Deep learning based system and method for predicting and optimizing thermal property parameters of composite compounds across scales

Abstract

This invention discloses a deep learning-based system and method for predicting and optimizing the multi-scale thermophysical parameters of composite compounds. The system includes: a data acquisition and preprocessing module for acquiring and preprocessing multi-scale data of composite compounds to obtain a standardized multi-scale feature dataset; a model training module for training a multi-scale thermophysical prediction model using the standardized multi-scale feature dataset to obtain the trained multi-scale thermophysical prediction model; a thermophysical parameter prediction module for inputting the multi-scale data of the composite compound to be tested into the trained multi-scale thermophysical prediction model to obtain the thermophysical prediction results; an intelligent optimization module for back-optimizing the composition ratio and microstructure parameters of the composite compound based on the thermophysical prediction results, and outputting the optimized material formulation and microstructure parameters; and a result visualization module for visualizing the thermophysical prediction results or the optimized material formulation and microstructure parameters.

Description

Technical Field

[0001] This invention belongs to the field of composite compound performance prediction and material design technology, and particularly relates to a system and method for predicting and optimizing cross-scale thermophysical parameters of composite compounds based on deep learning. Background Technology

[0002] Currently, the acquisition and optimization of thermophysical parameters of composite compounds mainly rely on traditional experimental tests and empirical models, which still have significant shortcomings: The research and development cycle is highly dependent on experimental testing. Traditional methods require the preparation of a large number of samples with different compositions and microstructures, and the testing of thermophysical properties is carried out by means of laser flare analysis (LFA), differential scanning calorimetry (DSC), thermomechanical analysis (TMA), etc. The preparation and testing cycle of a single sample can reach 1-2 weeks. When screening multiple parameters, the research and development cycle often reaches several months to several years. Moreover, the experimental-related costs, such as raw materials, equipment, and labor, account for more than 60% of the total research and development investment, which is difficult to meet the needs of rapid material iteration.

[0003] A key issue in the study of the thermal properties of composite compounds is the lack of cross-scale correlations. These thermal property parameters are jointly determined by multi-scale structures, including microscopic atomic structure, mesoscopic phase distribution, and macroscopic composition ratios. At the microscopic level, atomic bond lengths affect phonon transport efficiency, at the mesoscopic level, enhanced phase aggregation determines the heat conduction path, and at the macroscopic level, composition ratios form the basis of thermal properties.

[0004] However, most traditional models focus only on a single scale, such as relying solely on empirical formulas for macroscopic components to predict thermal conductivity, while neglecting the coupling effects between multiple scales. This simplification results in prediction errors generally exceeding 15%, failing to meet the stringent requirements of high-end manufacturing fields where the accuracy of thermal properties must not exceed 10%.

[0005] Currently, in the field of materials optimization, the problem of low optimization efficiency remains prominent. Traditional optimization methods, such as orthogonal experiments and single gradient descent, have obvious limitations. Orthogonal experiments can usually only handle three to four parameters, and are prone to missing the true optimal solution; while gradient descent is prone to getting trapped in local optima and is difficult to effectively deal with the complex high-dimensional nonlinear relationships between composition, microstructure, and thermophysical properties.

[0006] Taking the optimization of aluminum-based silicon carbide composite materials as an example, when only three parameters are considered, and each parameter has five levels, traditional orthogonal experiments require twenty-five sets of experiments. Even so, it may still be impossible to find the globally optimal solution. This inefficient optimization method severely restricts the research progress of new materials.

[0007] While deep learning technology has shown promise in current research on predicting the thermal properties of materials, it still faces several key challenges. Existing methods have significant shortcomings in multi-scale feature fusion, a cross-scale information transfer mechanism has not yet been effectively established, and there is a disconnect between material performance prediction and formulation optimization. To address these technical bottlenecks, there is an urgent need to construct an integrated system capable of integrating cross-scale experimental data and enabling collaborative work between performance prediction and formulation optimization, thereby driving substantial breakthroughs in materials research and development. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention proposes a deep learning-based system and method for predicting and optimizing the cross-scale thermal properties of composite compounds. This invention is applicable to the research and development and production of metal matrix composites, ceramic matrix composites, and polymer composites in fields such as aerospace, electronic packaging, and new energy devices. It can achieve high-precision prediction of key thermal properties of composite compounds, such as thermal conductivity, coefficient of thermal expansion, and specific heat capacity, as well as intelligent optimization of material composition ratios and microstructure parameters, thus solving the technical problems of long cycles, high costs, and low accuracy in traditional research and development processes.

[0009] To achieve the above objectives, this invention provides a deep learning-based system for predicting and optimizing the cross-scale thermal properties of composite compounds, comprising: a data acquisition and preprocessing module, a model training module, a thermal property prediction module, an intelligent optimization module, and a result visualization module. The data acquisition and preprocessing module is used to acquire multi-scale data of composite compounds and perform preprocessing to obtain a standardized multi-scale feature dataset. The model training module is used to train a cross-scale thermal property prediction model using the standardized multi-scale feature dataset, and obtain the trained cross-scale thermal property prediction model. The thermophysical parameter prediction module is used to input multi-scale data of the composite compound to be tested into the trained cross-scale thermophysical property prediction model to obtain thermophysical property prediction results. The intelligent optimization module is used to reverse-optimize the composition ratio and microstructure parameters of the composite compound based on the thermophysical property prediction results, and output the optimized material formulation and microstructure parameters. The results visualization module is used to visualize the predicted thermophysical properties or optimized material formulations and microstructure parameters, and generate exportable analysis reports.

[0010] Optionally, the multi-scale data of the composite compound includes: microstructure images, composition data, and experimental thermophysical property data of the composite compound; The microstructure images include scanning electron microscope images and transmission electron microscope images; The component data includes the mass fraction of each component; The experimental thermophysical data include thermal conductivity, coefficient of thermal expansion, and specific heat capacity.

[0011] Optionally, the cross-scale thermal property prediction model is an improved multi-scale attention convolutional neural network, which includes an input layer, a three-level attention layer, a cross-scale feature fusion unit, a convolutional layer, and a fully connected layer. The input layer is used to receive a microscopic feature vector, a mesoscopic feature vector, and a first macroscopic feature vector, wherein the microscopic feature vector is obtained based on the microstructure image, the mesoscopic feature vector is obtained based on the composition data, and the first macroscopic feature vector is obtained based on the experimental thermophysical property data. The three-level attention layer is used to calculate attention weights based on the input feature vectors at each scale. The cross-scale feature fusion unit is used to perform weighted fusion of multi-scale features according to the calculated attention weights to obtain fused features; The convolutional layer is used to extract local correlation features and global correlation features based on the fused features; The fully connected layer is used to output predicted values ​​of thermal property parameters based on the local correlation features and the global correlation features.

[0012] Optionally, the attention weights can be calculated based on the input feature vectors at various scales, including: ; in, For microscopic eigenvectors, For mesoscopic eigenvectors, This is the first macroscopic eigenvector. , For attention layer parameters, , , These are the attention weights for microscopic, mesoscopic, and macroscopic features, respectively. , For the micro-feature attention layer bias parameters, For the bias parameters of the mesoscopic feature attention layer, These are the bias parameters for the macroscopic feature attention layer. As a scale identifier variable, For microscale identification, For mesoscale identification, For macroscopic scale identification, This is a general-scale feature vector.

[0013] Optionally, the multi-scale features are weighted and fused based on the calculated attention weights to obtain fused features, including: The feature vectors at each scale are weighted according to the attention weights output by the three-level attention layer to obtain the weighted micro feature vector, meso feature vector and first macro feature vector; The weighted three-scale feature vectors are concatenated to obtain the concatenated feature vector; The concatenated feature vector is residually concatenated with the original features of the input layer to obtain a residual feature vector; The residual feature vectors are batch normalized to obtain the fused features.

[0014] Optionally, the system further includes an adaptive multi-scale coupling algorithm; Phonon scattering rate was obtained from molecular dynamics simulations; The phonon scattering rate is converted into phonon relaxation time, and the mesoscopic thermal conductivity is calculated based on the phonon relaxation time. A second macroscopic feature vector is generated based on mesoscopic thermal conductivity and macroscopic composition data. The second macroscopic feature vector is then input into the cross-scale thermal property prediction model to obtain the predicted macroscopic thermal property values. The prediction error is calculated based on the macroscopic thermophysical property predictions and experimental reference values. When the prediction error exceeds the preset threshold, the interfacial thermal resistance coefficient in the micro-mesoscopic coupling parameters is adjusted according to the error value, and the molecular dynamics simulation and mesoscopic analysis are re-executed until the prediction error is no greater than the preset threshold.

[0015] Optionally, adjusting the interfacial thermal resistance coefficient in the micro-mesoscopic coupling parameters based on the error value includes: ; in, The interfacial thermal resistance coefficient before correction. This is the corrected interfacial thermal resistance coefficient. The error between the macroscopic experimental thermophysical parameters and the predicted values, and , For macroscopic experimental thermal conductivity, This is the predicted value for macroscopic thermal conductivity.

[0016] Optionally, the intelligent optimization module employs a hybrid optimization strategy of genetic algorithm-gradient descent for optimization; The optimization strategy employing a hybrid approach of genetic algorithm and gradient descent includes: An optimization objective function is constructed based on the predicted thermal properties and constraints. The population is then initialized based on the optimization objective function, and the fitness values ​​of each parameter combination in the population are calculated. Based on the fitness value, selection, crossover, and mutation operations are performed on the population to generate a new generation of population; The global optimal candidate parameter set is selected based on the fitness values ​​of the new generation population; Based on the optimal candidate parameter set, the gradient descent method is used to locally fine-tune each parameter, and the gradient value of the optimization objective function with respect to each parameter is calculated. Update the parameters based on the gradient value until the convergence condition is met, and output the optimized material formulation and microstructure parameters.

[0017] Optionally, the optimization objective function is: ; in, For the target thermal conductivity, For the target coefficient of thermal expansion, , Parameter combinations The corresponding predicted values ​​of thermal conductivity and coefficient of thermal expansion, To increase the phase volume fraction, To enhance phase grain size, For the thickness of the matrix-reinforcement phase interface, , The target weight.

[0018] This invention also provides a method for predicting and optimizing the cross-scale thermal property parameters of composite compounds based on deep learning, including: Collect multi-scale data of composite compounds and preprocess them to obtain a standardized multi-scale feature dataset. A cross-scale thermal property prediction model is trained using the standardized multi-scale feature dataset to obtain the trained cross-scale thermal property prediction model, wherein the cross-scale thermal property prediction model is an improved multi-scale attention convolutional neural network. Acquire multi-scale data of the composite compound to be tested, input the multi-scale data of the composite compound to be tested into the trained cross-scale thermophysical property prediction model, and obtain the thermophysical property prediction results; Based on the predicted thermophysical properties, the composition ratio and microstructure parameters of the composite compound are optimized in reverse, and the optimized material formulation and microstructure parameters are output. The predicted thermophysical properties or optimized material formulations and microstructure parameters are visualized, and an exportable analysis report is generated.

[0019] Compared with the prior art, the present invention has the following advantages and technical effects: 1. High prediction accuracy: This invention dynamically allocates multi-scale weights through the three-level attention layer of MSA-CNN, and combines cross-scale feature fusion and adaptive coupling algorithms to significantly improve the prediction accuracy of thermal property parameters and effectively shorten the prediction bias. Compared with the prediction effect of traditional single-scale models, it has a significant improvement.

[0020] 2. High optimization efficiency: This invention adopts a hybrid optimization strategy that combines global search using genetic algorithms with local fine-tuning using gradient descent. It eliminates the need for tedious repeated experiments, quickly locks in the optimal parameter combination, significantly improves the optimization efficiency of composite compounds, and also helps to improve thermal conductivity and optimize the coefficient of thermal expansion.

[0021] 3. Low R&D cost and short cycle: With data-driven approach as the core, model prediction significantly reduces reliance on actual experimental testing. This not only effectively reduces various costs in the R&D process but also significantly shortens the R&D cycle of composite compounds, accelerating the application of new materials.

[0022] 4. Wide range of applications: Supports various composite compound types such as metal-based, ceramic-based, and polymer-based compounds, and is suitable for aerospace, electronic packaging, new energy and other fields. Parameters can be adjusted to meet personalized needs. 5. High interpretability of results: Attention weight distribution and multi-scale parameter correlation heatmap clearly show the parameter influence mechanism, avoiding the "black box model" problem and providing clear optimization direction for materials research and development. Attached Figure Description

[0023] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of a deep learning-based multi-scale thermophysical property parameter prediction and optimization system for composite compounds, according to an embodiment of the present invention. Figure 2 This is a flowchart of a deep learning-based method for predicting and optimizing the cross-scale thermophysical property parameters of composite compounds according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of the improved multi-scale attention convolutional neural network (MSA-CNN) according to an embodiment of the present invention. Detailed Implementation

[0024] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0026] This embodiment proposes a deep learning-based system for predicting and optimizing the cross-scale thermal property parameters of composite compounds, such as... Figure 1As shown, it specifically includes: a data acquisition and preprocessing module, a model training module (deep learning model training module), a thermal property parameter prediction module, an intelligent optimization module, and a result visualization module (result visualization and output module). The data acquisition and preprocessing module is used to acquire multi-scale data of composite compounds and perform preprocessing to obtain a standardized multi-scale feature dataset. The model training module is used to train a cross-scale thermal property prediction model using the standardized multi-scale feature dataset, and obtain the trained cross-scale thermal property prediction model. The thermophysical parameter prediction module is used to input multi-scale data of the composite compound to be tested into the trained cross-scale thermophysical property prediction model to obtain thermophysical property prediction results. The intelligent optimization module is used to reverse-optimize the composition ratio and microstructure parameters of the composite compound based on the thermophysical property prediction results, and output the optimized material formulation and microstructure parameters. The results visualization module is used to visualize the predicted thermophysical properties or optimized material formulations and microstructure parameters, and generate exportable analysis reports.

[0027] Specifically, the modules interact collaboratively through data interfaces. The data acquisition and preprocessing module is used to collect multi-scale data of composite compounds and perform cleaning, labeling and feature extraction, outputting a standardized multi-scale feature dataset. The deep learning model training module trains a cross-scale prediction model based on the standardized multi-scale feature dataset and outputs the trained cross-scale thermal property prediction model. The thermophysical property parameter prediction module takes into account the multi-scale parameters of the composite compound to be predicted, and calls the trained cross-scale thermophysical property prediction model to output the thermophysical property prediction results. The intelligent optimization module optimizes the composition ratio and microstructure parameters of the composite compound based on the target thermophysical parameters, and outputs the optimized material formulation and microstructure parameters. The results visualization and output module displays the predicted thermal properties and optimization paths in chart form, and generates an exportable analysis report.

[0028] Furthermore, the multi-scale data of the composite compound includes: microstructure images, composition data, and experimental thermophysical property data of the composite compound; The microstructure images include scanning electron microscope images and transmission electron microscope images; The component data includes the mass fraction of each component; The experimental thermophysical data include thermal conductivity, coefficient of thermal expansion, and specific heat capacity.

[0029] Specifically, the composite compound includes at least one of metal matrix composites, ceramic matrix composites, and polymer composites; the thermophysical parameters include at least one of thermal conductivity, coefficient of thermal expansion, and specific heat capacity; the multi-scale refers to the micro-meso-macro scale. The multi-scale data acquired by the data acquisition and preprocessing module includes microscopic structural images, compositional data, and experimental thermophysical data of the composite compound; the sources of the multi-scale data include literature data, experimental data, and microscopic imaging data.

[0030] Furthermore, such as Figure 3 As shown, the cross-scale thermal property prediction model is an improved multi-scale attention convolutional neural network, which includes an input layer, a three-level attention layer, a cross-scale feature fusion unit, a convolutional layer, and a fully connected layer. The input layer is used to receive microscopic feature vectors, mesoscopic feature vectors, and a first macroscopic feature vector. The microscopic feature vector is obtained based on the microstructure image, the mesoscopic feature vector is obtained based on the composition data, and the first macroscopic feature vector is obtained based on the experimental thermophysical property data. The first macroscopic feature vector refers to an experimental feature at a single macroscopic scale, which is used as the initial input of the model. The first macroscopic feature vector is one of the basic inputs of the second macroscopic feature vector.

[0031] The three-level attention layer is used to calculate attention weights based on the input feature vectors at each scale. The cross-scale feature fusion unit is used to perform weighted fusion of multi-scale features according to the calculated attention weights to obtain fused features; The convolutional layer is used to extract local correlation features and global correlation features based on the fused features; The fully connected layer is used to output predicted values ​​of thermal property parameters based on the local correlation features and the global correlation features.

[0032] Specifically, the cross-scale prediction model in the deep learning model training module is an improved multi-scale attention convolutional neural network (MSA-CNN), which includes a three-level attention layer of "micro-meso-macro" and a cross-scale feature fusion unit.

[0033] Furthermore, calculating the attention weights based on the input feature vectors at each scale includes: ; in, For microscopic eigenvectors, For mesoscopic eigenvectors, This is the first macroscopic eigenvector. , For attention layer parameters, , , These are the attention weights for microscopic, mesoscopic, and macroscopic features, respectively. , For the micro-feature attention layer bias parameters, For the bias parameters of the mesoscopic feature attention layer, These are the bias parameters for the macroscopic feature attention layer. As a scale identifier variable, For microscale identification, For mesoscale identification, For macroscopic scale identification, This is a general-scale feature vector.

[0034] Furthermore, the multi-scale features are weighted and fused based on the calculated attention weights to obtain the fused features, including: The feature vectors at each scale are weighted according to the attention weights output by the three-level attention layer to obtain the weighted micro feature vector, meso feature vector and first macro feature vector; The weighted three-scale feature vectors are concatenated to obtain the concatenated feature vector; The concatenated feature vector is residually concatenated with the original features of the input layer to obtain a residual feature vector; The residual feature vectors are batch normalized to obtain the fused features.

[0035] Specifically, the cross-scale feature fusion unit deeply integrates microscopic atomic structure features, mesoscopic phase distribution features, and macroscopic composition features through residual connection and feature splicing; the atomic structure features include bond length and coordination number, the phase distribution features include enhanced phase aggregation degree, and the composition features include the mass fraction of each component.

[0036] The deep learning model training module uses weighted mean squared error (W-MSE) as the loss function, and the formula for calculating the loss function is as follows: ; in, For the sample size, For the i-th sample, the p-th thermal property parameter is the true value. For the predicted value of the p-th thermal property parameter of the i-th sample, Here, k is the weighting coefficient, and k is the thermal conductivity. The coefficient of thermal expansion is Specific heat capacity.

[0037] The weighting coefficient The value of satisfies Preferably, , , .

[0038] Furthermore, the system also includes an adaptive multi-scale coupling algorithm; Phonon scattering rate was obtained from molecular dynamics simulations; The phonon scattering rate is converted into phonon relaxation time, and the mesoscopic thermal conductivity is calculated based on the phonon relaxation time. A second macroscopic feature vector is generated based on mesoscopic thermal conductivity and macroscopic composition data. The second macroscopic feature vector is then input into a cross-scale thermal property prediction model to obtain the predicted macroscopic thermal property values. The second macroscopic feature vector generated from mesoscopic thermal conductivity and macroscopic composition refers to a cross-scale coupled computational feature used for error correction after model prediction. The second macroscopic feature vector is an enhancement and optimization of the first macroscopic feature vector, supplementing mesoscopic computational data to improve accuracy. The prediction error is calculated based on the macroscopic thermophysical property predictions and experimental reference values. When the prediction error exceeds the preset threshold, the interfacial thermal resistance coefficient in the micro-mesoscopic coupling parameters is adjusted according to the error value, and the molecular dynamics simulation and mesoscopic analysis are re-executed until the prediction error is no greater than the preset threshold.

[0039] Specifically, the system also includes an adaptive multi-scale coupling algorithm for efficient information transfer between micro-, meso-, and macro-scales; the adaptive multi-scale coupling algorithm includes micro-scale calculation, meso-scale analysis, macro-scale mapping, and coupling verification and correction steps.

[0040] The microscale calculations are based on molecular dynamics simulations to obtain atomic-level thermal transport parameters; these atomic-level thermal transport parameters include phonon scattering rate. The mesoscale analysis calculates the thermal properties of the mesoscale region using the finite element method, generating a mesoscale characteristic vector; the mesoscale region is a representative volumetric unit (RVU).

[0041] The parameter transfer between the micro- and meso-scales is achieved through the following formula: ; in, The thermal conductivity of the mesoscopic region, Specific heat capacity per unit volume at constant volume For the phonon group velocity, Let be the phonon relaxation time, and .

[0042] In the coupling verification and correction step, if the error between the macroscopic experimental thermophysical parameters and the predicted values ​​exceeds a preset threshold, the micro-mesoscopic coupling parameters are adjusted and the calculation is iterated again; the preset threshold is 5%; the micro-mesoscopic coupling parameters include the interfacial thermal resistance coefficient. .

[0043] Furthermore, adjusting the interfacial thermal resistance coefficient in the micro-meta-coupling parameters based on the error value includes: ; in, The interfacial thermal resistance coefficient before correction. This is the corrected interfacial thermal resistance coefficient. The error between the macroscopic experimental thermophysical parameters and the predicted values, and , For macroscopic experimental thermal conductivity, This is the predicted value for macroscopic thermal conductivity.

[0044] Furthermore, the intelligent optimization module employs a hybrid optimization strategy of genetic algorithm-gradient descent for optimization; The optimization strategy employing a hybrid approach of genetic algorithm and gradient descent includes: An optimization objective function is constructed based on the predicted thermal properties and constraints. The population is then initialized based on the optimization objective function, and the fitness values ​​of each parameter combination in the population are calculated. Based on the fitness value, selection, crossover, and mutation operations are performed on the population to generate a new generation of population; The global optimal candidate parameter set is selected based on the fitness values ​​of the new generation population; Based on the optimal candidate parameter set, the gradient descent method is used to locally fine-tune each parameter, and the gradient value of the optimization objective function with respect to each parameter is calculated. Update the parameters based on the gradient value until the convergence condition is met, and output the optimized material formulation and microstructure parameters.

[0045] Specifically, this module addresses the high-dimensional, nonlinear, and multi-constraint characteristics of thermal property optimization for composite compounds. Through a collaborative mechanism of global search and local fine-tuning, it effectively overcomes the shortcomings of traditional single optimization methods. The genetic algorithm efficiently explores potential optimal parameter combinations across the entire constraint space, avoiding getting trapped in local optima. The gradient descent method, based on the globally optimal candidate set selected by the genetic algorithm, finely adjusts the objective function by calculating the gradient direction, thus solving the problem of insufficient local search accuracy in the genetic algorithm.

[0046] Furthermore, the optimization objective function is: ; in, For the target thermal conductivity, For the target coefficient of thermal expansion, , Parameter combinations The corresponding predicted values ​​of thermal conductivity and coefficient of thermal expansion, To increase the phase volume fraction, To enhance phase grain size, For the thickness of the matrix-reinforcement phase interface, , The target weight.

[0047] Specifically, the constraints of the optimization objective function are as follows: , , The target weights satisfy Preferably, , .

[0048] In the fitness evaluation step, the fitness value is calculated using the following formula: Where F is the fitness value, and J represents the objective function value; the larger the fitness value, the better the optimization effect of the corresponding parameter combination.

[0049] The genetic operations include selection, crossover, and mutation. The selection operation retains the individual with the best fitness value, the crossover operation generates offspring individuals through gene recombination, and the mutation operation introduces new genetic diversity by randomly changing parameter values.

[0050] The gradient descent fine-tuning step fine-tunes the local parameters of the optimal individual output by the genetic algorithm; in the convergence judgment step, if the fitness value of the population changes by less than 1% for 5 consecutive generations, the optimization result is output; otherwise, the iteration returns to the fitness evaluation step.

[0051] The charts generated by the results visualization and output module include a trend chart of predicted thermophysical parameters and an iterative curve of optimized parameters; the analysis report is in PDF format and includes data sources, model parameters, prediction results, optimization schemes, and verification conclusions.

[0052] The training hyperparameters of the deep learning model training module include learning rate, batch size, and number of iterations; the learning rate is 0.001, the batch size is 32, and the number of iterations is 200; the deep learning model training module uses the Adam optimizer for model training.

[0053] The data acquisition and preprocessing module extracts the microstructural features of the composite compound using an image segmentation algorithm; the microstructural features include the volume fraction of the reinforcing phase and the aggregation degree of the reinforcing phase; the data acquisition and preprocessing module converts the component data into a standardized mass fraction vector, and the generated standardized dataset is divided into a training set and a test set in an 8:2 ratio.

[0054] The system is applied to the research and development and production process of composite compounds in the fields of aerospace, electronic packaging and new energy devices; the system can shorten the research and development cycle of composite compounds by 60%-80% and reduce the research and development cost by 30%-50%.

[0055] The following is a detailed description of this embodiment: This embodiment adopts a modular, layered architecture design, integrating five core functional modules: a data acquisition and preprocessing module, a deep learning model training module, a thermophysical parameter prediction module, an intelligent optimization module, and a result visualization and output module. The modules interact efficiently through standardized data interfaces, ensuring both data format consistency and system stability. The entire system constructs a complete closed-loop process from data input to model training, predictive analysis, optimization output, and final result feedback.

[0056] The standardized multi-scale feature dataset output by the data acquisition and preprocessing module serves as the input to the deep learning model training module; the cross-scale thermal property prediction model generated by the deep learning model training module provides support for the thermal property parameter prediction module and the intelligent optimization module; the thermal property prediction results and the intelligent optimization scheme are jointly input to the result visualization and output module; the analysis report can be fed back to the data acquisition and preprocessing module for adjusting the data processing strategy.

[0057] The various modules within the system communicate and exchange data through standardized API interfaces. This design ensures that the thermophysical property prediction model can be accurately loaded and that the optimization module can obtain prediction data in real time, thereby guaranteeing the smooth operation of the system and the consistency of the data.

[0058] Data acquisition and preprocessing module: This module serves as the system's data input source, responsible for collecting multi-scale data on composite compounds. Through standardized processing such as cleaning, labeling, and feature extraction, it generates a high-quality, standardized multi-scale feature dataset, providing reliable input for subsequent model training.

[0059] Multi-scale data collection scope, including but not limited to the following scope.

[0060] Based on the factors influencing the thermophysical properties of composite compounds, the collected data were divided into three scales: microscopic, mesoscopic, and macroscopic.

[0061] The microscale data mainly comprises two parts: microstructure images and atomic structure parameters. Microstructure images were acquired using scanning electron microscopy and transmission electron microscopy, with resolutions reaching 10 nm and 0.1 nm, respectively. Atomic structure parameters, including bond lengths and coordination numbers, were obtained through molecular dynamics simulations. These simulations covered a temperature range from -50°C to 500°C, ensuring that the obtained parameters fully reflect the material's properties under real-world application conditions.

[0062] Mesoscopic-scale data mainly includes phase distribution data and mesoscopic thermophysical property data. The reinforcing phase aggregation degree in the phase distribution data ranges from [0,1], corresponding to the two extreme states of complete dispersion and complete aggregation, respectively. This data is obtained through thresholding and particle analysis of mesoscopic structure images with a resolution of at least 100 nm. Mesoscopic thermophysical property data, namely the thermal conductivity and coefficient of thermal expansion of a representative volumetric unit, are obtained through finite element analysis. To ensure that this volumetric unit fully represents the overall material properties, its size is set to five to ten times the average particle size of the reinforcing phase.

[0063] Macroscopic data primarily includes material composition data and thermophysical reference data. Composition data involves the mass and volume fractions of each component, obtained through compositional analysis techniques with a detection accuracy of 0.1%. Thermophysical reference data includes thermal conductivity, coefficient of thermal expansion, and specific heat capacity, measured using laser scintillation, thermomechanical analysis, and differential scanning calorimetry, respectively, with measurement accuracies better than 3%, 5%, and 2%, respectively. All thermophysical property tests cover four characteristic temperature points: 25 °C, 100 °C, 200 °C, and 300 °C, to characterize the temperature dependence of material properties.

[0064] Data preprocessing workflow: The system employs appropriate outlier detection methods for different types of experimental data. For numerical data, a dual screening method is used, employing both statistical tests based on the three-standard-deviation principle and box plots based on interquartile range. For image data, outlier images are screened by evaluating their sharpness and detecting noise levels.

[0065] For missing values ​​in the dataset, the system adopts a tiered processing strategy based on the missing proportion. When the missing proportion is less than 5%, the mean feature of the same type of material is used for imputation; if the missing proportion is between 5% and 20%, K-nearest neighbor interpolation based on the similarity of material composition and microstructure is used for imputation; and when the missing proportion exceeds 20%, the sample will be directly removed to ensure data quality.

[0066] To eliminate duplicate samples in the dataset, the system performs precise comparisons based on the sample's unique identifier. This identifier is typically composed of a combination of material type and key parameters. Once a duplicate is identified, the system automatically deletes subsequent duplicate samples and retains the first occurrence as the standard data.

[0067] To achieve standardized data management, the system automatically labels various types of collected data. For microstructure images, the system uses an image segmentation model to automatically identify and label the enhancement phase, matrix, and interface regions, while generating standardized annotation files. For numerical data, the system automatically adds labels in the format of "parameter name and unit" according to unified specifications, thereby ensuring that all data have clear physical meaning and consistent format.

[0068] To ensure data annotation quality, the system has a manual review mechanism. The specific process involves randomly selecting 10% of the automatically annotated data as a sample for manual review. Only when the annotation accuracy reaches or exceeds 95% is the batch of data approved. If the review result does not meet this standard, the model parameters must be adjusted and the annotation process must be repeated.

[0069] In the image processing stage, the system quantitatively characterizes the morphology and distribution of the enhancement phase. The key features extracted specifically include equivalent diameter, roundness, aspect ratio, volume fraction, and aggregation degree. Among these, the aggregation degree is defined by calculating the ratio of the actual aggregated region area to the theoretically completely dispersed region area, thus scientifically characterizing the particle aggregation state.

[0070] In the numerical feature extraction stage, the system standardizes and transforms different types of raw data. Specifically, atomic bond lengths are converted into relative bond lengths, i.e., the ratio of actual measured values ​​to standard reference values; material composition data are constructed into standardized mass fraction vectors. In addition, to further characterize the thermal response properties of materials, the system also derives the rate of change of thermophysical properties in different temperature ranges as supplementary features, such as calculating the relative change in thermal conductivity between 100 degrees Celsius and 25 degrees Celsius.

[0071] To eliminate the potential bias caused by the difference in magnitude between different numerical features during model training, this system standardizes all numerical features. This standardization process ensures that the mean of the numerical distribution of each feature is zero and the standard deviation is one, thereby guaranteeing that each feature has a balanced weight contribution during model training.

[0072] During the dataset partitioning phase, a stratified sampling method was used to randomly divide the overall data into a training set and a test set at an 8:2 ratio to ensure that the distribution of key attributes such as material type and temperature range remained consistent across the two sets. To further optimize the model training process, 10% of the data was extracted from the training set as a validation set. This validation set was primarily used for adjusting model hyperparameters and implementing an early stopping mechanism to prevent overfitting. Specifically, training was automatically terminated when the validation set loss function no longer decreased for ten consecutive training epochs.

[0073] Deep learning model training module: This module is the core computing unit of the system. It trains an improved multi-scale attention convolutional neural network (MSA-CNN) based on a standardized multi-scale feature dataset to generate a cross-scale thermal property prediction model. The core innovations of the module include a three-level attention layer, a cross-scale feature fusion unit, and an adaptive multi-scale coupling algorithm, which ensure that the model effectively utilizes multi-scale information and improves prediction accuracy.

[0074] Improved Multi-Scale Attention Convolutional Neural Network (MSA-CNN): The model is structurally built as a deep architecture comprising eight consecutive network layers. Its core design follows a technical roadmap of multi-scale data input, weighted processing via attention mechanisms, deep feature fusion, and convolutional processing, ultimately achieving accurate predictions. The specific parameter configurations for each layer are shown in Table 1. Table 1 The model introduces a dedicated three-level attention mechanism layer designed to dynamically evaluate and assign importance weights to features at the micro, meso, and macro scales. This mechanism adaptively highlights the scale information that has the most significant impact on the prediction of thermal properties, and its core weights are calculated using the following mathematical formula: ; in, For microscopic eigenvectors, For mesoscopic eigenvectors, For macroscopic eigenvectors; , These are the learnable parameters for the attention layer; , , The attention weights are for microscopic, mesoscopic, and macroscopic features, respectively, and satisfy the following conditions. During training, the weights are dynamically adjusted according to the prediction error; for example, when the prediction error of thermal conductivity is large, the weight of the microscopic scale is automatically increased.

[0075] The model incorporates a cross-scale feature fusion unit, which combines residual connectivity with feature concatenation to effectively address the information loss issue that may occur during multi-scale feature fusion. The specific processing flow is as follows: Step 1: Calculate weighted features, the formula is: ; Step 2: Feature concatenation, the formula is: ; Step 3: Residual connection, the formula is: ; in This is the multi-scale feature matrix of the input layer.

[0076] Step 4: Batch Normalization (BN) processing, the formula is: ; in, The batch characteristic mean, For batch characteristic variance, , These are the learnable parameters of the BN layer. (To prevent the denominator from being 0).

[0077] Loss function and training hyperparameters: The weighted mean square error (W-MSE) is used, and weights are assigned according to the importance of the thermophysical parameters, as shown in the formula: ; in, The number of samples in the training set; For the first i The first sample p Reference values ​​for various thermophysical parameters; For the first i The first sample p Predicted values ​​of various thermophysical parameters; For the first p The weighting coefficients of various thermal property parameters, and satisfying The default value is (thermal conductivity) (Coefficient of thermal expansion) (Specific heat capacity) Users can adjust the weight according to their actual needs.

[0078] Training hyperparameters: The Adam optimizer is configured with the following parameters: learning rate, first-order momentum coefficient, second-order momentum coefficient, and numerical stability parameter. Batch size 32, balancing training stability and computational resource consumption; The maximum number of iterations for model training is set to 200 epochs. Meanwhile, to effectively prevent overfitting, an early stopping mechanism is implemented as an intervention strategy during training. This mechanism continuously monitors the loss function on the validation set, and automatically terminates the training process if its value does not decrease for ten consecutive training epochs.

[0079] The fully connected layer 1 introduces Dropout regularization (Dropout probability = 0.5) and also uses L2 regularization (regularization coefficient) to constrain the model weight parameters, as shown in the formula: ; in, These are all the learnable weight parameters for the model.

[0080] Adaptive multi-scale coupling algorithm: This embodiment constructs a closed-loop workflow comprising five key stages: microscopic computation, mesoscopic transfer, macroscopic mapping, coupled verification, and parameter correction. The aim is to achieve efficient transfer of multi-scale information and progressive correction of prediction biases. The specific execution steps and related mathematical models of this workflow are described below.

[0081] Step 1 involves microscale calculations. First, an atomic model of the composite compound is constructed using molecular dynamics simulations to obtain the phonon scattering rate in atomic-level thermal transport parameters. This simulation uses the COMPASS force field to describe interatomic interactions, with a room temperature of 25°C and a total duration of 100 ps. This duration is divided into two phases: the first 50 ps are used to allow the system to reach equilibrium, and the latter 50 ps are used to collect effective thermal transport data.

[0082] Step 2 is the mesoscale analysis. In this step, a representative volume element model at the mesoscale is first constructed using the finite element method. Then, the phonon scattering rate calculated at the microscale is converted into phonon relaxation time, which is then substituted into the theoretical formula as a key parameter to finally calculate the equivalent thermal conductivity of the representative volume element, i.e., the mesoscale thermal conductivity.

[0083] Step 3 is the macroscopic-scale mapping. In this step, key features such as the thermal conductivity of representative volumetric units and the agglomeration degree of the reinforcing phase obtained from the mesoscopic-scale analysis are first fused with macroscopic-scale material composition data, such as the mass fraction of each component, to generate a comprehensive macroscopic feature vector. Subsequently, this feature vector is input into the MSA-CNN model for calculation, and finally, the predicted values ​​of the material's macroscopic thermal properties, including thermal conductivity, coefficient of thermal expansion, and specific heat capacity, are output.

[0084] Step 4 involves coupling verification and correction, calculating the error between the predicted macroscopic thermal properties and the reference values, using the following formula: ; in, This is a reference value for macroscopic thermal properties. These are predicted values ​​for macroscopic thermal properties.

[0085] like (Preset error threshold) then corrects the core parameters of the micro-mesoscopic coupling, including the interfacial thermal resistance coefficient. (Unit: m) 2 (⋅K / W), the corrected formula is: ; in, The interfacial thermal resistance coefficient before correction. This is the corrected interfacial thermal resistance coefficient. After correction, repeat steps 1-3 until... .

[0086] Thermophysical property parameter prediction module: This module serves as the system's application entry point. Users input multi-scale parameters of the composite compound to be predicted, and the module automatically calls the trained cross-scale thermophysical property prediction model, outputting the prediction results and error ranges for thermal conductivity, coefficient of thermal expansion, and specific heat capacity. Real-time prediction is supported, with a single response time ≤ 1 s.

[0087] Prediction process, parameter input: The system interface adopts a responsive design, ensuring accessibility across various terminal devices. The input interface layout is primarily divided into four functional areas: basic material information, microscopic parameters, mesoscopic parameters, and macroscopic parameters. To enhance user experience and data input accuracy, each parameter is accompanied by clear explanations and its value range; for example, the input range for the volume fraction of the reinforcing phase is explicitly limited to 0 < 0. f <0.6.

[0088] This embodiment provides a standardized interface input, supporting third-party systems to call it through a unified API. The request data format must explicitly include key fields such as material type, microscopic parameters, mesoscopic parameters, and macroscopic parameters. Among them, microscopic parameters cover characteristics such as grain size and interface thickness, mesoscopic parameters include structural information such as the agglomeration degree of the reinforcing phase, and macroscopic parameters involve the overall material composition such as the mass fraction of each component.

[0089] Next is parameter standardization. The numerical input parameters are standardized based on the feature mean and standard deviation of the model training set, using the following formula: ; The input classification parameters, such as material type, are converted into numerical vectors using one-hot encoding. For example, metal matrix composites are encoded as follows: Ceramic matrix composites are coded as .

[0090] The process involves several steps: first, loading the trained MSA-CNN model via the model service interface; second, inputting the standardized parameter vector into the model; and finally, obtaining the standardized predicted values ​​output by the model. .

[0091] Finally, the results are denormalized, restoring the standardized predicted values ​​to thermophysical property parameters with practical physical meaning. The formula is as follows: ; in, , These represent the mean and standard deviation of the corresponding thermophysical parameters in the training set, respectively.

[0092] Based on the prediction error distribution of the model test set, the error range under the 95% confidence interval is calculated using the following formula: ; in, This represents the standard deviation of the prediction error for the corresponding thermophysical parameters of the test set.

[0093] Output results: The interface outputs the predicted results of thermal property parameters in a clear tabular format, including key information such as the parameter name, predicted value, corresponding unit, and error range. This table supports exporting to common office software formats such as Excel, facilitating subsequent data analysis and archiving.

[0094] The interface outputs prediction results in a standardized format, mainly including fields such as prediction timestamp, material type identifier, and prediction result list. The prediction result list is an array structure, and each record contains complete information such as parameter name, predicted value, unit of measurement, and error range, which facilitates structured parsing and subsequent processing by third-party systems.

[0095] Intelligent optimization module: This module, based on user-defined target thermophysical parameters, employs a hybrid optimization strategy combining genetic algorithms and gradient descent to inversely optimize the composition ratio of composite compounds (such as the volume fraction of the reinforcing phase). f ) and microstructure parameters (such as grain size D, interface thickness) t It outputs the globally optimal material formulation and microstructure scheme.

[0096] Optimization objectives and constraints: (1) Taking thermal conductivity close to the target value and thermal expansion coefficient lower than the target value as the core objectives, the objective function is optimized by introducing target weights to balance priorities. The formula is: ; in, To enhance the phase volume fraction (optimization variable 1); To enhance phase grain size (unit: nm, optimization variable 2); Thickness of the matrix-reinforcement phase interface (unit: nm, optimization variable 3); Target thermal conductivity; The target thermal expansion coefficient; , Parameter combinations The corresponding predicted values ​​for thermal conductivity and coefficient of thermal expansion; , The target weight is, and satisfies The default value is (Prioritize thermal conductivity) (Taking into account the coefficient of thermal expansion), users can adjust it as needed.

[0097] (2) Based on the feasibility of the composite compound preparation process and the stability of material properties, the constraint range of the optimization variables is set: ; Hybrid optimization strategy process: (1) Global search using genetic algorithm (number of iterations: 50 generations): Step 1: Initialize the population. Randomly generate 100 parameter combinations that meet the constraints. As the initial population, the generation process uses uniform random sampling to ensure that the parameters are evenly distributed within the constraints.

[0098] Step 2: Fitness Assessment. For each parameter combination in the population, calculate the fitness value (to assess the optimization effect; a higher value indicates better performance). The formula is: ; in, For fitness value, To optimize the objective function value, the thermophysical property parameter prediction module needs to be called during the evaluation to obtain... and .

[0099] Step 3: Genetic manipulation.

[0100] First, a roulette wheel selection method is used, where the probability of an individual being selected is positively correlated with its fitness value, as shown in the formula: ; in, For the first i The probability of an individual's choice For the first i The fitness value of each individual was used to select 100 individuals to enter the parent population.

[0101] Secondly, a single-point crossover method is used. The parent population is randomly paired (50 pairs in total), and a crossover point such as "enhancing phase volume fraction" is randomly selected. The parameters after the crossover point of the paired individuals are exchanged to generate offspring individuals. The crossover probability is set to 0.8.

[0102] Finally, for the offspring individuals after crossover, each optimization variable undergoes random mutation with a probability of 0.05. Within the constraints, the parameter values ​​are fine-tuned to introduce population diversity and avoid local optima.

[0103] Step 4: Population Renewal. Replace the initial population with the offspring population after genetic manipulation, and repeat steps 2-3.

[0104] Step 5: Global Optimal Candidate Selection. After 50 generations of iteration, the top 10 individuals with the highest fitness values ​​are selected from the last generation of the population as the global optimal candidate set, and then the local fine-tuning stage begins.

[0105] (2) Local fine-tuning of gradient descent (number of iterations: 50): Step 1: Initial Parameter Setting. Using each individual in the global optimal candidate set as the initial parameter, set the learning rate. To ensure the stability of fine-tuning.

[0106] Step 2: Gradient Calculation. The gradient of the objective function with respect to each optimization variable is calculated using the numerical difference method, where the gradient direction is... J The formula for the direction of fastest descent is: ; in, To optimize variables ( f , D or t ), For small perturbation values ​​(take) 1% of the constraint range, such as f of ).

[0107] Step 3: Parameter Update. Update the optimization variables according to the gradient descent direction, using the following formula: ; If the updated parameters exceed the constraint range, they will be automatically adjusted to the constraint boundary values ​​(e.g., ...). When, set as ).

[0108] Step 4: Iterative fine-tuning. Repeat steps 2-3 for 50 iterations, and calculate the fitness value of the fine-tuned parameter combination.

[0109] Step 5: Determine the optimal parameters. Compare the fitness values ​​of the 10 candidate individuals after fine-tuning, and select the parameter combination with the highest fitness value as the final optimization scheme.

[0110] (3) Convergence Judgment: During the global search phase of the genetic algorithm, if the average fitness value of the population changes by less than 1% over 5 consecutive generations, that is: ; in For the first If the average fitness value of the generation population is reached, the genetic algorithm iteration will end prematurely and enter the gradient descent local fine-tuning stage.

[0111] Results visualization and output module: This module presents the predicted thermal properties and optimization process in intuitive charts and graphs, and generates standardized analysis reports to facilitate user understanding, verification, and application.

[0112] This embodiment also provides a method for predicting and optimizing the cross-scale thermal property parameters of composite compounds based on deep learning, such as... Figure 2 As shown, it specifically includes: Collect multi-scale data of composite compounds and preprocess them to obtain a standardized multi-scale feature dataset. A cross-scale thermal property prediction model is trained using the standardized multi-scale feature dataset to obtain the trained cross-scale thermal property prediction model, wherein the cross-scale thermal property prediction model is an improved multi-scale attention convolutional neural network. Acquire multi-scale data of the composite compound to be tested, input the multi-scale data of the composite compound to be tested into the trained cross-scale thermophysical property prediction model, and obtain the thermophysical property prediction results; Based on the predicted thermophysical properties, the composition ratio and microstructure parameters of the composite compound are optimized in reverse, and the optimized material formulation and microstructure parameters are output. The predicted thermophysical properties or optimized material formulations and microstructure parameters are visualized, and an exportable analysis report is generated.

[0113] Specifically, S1: Collect multi-scale data of composite compounds through the data acquisition and preprocessing module, perform cleaning, labeling and feature extraction, and generate a standardized multi-scale feature dataset; S2: Train an improved multi-scale attention convolutional neural network (MSA-CNN) based on a standardized multi-scale feature dataset using a deep learning model training module to generate a cross-scale thermal property prediction model; S3: Input the multi-scale parameters of the composite compound to be predicted through the thermophysical parameter prediction module, and call the cross-scale thermophysical property prediction model to output the thermophysical property prediction results. S4: Input the target thermophysical property parameters through the intelligent optimization module, and output the optimized material formulation and microstructure parameters using a hybrid optimization strategy of genetic algorithm-gradient descent. S5: Display prediction results and optimization paths through the results visualization and output module, and generate analysis reports; S6: Prepare composite compound samples based on the analysis report and conduct experimental verification. If the verification results meet the target requirements, the research and development is completed; if not, return to S1 to adjust the data acquisition range or S2 to optimize the model parameters, and repeat steps S1-S5.

[0114] More specifically, such as Figure 2 As shown: S1. Data Acquisition and Preprocessing: The data acquisition and preprocessing module is activated. The multi-scale data acquisition range is determined according to the type of target compound. After the data is acquired, it is cleaned, labeled, feature extracted, and standardized to generate training set, test set, and validation set, which are divided in an 8:2:0.8 ratio.

[0115] S2, Model Training: Load the dataset generated by S1, configure the MSA-CNN model structure parameters and training hyperparameters, and train the model using the W-MSE loss function and Adam optimizer. During training, enable the adaptive multi-scale coupling algorithm to dynamically correct cross-scale parameter bias. When the validation set loss does not decrease for 10 consecutive generations or the number of iterations reaches 200 generations, stop training. Validate the model performance on the test set. If the prediction error is ≤8%, it is considered qualified. Output the trained cross-scale thermal property prediction model.

[0116] S3. Prediction of thermal properties: The thermal property parameter prediction module is started. The user inputs the multi-scale parameters of the material to be predicted. After standardizing the input parameters, the module calls the model generated in S2 and outputs the predicted thermal property values ​​and error range. The user judges whether the prediction results meet the target requirements: if they do, proceed to S5; otherwise, proceed to S4.

[0117] S4, Intelligent Optimization: Start the intelligent optimization module and input the target thermal property parameters and weights; adopt the genetic algorithm-gradient descent hybrid optimization strategy, first use the genetic algorithm (50 generations of iterations) to screen the globally optimal candidate parameters, and then use gradient descent (50 iterations) to fine-tune the local parameters; call the prediction interface of S3 to verify the optimization scheme. If the target requirements are met, proceed to S5; otherwise, adjust the target weights or constraints and re-execute S4.

[0118] S5. Results Display and Report Generation: The system launches the results visualization and output module, generating thermal property prediction trend charts, optimization iteration curves, and multi-scale parameter correlation thermograms; it also automatically generates PDF analysis reports, which users can download or send to a specified email address.

[0119] S6, Subsequent Iterations: If you need to adjust the optimization direction or improve the model accuracy, you can return to S1 to adjust the data processing strategy, such as supplementing data for a specific parameter range, or return to S2 to optimize the model parameters, such as adjusting the number of network layers and the learning rate, and then re-execute S1-S5.

[0120] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A deep learning-based system for predicting and optimizing the cross-scale thermal properties of composite compounds, characterized in that, include: The system includes a data acquisition and preprocessing module, a model training module, a thermophysical parameter prediction module, an intelligent optimization module, and a results visualization module. The data acquisition and preprocessing module is used to acquire multi-scale data of composite compounds and perform preprocessing to obtain a standardized multi-scale feature dataset. The model training module is used to train a cross-scale thermal property prediction model using the standardized multi-scale feature dataset, and obtain the trained cross-scale thermal property prediction model. The thermophysical parameter prediction module is used to input multi-scale data of the composite compound to be tested into the trained cross-scale thermophysical property prediction model to obtain thermophysical property prediction results. The intelligent optimization module is used to reverse-optimize the composition ratio and microstructure parameters of the composite compound based on the thermophysical property prediction results, and output the optimized material formulation and microstructure parameters. The results visualization module is used to visualize the predicted thermophysical properties or optimized material formulations and microstructure parameters, and generate exportable analysis reports.

2. The deep learning-based multi-scale thermophysical property parameter prediction and optimization system for composite compounds according to claim 1, characterized in that, The multi-scale data of the composite compound includes: microstructure images, composition data, and experimental thermophysical property data of the composite compound; The microstructure images include scanning electron microscope images and transmission electron microscope images; The component data includes the mass fraction of each component; The experimental thermophysical data include thermal conductivity, coefficient of thermal expansion, and specific heat capacity.

3. The deep learning-based multi-scale thermophysical property parameter prediction and optimization system for composite compounds according to claim 2, characterized in that, The cross-scale thermal property prediction model is an improved multi-scale attention convolutional neural network, which includes an input layer, a three-level attention layer, a cross-scale feature fusion unit, a convolutional layer, and a fully connected layer. The input layer is used to receive a microscopic feature vector, a mesoscopic feature vector, and a first macroscopic feature vector, wherein the microscopic feature vector is obtained based on the microstructure image, the mesoscopic feature vector is obtained based on the composition data, and the first macroscopic feature vector is obtained based on the experimental thermophysical property data. The three-level attention layer is used to calculate attention weights based on the input feature vectors at each scale. The cross-scale feature fusion unit is used to perform weighted fusion of multi-scale features according to the calculated attention weights to obtain fused features; The convolutional layer is used to extract local correlation features and global correlation features based on the fused features; The fully connected layer is used to output predicted values ​​of thermal property parameters based on the local correlation features and the global correlation features.

4. The deep learning-based multi-scale thermophysical property parameter prediction and optimization system for composite compounds according to claim 3, characterized in that, The attention weights are calculated based on the input feature vectors at various scales, including: ； in, For microscopic eigenvectors, For mesoscopic eigenvectors, This is the first macroscopic eigenvector. , For attention layer parameters, , , These are the attention weights for microscopic, mesoscopic, and macroscopic features, respectively. , For the micro-feature attention layer bias parameters, For the bias parameters of the mesoscopic feature attention layer, These are the bias parameters for the macroscopic feature attention layer. As a scale identifier variable, For microscale identification, For mesoscale identification, For macroscopic scale identification, This is a general-scale feature vector.

5. The deep learning-based multi-scale thermophysical property parameter prediction and optimization system for composite compounds according to claim 4, characterized in that, The multi-scale features are weighted and fused based on the calculated attention weights to obtain the fused features, including: The feature vectors at each scale are weighted according to the attention weights output by the three-level attention layer to obtain the weighted micro-feature vector, meso-feature vector and first macro-feature vector; The weighted three-scale feature vectors are concatenated to obtain the concatenated feature vector; The concatenated feature vector is residually concatenated with the original features of the input layer to obtain a residual feature vector; The residual feature vectors are batch normalized to obtain the fused features.

6. The deep learning-based multi-scale thermophysical property parameter prediction and optimization system for composite compounds according to claim 3, characterized in that, The system also includes an adaptive multi-scale coupling algorithm; Phonon scattering rate was obtained from molecular dynamics simulations; The phonon scattering rate is converted into phonon relaxation time, and the mesoscopic thermal conductivity is calculated based on the phonon relaxation time. A second macroscopic feature vector is generated based on mesoscopic thermal conductivity and macroscopic composition data. The second macroscopic feature vector is then input into the cross-scale thermal property prediction model to obtain the predicted macroscopic thermal property values. The prediction error is calculated based on the macroscopic thermophysical property predictions and experimental reference values. When the prediction error exceeds the preset threshold, the interfacial thermal resistance coefficient in the micro-mesoscopic coupling parameters is adjusted according to the error value, and the molecular dynamics simulation and mesoscopic analysis are re-executed until the prediction error is no greater than the preset threshold.

7. The deep learning-based multi-scale thermophysical property parameter prediction and optimization system for composite compounds according to claim 6, characterized in that, Adjusting the interfacial thermal resistance coefficient in the micro-mesoscopic coupling parameters based on the error value includes: ； in, The interfacial thermal resistance coefficient before correction. This is the corrected interfacial thermal resistance coefficient. The error between the macroscopic experimental thermophysical parameters and the predicted values, and , For macroscopic experimental thermal conductivity, This is the predicted value for macroscopic thermal conductivity.

8. The deep learning-based multi-scale thermophysical property parameter prediction and optimization system for composite compounds according to claim 1, characterized in that, The intelligent optimization module employs a hybrid optimization strategy of genetic algorithm and gradient descent for optimization. The optimization strategy employing a hybrid approach of genetic algorithm and gradient descent includes: An optimization objective function is constructed based on the predicted thermal properties and constraints. The population is then initialized based on the optimization objective function, and the fitness values ​​of each parameter combination in the population are calculated. Based on the fitness value, selection, crossover, and mutation operations are performed on the population to generate a new generation of population; The global optimal candidate parameter set is selected based on the fitness values ​​of the new generation population; Based on the optimal candidate parameter set, the gradient descent method is used to locally fine-tune each parameter, and the gradient value of the optimization objective function with respect to each parameter is calculated. Update the parameters based on the gradient value until the convergence condition is met, and output the optimized material formulation and microstructure parameters.

9. The deep learning-based multi-scale thermophysical property parameter prediction and optimization system for composite compounds according to claim 8, characterized in that, The optimization objective function is: ； in, For the target thermal conductivity, The target thermal expansion coefficient, , Parameter combinations The corresponding predicted values ​​of thermal conductivity and coefficient of thermal expansion, To increase the phase volume fraction, To enhance phase grain size, For the thickness of the matrix-reinforcement phase interface, , The target weight.

10. A deep learning-based method for predicting and optimizing the cross-scale thermophysical parameters of composite compounds, used to implement the system as described in any one of claims 1-9, characterized in that, include: Collect multi-scale data of composite compounds and preprocess them to obtain a standardized multi-scale feature dataset. A cross-scale thermal property prediction model is trained using the standardized multi-scale feature dataset to obtain the trained cross-scale thermal property prediction model, wherein the cross-scale thermal property prediction model is an improved multi-scale attention convolutional neural network. Acquire multi-scale data of the composite compound to be tested, input the multi-scale data of the composite compound to be tested into the trained cross-scale thermophysical property prediction model, and obtain the thermophysical property prediction results; Based on the predicted thermophysical properties, the composition ratio and microstructure parameters of the composite compound are optimized in reverse, and the optimized material formulation and microstructure parameters are output. The predicted thermophysical properties or optimized material formulations and microstructure parameters are visualized, and an exportable analysis report is generated.

Images