An epoxy resin material preparation method based on a generative adversarial network and a multi-objective optimization method
By constructing a hybrid constrained generative adversarial network and a multi-objective optimization method, the problems of insufficient data and multi-objective optimization in epoxy resin material preparation were solved, achieving efficient and reliable performance prediction and parameter optimization, and improving the efficiency of material research and development.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTHEAST UNIV
- Filing Date
- 2026-05-09
- Publication Date
- 2026-07-31
AI Technical Summary
Existing technologies in epoxy resin material preparation suffer from long experimental cycles, high costs, limited parameter space coverage, and difficulty in achieving a systematic trade-off between multiple performance objectives, resulting in insufficient data, unstable model training, and poor predictive generalization ability.
We employ a generative adversarial network (GAN) and multi-objective optimization approach. We construct a hybrid constrained GAN (HCC-GAN) extended dataset and combine it with models such as multilayer perceptron, support vector regression, and extreme gradient boosting machine for performance prediction. We utilize a non-dominated sorting genetic algorithm for multi-objective optimization to achieve automated search and optimization of preparation parameters.
It improves the quality of data augmentation and the accuracy of performance prediction under conditions of scarce experimental data, reduces the number of experiments and costs, achieves multi-objective optimization, outputs the optimal set of formulations for engineering selection, and improves the efficiency of epoxy resin material research and development.
Smart Images
Figure CN122494077A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of materials science and intelligent manufacturing technology, specifically relating to a method for preparing epoxy resin materials based on generative adversarial networks and multi-objective optimization methods. Background Technology
[0002] Epoxy resins are widely used in structural adhesives, composite matrixes, coatings, and electronic packaging due to their excellent bonding, mechanical, and environmental resistance properties. The mechanical properties of epoxy resin materials, such as tensile strength and elongation at break, are typically influenced by a variety of preparation parameters, including the ratio of curing agent, the amount of toughening agent, and the temperature and time of multi-stage curing. Their performance formation mechanism is complex, exhibiting significant nonlinearity and multi-factor coupling characteristics.
[0003] In practical materials research and development, the optimization of epoxy resin formulations and curing processes mainly relies on extensive physical experiments. These experiments typically suffer from long experimental cycles, high costs, and limited parameter space coverage, resulting in a limited and sparsely distributed number of experimental data samples. Under small sample conditions, directly using traditional machine learning or deep learning methods to build performance prediction models often leads to insufficient training, poor generalization ability, and insufficient prediction stability, making it difficult to meet the needs of engineering applications.
[0004] To alleviate the problem of insufficient experimental data, existing techniques have attempted to expand material datasets by introducing data augmentation methods such as generative adversarial networks (GANs). However, most existing methods are based on traditional GAN structures, which are prone to problems such as pattern collapse and gradient instability during training. Furthermore, they typically lack constraints on the physical boundaries of preparation parameters and the reasonable range of performance, making it difficult to guarantee the engineering feasibility and numerical reliability of the generated data.
[0005] Furthermore, epoxy resin materials commonly face the problem of mutual constraints among multiple performance objectives in engineering applications. For example, increasing tensile strength often reduces elongation at break. Existing research mostly focuses on the prediction or optimization of single performance indicators, or relies on empirical rules for parameter adjustment, making it difficult to achieve a systematic trade-off among multiple objectives and output an optimal set of formulations for engineering selection.
[0006] Therefore, there is an urgent need for a method that can achieve highly reliable data augmentation, rapid prediction of material properties, and intelligent optimization of multi-objective formulations under limited sample size conditions, so as to improve the efficiency of epoxy resin material research and development and reduce experimental costs. Summary of the Invention
[0007] Purpose of the Invention: To address the problems that existing methods are mostly based on traditional generative adversarial networks (GANs), which are prone to pattern collapse and gradient instability during training, and often lack constraints on the physical boundaries and reasonable performance range of preparation parameters, making it difficult to guarantee the engineering feasibility and numerical reliability of the generated data; and to solve the problem that existing research often focuses on the prediction or optimization of a single performance index or relies on empirical rules for parameter adjustment, making it difficult to achieve a systematic trade-off among multiple objectives and output an optimal set of formulations that can be selected for engineering purposes, this invention proposes a method for preparing epoxy resin materials based on generative adversarial networks and multi-objective optimization methods.
[0008] Technical solution: A method for preparing epoxy resin materials based on generative adversarial networks and multi-objective optimization, comprising:
[0009] Step 1: Obtain and preprocess the performance data of epoxy resin materials under different preparation parameters to obtain the original experimental samples; based on prior knowledge of material processes, form a feasible domain for preparation parameters, which include process parameters and formulation; the performance data are obtained from multiple mutually restrictive performance indicators.
[0010] Step 2: Construct a Generative Adversarial Network (GAN). The generator in the GAN takes preprocessed preparation parameters and random noise as input and outputs a performance prediction value. The discriminator in the GAN takes preparation parameters and performance data as input and outputs a score value. The GAN is trained using an adversarial loss function.
[0011] Step 3: Extend the preparation parameters of the original experimental samples by convex combination interpolation to construct new preparation parameters. Perform physical boundary screening on the new preparation parameters according to the feasible region of the preparation parameters. Input the screened preparation parameters into the trained generator to obtain the corresponding performance data, thereby constructing the augmented dataset.
[0012] Step 4: Train multiple pre-built learning agent models using the augmented dataset, and select the learning agent model with the best prediction performance through cross-validation and prediction accuracy evaluation.
[0013] Step 5: Based on the machine learning agent model with the best prediction performance, construct a multi-objective optimization problem with multiple mutually constraining performance index values as optimization objectives. Within the feasible region of the preparation parameters, solve the multi-objective optimization problem to obtain the optimal preparation parameters that satisfy the optimization objectives.
[0014] Step 6: Based on the optimal preparation parameters, prepare the epoxy resin material.
[0015] Furthermore, the preparation parameters include: curing agent ratio, toughening agent dosage, multi-stage curing temperature and curing time; the multiple mutually restrictive performance indicators include at least tensile strength and elongation at break.
[0016] Furthermore, the loss function of the discriminator includes the difference between the real sample score and the generated sample score, as well as a gradient penalty term, to satisfy the 1-Lipschitz condition, and the loss function of the generator includes adversarial loss and L1 supervised loss.
[0017] Furthermore, the process of extending the preparation parameters of the original experimental samples through convex combination interpolation to construct new preparation parameters includes:
[0018] Preparation parameters for k randomly selected original experimental samples are used to generate a weight vector using a Dirichlet distribution. The new preparation parameters are obtained by constructing the following formula:
[0019] .
[0020] Furthermore, the generator and discriminator of the generative adversarial network both adopt a multilayer perceptron structure, each containing three hidden layers, with the number of neurons in each layer being 128, 128, and 64 respectively, and the activation function being LeakyReLU.
[0021] Furthermore, the learning agent model employs one or more of the following: multilayer perceptron, support vector regression, extreme gradient boosting machine, and lightweight gradient boosting machine.
[0022] Furthermore, the step of solving the multi-objective optimization problem within the feasible region of the preparation parameters to obtain the optimal preparation parameters that satisfy the optimization objectives includes:
[0023] Within the feasible region of the preparation parameters, the multi-objective optimization problem is solved using a non-dominated sorting genetic algorithm to obtain the optimal preparation parameters that satisfy the optimization objectives.
[0024] Among them, the hyperparameters of the non-dominated sorting genetic algorithm are adaptively adjusted by the Bayesian optimization method. The optimized hyperparameters include population size, number of iterations, simulated binary crossover probability, crossover distribution index, and mutation distribution index. The Bayesian optimization method uses expectation boosting as the acquisition function and hypervolume index as the optimization objective to optimize the hyperparameters of the non-dominated sorting genetic algorithm.
[0025] Beneficial Effects: This invention includes steps such as data augmentation, performance prediction model construction, multi-objective optimization, and evaluation. In the data augmentation stage, a Hybrid Constrained Generative Adversarial Network (HCC-GAN) algorithm is proposed to achieve high-fidelity generation and expansion of epoxy resin test data. Using the data generated by the augmentation algorithm, various machine learning models such as MLP, SVR, XGBoost, and LightGBM are built to predict epoxy resin performance. The predicted epoxy resin material properties are validated and compared with real test data to select the optimal surrogate model. The selected machine learning model is used as the evaluation surrogate model in the multi-objective optimization algorithm, combined with the Non-Dominated Sorting Genetic Algorithm (NSGA-II) to achieve epoxy resin material formulation and process optimization design under multi-objective constraints. Through the above technical solutions, this invention has the following advantages compared with existing technologies:
[0026] (1) Under the condition of scarce experimental data, this invention introduces a hybrid constraint condition generator adversarial network (HCC-GAN) with stable training characteristics and superimposes gradient penalty and performance supervision constraints, which can improve the quality of material data augmentation in terms of both distribution consistency and numerical reliability.
[0027] (2) This invention combines Dirichlet convex combination interpolation to expand the process space and generates samples through physical boundary screening and extreme value filtering constraints, so that the augmented data meets the requirements of engineering feasibility, thereby providing a more sufficient and reliable data foundation for surrogate model training;
[0028] (3) The surrogate model obtained by training and screening with augmented data in this invention can more stably characterize the nonlinear mapping relationship between preparation parameters and material properties, and improve the prediction and generalization ability of real experimental data.
[0029] (4) The multi-objective optimization of the present invention can systematically reveal the trade-off between strength and toughness and output a set of Pareto optimal formulations that can be selected for engineering. When local sensitivity index is introduced, the sensitivity of the optimization solution to process disturbances can be reduced and the robustness can be improved.
[0030] (5) By combining the experimental verification process, the present invention can reduce the number of experiments and trial and error costs, and realize the automated search and optimization of formula and curing preparation parameters, thereby improving the efficiency of epoxy resin material research and development and reducing research and development risks.
[0031] (6) This invention improves the reliability of data enhancement and the accuracy of performance prediction of epoxy resin materials under the condition of scarce experimental data, realizes multi-objective optimization of material formulation and preparation parameters, and provides an effective method for intelligent design of epoxy resin materials. Attached Figure Description
[0032] Figure 1This is a flowchart of a method for preparing epoxy resin materials based on generative adversarial networks and multi-objective optimization proposed in this invention;
[0033] Figure 2 This is a schematic diagram of the data augmentation method framework based on HCC-GAN;
[0034] Figure 3 Flowchart of a method for predicting and optimizing the properties of epoxy resin materials using multiple objectives;
[0035] Figure 4 A schematic diagram of epoxy resin material preparation, performance testing, and dataset structure;
[0036] Figure 5 Pareto frontiers under different proxy models;
[0037] Figure 6 This serves as an experimental verification of the extrapolation results. Detailed Implementation
[0038] To make the objectives, technical solutions, and advantages of the present invention clearer, the design method proposed in this invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0039] Example 1:
[0040] This invention proposes a method for preparing epoxy resin materials based on generative adversarial networks and multi-objective optimization methods, such as... Figure 1 As shown, it includes the following steps:
[0041] Step 1: Organize and preprocess the performance data of epoxy resin materials under different formulations and process conditions. In this embodiment of the invention, different formulations include the proportion of curing agent and the amount of toughening agent. For example, this includes the proportion of curing agent (CAR, unit: %) and the amount of hyperbranched polymer (HBP, unit: %). Different process conditions include multi-stage curing temperatures and times, for example, including: low-temperature curing temperature (LT, unit: °C), low-temperature curing time (Lt, unit: d), high-temperature curing temperature (HT, unit: °C), and high-temperature curing time (Ht, unit: h). Performance indicators include at least tensile strength and elongation at break.
[0042] The preparation parameters and performance data are Z-score standardized according to the following formula. Based on prior knowledge of materials engineering, a physically reasonable range is set for the preparation parameters as a constraint for subsequent data generation and optimization.
[0043]
[0044] Where x is the original variable, For standardized variables, These are the sample mean and standard deviation, respectively.
[0045] As one implementation method, see Figure 4 The performance data of epoxy resin materials under different formulations and processes are obtained through material preparation, curing treatment and mechanical property testing. The input variable is the preparation parameter vector X, whose components may include curing agent ratio, HBP dosage, low temperature curing temperature, low temperature curing time, high temperature curing temperature and high temperature curing time. The output variable is y, which includes tensile strength (TS, unit MPa) and elongation at break (EB, %).
[0046] To eliminate the influence of different units of measurement on model training, the input and output variables are standardized. Simultaneously, based on prior knowledge of materials and processes, physically reasonable ranges are defined for each input variable, forming a feasible parameter region for subsequent sample selection, constraint modeling, and multi-objective optimization.
[0047] Step 2: As Figure 2 As shown, a hybrid constrained generative adversarial network (HCC-GAN) is constructed for material property data augmentation. It includes a generator G and a discriminator C. The generator takes the standardized preparation parameters X and random noise z as inputs to generate corresponding performance prediction values. , represented as:
[0048] = G (X, z)
[0049] The discriminator takes the preparation parameters and performance data (X, y) as input and outputs a score value C(X, y).
[0050] A Hybrid Constrained Generative Adversarial Network (HCC-GAN) is trained using an adversarial loss function based on Wasserstein distance, where the discriminator loss function includes the difference between the real sample score and the generated sample score, plus a gradient penalty term, satisfying the 1-Lipschitz condition, and is expressed as:
[0051]
[0052] in, This indicates the distribution of real data; To generate samples; This represents the interpolated sample used for gradient penalty; Represents the 2-norm; represents the weight coefficients of the gradient penalty term. A gradient penalty mechanism is introduced during training to constrain the discriminator to satisfy the Lipschitz continuity condition.
[0053] In addition to the adversarial loss, the generator loss function also incorporates L1 supervised loss to constrain the deviation between the generated performance and the real performance. Training is performed by minimizing the loss function, which is expressed as:
[0054]
[0055] in, These are supervisory weighting coefficients; It represents the 1-norm.
[0056] After training, the formulation and preparation parameters of the original experimental samples are extended by convex combination interpolation based on the Dirichlet distribution. For example, k original samples are randomly selected, and weight vectors are generated by the Dirichlet distribution. And construct new preparation parameter samples:
[0057]
[0058] For the generated candidate preparation parameter samples Physical boundary screening is performed according to the feasible region. The selected preparation parameters are input into the trained generator to obtain the corresponding material property data. The generated results are then filtered through inverse normalization and performance extrema to construct an enhanced dataset that meets engineering feasibility requirements.
[0059] The generator and discriminator of the Hybrid Constrained Generative Adversarial Network (HCC-GAN) constructed in this embodiment of the invention both adopt a multilayer perceptron (MLP) structure, which contains three hidden layers, with the number of neurons in each layer being 128, 128, and 64 respectively, and the activation function is LeakyReLU.
[0060] Step 3: As Figure 3 As shown, a machine learning surrogate model for predicting material properties is constructed using an augmented dataset. Through cross-validation and prediction accuracy evaluation, the machine learning surrogate model with the best prediction performance is selected to characterize the mapping relationship between preparation parameters and material properties, expressed as:
[0061]
[0062] Where f(·) is the proxy model, These are the predicted values for tensile strength and elongation at break, respectively.
[0063] The machine learning proxy model for predicting material properties in this invention can employ, but is not limited to, multilayer perceptron (MLP), support vector regression (SVR), extreme gradient boosting machine (XGBoost), and lightweight gradient boosting machine (LightGBM). The hyperparameter optimization method uses three-fold cross-validation to determine the coefficients ( The mean absolute error (MAE), root mean square error (RMSE), and normalized root mean square error (NRMSE) were used as evaluation metrics, and the model with the best generalization performance was selected as the surrogate model for subsequent multi-objective optimization.
[0064] For example, when using a multilayer perceptron model as a machine learning proxy model for predicting material properties, its forward propagation process is represented as follows:
[0065]
[0066] in, This represents the output of the Lth layer neuron. and These are the weight matrix and the bias vector, respectively. (·) is a non-linear activation function.
[0067] A local sensitivity index of the output to input perturbation is introduced as a robustness constraint to obtain a Pareto optimal formulation set that balances performance and stability.
[0068] For example, when using a support vector regression model as a machine learning proxy model for predicting material properties, its regression function is:
[0069]
[0070] Its optimization objective function is:
[0071]
[0072] The objective function satisfies the following constraints:
[0073]
[0074] Where C is the penalty coefficient. , For parameters insensitive to intervals, For feature mapping. Step 4: Using the model with the best generalization performance as the performance discriminator, construct a multi-objective optimization model with the objectives of maximizing tensile strength and maximizing elongation at break, expressed as:
[0075]
[0076] in, The physically feasible region is determined by prior technological knowledge. and The predicted value is the value predicted by the surrogate model.
[0077] The above multi-objective optimization problem is solved using the non-dominated sorting genetic algorithm (NSGA-II) within the physical boundary of the preset preparation parameters to obtain the Pareto optimal formulation set.
[0078] To improve algorithm performance, Bayesian optimization is used to adaptively adjust the hyperparameters of the Non-Dominated Sorting Genetic Algorithm (NSGA-II). These hyperparameters include population size (pop_size), number of generations (n_gen), simulated binary crossover probability (sbx_prob), and crossover distribution index (…). ) and the variation distribution index ( );
[0079] Table 1. Effects of each hyperparameter
[0080]
[0081] In the Bayesian optimization process, the hypervolume index is used as the objective function. This hypervolume index measures the volume of the target space covered by the Pareto front relative to the reference point, and its expression is:
[0082]
[0083] Where P is the Pareto solution set. Let r be the value of the i-th objective function. i Let λ(·) be the reference point, and let λ(·) denote the measure function. By maximizing the hypervolume index, the hyperparameters of NSGA-II are optimized, thereby simultaneously improving the convergence and distribution of the Pareto front.
[0084] Step 5: Select representative combinations of preparation parameters from the Pareto optimal formulation set, conduct experimental preparation according to the established material preparation process, and test its tensile strength and elongation at break. Compare the experimental test results with the prediction results of the surrogate model to verify the feasibility and reliability of the method of the present invention under actual engineering conditions.
[0085] Example 2:
[0086] This invention proposes a method for preparing epoxy resin materials based on generative adversarial networks and multi-objective optimization methods, comprising the following steps:
[0087] Step 1: By changing the combination of preparation parameters, 35 sets of samples were obtained. Each set was subjected to 6-10 parallel experiments, and the fluctuation of parallel samples was controlled within 10%. After removing experimental losses, 250 valid raw data points were obtained. Z-score standardization was performed on the input and output variables. When the generated data and model prediction results needed to be returned to the original scale, inverse standardization was used.
[0088] Define the physical boundary feasible region See Table 2.
[0089] Table 2 Physical boundary range of each feature
[0090]
[0091] Step 2: Construct a Hybrid Constrained Generative Adversarial Network (HCC-GAN), generator Input is The output is Discriminator Input is The output is a scalar score. The discriminator objective function is defined in the WGAN-GP form, and the generator objective function is "adversarial term + L1 supervision term".
[0092] Both the generator and discriminator employ a multilayer perceptron (MLP) architecture, each containing three hidden layers with 128, 128, and 64 neurons respectively. The LeakyReLU activation function is used. Training settings: batch size = 16, Adam optimizer, learning rate 1×10⁻⁶. -4 With β1=0.5 and β2=0.9, the discriminator to generator update ratio is 5:1; two augmentation networks are trained for TS and EB respectively, with each network trained for 4000 steps.
[0093] To expand the coverage of the input space, a convex combination interpolation mechanism based on Dirichlet distribution is adopted: randomly selecting... Standardized input samples Generate weight vector ,structure
[0094]
[0095] After mapping back to the original scale, only those satisfying the physical boundaries are retained. The candidate inputs. Qualified... The input generator gets And obtain by destandardization .
[0096] Extreme value filtering is applied to the generated results: samples exceeding ±20% of the experimental range are removed, i.e., only those samples are retained. The samples were then processed. Ultimately, 1000 high-quality synthetic samples were obtained, forming an augmented dataset.
[0097] Step 3: Train four proxy models (MLP, SVR, XGBoost, and LightGBM) based on the augmented dataset. Use three-fold cross-validation to determine the model hyperparameters and use R... 2Generalization ability is evaluated using metrics such as MAE, RMSE, and NRMSE. Let's take a set of feasible hyperparameter configurations as an example:
[0098] Table 3 Hyperparameter settings for candidate agent models
[0099]
[0100] Based on the comprehensive evaluation results across the cross-domain, MLP was selected as the surrogate model for subsequent multi-objective optimization.
[0101] Step 4: Construct a dual-objective maximization problem:
[0102]
[0103] The NSGA-II algorithm was used for the solution. To improve the performance of NSGA-II in continuous search spaces, Bayesian optimization was employed to optimize key hyperparameters, including population size `pop_size`, number of iterations `n_gen`, simulated binary crossover probability `sbx_prob`, and crossover exponent. and the variation distribution index The optimized set of parameters is: pop_size=783, n_gen=361, sbx_prob=0.6385. , We obtain the Pareto front and the corresponding Pareto solution set.
[0104] To demonstrate the impact of different surrogate models on the Pareto front, multi-objective optimization problems were solved using MLP, SVR, LightGBM, and XGBoost as surrogate models, and the obtained Pareto fronts were compared with empirical Pareto fronts. The results are as follows: Figure 5 .
[0105] Step 5: Select the Top-50 candidate points from the Pareto solution set, and use a combination strategy of geometric mean scoring and stratified sampling to screen representative formulations and process conditions. In the geometric mean scoring, normalized TS and EB are comprehensively considered, and a local sensitivity penalty term is introduced with a penalty coefficient of 0.3 to improve the robustness of candidate solutions.
[0106] Representative candidate points were selected from the high ductility region (HR), high strength region (HS), and comprehensive region (Com) for experimental reproduction. Examples of candidate parameters are as follows:
[0107] Table 4. Representative preparation parameter combinations in the Pareto optimal solution set
[0108]
[0109] Samples were prepared for the above candidate sites, and TS and EB were tested. The experimental results were compared with the prediction results of the surrogate model. The experimental verification results are as follows: Figure 6 As shown, the results indicate that the experimental points generally fall near or close to the Pareto front obtained by extrapolation from the surrogate model, and the relative deviations of TS and EB are controlled within 15% and 17%, respectively, thus verifying the feasibility and reliability of the proposed method for optimizing preparation parameters.
Claims
1. A method for preparing epoxy resin materials based on generative adversarial networks and multi-objective optimization, characterized in that: include: Step 1: Obtain and preprocess the performance data of epoxy resin materials under different preparation parameters to obtain the original experimental samples; Based on prior knowledge of materials and processes, a feasible domain for preparation parameters is formed, wherein the preparation parameters include process parameters and formulation; the performance data are obtained from multiple mutually constraining performance indicators. Step 2: Construct a Generative Adversarial Network (GAN). The generator in the GAN takes preprocessed preparation parameters and random noise as input and outputs a performance prediction value. The discriminator in the GAN takes preparation parameters and performance data as input and outputs a score value. The GAN is trained using an adversarial loss function. Step 3: Extend the preparation parameters of the original experimental samples by convex combination interpolation to construct new preparation parameters. Perform physical boundary screening on the new preparation parameters according to the feasible region of the preparation parameters. Input the screened preparation parameters into the trained generator to obtain the corresponding performance data, thereby constructing the augmented dataset. Step 4: Train multiple pre-built learning agent models using the augmented dataset, and select the learning agent model with the best prediction performance through cross-validation and prediction accuracy evaluation. Step 5: Based on the machine learning agent model with the best prediction performance, construct a multi-objective optimization problem with multiple mutually constraining performance index values as optimization objectives. Within the feasible region of the preparation parameters, solve the multi-objective optimization problem to obtain the optimal preparation parameters that satisfy the optimization objectives. Step 6: Based on the optimal preparation parameters, prepare the epoxy resin material.
2. The method for preparing epoxy resin material based on generative adversarial networks and multi-objective optimization methods according to claim 1, characterized in that: The preparation parameters include: curing agent ratio, toughening agent dosage, multi-stage curing temperature and curing time; the multiple mutually restrictive performance indicators include at least tensile strength and elongation at break. 3.The method of claim 1, wherein the method comprises: The loss function of the discriminator includes the difference between the real sample score and the generated sample score, as well as a gradient penalty term, to satisfy the 1-Lipschitz condition. The loss function of the generator includes adversarial loss and L1 supervised loss. 4.The method of claim 1, wherein the method is characterized by: The process of extending the preparation parameters of the original experimental samples through convex combination interpolation to construct new preparation parameters includes: Randomly select k original experimental sample preparation parameters, generate a weight vector from a Dirichlet distribution , construct a new preparation parameter according to the following formula: 。 5. The method for preparing epoxy resin material based on generative adversarial network and multi-objective optimization method according to claim 1, characterized in that: The generator and discriminator of the generative adversarial network both adopt a multilayer perceptron structure, each containing three hidden layers, with 128, 128 and 64 neurons in each layer respectively, and LeakyReLU is used as the activation function.
6. The method for preparing epoxy resin material based on generative adversarial network and multi-objective optimization method according to claim 1, characterized in that: The learning agent model employs one or more of the following: multilayer perceptron, support vector regression, extreme gradient boosting machine, and lightweight gradient boosting machine.
7. The method for preparing epoxy resin material based on generative adversarial network and multi-objective optimization method according to claim 1, characterized in that: Solving the multi-objective optimization problem within the feasible region of the preparation parameters to obtain the optimal preparation parameters that satisfy the optimization objectives includes: Within the feasible region of the preparation parameters, the multi-objective optimization problem is solved using a non-dominated sorting genetic algorithm to obtain the optimal preparation parameters that satisfy the optimization objectives. Among them, the hyperparameters of the non-dominated sorting genetic algorithm are adaptively adjusted by the Bayesian optimization method. The optimized hyperparameters include population size, number of iterations, simulated binary crossover probability, crossover distribution index, and mutation distribution index. The Bayesian optimization method uses expectation boosting as the acquisition function and hypervolume index as the optimization objective to optimize the hyperparameters of the non-dominated sorting genetic algorithm.