean squared error (MSE) is used as the loss function to measure the difference between the model's predicted values ​​and the true values. Its expression is as follows:

[0018]

[0019] In the formula, N represents the total number of samples, and n represents the number of output variables for each sample. Let represent the predicted value of the i-th output variable in the h-th sample. This represents the true value corresponding to the i-th output variable in the h-th sample.

[0020] Gradient descent is used to optimize model parameters to minimize the loss function, thereby improving model prediction accuracy. The descent direction for the current weight parameter point is determined by calculating the partial derivatives of the loss function with respect to each neuron parameter (including weights and biases), and the weight parameters are updated according to the learning rate, causing the loss function value to gradually converge. The weight parameter update formula is as follows:

[0021]

[0022] In the formula, ω represents the weight parameter, and η is the learning rate. t+1 represents the gradient of the loss function with respect to the weight parameters, and t+1 represents the current iteration number.

[0023] Training is considered complete when the loss function L converges to below a set threshold or reaches the set maximum number of training rounds. The trained model is then used to quickly predict the performance of composite materials with any combination of variable parameters.

[0024] S4: Model parameter optimization;

[0025] To further improve the model's prediction accuracy, an automatic hyperparameter optimization method is employed, systematically searching for each model hyperparameter to be optimized within its corresponding multidimensional hyperparameter search space. During the search process, a Bayesian optimization algorithm is used to automatically sample hyperparameter combinations. After training the model on the training set, performance prediction results are output on the validation set. The mean squared error (MSE) is used as the objective function to calculate the deviation between the predicted and true values, and this MSE is used as the fitness evaluation metric. The model hyperparameters to be optimized include: learning rate, number of hidden layers, number of neurons per layer, activation function type, number of iterations, and batch size.

[0026] As the iteration process progresses, the sampling strategy is dynamically adjusted based on the mean square error value to gradually approach the optimal solution. When the mean square error no longer decreases significantly or reaches the maximum number of iterations, the search process terminates, and the current optimal hyperparameter combination is output. The optimized final model is constructed using the optimal hyperparameters and can be used as input for any set of composite material process parameters to achieve high-precision prediction of target performance indicators, significantly improving parameter configuration efficiency and performance control capabilities.

[0027] S5: Process parameter optimization;

[0028] To achieve automated optimization of composite material processing parameters, a reverse optimization strategy based on genetic algorithms is introduced to guide the configuration of process parameters in actual production, thereby achieving optimal performance. The process includes:

[0029] S5.1 The process parameters of the composite material to be optimized are encoded to construct an initial parameter population containing multiple individual parameter combinations. Each "individual" in the parameter population is a set of process parameter combinations, including the key feature variables selected in S2.

[0030] S5.2 invokes the trained feedforward artificial neural network model to predict performance indicators for each individual parameter combination in the parameter population. The predicted output includes the tensile strength of the target composite material. elastic modulus and damage probability P damage This is used as a fitness evaluation parameter. To achieve collaborative optimization of multi-objective performance, the following weighted fitness function is constructed as an evaluation criterion:

[0031]

[0032] Among them, ω1, ω2, and ω3 are the weighting factors of the performance indicators.

[0033] S5.3 Based on the natural selection principle of genetic algorithms, selection, crossover, and mutation operations are performed on the process parameters to be optimized in each individual parameter combination to generate a new generation of individual parameter combinations and form a new generation of parameter population. The trained feedforward artificial neural network model is then invoked to predict the performance indicators of the new generation of individual parameter combinations and calculate the corresponding fitness function values. The selection operation prioritizes retaining high-fitness individuals, the individual crossover operation involves randomly exchanging sites between two individual parameters, and the mutation operation involves slightly perturbing some parameter values ​​to increase diversity.

[0034] S5.4 Determine if the current iteration meets the termination condition: If the maximum value of the fitness function converges within a certain number of generations, or reaches the set maximum generation limit, then the optimization process is considered to have ended, and the optimal combination of process parameters with the current fitness is output; otherwise, return to S5.3.

[0035] Finally, an optimal combination of process parameters is obtained.

[0036] The beneficial effects of this invention are as follows:

[0037] (1) The model of this invention can handle high-dimensional, complex, and significantly coupled process variables, accurately predict the mechanical strength, modulus and damage mode of composite materials. Compared with traditional models (such as traditional random forest models), it has higher fitting accuracy and stronger generalization ability, and significantly reduces the dependence of performance prediction on the experimental sample size.

[0038] (2) This invention combines a genetic algorithm to perform a global search of the process parameter space, with optimal performance as the objective function, effectively solving the problem of “inferring process configuration from performance target”, providing a quantitative and feasible optimization path for composite material process parameter design, and supporting process combination optimization under multiple objectives and constraints.

[0039] (3) By using the Optuna platform to perform automated hyperparameter optimization of neural networks, the system can quickly converge to the optimal model structure and training parameters without human intervention, which significantly improves the efficiency of model building and reduces the number of experimental iterations.

[0040] (4) This method is not only applicable to composite materials, but can also be extended to the prediction of various material properties and process optimization in other fields, and has strong versatility and applicability.

[0041] Therefore, this invention provides a scientific, efficient, and reliable tool for the design and optimization of composite materials, and has broad application prospects, especially in demanding industries such as aerospace. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating the framework of the present invention.

[0043] Figure 2 This is a diagram of the architecture of a feedforward artificial neural network model.

[0044] Figure 3 This is a graph showing the importance of parameter features.

[0045] Figure 4 The following are comparison charts showing the predicted and actual values ​​of the example model and the traditional random forest model: (a) is a comparison chart of the predicted and actual values ​​of the feedforward artificial neural network model provided in the example, and (b) is a comparison chart of the predicted and actual values ​​of the traditional random forest model. Detailed Implementation

[0046] The following detailed description, in conjunction with the accompanying drawings, illustrates a specific implementation of the neural network-based composite material performance prediction and process optimization method proposed in this invention. It should be understood that the described embodiments are merely illustrative of the technical solutions of this invention and do not constitute a limitation on the scope of protection of this invention. Figure 1 The implementation flowchart shown includes the following specific contents:

[0047] S1: Data collection and preprocessing;

[0048] First, multi-source data were collected during the preparation and testing of composite materials, mainly including four categories: process parameters, material properties, external operating conditions, and mechanical performance indicators. Among them, process parameters, material properties, and external operating conditions are used as input variables, and mechanical performance indicators are used as output variables. Process parameters include fiber volume fraction, hot-pressing temperature, holding time, hot-pressing pressure, and cooling rate; material properties include resin type, fiber type, interface modification method, and resin content; operating condition parameters include loading rate, loading temperature, damp heat aging time, and ultraviolet irradiation time; performance indicators include tensile strength (MPa), elastic modulus (GPa), and main damage modes (such as interface debonding, resin cracking, and fiber breakage).

[0049] The collected raw data was standardized and normalized, and missing and outlier values ​​were removed to improve data quality. The processed data was used to construct a dataset sample, where each sample consists of a set of input variables and their corresponding output variables. The total number of samples is denoted as N, forming a complete training dataset that provides foundational data support for subsequent feature analysis and neural network modeling training.

[0050] For continuous variables, the Z-score standardization method is used for standardization:

[0051]

[0052] Where, xi Here are the original data, μ is the sample mean, and σ is the standard deviation. This is the standardized data.

[0053] The normalization process uses the Min-Max normalization method:

[0054]

[0055] In the formula, x max x is the maximum value in the dataset. min x is the minimum value in the dataset. i The data before normalization. This is the normalized data.

[0056] Categorical variables (such as fiber type, interface modification method, etc.) are handled using one-hot encoding (0 / 1). Missing values ​​are filled using mean imputation, and outliers are cleaned using a combination of box plots and the 3σ criterion to ensure that the input data quality meets the modeling requirements.

[0057] S2: Feature extraction and selection;

[0058] After data preprocessing, key features affecting the performance of composite materials are extracted and screened based on the training dataset and used as input features. In this embodiment, redundant features are initially removed using Pearson correlation coefficient analysis, and then a gradient boosting tree-based feature importance ranking method is used to identify key influencing variables.

[0059] Figure 3 The importance analysis results of the selected input features are presented. It can be seen that fiber volume fraction, fiber type, layup angle, and hot-pressing temperature are key input parameters, significantly impacting the predicted material properties. Other variables such as hot-pressing pressure, resin type, and loading temperature also have certain weights, while low-weight variables such as hot-pressing cooling rate and ultraviolet radiation conditions are retained to enhance the model's generalization ability. These results further validate the rationality of the feature selection strategy of this invention and provide clear input criteria for subsequent neural network modeling.

[0060] Finally, 30 key feature variables were selected from the original approximately 50 variables. The final data sample set was used for model training, which comprehensively covered the process, material and environmental conditions of composite materials. The data sample set was divided into training set and test set in an 8:2 ratio.

[0061] S3: Model building and training;

[0062] A feedforward artificial neural network model is constructed for predicting the performance of composite materials. The model includes an input layer, hidden layers, and an output layer. The input layer receives input features and contains 30 nodes, corresponding to 30 selected key feature variables. The hidden layers are configured with a multi-layer structure; in this embodiment, the initial model parameters are set to two hidden layers with 64 and 32 neurons respectively, and the ReLU activation function is used. During model training, each input feature is fed into the network through the input layer and linearly combined with corresponding weight parameters in each hidden layer. After processing by the activation function, the feature is output to the next layer of neurons, achieving nonlinear feature extraction and mapping. The neurons receive multiple input signals from the previous layer, generate intermediate features through weighted summation and bias adjustment, and the activation function introduces nonlinear transformations, enabling the model to handle complex coupling relationships, preventing the network from degenerating into a linear mapping, and enhancing the model's expressiveness and convergence efficiency. This training and forward propagation process is executed layer by layer in the hidden layers, ensuring that the input parameters are abstracted and reconstructed into higher-order features within the network. The output layer outputs target performance indicators, including tensile strength, elastic modulus, and damage type.

[0063] like Figure 2 As shown, in the model, let the weight of the j-th neuron in the l-th hidden layer be... The bias term is Where l represents the hidden layer number, and j represents the j-th neuron in that layer. This is used to connect the k-th neuron in layer (l-1) to the j-th neuron in layer l. Therefore, the input received by the j-th neuron in layer l from the previous layer can be represented as:

[0064]

[0065] In the formula, K represents the total number of neurons in the (l-1)th layer. Defined as:

[0066]

[0067] The model is trained, and during the training process, the mean squared error (MSE) is used as the loss function to measure the difference between the model's predicted values ​​and the true values. Its expression is as follows:

[0068]

[0069] In the formula, N represents the total number of training samples, and n represents the number of output variables for each sample. Let represent the predicted value of the i-th output variable in the h-th sample. This represents the true value corresponding to the i-th output variable in the h-th sample.

[0070] To minimize the loss function and thus improve the model's prediction accuracy, gradient descent is used to optimize the model parameters. By calculating the partial derivatives of the loss function with respect to each neuron's parameters (including weights and biases), the descent direction for the current weight parameter point is determined, and the weight parameters are updated according to the learning rate, causing the loss function value to gradually converge. The weight parameter update formula is as follows:

[0071]

[0072] In the formula, ω represents the weight parameter, and η is the learning rate. This represents the gradient of the loss function with respect to that parameter.

[0073] In the implementation, the Adam optimizer is used for iterative training of the neural network. This optimizer combines momentum and adaptive learning rate mechanisms, which can accelerate convergence in the early stages of training and maintain stability in the later stages. During training, the network adjusts the weights and biases of each neuron in each layer according to error backpropagation, causing the loss function L to continuously decrease.

[0074] When the loss function L MSE When the model converges to below a set threshold or reaches the set maximum number of training rounds (500 rounds), the training is considered complete. The trained model can be used to quickly predict the performance of composite materials with any combination of parameters.

[0075] S4: Model parameter optimization;

[0076] To further improve the prediction accuracy of the neural network model, an automatic hyperparameter optimization method based on the Optuna platform is adopted. A systematic search is conducted in the multidimensional hyperparameter search space to obtain the optimal model structure and training strategy, thereby enhancing the model's fitting and generalization abilities to the composite material properties. The model hyperparameters to be optimized include: learning rate, number of hidden layers, number of neurons per layer, activation function type, and batch size. Specifically, the learning rate search range is set to 1×10⁻⁶. -5 ~1×10 -1 The number of hidden layers is selected between 1 and 4, the number of neurons in each layer is set to 16 to 256, the activation function types include ReLU, Sigmoid and Tanh, and the batch size varies between 16 and 128.

[0077] During the parameter search process, the Optuna platform automatically samples hyperparameter combinations based on the Bayesian optimization algorithm. After training the neural network model on the training set, it outputs performance prediction results on the validation set. The mean squared error (MSE) is used as the objective function to calculate the deviation between the predicted and actual values, and this MSE is used as the fitness evaluation metric. As the iteration progresses, Optuna dynamically adjusts its sampling strategy based on the MSE value calculated after the previous iteration, gradually approaching the optimal solution. When the stopping condition is met (the validation error no longer decreases significantly or the maximum number of iterations is reached), the search process terminates, and the current optimal hyperparameter combination is output.

[0078] The final optimal hyperparameters will be used for the final training and deployment of the neural network model. The completed prediction model can be input with any set of composite material process parameters to achieve high-precision prediction of target performance indicators, significantly improving parameter configuration efficiency and performance control capabilities.

[0079] S5: Process parameter optimization;

[0080] To achieve automated optimization of composite material processing parameters, a reverse optimization strategy based on genetic algorithms is introduced to guide the configuration of process parameters in actual production, thereby achieving optimal performance. Its basic process includes:

[0081] S5.1 The process parameters of the composite material to be optimized are encoded to construct an initial parameter population containing multiple individual parameter combinations. Each "individual" in the parameter population is a set of process parameter combinations, including the key feature variables selected in S2.

[0082] S5.2 invokes the trained feedforward artificial neural network model to predict performance indicators for each individual parameter combination in the parameter population. The predicted output includes the tensile strength of the target composite material. elastic modulus and damage probability P damage As a fitness evaluation parameter, a weighted fitness function is constructed as the evaluation criterion to achieve collaborative optimization of multi-objective performance:

[0083]

[0084] Among them, ω1, ω2, and ω3 are the weighting factors of the performance indicators.

[0085] S5.3 Based on the natural selection principle of genetic algorithms, selection, crossover, and mutation operations are performed on the process parameters to be optimized in each individual parameter combination to generate a new generation of individual parameter combinations and form a new generation of parameter population. The trained feedforward artificial neural network model is then invoked to predict the performance indicators of the new generation of individual parameter combinations and calculate the corresponding fitness function values. The selection operation prioritizes retaining high-fitness individuals, the individual crossover operation involves randomly exchanging sites between two individual parameters, and the mutation operation involves slightly perturbing some parameter values ​​to increase diversity.

[0086] S5.4 Determine if the current iteration meets the termination condition: If the maximum value of the fitness function converges within a certain number of generations, or reaches the set maximum generation limit, then the optimization process is considered to have ended, and the optimal combination of process parameters with the current fitness is output; otherwise, return to S5.3.

[0087] The final set of optimal process parameters can be used to guide the actual composite material production process, thereby achieving the goal of performance optimization, improving the performance of composite materials, reducing production costs, and increasing production efficiency.

[0088] Result verification:

[0089] The performance prediction results of the feedforward artificial neural network model constructed and trained in this embodiment are compared with those of the traditional random forest model, both using the same test set data. Figure 4 The comparison results of the two models in performance prediction are shown. As can be seen from the figure, the neural network model constructed in this embodiment exhibits higher accuracy in the performance prediction task. Its coefficient of determination R0 2 The mean squared error (MSE) is 0.93, significantly higher than the 0.87 of the random forest model, indicating that the neural network model has a stronger interpretability of the sample data and a better fitting effect. The mean squared error (MSE) is 3.52, much lower than the 6.01 of the random forest model, indicating that the neural network model has smaller error fluctuations and lower squared error values ​​in the overall prediction process. The mean absolute error (MAE) is 1.51, also lower than the 1.73 of the random forest model, further indicating that the neural network model has a smaller average prediction bias at each sample point and has higher prediction stability. In summary, the neural network model is superior to the traditional random forest model in terms of accuracy, stability, and generalization ability, and can efficiently predict the mechanical strength, modulus, and damage behavior of composite materials under different process conditions. This invention has significant advantages in the field of composite material modeling and simulation and can be widely applied in research and practical engineering in related fields.

[0090] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the claims of this application.

Claims

1. A method for predicting the properties and optimizing the process of composite materials based on neural networks, characterized in that, Includes the following steps: S1: Data collection and preprocessing, and construction of a sample set; S2: Identify key feature variables through feature extraction and selection; S3: Model building and training; S4: Model parameter optimization; S5: Process parameter optimization.

2. The method for predicting composite material properties and optimizing processes based on neural networks according to claim 1, characterized in that, S1 specifically refers to: collecting multi-source data during the preparation and testing process, mainly including four categories: process parameters, material properties, external working conditions, and mechanical performance indicators. Among them, process parameters, material properties, and external working conditions are used as input variables, and mechanical performance indicators are used as output variables. The collected raw data is preprocessed to improve data quality. The processed data is then used to construct dataset samples to form a training dataset, where each sample consists of a set of input variables and their corresponding output variables.

3. The method for predicting composite material properties and optimizing processes based on neural networks according to claim 2, characterized in that, For continuous variables in the original data, Z-score standardization was used for standardization, and Min-Max normalization was used for normalization. For categorical variables in the original data, one-hot encoding was used. Missing values ​​were filled with mean imputation, and outliers were cleaned using a combination of box plots and the 3σ criterion.

4. The method for predicting composite material properties and optimizing processes based on neural networks according to claim 1, characterized in that, Step 2 specifically involves: extracting and filtering the preprocessed data; using Pearson correlation coefficient analysis to initially remove redundant features; then using a gradient boosting tree-based feature importance ranking method to identify key influencing variables; determining key feature variables that affect performance as input features; and deleting non-key feature variables.

5. The method for predicting composite material properties and optimizing processes based on neural networks according to claim 1, characterized in that, S3 specifically involves: constructing a feedforward artificial neural network model, including an input layer, hidden layers, and an output layer. The input layer receives input features and contains several nodes, the number of which corresponds to the number of key feature variables after filtering. The hidden layers are configured with a multi-layer structure, each containing several neurons and an activation function. Each input feature is linearly combined with its corresponding weight parameters in each hidden layer, processed by the activation function, and then output to the next layer of neurons, achieving non-linear feature extraction and mapping. The output layer outputs the target performance index. Initial model parameters are determined, including the number of hidden layers, the number of neurons in each layer, and the types of activation functions. The mean squared error is used as the loss function to measure the difference between the model's predicted values ​​and the true values. Its expression is as follows: In the formula, N represents the total number of samples, and n represents the number of output variables for each sample. Let represent the predicted value of the i-th output variable in the h-th sample. This represents the true value corresponding to the i-th output variable in the h-th sample; The model parameters are optimized using gradient descent. The model is considered to be trained when the loss function L converges to below a set threshold or reaches the set maximum number of training rounds.

6. The method for predicting composite material properties and optimizing processes based on neural networks according to claim 5, characterized in that, By calculating the partial derivatives of the loss function with respect to the parameters of each neuron, i.e., the weight parameters, the descent direction of the current weight parameter point is determined, and the weight parameters are updated according to the learning rate, so that the loss function value gradually converges. The formula for updating the weight parameters is: In the formula, ω represents the weight parameter, and η is the learning rate. t+1 represents the gradient of the loss function with respect to the weight parameters, and t+1 represents the current iteration number.

7. The method for predicting composite material properties and optimizing processes based on neural networks according to claim 5, characterized in that, Specifically, S4 involves: employing an automatic hyperparameter optimization method to systematically search for each model hyperparameter to be optimized in its corresponding multidimensional hyperparameter search space; during the search process, automatically sampling hyperparameter combinations based on the Bayesian optimization algorithm, training the model on the training set, and outputting performance prediction results on the validation set; using the mean squared error as the objective function to calculate the deviation between the predicted value and the true value, and using the mean squared error as the fitness evaluation index. The sampling strategy is dynamically adjusted based on the mean square error value. When the mean square error no longer decreases significantly or reaches the maximum number of iterations, the search process terminates, the current optimal hyperparameter combination is output, and the optimized final model is constructed using the optimal hyperparameters.

8. The method for predicting the performance of composite materials and optimizing processes based on neural networks according to claim 7, characterized in that, The model hyperparameters to be optimized include: learning rate, number of hidden layers, number of neurons per layer, activation function type, number of iterations, and batch size.

9. The method for predicting the properties and optimizing the process of composite materials based on neural networks according to claim 1, characterized in that, Specifically, S5 involves introducing a reverse optimization strategy based on a genetic algorithm to achieve optimal performance, including: S5.1 Encode the process parameters of the composite material to be optimized and construct an initial parameter population containing multiple individual parameter combinations; each "individual" in the parameter population is a set of process parameter combinations, including the key feature variables screened in S2; S5.2 calls the final trained model, predicts the performance index of each individual parameter combination in the parameter population, and constructs a weighted fitness function as the evaluation standard; S5.3 Based on the natural selection principle of genetic algorithm, selection, crossover, and mutation operations are performed on the process parameters to be optimized in each individual parameter combination to generate a new generation of individual parameter combinations and form a new generation of parameter population; the trained feedforward artificial neural network model is called to predict the performance index of the new generation of individual parameter combinations and calculate the corresponding fitness function value; the selection operation prioritizes the retention of high fitness, the individual crossover operation randomly swaps the sites of two individual parameters, and the mutation operation slightly perturbs some parameter values ​​to increase diversity; S5.4 Determine if the current iteration meets the termination condition: If the maximum value of the fitness function converges within a certain number of generations, or reaches the set maximum generation limit, then the optimization process is considered to have ended, and the optimal combination of process parameters with the current fitness is output; otherwise, return to S5.

3. Finally, an optimal combination of process parameters is obtained.

10. The method for predicting the properties and optimizing the process of composite materials based on neural networks according to claim 1, characterized in that, The weighted fitness function is as follows: in, For tensile strength, P is the elastic modulus. damage ω1, ω2, and ω3 represent the damage probability, and ω3 are the weighting factors for the performance indicators.

Images