Regenerated composite material reverse design method and system based on generative adversarial network

Abstract

The invention relates to the technical field of regenerative composite material reverse design, in particular to a regenerative composite material reverse design method and system based on a generative adversarial network. The method comprises the following steps: firstly, constructing a database containing various regenerated composite material formulas and corresponding actually measured electrical properties thereof; then, a conditional generative adversarial network cGAN is trained based on the database, in training, the target electrical performance serves as condition input, and the virtual microstructure features meeting the condition and the raw material ratio serve as a generation target; and inputting the electrical performance requirement set for the target electrical component into the trained cGAN network, and reversely generating a raw material ratio suggestion meeting the performance requirement by the network. According to the method, direct and rapid reverse mapping from a complex performance target to a multi-component formula is realized, and through'design-verification-feedback 'closed-loop iterative optimization, the research and development efficiency and success rate of the regenerated composite material for specific electrical application are greatly improved, and the trial and error cost is remarkably reduced.

Description

Technical Field

[0001] This invention relates to the field of reverse engineering technology for recycled composite materials, and specifically to a method and system for reverse engineering of recycled composite materials based on generative adversarial networks. Background Technology

[0002] With the increasing global emphasis on sustainable development and the circular economy, the high-value utilization of recycled plastics has become a key issue for the plastics industry. Blending various recycled plastics with inorganic fillers and reinforcing fibers to prepare recycled composite materials is an important way to increase the added value of recycled materials and expand their application areas, especially in the electrical and electronic fields, where they can be used to produce various components such as grounding parts and insulating supports. However, the development of such recycled composite material systems faces severe challenges. The raw material sources are complex, and the composition, aging degree, and performance baseline of different batches of recycled materials vary. Furthermore, the interfacial interactions between recycled materials and fillers / fibers are complex, resulting in a highly nonlinear and non-monotonic relationship between the key electrical properties of the final product, such as conductivity and dielectric strength, and the formulation composition. This makes accurate prediction and design difficult using traditional physical models or simple empirical formulas.

[0003] Currently, the development of recycled composite material formulations for specific electrical performance requirements heavily relies on the experience of researchers and extensive trial-and-error experiments. Researchers must initially formulate a formula based on limited experience, and then go through lengthy processes such as ingredient mixing, blending, granulation, sample preparation, and performance testing before obtaining the performance data. If the results are unsatisfactory, the formula must be adjusted and the process repeated. This method is not only time-consuming and costly in terms of manpower and materials, but also, due to the vast design space, often limits the search to a localized area, making it difficult to find globally optimal or innovative formulation solutions. This severely restricts the rapid, targeted development and application of high-performance recycled composite materials.

[0004] In recent years, data-driven methods such as machine learning have shown potential in materials research and development, but existing research has largely focused on establishing "forward" predictive models from composition to performance. However, in actual research and development, a more pressing need is to derive feasible "composition and processes" from specified "performance targets," i.e., reverse design. Traditional inverse problem solving typically requires complex optimization algorithms to iteratively search within a vast design space, resulting in high computational costs and heavy reliance on the accuracy of the forward predictive model. How to efficiently and accurately achieve the inverse mapping from complex electrical performance targets to multi-component recycled composite material formulations remains an unresolved technical challenge in this field.

[0005] Therefore, existing technologies still need further development. Summary of the Invention

[0006] The purpose of this invention is to overcome the above-mentioned technical deficiencies and provide a reverse design method and system for regenerative composite materials based on generative adversarial networks, so as to solve the problems existing in the prior art.

[0007] To achieve the above-mentioned technical objectives, according to a first aspect of the present invention, the present invention provides a reverse design method for regenerative composite materials based on generative adversarial networks, comprising: S1. Obtain known formulations of various recycled composite materials and their corresponding measured electrical performance data, and construct a formulation-performance sample database; S2. Based on the formula-performance sample database, construct and train a conditional generative adversarial network (cGAN), wherein the target electrical performance is used as the conditional input, and the virtual microstructure characteristics and raw material ratio of the recycled composite material that meets the target electrical performance are used as the generation target. S3. Input the target electrical performance set for the target electrical component into the trained conditional generative adversarial network cGAN. S4. Using the trained Conditional Generative Adversarial Network (cGAN), reverse-generate virtual microstructure features and corresponding raw material ratio suggestions that match the target electrical performance.

[0008] Specifically, the generator network of the conditional generative adversarial network (cGAN) is used to receive the noise vector and the conditional vector of the target electrical performance, and output the virtual microstructure feature vector and the raw material ratio vector.

[0009] Specifically, the discriminator network of the conditional generative adversarial network (cGAN) is used to receive real or generated virtual microstructure feature vectors and raw material ratio vectors, and combine them with the conditional vector of the target electrical performance to determine the authenticity of the virtual microstructure feature vectors and raw material ratio vectors.

[0010] Specifically, in step S2, the loss function for training the conditional generative adversarial network (cGAN) includes at least an adversarial loss to measure the difference between the generated data and the real data distribution, and a conditional matching loss to constrain the generated data to meet the target electrical performance.

[0011] Specifically, the conditional matching loss is calculated by inputting the virtual microstructure feature vector and raw material ratio vector generated by the generator into a pre-trained performance prediction model, and comparing the predicted electrical performance of the model with the target electrical performance.

[0012] Specifically, the performance prediction model is a neural network model, which is trained through the formula-performance sample database and is used to predict electrical performance based on the input raw material ratio.

[0013] Specifically, after step S4, the process further includes: S5, preparing a recycled composite material sample based on the raw material ratio recommendation and testing its actual electrical properties; S6, adding the raw material ratio recommendation and its corresponding actual electrical performance data as new samples to the formula-performance sample database to iteratively optimize the conditional generative adversarial network cGAN.

[0014] Specifically, the target electrical properties include at least one of conductivity, resistivity, dielectric constant, and dielectric strength.

[0015] Specifically, the raw materials of the recycled composite material include various recycled plastics, inorganic fillers, and reinforcing fibers.

[0016] According to a second aspect of the present invention, a reverse design system for regenerative composite materials based on generative adversarial networks is provided, comprising: The database module is used to store and manage the recipe-performance sample database; The network training module is used to construct and train a conditional generative adversarial network (cGAN) based on the recipe-performance sample database. The input module is used to receive the target electrical performance settings for the target electrical component; The reverse generation module is used to input the target electrical properties into the trained conditional generative adversarial network (cGAN) and obtain the virtual microstructure features generated by the reverse generation and the corresponding raw material ratio suggestions.

[0017] Beneficial effects: The reverse design method and system for regenerative composite materials based on generative adversarial networks provided by this invention have the following significant advantages compared to existing technologies: First, this invention innovatively applies Conditional Generative Adversarial Networks (cGANs) to the reverse design problem of recycled composite materials, constructing an end-to-end generative model of "target performance → virtual structure → raw material ratio". This method fundamentally changes the traditional R&D model that relies on trial and error. By using the performance target as a conditional input to a pre-trained cGAN model, a series of virtual microstructural features and raw material ratio suggestions that meet the conditions can be directly and quickly generated. This achieves a direct and efficient mapping from the performance space to the formulation space, transforming reverse design from a complex iterative optimization problem into an efficient forward generative problem, greatly shortening the initial formulation design cycle and reducing the blind spots and initial costs of R&D.

[0018] Second, this invention ensures the accuracy and reliability of the reverse design results by designing a composite loss function that includes conditional matching loss and introducing a pre-trained high-precision performance prediction model as a constraint. The adversarial loss of the generative adversarial network ensures the rationality of the generated recipe in the historical data distribution; while the conditional matching loss, through the performance prediction model, forces the predicted performance of the generated recipe to closely approximate the target performance set by the user. This dual constraint mechanism makes the generated recipe not only "like" a real and feasible recipe, but also "accurately" points to the preset performance indicators, thereby significantly improving the success rate and practical value of the reverse design results.

[0019] Third, the method proposed in this invention possesses excellent capabilities in handling multi-objective and nonlinear complex constraints. Complex trade-offs often exist between multiple electrical performance indicators of recycled composite materials (such as high resistivity and high dielectric strength). The cGAN model of this invention can automatically learn these complex, nonlinear relationships and constraints from training data. When a user inputs a multi-dimensional performance objective vector, the model can generate a formulation scheme that coordinates and balances multiple performance constraints, solving the pain point that traditional single-objective optimization methods or simple weighted methods struggle to handle complex trade-offs.

[0020] Fourth, this invention introduces a closed-loop iterative optimization process of "computational design - experimental verification - data feedback," enabling the system to continuously learn and self-evolve. The initial model is trained based on limited historical data. After its suggestions are experimentally verified, new "recipe-performance" data pairs are fed back to the database for periodic retraining of the model. This process continuously expands the system's knowledge base, and its prediction and design capabilities continuously improve with increased usage. It can gradually explore a broader design space and even discover novel, high-performance recipes that surpass existing experience, achieving an intelligent and automated upgrade of the R&D process.

[0021] Fifth, this invention integrates the above methods into a complete system, covering the entire process of data management, model training, human-computer interaction, and result generation. This system encapsulates complex machine learning algorithms into easy-to-use tools, enabling materials engineers to operate without in-depth knowledge of artificial intelligence details. This significantly lowers the barrier to entry for advanced technologies, facilitating their rapid promotion and application in industry and promoting the transformation of the recycled composite materials industry towards digital and intelligent R&D. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the reverse design method for regenerative composite materials based on generative adversarial networks provided in a specific embodiment of the present invention. Figure 2 This is a schematic diagram of the system composition of the reverse design system for regenerative composite materials based on generative adversarial networks provided in a specific embodiment of the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments in this application, other similar embodiments obtained by those skilled in the art without creative effort should all fall within the scope of protection of this application. Furthermore, directional terms mentioned in the following embodiments, such as "up," "down," "left," and "right," are only for reference to the directions in the accompanying drawings; therefore, the directional terms used are for illustrative purposes and not for limiting the invention.

[0024] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments.

[0025] Please see Figure 1 This invention provides a reverse design method for regenerative composite materials based on generative adversarial networks, comprising: S1. Obtain known formulations of various recycled composite materials and their corresponding measured electrical performance data, and construct a formulation-performance sample database; It should be further noted that step S1 is the foundation of the data, requiring the systematic collection of experimentally validated recycled composite material sample data. Each sample data point must include complete raw material mass ratios (the sum of the percentages of each component is 100%) and at least one electrical property data measured according to national standards (such as GB / T1410-2006 for volume resistivity measurement and GB / T1409-2006 for dielectric constant measurement). The database is preferably stored in a structured format (such as CSV or SQL tables), with fields including at least: formulation identifier, recycled PP content (wt%), recycled ABS content (wt%), CaCO3 filler content (wt%), chopped glass fiber content (wt%), other additive content (wt%), volume resistivity (Ω·cm), dielectric constant (at 1kHz), and dielectric strength (kV / mm). The initial database sample size should be no less than 300 sets to ensure statistical significance for subsequent model training. The samples should cover as wide a formulation space and performance range as possible.

[0026] S2. Based on the formula-performance sample database, construct and train a conditional generative adversarial network (cGAN), wherein the target electrical performance is used as the conditional input, and the virtual microstructure characteristics and raw material ratio of the recycled composite material that meets the target electrical performance are used as the generation target. It is worth further elaboration that step S2 is the core of model construction. cGAN consists of a generator G and a discriminator D, and its unique feature is that it incorporates the "target electrical properties" as conditional information into the generation and discrimination processes. The generator's task is to learn a mapping G(z,c), (M,F) from random noise z and the target condition c to the "virtual material design" (including microstructural features M and raw material ratio F). Here, the virtual microstructural features M are a multi-dimensional vector used to quantify the microstate of the material's interior, such as: continuous phase matrix type (one-hot encoded), average particle size (normalized value) of the dispersed phase (filler / glass fiber), aspect ratio of the dispersed phase, volume fraction of the dispersed phase, and approximate aggregation index of the dispersed phase in the matrix (a value between 0 and 1, where 0 represents complete uniform dispersion and 1 represents severe aggregation). Although these features cannot be directly obtained from the formulation, they can be estimated through correlation analysis of historical data or empirically simplified models, and paired with formulation data as part of the supervision information to guide cGAN in learning the more fundamental "structure-property" relationship. The raw material ratio F is a K-dimensional vector (K is the number of raw material types), for example F=[w1,w2,…,wK], representing the mass percentage of each raw material. The sum of these ratios is guaranteed to be 1 through the Softmax output layer.

[0027] S3. Input the target electrical performance set for the target electrical component into the trained conditional generative adversarial network cGAN. It should be further explained that step S3 is the application interface. The target electrical performance c is set by the user according to the design requirements of the specific electrical component, and can be a specific value (such as dielectric constant ε=3.5) or a range (such as volume resistivity ρ>1×10^12Ω·cm). The system performs the same normalization processing on c as the training data to form a conditional vector.

[0028] S4. Using the trained Conditional Generative Adversarial Network (cGAN), reverse-generate virtual microstructure features and corresponding raw material ratio suggestions that match the target electrical performance.

[0029] It should be further explained that step S4 is the reverse design output. The normalized conditional vector c and the noise vector z (dimension 100) sampled from the standard normal distribution are input into the trained generator G. The generator outputs a candidate design scheme (M', F'). Due to the randomness of the noise z, multiple different candidate schemes can be generated each time, providing users with diverse choices. After inverse normalization, the raw material ratio F' yields a specific raw material input ratio suggestion that can be experimentally verified.

[0030] Understandably, this method transforms the inverse design problem of complex multiphase recycled composite material systems, which are difficult to model precisely, into a data-driven, controllable generative modeling problem. Leveraging the powerful distributed learning and generative capabilities of cGAN, it directly establishes an inverse mapping from the performance space to the formulation and structural space, skipping the tedious forward modeling and optimization iteration process. This enables rapid end-to-end generation of "performance requirements and design schemes," significantly accelerating the development process of recycled composite materials with specific electrical performance requirements and reducing R&D costs and resource consumption.

[0031] Specifically, the generator network of the conditional generative adversarial network (cGAN) is used to receive the noise vector and the conditional vector of the target electrical performance, and output the virtual microstructure feature vector and the raw material ratio vector.

[0032] It should be further explained that the generator G is a deep neural network, and its specific architecture and data flow are described below. The input consists of two concatenated parts: the first part is a noise vector z, whose dimension is preferably 100. Choosing 100 dimensions is to strike a balance between ensuring generation diversity and avoiding the curse of dimensionality; too low a dimension (e.g., 10) limits the diversity of generated samples, while too high a dimension (e.g., 500) increases training difficulty and may lead to overfitting. z is sampled from a multidimensional normal distribution N(0,I) with a mean of 0 and a standard deviation of 1, i.e. The second part is the condition vector c, whose dimension L equals the number of types of target electrical properties. For example, if both volume resistivity (ρ) and dielectric constant (ε) are optimized simultaneously, then c is a 2-dimensional vector that needs to be normalized before input. , where μ and σ are the mean and standard deviation of the logarithmic or raw values ​​of the corresponding performance parameters in the training set, respectively.

[0033] Furthermore, the preferred network structure of generator G is as follows: ① Input layer: receiving Concatenate vectors with a dimension of 100+L.

[0034] ② Fully connected layer 1 (Dense1): 256 neurons, using the Leaky ReLU activation function (negative slope α=0.2). This layer is used for the initial fusion of noise and conditional information.

[0035] ③ Fully connected layer 2 (Dense2): 512 neurons, using the Leaky ReLU activation function (α=0.2). This layer is used to learn higher-level feature representations.

[0036] ④ Fully connected layer 3 (Dense3): 256 neurons, using the Leaky ReLU activation function (α=0.2).

[0037] ⑤ Output layer: Consists of two parallel sub-output layers: (1) Sub-output layer_M (microstructure): The number of neurons is (Number of microstructural features, for example, 8), using the Tanh activation function, the output value is in the range [-1, 1], corresponding to the normalized value of each microstructural feature.

[0038] (2) Sub-output layer _F (raw material ratio): The number of neurons is K (the number of raw material groups). The Softmax activation function is used to output the mass percentage of each raw material, which automatically satisfies the constraint that the sum is 1.

[0039] Understandably, this generator structure, through multiple fully connected layers and nonlinear activation functions, gradually fuses random noise with deterministic conditional transformations, decoupling and ultimately mapping them to microstructural features and raw material proportions with clear physical meaning. The separate output layer design enables the network to simultaneously learn and generate two different but related types of outputs, providing more comprehensive information for the final material design.

[0040] Specifically, the discriminator network of the conditional generative adversarial network (cGAN) is used to receive real or generated virtual microstructure feature vectors and raw material ratio vectors, and combine them with the conditional vector of the target electrical performance to determine the authenticity of the virtual microstructure feature vectors and raw material ratio vectors.

[0041] It should be further explained that the discriminator D is a binary classification neural network used to determine whether the input triple "(microstructure, raw material ratio, performance conditions)" comes from the real dataset or the generator. Its input is real data. Or generate data To ensure the effective use of conditional information, the input is processed as follows: First, the microstructure vector M and the raw material ratio vector F are concatenated to obtain the "design vector". Its dimensions are Then, this design vector x and the condition vector c are concatenated along the channel dimension to form the total input of the discriminator. .

[0042] Furthermore, the preferred network structure of the discriminator D is as follows: ① Input layer: Receives the concatenated vector , dimension .

[0043] ② Fully connected layer 1 (Dense1): The number of neurons is 512, using the Leaky ReLU activation function (α=0.2), followed by a Dropout layer with a dropout rate of 0.3 to prevent overfitting.

[0044] ③ Fully connected layer 2 (Dense2): The number of neurons is 256, using the Leaky ReLU activation function (α=0.2), followed by a Dropout layer (dropout rate 0.3).

[0045] ④ Fully connected layer 3 (Dense3): 128 neurons, using the Leaky ReLU activation function (α=0.2).

[0046] ⑤ Output layer: One neuron, using the Sigmoid activation function, outputs a scalar between 0 and 1, representing the probability that the input data is "real".

[0047] Understandably, the discriminator's training objective is to maximize its ability to distinguish between real and fake data. This forces the generator to produce samples that are not only reasonably distributed in the design space (M,F) but also highly consistent with the specified performance condition c. This "conditional discrimination" mechanism is key to the directional and controllable inverse design capabilities of the cGAN in this invention, ensuring that the generation process is strictly constrained to the user-specified performance objectives.

[0048] Specifically, in step S2, the loss function for training the conditional generative adversarial network (cGAN) includes at least an adversarial loss to measure the difference between the generated data and the real data distribution, and a conditional matching loss to constrain the generated data to meet the target electrical performance.

[0049] It should be further explained that cGAN training is accomplished by minimizing a composite loss function, which is the discriminator loss. and generator loss The alternating optimization process. Discriminator loss. The purpose of using binary cross-entropy is to correctly distinguish between real data and generated data. in, For small batch sample sizes, 32 or 64 is preferred. and The first sampled from the real dataset A true design vector and its corresponding true performance conditions. It is the sampled noise vector. It is the design vector generated by the generator under the current parameters. This represents the true probability output by the discriminator. The goal of the discriminator is to maximize... That is to let Approaching 1, Close to 0.

[0050] Furthermore, generator loss It is the loss that it hopes to minimize, which is a weighted sum of two parts: in, It is designed to counteract losses and deceive the discriminator: The goal of the generator is to make the discriminator believe that the data it generates is real, that is, to make Approaching 1, thus minimizing .

[0051] Furthermore, This is the conditional matching loss, used to force the generated data to meet the target performance conditions. It is preferably calculated using mean squared error (MSE). here, This represents the raw material ratio vector F output by the generator. It is a pre-trained performance prediction model that predicts electrical performance based on formula F. This represents the square of the L2 norm. This loss term directly measures the difference between the predicted performance of the generated formulation and the target performance.

[0052] Furthermore, This is a hyperparameter that balances the two losses; the optimal value is 5. (Select...) The reason is: initial experiments showed that if If the value is too small (e.g., 0.1), the constraint is insufficient, and the resulting formulation may deviate from the target performance; if... If the value is too large (e.g., 50), the adversarial loss effect is suppressed, the diversity of generated samples decreases, and training may become unstable. It can maintain a certain degree of generation diversity while ensuring performance accuracy, achieving a good balance in practice.

[0053] Furthermore, the entire training process employs an alternating update strategy: first, the generator G is fixed, and then the parameters of the discriminator D are updated to minimize... (Actually calculate gradient ascent); then fix the discriminator D and update the parameters of the generator G to minimize Using the Adam optimizer, with an initial learning rate of 0.0002, the exponential decay rate estimated by the first moment is... The exponential decay rate estimated by the second moment Small batch size The training cycle (epoch) usually requires 20,000-50,000 times until the loss function converges and the quality of the generated samples is stable.

[0054] Specifically, the conditional matching loss is calculated by inputting the virtual microstructure feature vector and raw material ratio vector generated by the generator into a pre-trained performance prediction model, and comparing the predicted electrical performance of the model with the target electrical performance.

[0055] It should be further noted that the present invention specifically preferably uses conditional matching loss. The computational path. The performance prediction model P is an independent, differentiable function approximator, which is trained and its parameters are frozen before the start of cGAN training. In computation... The specific steps are as follows: 1. Forward propagation: For a mini-batch of noise and conditions The corresponding raw material ratio vector is obtained through generator G. ,in .

[0056] 2. Performance Prediction: This batch of generated allocation vectors... Input the performance prediction model P to obtain the predicted performance vector. ,Right now .

[0057] 3. Loss Calculation: Calculate the target conditions With predictive performance The mean squared error (MSE) between the two sides is used as the conditional matching loss: in, If it's a dimension of performance conditions (such as resistivity and dielectric constant, then L=2), and They are the first The first sample The true target value (after normalization) and the model prediction value for each performance condition.

[0058] Understandably, through this differentiable model P, the gradient of the performance constraint can be derived from the loss function. The backpropagation proceeds smoothly to the generator G, guiding it to adjust its parameters so that the generated recipe F, after being predicted by P, exhibits good performance. The goal is to get as close as possible to the target c. This design cleverly transforms a complex physical constraint into a differentiable mathematical constraint, enabling gradient-based optimization algorithms to be directly applied to inverse design problems. This is the key to the efficiency and accuracy of this method.

[0059] Specifically, the performance prediction model is a neural network model, which is trained through the formula-performance sample database and is used to predict electrical performance based on the input raw material ratio.

[0060] It should be further noted that the performance prediction model P preferably adopts a multilayer perceptron (MLP) structure, and its specific construction and training process is as follows: ① Network structure: ① Input layer: The number of neurons is K (number of raw material types), for example, K=5 (recycled PP, recycled ABS, CaCO3, glass fiber, compatibilizer). The input vector is the normalized percentage of raw material mass.

[0061] ②Hidden layer 1: The number of neurons is 256, using the ReLU activation function, followed by a Batch Normalization layer and a Dropout layer (dropout rate 0.2).

[0062] ③Hidden layer 2: 128 neurons, using the ReLU activation function, followed by a Batch Normalization layer and a Dropout layer (dropout rate 0.2).

[0063] ④ Hidden layer 3: 64 neurons, using the ReLU activation function.

[0064] ⑤ Output layer: The number of neurons is L (the number of electrical properties to be predicted, for example, L=2, corresponding to volume resistivity and dielectric constant respectively), and a linear activation function is used.

[0065] The rationale for choosing 3 hidden layers and the aforementioned number of neurons is that, for moderately nonlinear problems such as material formulation-performance, a 3-layer network has sufficient expressive power to capture complex relationships. The first layer of 256 neurons is used to capture high-dimensional features, followed by 128 and 64 neurons to progressively refine and compress information, preventing overfitting. Batch Normalization accelerates training and improves stability, while Dropout is used for regularization.

[0066] ② Data preprocessing: For the input formula F, perform min-max normalization to the [0,1] interval. For the output performance y, since resistivity usually spans multiple orders of magnitude, it is necessary to first take the logarithm to base 10. Then perform z-score standardization (subtract the mean and divide by the standard deviation).

[0067] ③ Training configuration: (1) Loss function: mean squared error (MSE); (2) Optimizer: Adam, learning rate lr=0.001, , ; (3) Batch size: 32; (4) Training cycles: 500. Evaluation is performed on the validation set every 50 cycles. If the validation loss does not decrease for 20 consecutive cycles, training is stopped early. (5) Data set partitioning: The recipe-performance database is randomly divided into training set, validation set and test set according to 70%:15%:15%.

[0068] Understandably, after training, the prediction accuracy of model P is evaluated on the test set, for example, by requiring... A score greater than 0.85 and a mean absolute percentage error (MAPE) less than 15% are required to ensure that the predictive reliability is sufficient to guide cGAN training. Only a high-precision predictive model P can ensure the effectiveness of the conditional matching loss. The guidance is accurate, which is the foundation of the effectiveness of the entire reverse engineering system.

[0069] Specifically, after step S4, the process further includes: S5, preparing a recycled composite material sample based on the raw material ratio recommendation and testing its actual electrical properties; S6, adding the raw material ratio recommendation and its corresponding actual electrical performance data as new samples to the formula-performance sample database to iteratively optimize the conditional generative adversarial network cGAN.

[0070] It should be further noted that this invention provides a closed-loop optimization process for achieving model self-evolution and performance improvement. In step S4, one or more candidate formulation suggestions are obtained. Following this, the S5 experimental verification stage begins. One to three formulations with the most predictable performance close to the target or showing the greatest potential are selected for experimentation. Each component (e.g., recycled PP granules, ABS crushed material, CaCO3 powder, chopped glass fiber, coupling agent, etc.) is precisely weighed according to the specified ratio, melt-blended and granulated in a twin-screw extruder, and then made into standard test specimens (e.g., dumbbell-shaped specimens as specified in GB / T 1040) on an injection molding machine. The actual electrical properties of the specimens are tested strictly according to relevant national standards (e.g., GB / T1410) to obtain measured performance data. Furthermore, step S6 is the knowledge accumulation and model update stage. The "recipe" obtained in this experiment will be used... - Actual performance "Data pairs are added as new samples to the existing recipe-performance sample database. When the number of new samples accumulates to a certain level (e.g., 50 sets) or the model performance shows a significant deviation, the model retraining process is triggered. Using an updated and larger database, the performance prediction model P and cGAN are retrained. Retraining model P allows it to predict a wider range of recipes and make more accurate predictions. Retraining cGAN enables it to learn the design rules of newly explored regions, thereby generating more reliable and accurate recipe suggestions in subsequent reverse design tasks."

[0071] Understandably, this closed-loop process tightly integrates data-driven computational design with physical experiments, forming a complete R&D iteration loop of "computational design, experimental verification, data feedback, and model updates." Each iteration effectively expands the boundaries of the database and the cognitive scope of the model, enabling the system to continuously learn from practice, gradually reducing its dependence on the initial data quantity and quality, and potentially discovering new formulas with better performance. This gives the system the ability to continuously learn and self-improve, greatly enhancing its practical value and long-term benefits.

[0072] Specifically, the target electrical properties include at least one of conductivity, resistivity, dielectric constant, and dielectric strength.

[0073] It should be further noted that the present invention preferably targets specific electrical performance indicators for its method, which are key parameters for evaluating the use of recycled composite materials in electrical and electronic components. In practical applications, one or more performance indicators can be selected as design objectives based on the component's function, specifically including: ① Resistivity / Conductivity: Used to distinguish between insulating materials, antistatic materials, and conductive materials. For example, grounding components require a volume resistivity lower than [a certain value]. The requirements for insulating supports are higher than those for insulating supports. When used as a conditional input, the logarithm of resistivity is typically used. To smooth its range across multiple orders of magnitude.

[0074] ② Dielectric constant ( This affects the capacitance characteristics of components. In high-frequency applications or capacitors, precise control is required. Typically, measured values ​​are used directly as input conditions.

[0075] ③ Dielectric strength ( (Unit: kV / mm): Characterizes the insulation withstand voltage of a material and is a key indicator of insulating materials. Use measured values ​​directly.

[0076] It is understandable that when constructing the condition vector c, if multiple performance targets are set simultaneously, such as "volume resistivity",... and dielectric strength If c is a multidimensional vector, then c is a multidimensional vector. The cGAN model can learn the complex, nonlinear trade-offs between these performance metrics. For example, adding certain fillers to improve dielectric strength may lead to an increase in dielectric constant or a change in resistivity. cGAN learns these relationships from data during training, so when given a multi-performance objective, it can generate the most optimal formulation possible under these constraints, rather than satisfying a single objective in isolation. This ability to handle multi-objective, multi-constraint optimization problems is a significant advantage of this invention compared to traditional single-objective optimization or trial-and-error methods.

[0077] Specifically, the raw materials of the recycled composite material include various recycled plastics, inorganic fillers, and reinforcing fibers.

[0078] It should be further noted that this invention provides a complex material system that can be processed by the method of this invention, which is the real-world context in which this method solves the technical problem. Specific examples of raw materials are as follows: 1. Various recycled plastics: These constitute the matrix phase, such as recycled polypropylene (rPP), recycled acrylonitrile-butadiene-styrene copolymer (rABS), recycled polycarbonate (rPC), and recycled polyethylene (rPE). Their diverse origins, varying degrees of performance degradation, and poor compatibility are the main reasons why the final material properties are difficult to predict.

[0079] 2. Inorganic fillers: Commonly used to adjust cost, rigidity, heat resistance, and certain electrical properties. Examples include calcium carbonate (CaCO3, which can reduce cost and improve rigidity, but excessive amounts may reduce toughness), talc (which can improve rigidity and heat resistance, and may affect the dielectric constant), and barium sulfate (high density, used for weighting, with a relatively high dielectric constant). The particle size, morphology, and surface treatment (whether or not they are coupled) of the filler significantly affect the final performance.

[0080] 3. Reinforcing fibers: Primarily used to significantly improve the mechanical strength and stiffness of materials, such as chopped glass fiber (GF) and carbon fiber (CF). Glass fiber is an insulator, while carbon fiber is a conductor. The addition of even a small amount of carbon fiber can transform a material from an insulator into a semiconductor or even a conductor, having a revolutionary impact on its electrical properties.

[0081] Understandably, the core value of this method lies in its ability to handle systems with multiple components, multiple phases, complex interfaces, and intertwined performance-influencing factors. The cGAN model does not rely on any prior physical equations regarding phase morphology or interfacial interactions; instead, it learns end-to-end from historical experimental data the hidden, high-dimensional, nonlinear mapping relationships and their inverse mappings between complex formulations (including the types and amounts of each component) and the final macroscopic electrical properties. As long as the training database contains a sufficiently large and widely distributed range of formulation combinations and their performance results, the model can internalize these complex patterns and use them for reverse reasoning towards new performance targets. This provides an efficient and feasible technical path for developing high-value-added, high-performance electrical components using recycled materials with complex compositions and large performance fluctuations.

[0082] Please see Figure 2 The present invention provides another embodiment, which provides a reverse design system for regenerated composite materials based on generative adversarial networks. The reverse design system for regenerated composite materials based on generative adversarial networks includes: Database module 100 is used to store and manage the recipe-performance sample database; The network training module 200 is used to construct and train a conditional generative adversarial network (cGAN) based on the recipe-performance sample database. Input module 300 is used to receive the target electrical performance settings for the target electrical component; The reverse generation module 400 is used to input the target electrical performance into the trained conditional generative adversarial network cGAN and obtain the virtual microstructure features generated by its reverse generation and the corresponding raw material ratio suggestions.

[0083] It should be further noted that this invention protects the hardware / software system for implementing the above method, which is typically deployed on a computing server or cloud platform and includes the following specific modules: 1. Database Module 100: Implemented using a relational database (such as MySQL) or NoSQL database. It is responsible for storing and managing all historical and iterative data. The data is structured and stored in multiple tables, including a "Raw Material Information Table," "Formula Table," "Performance Test Table," and "Microstructure Characteristics Table." This module provides interfaces for data entry, querying, modification, and deletion, and can export data to formats required for model training (such as NumPy arrays or TensorFlow Datasets). It also includes a built-in data preprocessing submodule that automatically cleans, normalizes, and standardizes newly entered data.

[0084] 2. Network Training Module 200: This is the computational core of the system, implemented using a deep learning framework (such as PyTorch 1.9.0). This module contains the following sub-modules: ① Model definition submodule: This module procedurally defines the network architecture, number of layers, number of neurons, activation functions, etc. of the generator G, discriminator D, and performance prediction model P.

[0085] ② Training Configuration Submodule: Allows users to set or use default hyperparameters, including: learning rate (0.0002), optimizer (Adam), loss function weights (λ=5), batch size (64), number of training epochs, etc.

[0086] ③ Training Execution Submodule: Reads preprocessed data from the database module and divides it into training and validation sets proportionally. Executes the adversarial training process, alternately updating D and G. Monitors the loss function. , , It also supports changes in the quality of generated samples and supports saving checkpoints during the training process so that it can resume from where it was interrupted.

[0087] ④ Model Evaluation and Saving Submodule: After training, evaluate the quality of the generated samples on an independent test set (e.g., by calculating the distance between the generated samples and the real samples in terms of statistical features) and the accuracy of performance prediction. Save the parameters of the well-trained generator model G as a file (e.g., .pt or .h5 format) for use by the reverse generation module.

[0088] 3. Input Module 300: Provides a human-computer interaction interface, which can be a graphical user interface (GUI) or an application programming interface (API). In the GUI, users can set the type, target value, or range of the target electrical performance through form input or slider selection (e.g., selecting "volume resistivity" from the drop-down menu and entering ">1.0E+12" in the input box). This module is responsible for converting the performance target input by the user into a condition vector c required by the model, which has been normalized to be consistent with the training data.

[0089] 4. Reverse Generation Module 400: Loads the optimal generator model file saved by the network training module. Receives the condition vector c from the input module. Samples different noise vectors z from the standard normal distribution multiple times (e.g., 100 times) and combines them with the same condition vector c, inputting them into the generator G to generate multiple candidate design schemes (M', F') in batches. This module also includes a simple sorting or filtering sub-function. For example, it can call the performance prediction model P to calculate the predicted performance of each candidate formulation F', sort them according to their proximity to the target performance c, and output the top N (e.g., the top 5) optimal schemes along with their predicted microstructural features M' to the user as a detailed R&D recommendation report.

[0090] Understandably, this system encapsulates the aforementioned complex algorithm process into an easy-to-use software tool, enabling materials engineers to utilize advanced generative adversarial network technology for efficient reverse design of recycled composite materials without needing to master deep learning expertise, thus significantly improving the intelligence and efficiency of R&D.

[0091] In a preferred embodiment, this application also provides an electronic device, the electronic device comprising: The computer device includes a memory and a processor, wherein the memory stores computer-readable instructions that, when executed by the processor, implement the described reverse design method for regenerative composite materials based on generative adversarial networks. The computer device can be broadly categorized as a server, terminal, or any other electronic device with the necessary computing and / or processing capabilities. In one embodiment, the computer device may include a processor, memory, network interface, communication interface, etc., connected via a system bus. The processor of the computer device can be used to provide the necessary computing, processing, and / or control capabilities. The memory of the computer device may include a non-volatile storage medium and internal memory. The non-volatile storage medium may store an operating system, computer programs, etc. The internal memory can provide an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface and communication interface of the computer device can be used to connect and communicate with external devices via a network. When the computer program is executed by the processor, it performs the steps of the method of the present invention.

[0092] This invention can be implemented as a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the steps of the methods of embodiments of the invention to be performed. In one embodiment, the computer program is distributed across multiple network-coupled computer devices or processors, such that the computer program is stored, accessed, and executed in a distributed manner by one or more computer devices or processors. A single method step / operation, or two or more method steps / operations, may be executed by a single computer device or processor or by two or more computer devices or processors. One or more method steps / operations may be executed by one or more computer devices or processors, and one or more other method steps / operations may be executed by one or more other computer devices or processors. One or more computer devices or processors may execute a single method step / operation, or execute two or more method steps / operations.

[0093] Those skilled in the art will understand that the method steps of this invention can be performed by a computer program instructing related hardware, such as a computer device or processor, to perform the steps of this invention when executed. Depending on the context, any references herein to memory, storage, databases, or other media may include non-volatile and / or volatile memory. Examples of non-volatile memory include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, etc. Examples of volatile memory include random access memory (RAM), external cache memory, etc.

[0094] The technical features described above can be combined arbitrarily. Although not all possible combinations of these technical features are described, any combination of these technical features should be considered to be covered by this specification, provided that such combination does not contain contradictions.

[0095] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A reverse design method for regenerative composite materials based on generative adversarial networks, characterized in that, Includes the following steps: S1. Obtain known formulations of various recycled composite materials and their corresponding measured electrical performance data, and construct a formulation-performance sample database; S2. Based on the formula-performance sample database, construct and train a conditional generative adversarial network (cGAN), wherein the target electrical performance is used as the conditional input, and the virtual microstructure characteristics and raw material ratio of the recycled composite material that meets the target electrical performance are used as the generation target. S3. Input the target electrical performance set for the target electrical component into the trained conditional generative adversarial network cGAN. S4. Using the trained Conditional Generative Adversarial Network (cGAN), reverse-generate virtual microstructure features and corresponding raw material ratio suggestions that match the target electrical performance.

2. The method according to claim 1, characterized in that, The generator network of the conditional generative adversarial network (cGAN) is used to receive the noise vector and the conditional vector of the target electrical performance, and output the virtual microstructure feature vector and the raw material ratio vector.

3. The method according to claim 2, characterized in that, The discriminator network of the conditional generative adversarial network (cGAN) is used to receive real or generated virtual microstructure feature vectors and raw material ratio vectors, and combine them with the conditional vector of the target electrical performance to determine the authenticity of the virtual microstructure feature vectors and raw material ratio vectors.

4. The method according to claim 3, characterized in that, In step S2, the loss function for training the conditional generative adversarial network (cGAN) includes at least an adversarial loss to measure the difference between the generated data and the real data distribution, and a conditional matching loss to constrain the generated data to meet the target electrical performance.

5. The method according to claim 4, characterized in that, The conditional matching loss is calculated by inputting the virtual microstructure feature vector and raw material ratio vector generated by the generator into a pre-trained performance prediction model, and comparing the predicted electrical performance of the model with the target electrical performance.

6. The method according to claim 5, characterized in that, The performance prediction model is a neural network model, which is trained through the formula-performance sample database and is used to predict electrical performance based on the input raw material ratio.

7. The method according to claim 1, characterized in that, After step S4, the method further includes: S5, preparing a recycled composite material sample based on the raw material ratio recommendation and testing its actual electrical properties; S6, adding the raw material ratio recommendation and its corresponding actual electrical performance data as new samples to the formula-performance sample database to iteratively optimize the conditional generative adversarial network cGAN.

8. The method according to claim 1, characterized in that, The target electrical properties include at least one of conductivity, resistivity, dielectric constant, and dielectric strength.

9. The method according to claim 1, characterized in that, The raw materials for the recycled composite material include various recycled plastics, inorganic fillers, and reinforcing fibers.

10. A reverse design system for regenerative composite materials based on generative adversarial networks, characterized in that, The system for implementing the method as described in any one of claims 1 to 9 comprises: The database module is used to store and manage the recipe-performance sample database; The network training module is used to construct and train a conditional generative adversarial network (cGAN) based on the recipe-performance sample database. The input module is used to receive the target electrical performance settings for the target electrical component; The reverse generation module is used to input the target electrical properties into the trained conditional generative adversarial network (cGAN) and obtain the virtual microstructure features generated by the reverse generation and the corresponding raw material ratio suggestions.

Images