All-silicon carbide double-sided heat dissipation module packaging method based on gradient thermal resistance optimization

Abstract

The invention provides an all-silicon carbide double-sided heat dissipation module packaging method based on gradient thermal resistance optimization, and belongs to the technical field of electronic component packaging. Establishing a thermal resistance characteristic curve, and calculating a gradient thermal resistance vector to determine a heat flow path; constructing a heat dissipation temperature distribution matrix; obtaining thermal diffusion data; heat dissipation path thermal resistance is calculated by combining the thermal diffusion matrix and the temperature distribution matrix, and a thermal resistance sparse matrix is established and is dimensionally reduced into a thermal resistance low-dimensional matrix; calculating an optimal heat dissipation path parameter set; inputting the parameter set into a thermal resistance prediction deep attention network model for verification and fine tuning, and generating a final thermal resistance distribution scheme; a module structure and a material ratio are designed according to the scheme to form a sandwich structure; a silver paste sintering process is adopted to realize low thermal resistance connection and complete packaging, and the technical problems of non-uniform thermal resistance distribution and difficulty in optimizing a heat flow path in a double-sided heat dissipation structure of a silicon carbide power module in the prior art are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of electronic component packaging, and more particularly, relates to a packaging method for a fully silicon carbide double-sided heat dissipation module based on gradient thermal resistance optimization. Background Art

[0002] Silicon carbide power modules are widely used in high power density application scenarios such as electric vehicles and new energy power conversion due to their high temperature tolerance and excellent thermal conductivity. The heat dissipation packaging of traditional silicon carbide power modules generally adopts a single-sided heat dissipation structure, where the chip is directly welded or sintered onto a copper substrate, and then heat is dissipated through radiators, air cooling or liquid cooling. With the continuous increase in power density, the single-sided heat dissipation structure has become difficult to meet the thermal management requirements of highly integrated modules, and thus the double-sided heat dissipation structure has become the focus of research.

[0003] However, the existing silicon carbide double-sided heat dissipation module packaging technology faces problems such as uneven thermal resistance distribution, complex heat dissipation paths, and uncontrollable heat transfer in practical applications. Since the heat flow in the double-sided heat dissipation structure transfers in both the upward and downward directions simultaneously, a complex heat conduction network is formed inside the module, greatly increasing the difficulty of optimizing the heat flow path. Traditional thermal resistance analysis methods are usually based on simplified models, ignoring the mutual influence between interface thermal resistance and internal heat flow distribution, and unable to accurately describe the dynamic characteristics of heat flow in a double-sided heat dissipation environment.

[0004] Currently, the industry lacks a systematic optimization method that can simultaneously consider the internal heat transfer conduction and interface heat transfer characteristics of the module, making it difficult to achieve global optimization and precise control of the thermal resistance distribution, which has become a key technical bottleneck restricting the improvement of the packaging performance of silicon carbide double-sided heat dissipation modules. Especially in high power density applications, the problem of local overheating caused by uneven heat flux density is more prominent, and there is an urgent need for a global optimization packaging method based on the gradient thermal resistance theory. That is to say, there are technical problems in the prior art such as uneven thermal resistance distribution and difficult optimization of heat flow paths in the double-sided heat dissipation structure of silicon carbide power modules. Summary of the Invention

[0005] In view of this, the present invention provides a packaging method for a fully silicon carbide double-sided heat dissipation module based on gradient thermal resistance optimization, which can solve the technical problems in the prior art such as uneven thermal resistance distribution and difficult optimization of heat flow paths in the double-sided heat dissipation structure of silicon carbide power modules.

[0006] The present invention is implemented as follows: The present invention provides a packaging method for a fully silicon carbide double-sided heat dissipation module based on gradient thermal resistance optimization, including: constructing an upper-layer function and a lower-layer function of heat transfer conduction for the silicon carbide double-sided heat dissipation module to form a double-layer heat transfer conduction function game model; establishing a thermal resistance characteristic curve based on this model, calculating the gradient thermal resistance vector, and determining the main heat transfer conduction path and the secondary path; measuring the key point temperatures according to the heat flow path to construct a heat dissipation temperature distribution matrix; obtaining surface heat diffusion data by using thermal infrared imaging to construct a heat diffusion distribution matrix; calculating the thermal resistance of the heat dissipation path by combining the two matrices, establishing a thermal resistance sparse matrix, and screening key heat dissipation nodes; reducing the dimensionality of the thermal resistance sparse matrix to a low-dimensional thermal resistance matrix by singular value decomposition to extract key heat dissipation characteristics; calculating the optimal heat dissipation path parameter set by applying the thermal resistance gradient optimization function; inputting the parameter set into the thermal resistance prediction deep attention network model for verification and fine-tuning to generate the final thermal resistance distribution scheme; designing the module structure and material ratio according to the scheme to form a sandwich structure; and realizing low thermal resistance connection by using a silver paste sintering process to complete the packaging.

[0007] Among them, the upper-layer function of heat transfer conduction specifically refers to a mathematical function that describes the heat transfer conduction behavior inside the silicon carbide double-sided heat dissipation module, including expressions of heat flux density distribution and temperature gradient relationship; the lower-layer function of heat transfer conduction specifically refers to a mathematical function that describes the heat transfer conduction behavior at the interface between the silicon carbide double-sided heat dissipation module and the heat dissipation substrate, reflecting the relationship between the interface thermal resistance and the heat flux density.

[0008] Among them, the double-layer heat transfer conduction function game model specifically refers to an optimization model formed by combining the upper-layer function and the lower-layer function of heat transfer conduction, which is a mathematical expression for finding the global optimal heat dissipation path through iterative calculation.

[0009] Among them, the gradient thermal resistance vector specifically refers to a vector that describes the gradient of the thermal resistance in the spatial distribution, and its direction points to the direction where the thermal resistance increases fastest, which is used to guide the heat flow optimization path.

[0010] Among them, the heat dissipation temperature distribution matrix specifically refers to a two-dimensional data matrix that characterizes the temperature distribution of each point on the surface of the silicon carbide double-sided heat dissipation module, which is used for hot spot identification and temperature field analysis.

[0011] Among them, the heat diffusion distribution matrix specifically refers to a two-dimensional data matrix that characterizes the spatial distribution of the heat diffusion coefficient on the surface of the silicon carbide double-sided heat dissipation module, reflecting the spatial variation of the material's thermal conductivity; the thermal resistance sparse matrix specifically refers to a matrix that describes the internal thermal network structure of the silicon carbide double-sided heat dissipation module, with most elements being zero, and only the thermal resistance values between the key heat conduction path nodes are recorded.

[0012] Among them, the thermal resistance gradient optimization function is used to perform non-linear optimization calculations on the low-dimensional thermal resistance matrix, find the optimal thermal resistance distribution scheme. The inputs include the low-dimensional thermal resistance matrix, the temperature gradient constraint parameters obtained from the heat dissipation temperature distribution matrix, the upper limit parameters of the heat flux density obtained from the upper-layer heat conduction function, the material thermal conductivity vector of the silicon carbide double-sided heat dissipation module, and the interface contact thermal resistance parameters obtained from the lower-layer heat conduction function. The output is the optimal heat dissipation path parameter set, which contains the optimized thermal resistance distribution scheme and the corresponding temperature uniformity index.

[0013] Among them, the specific structure of the deep attention network model for thermal resistance prediction is a deep convolutional neural network based on the multi-head self-attention mechanism, including an encoder-decoder architecture. The encoder consists of three layers of multi-head self-attention layers and two layers of feed-forward neural networks, which are used to extract the thermal resistance distribution features. The weight dimension of each self-attention head is determined by the eigenvalues of the heat dissipation temperature distribution matrix. The decoder consists of two layers of cross-attention layers and one layer of feed-forward neural network, which is used to generate the optimized thermal resistance distribution scheme.

[0014] Among them, the low-dimensional thermal resistance matrix specifically refers to the key feature matrix extracted from the thermal resistance sparse matrix through singular value decomposition technology, which reduces the computational complexity while retaining the main heat conduction information.

[0015] Among them, the sandwich structure specifically refers to a stacked heat dissipation structure formed by placing the silicon carbide double-sided heat dissipation module between the upper and lower heat dissipation substrates, realizing the two-way heat dissipation function.

[0016] Compared with the prior art, a packaging method for a fully silicon carbide double-sided heat dissipation module based on gradient thermal resistance optimization provided by the present invention. The packaging method for a fully silicon carbide double-sided heat dissipation module based on gradient thermal resistance optimization proposed by the present invention realizes the precise control and optimized design of the heat flow path by constructing a double-layer heat conduction function game model, combining gradient thermal resistance vector analysis and thermal resistance sparse matrix dimensionality reduction technology, and solves the problem of uneven thermal resistance distribution of the silicon carbide double-sided heat dissipation module.

[0017] This method first forms a double-layer heat conduction game model by establishing the upper and lower layer heat conduction functions to accurately capture the dynamic characteristics of the heat flow distribution; secondly, it realizes the quantitative characterization of the main and secondary heat flow paths through the calculation of the gradient thermal resistance vector and the construction of the temperature distribution matrix; then, it extracts the key heat dissipation features and reduces the computational complexity through the construction of the thermal resistance sparse matrix and singular value decomposition; finally, it combines the deep attention network model for thermal resistance prediction for verification and optimization to generate the optimal heat dissipation structure design scheme, effectively solving the problem that the heat flow path is difficult to accurately control in the traditional method.

[0018] Through a multi-level thermal resistance analysis and optimization method, the present invention successfully solves the core technical problems of uneven thermal resistance distribution and difficult optimization of heat flow paths in a double-sided heat dissipation structure, realizes precise control and global optimization of heat conduction, and provides a systematic solution for the high-performance heat dissipation packaging of silicon carbide power modules. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flowchart of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0021] As Figure 1 shown, it is a flowchart of a method for packaging a fully silicon carbide double-sided heat dissipation module based on gradient thermal resistance optimization provided by the present invention. The method includes the following steps:

[0022] S01. Construct an upper-layer function of heat conduction and a lower-layer function of heat conduction for the silicon carbide double-sided heat dissipation module to form a two-layer heat conduction function game model;

[0023] S02. Based on the two-layer heat conduction function game model, establish a thermal resistance characteristic curve of the silicon carbide module, calculate the gradient thermal resistance vector, and determine the main heat conduction path and the secondary path;

[0024] S03. Measure the temperatures of each key point of the silicon carbide module according to the main heat conduction path and the secondary path, construct a heat dissipation temperature distribution matrix, and quantify the temperature gradient field;

[0025] S04. Based on the heat dissipation temperature distribution matrix, use thermal infrared imaging technology to obtain the surface heat diffusion data of the silicon carbide module, and construct a heat diffusion distribution matrix;

[0026] S05. Combine the heat diffusion distribution matrix and the heat dissipation temperature distribution matrix, use the principle of heat flow conservation to calculate the thermal resistance of the heat dissipation path, establish a thermal resistance sparse matrix, and screen the key heat dissipation nodes;

[0027] S06. Based on the key heat dissipation nodes, reduce the dimension of the thermal resistance sparse matrix to a thermal resistance low-dimensional matrix by singular value decomposition, and extract the key heat dissipation features;

[0028] S07. According to the key heat dissipation features, apply a thermal resistance gradient optimization function to calculate the thermal resistance low-dimensional matrix, and obtain an optimal heat dissipation path parameter set;

[0029] S08. Input the optimal heat dissipation path parameter set into a pre-trained deep attention network model for thermal resistance prediction for verification and fine-tuning to generate a final thermal resistance distribution scheme;

[0030] S09. Design the structural dimensions and material ratios of the silicon carbide double-sided heat dissipation module according to the final thermal resistance distribution scheme, and combine the silicon carbide double-sided heat dissipation module with the upper and lower heat dissipation substrates respectively to form a sandwich structure;

[0031] S10. Based on the sandwich structure, use the silver paste sintering process to achieve low thermal resistance connection between the silicon carbide double-sided heat dissipation module and the heat dissipation substrate to complete the encapsulation.

[0032] Among them, the upper-layer heat transfer function specifically refers to a mathematical function that describes the internal heat transfer behavior of the silicon carbide double-sided heat dissipation module, including expressions of heat flux density distribution and temperature gradient relationship.

[0033] Among them, the lower-layer heat transfer function specifically refers to a mathematical function that describes the interfacial heat transfer behavior between the silicon carbide double-sided heat dissipation module and the heat dissipation substrate, reflecting the relationship between interfacial thermal resistance and heat flux density.

[0034] Among them, the double-layer heat conduction function game model specifically refers to an optimization model formed by combining the upper-layer heat transfer function and the lower-layer heat transfer function, and is a mathematical expression for finding the global optimal heat dissipation path through iterative calculation.

[0035] Among them, the gradient thermal resistance vector specifically refers to a vector that describes the spatial distribution gradient of thermal resistance, and its direction points to the direction where the thermal resistance increases fastest, and is used to guide the heat flow optimization path.

[0036] Among them, the heat dissipation temperature distribution matrix specifically refers to a two-dimensional data matrix that characterizes the temperature distribution of each point on the surface of the silicon carbide double-sided heat dissipation module, and is used for hot spot identification and temperature field analysis.

[0037] Among them, the heat diffusion distribution matrix specifically refers to a two-dimensional data matrix that characterizes the spatial distribution of the thermal diffusion coefficient on the surface of the silicon carbide double-sided heat dissipation module, reflecting the spatial variation of the material's thermal conductivity.

[0038] Among them, the thermal resistance sparse matrix specifically refers to a matrix that describes the internal thermal network structure of the silicon carbide double-sided heat dissipation module, most of whose elements are zero, and only the thermal resistance values between the key heat conduction path nodes are recorded.

[0039] Among them, the low-dimensional thermal resistance matrix specifically refers to a key feature matrix extracted from the thermal resistance sparse matrix through singular value decomposition technology, which reduces the computational complexity while retaining the main heat conduction information.

[0040] Among them, the thermal resistance gradient optimization function is used to perform non-linear optimization calculations on the low-dimensional thermal resistance matrix to find the optimal thermal resistance distribution scheme. The inputs include the low-dimensional thermal resistance matrix, the temperature gradient constraint parameters obtained from the heat dissipation temperature distribution matrix, the upper limit parameters of the heat flux density obtained from the upper layer function of heat transfer conduction, the material thermal conductivity vector of the silicon carbide double-sided heat dissipation module, and the interface contact thermal resistance parameters obtained from the lower layer function of heat transfer conduction. The output is the optimal heat dissipation path parameter set, which includes the optimized thermal resistance distribution scheme and the corresponding temperature uniformity index.

[0041] Among them, the specific structure of the deep attention network model for thermal resistance prediction is a deep convolutional neural network based on the multi-head self-attention mechanism, including an encoder-decoder architecture. The encoder consists of three layers of multi-head self-attention layers and two layers of feed-forward neural networks, which are used to extract the thermal resistance distribution features. The weight dimension of each self-attention head is determined by the eigenvalue of the heat dissipation temperature distribution matrix. The decoder consists of two layers of cross-attention layers and one layer of feed-forward neural networks, which are used to generate the optimized thermal resistance distribution scheme. Among them, the self-attention sparsity parameter is determined by the amplitude distribution function of the gradient thermal resistance vector. The final output layer of the deep convolutional neural network is a fully connected layer, and the activation function uses the rectified linear unit function. The loss function combines the mean square error and the heat flux non-uniformity penalty term.

[0042] Among them, the steps for establishing the training data set in the training process of the deep attention network model for thermal resistance prediction specifically include collecting the heat flux density distribution data of silicon carbide heat dissipation modules with various different structures, measuring the corresponding temperature field distribution through a high-precision thermal imager, calculating the actual thermal resistance distribution according to Fourier's law of heat conduction as the label data, normalizing the data and dividing it into a training set, a validation set, and a test set. The training set accounts for 70% of the total data, the validation set accounts for 15%, and the test set accounts for 15%. At the same time, data augmentation processing is performed, including rotation transformation, reflection transformation, and scale transformation, expanding the number of training samples to 4 times the original data set. Finally, all samples are converted into a tensor format.

[0043] Among them, the steps for training the deep attention network model for thermal resistance prediction specifically include initializing the network weight parameters, randomly initializing the weights of the convolutional layer using the standard normal distribution, initializing the weights of the self-attention layer using the uniform distribution, setting the batch size to 32, using the Adam optimizer with a learning rate of 0.001, using a learning rate warm-up strategy in the initial stage of training, gradually increasing the learning rate to the set value within the first 10 training epochs, and then using the cosine annealing learning rate scheduling strategy. A total of 200 epochs are trained. The performance of the model is evaluated on the validation set every 5 epochs and the best performing checkpoint is saved. When the performance of the validation set no longer improves for 20 consecutive epochs, the early stopping strategy is applied. Finally, the model with the smallest thermal resistance prediction error and the highest temperature uniformity on the test set is selected as the final model.

[0044] Among them, the final thermal resistance distribution scheme specifically refers to the optimal thermal resistance distribution design scheme inside the silicon carbide double-sided heat dissipation module and at the interface with the heat dissipation substrate, which is determined after verification and fine-tuning by the thermal resistance prediction deep attention network model, and is used to guide the subsequent packaging structure design.

[0045] Among them, the sandwich structure specifically refers to a stacked heat dissipation structure formed by placing the silicon carbide double-sided heat dissipation module between the upper and lower heat dissipation substrates, realizing the two-way heat dissipation function.

[0046] The following will describe the specific implementation manners of the above steps in detail.

[0047] The specific implementation manner of step S01 is to construct a heat transfer conduction function system for the silicon carbide double-sided heat dissipation module. First, based on Fourier's law of heat conduction, establish the upper-layer function of heat transfer conduction where k(T) is the thermal conductivity of the silicon carbide material, is the temperature gradient; at the same time, establish the lower-layer function of heat transfer conduction Q 下 (x, y, T) = h(T)(T 模块 - T 基板 ), where h(T) is the interfacial heat transfer coefficient, T 模块 is the module temperature, and T 基板 is the substrate temperature. Then combine the upper and lower layer functions to form a two-layer heat conduction function game model F(Q 上 , Q 下 ), and this model is solved by minimizing the objective function J = ∫(T max - T min ) 2 dxdy + α∫Q 2 dxdy, where α is the weight coefficient, usually taking a value of 0.01 - 0.05. The purpose of this step is to establish a mathematical description of heat transfer conduction and lay a theoretical foundation for subsequent thermal resistance optimization.

[0048] The specific implementation manner of step S02 is to establish a thermal resistance characteristic curve based on the two-layer heat conduction function game model. First, solve the model equation by the finite element method to obtain the temperature distribution T(x, y, z) and the heat flux density distribution Q(x, y, z); then calculate the thermal resistance R(x, y, z) = ΔT / Q and establish a characteristic curve of the thermal resistance varying with temperature and position R =

[0049] f(T, x, y, z); then calculate the gradient thermal resistance vector This vector points in the direction where the thermal resistance increases fastest; finally, based on the heat flow line density and the direction of the gradient thermal resistance vector, the main heat flow path (the path where the magnitude of the gradient thermal resistance vector is less than 0.02 K / (W·m)) and the secondary path (the path where the magnitude of the gradient thermal resistance vector is between 0.02 and 0.05 K / (W·m)) are determined. The purpose of this step is to quantify the characteristics of the thermal resistance distribution and provide a basis for subsequent hot spot identification.

[0050] The specific implementation of step S03 is to measure and construct a heat dissipation temperature distribution matrix. First, key measurement points are selected on the main path and the secondary path, and the spacing between key points is usually 3 - 5 mm; then, the temperature of each point is measured using a high-precision thermocouple (with an accuracy better than 0.1 °C) under different power loads; then, the bicubic spline interpolation algorithm is used to generate a complete temperature distribution matrix T 分布 [i][j], and the matrix size is usually 107×78 (covering the entire module surface); finally, the temperature gradient field is calculated for quantifying the temperature change rate. The purpose of this step is to obtain the actual temperature distribution data of the module and provide a basis for heat diffusion analysis.

[0051] The specific implementation of step S04 is to obtain heat diffusion data using thermal infrared imaging technology. First, an infrared thermal imager with a resolution of not less than 640×480 pixels and a temperature measurement accuracy better than 0.05 °C is used to scan the surface of the module under the condition of a constant ambient temperature (25 ± 1 °C); then, the pulse heat source response method is applied, and the module is excited by an excitation signal (a 10 - 50 ms square wave pulse), and the temperature transient response is recorded; then, the heat diffusion equation is used to analyze the temperature change rate, where α is the thermal diffusion coefficient; finally, a heat diffusion distribution matrix D[i][j] is constructed to characterize the thermal diffusion coefficient of each point. The purpose of this step is to obtain the spatial distribution of the material's thermal conductivity and provide parameters for thermal resistance calculation.

[0052] The specific implementation of step S05 is to construct a thermal resistance sparse matrix. First, based on the heat diffusion distribution matrix and the temperature distribution matrix, the principle of heat flow conservation is used to calculate the heat flow between each point; then, an equivalent thermal network model is established, and the module is discretized into a three-dimensional grid, and the thermal resistance between nodes is calculated as R ij =(T i -T j ) / Q ij ; then, a thermal resistance sparse matrix R 稀疏 [i][j][k] is constructed, and only the thermal resistance values between the nodes of the key heat conduction paths are recorded in the matrix, and the non-key paths are set to zero; finally, the threshold screening algorithm is applied to select the nodes with a thermal resistance value less than 0.1 K / W and a passing heat flow greater than 10% of the total heat flow as the key heat dissipation nodes. The role of this step is to identify and quantify the main heat dissipation paths and reduce the computational complexity of subsequent optimization.

[0053] The specific implementation of step S06 is to reduce the dimensionality of the thermal resistance sparse matrix to a low-dimensional thermal resistance matrix. First, perform singular value decomposition on the thermal resistance sparse matrix R 稀疏 = UΣV T , where U and V are orthogonal matrices, and Σ is a diagonal matrix of singular values; then, according to the magnitudes of the singular values, retain the main singular values with a cumulative contribution rate exceeding 95%, usually the first 5 - 8; then reconstruct to obtain the low-dimensional thermal resistance matrix where U r , Σ r , V r are the matrices after truncation respectively; finally, extract the key heat dissipation features, including the distribution of the main heat flow channels, the eigenvectors of the thermal resistance gradient, and the main modes of the thermal resistance spatial distribution. The purpose of this step is to reduce the problem complexity, extract key features, and improve the efficiency for optimization calculation.

[0054] The specific implementation of step S07 is to calculate the optimal heat dissipation path using the thermal resistance gradient optimization function. First, construct the optimization objective function f(R) = w1max(T) - w2min(T) + w3σ T + w4∑R i , where w1 to w4 are weight coefficients, usually taking values of 0.4, 0.1, 0.3, 0.2 respectively, and σ T is the standard deviation of temperature; then set the constraint conditions: the temperature gradient does not exceed 20°C / cm, the upper limit of the heat flux density does not exceed 150 W / cm 2 , the range of the material thermal conductivity vector is in [120, 500] W / (m·K), and the interfacial contact thermal resistance does not exceed 0.1 K / W; then use the conjugate gradient method to solve the optimization problem, set the iteration accuracy to 10 -5 , and the maximum number of iterations is 500 times; finally, generate the optimal heat dissipation path parameter set P 最优 , which includes the optimized thermal resistance distribution scheme and the corresponding temperature uniformity index (the standard deviation of temperature is less than 3°C). The purpose of this step is to find the theoretically optimal thermal resistance distribution through mathematical optimization.

[0055] The specific implementation of step S08 is to use a thermal resistance prediction deep attention network model for verification and fine-tuning. The specific structure of the thermal resistance prediction deep attention network model is a deep convolutional neural network based on the multi-head self-attention mechanism. The encoder part of this model contains three layers of multi-head self-attention layers and two layers of feed-forward neural networks. Each layer of the multi-head self-attention layer contains 8 attention heads, and the query, key, and value vector dimensions of each head are 64, with a total dimension of 512. The feed-forward network in the encoder consists of two fully connected layers, with a hidden layer dimension of 2048, and uses the rectified linear unit activation function ReLU(x) = max(0, x). The decoder part contains two layers of cross-attention layers and one layer of feed-forward neural network. The cross-attention layer uses the encoder output as the key and value, and the decoder input as the query. The self-attention sparsity parameter is determined by the amplitude distribution function of the gradient thermal resistance vector, and the calculation formula is where γ usually takes a value of 0.8 and β takes a value of 0.03. The convolutional layer uses a 3×3 convolutional kernel, with a total of 32 feature channels, a stride of 1, and zero padding at the edges to maintain the size of the feature map. The final output layer is a fully connected layer that outputs the thermal resistance distribution scheme. The model loss function is L = MSE + λ·U Q where MSE is the mean square error of thermal resistance prediction, U Q is the heat flux non-uniformity penalty term, and λ is the balance coefficient, taking a value of 0.25. First, input the optimal heat dissipation path parameter set into this network; then calculate the predicted thermal resistance distribution through forward propagation; then compare the prediction result with the theoretical optimal scheme to calculate the deviation; finally, based on the deviation, apply backpropagation to fine-tune the network parameters to obtain the final thermal resistance distribution scheme. The purpose of this step is to verify the theoretical optimization result and perform fine-tuning in combination with actual data to improve the reliability of the scheme.

[0056] The specific implementation of step S09 is to design the structural dimensions and material ratios of the silicon carbide double-sided heat dissipation module. First, according to the final thermal resistance distribution scheme, determine the overall size of the module, usually with a thickness of 0.3 - 0.5 mm, and the area is determined according to the power requirement; then optimize the material ratio, with the pure silicon carbide matrix accounting for more than 95%, and adding a small amount (<5%) of high thermal conductivity fillers such as carbon nanotubes and graphene to enhance the thermal conductivity; then design the structural details, including surface microstructures (etching 10 - 30 μm deep V-shaped grooves) to increase the heat dissipation area, and edge chamfers (chamfer radius 0.2 - 0.3 mm) to reduce stress concentration; finally, pre-assemble the silicon carbide double-sided heat dissipation module with the upper and lower heat dissipation substrates respectively to form a sandwich structure, and the pre-assembly pressure is controlled at 0.5 - 1 MPa. This step aims to transform the theoretical optimization result into an actual product design and realize the physical structure implementation.

[0057] The specific implementation of step S10 is to achieve low thermal resistance connection by using silver paste sintering process. First, select silver nanoparticle paste with a particle size of 15 - 25 nm, silver content of 85% - 92%, and organic carrier content of 8% - 15%; then evenly coat the silver paste on the surface of the heat dissipation substrate through screen printing process, with the thickness controlled at 50 - 80 μm; then carry out pre-drying at a temperature of 75 - 85 °C for 10 - 15 minutes to remove part of the organic solvent; then place the silicon carbide module on the silver paste layer and apply a pressure of 0.8 - 1.2 MPa; finally, carry out sintering at a temperature controlled at 240 - 260 °C for 20 - 30 minutes to form a dense silver layer to achieve low thermal resistance connection (interface thermal resistance is less than 0.03 K / W). The function of this step is to achieve efficient thermal coupling between the module and the heat dissipation substrate through high thermal conductivity connection materials, and finally complete the encapsulation.

[0058] The specific implementation of establishing the training dataset for the deep attention network model for thermal resistance prediction is to systematically collect data of various silicon carbide heat dissipation modules. First, select 10 - 15 silicon carbide module samples with different sizes, thicknesses, and surface treatment processes; then, in a standard experimental environment (temperature 25 ± 1 °C, humidity 45% - 55%), use a heat flow meter with an accuracy of 0.01 W / cm 2 and a high-precision thermal imager with an accuracy of 0.05 °C (resolution not less than 1024 × 768 pixels) to measure the heat flux density distribution and the corresponding temperature field distribution of each sample under different power loads (20 - 200 W / cm 2 , with a step of 10 W / cm 2 ); then, according to Fourier's law of heat conduction inversely calculate the actual thermal resistance distribution as the label data; then normalize all the data, using the Z-score normalization method Z = (X - μ) / σ for processing, where μ is the mean value and σ is the standard deviation; then randomly divide the dataset according to the ratio of 70% training set, 15% validation set, and 15% test set; finally, carry out data augmentation processing, including rotation transformation (0°, 90°, 180°, 270°), reflection transformation (horizontal, vertical), and scale transformation (scaling by 0.9 - 1.1 times), expand the number of samples to 4 times the original dataset, and convert all samples into tensor format. The purpose of this process is to establish a comprehensive and representative training dataset to provide a reliable learning basis for the network model.

[0059] The specific implementation of training the deep attention network model for thermal resistance prediction is to adopt a step-by-step training strategy. First, initialize the network weight parameters. The weights of the convolutional layer are randomly initialized using the standard normal distribution N(0, 0.02), and the weights of the self-attention layer are initialized using the uniform distribution where k = 1 / d in , d inis the input dimension; then set the batch size to 32 and use the Adam optimizer with a learning rate of 0.001, where β1 = 0.9, β2 = 0.999, and ∈ = 10 -8 ; then adopt a learning rate warm-up strategy in the first 10 training epochs, where the learning rate linearly increases from 0.0001 to 0.001; then implement a cosine annealing learning rate scheduling strategy, and the learning rate change formula is η t = η min + 0.5(η max - η min )(1 + cos(πt / T)), where T is the total number of epochs 200, η min is 0.00001, and η max is 0.001; then evaluate the model performance metrics on the validation set every 5 epochs, including the mean squared error of thermal resistance prediction (threshold 0.005 K / W) and temperature uniformity (threshold 3°C); when the performance on the validation set does not improve for 20 consecutive epochs, apply the early stopping strategy to avoid overfitting; finally, evaluate all saved checkpoints on the test set and select the model with the smallest thermal resistance prediction error and the highest temperature uniformity as the final model. Usually, the relative error of thermal resistance prediction should be less than 3%, and the temperature uniformity index (the difference between the highest temperature and the lowest temperature) should be less than 5°C. This process aims to obtain a high-precision thermal resistance prediction model through a scientific training method and provide a reliable tool for module structure optimization.

[0060] The following details the mathematical models or calculation processes involved in the present invention.

[0061] The upper-layer function of heat conduction constructed in step S01 is specifically expressed as follows:[[]]

[0062]

[0063] In the formula, Q 上 is the upper-layer function of heat conduction, with the unit of W / m 2 ; k(T) is the thermal conductivity of the silicon carbide material, which is a function of temperature, with the unit of W / (m·K); is the temperature gradient, with the unit of K / m; x, y, z are spatial coordinates, with the unit of m; T is the temperature, with the unit of K.

[0064] This equation is based on Fourier's law of heat conduction and describes the heat conduction behavior inside the silicon carbide material, considering the characteristic that the thermal conductivity varies with temperature. The method for obtaining the thermal conductivity k(T) is: experimentally measure the thermal conductivity of the silicon carbide material at different temperatures and then fit it as a function of temperature, usually expressed as k(T) = k0(1 + α T ·(T - T0) + β T ·(T - T0) 2 ), where k0 is the thermal conductivity at the reference temperature T0, and αT is the first-order temperature coefficient, β T is the second-order temperature coefficient.

[0065] The lower-layer function of heat transfer conduction is specifically expressed as follows:

[0066] Q 下 (x, y, T) = h(T)·(T 模块 (x, y) - T 基板 (x, y));

[0067] In the formula, Q 下 is the lower-layer function of heat transfer conduction, with the unit of W / m 2 ; h(T) is the interfacial heat transfer coefficient, which is a function of temperature, with the unit of W / (m 2 ·K); T 模块 is the module temperature, with the unit of K; T 基板 is the substrate temperature, with the unit of K; x, y are the interfacial coordinates, with the unit of m.

[0068] This equation describes the heat transfer behavior at the interface between the silicon carbide module and the heat dissipation substrate. The method for obtaining h(T) is: measure the interfacial heat transfer coefficient at different temperatures through an interfacial thermal resistance tester, and fit it to h(T) = h0·(1 + γ T ·(T - T0)), where h0 is the interfacial heat transfer coefficient at the reference temperature T0, and γ T is the temperature coefficient.

[0069] The objective function of the double-layer heat conduction function game model is specifically expressed as follows:

[0070] J = ∫ A (T max (x, y) - T min (x, y)) 2 dxdy + α∫ A |Q(x, y)| 2 dxdy;

[0071] In the formula, J is the objective function, dimensionless; T max (x, y) is the highest temperature in the region, with the unit of K; T min (x, y) is the lowest temperature in the region, with the unit of K; Q(x, y) is the heat flux density, with the unit of W / m 2 ; α is the weight coefficient, with a value range of 0.01 to 0.05, dimensionless; A is the integration region, representing the module surface.

[0072] The objective function combines two aspects: temperature uniformity and heat flux density. The first term is used to minimize the temperature difference, and the second term is used to optimize the heat flux distribution. The weight coefficient α is obtained through parameter scanning, and the value that can optimize the heat flux distribution is selected on the premise of meeting the temperature uniformity.

[0073] The thermal resistance calculated in step S02 is specifically expressed as follows:

[0074]

[0075] In the formula, R(x, y, z) is the thermal resistance, with the unit of K / W; ΔT(x, y, z) is the temperature difference, with the unit of K;

[0076] Q(x, y, z) is the heat flux, with the unit of W; T(x, y, z) is the temperature at the position (x, y, z), with the unit of K; T ref is the reference temperature, usually taken as the ambient temperature or the heat sink temperature, with the unit of K.

[0077] Based on the definition of thermal resistance, this equation describes the ratio relationship between the temperature difference and the heat flux. The calculation method of thermal resistance is as follows: First, solve the heat conduction equation by the finite element method to obtain the temperature field T(x, y, z) and the heat flux field Q(x, y, z), and then calculate the thermal resistance at each point according to the above formula.

[0078] The thermal resistance characteristic curve is specifically expressed as follows:

[0079] R = f(T, x, y, z) = R0(x, y, z)·(1 + δ T ·(T - T0) + η T ·(T - T0) 2 )·(1 + ξ x ·x + ξ y ·y + ξ z ·z);

[0080] In the formula, R is the thermal resistance, with the unit of K / W; R0(x, y, z) is the reference thermal resistance value at the reference temperature T0 at the position (x, y, z), with the unit of K / W; δ T is the first-order temperature coefficient of thermal resistance, with the unit of K -1 ; η T is the second-order temperature coefficient of thermal resistance, with the unit of K -2 ; ξ x 、ξ y 、ξ z are the position-dependent coefficients of thermal resistance in the x, y, and z directions, respectively, with the unit of m -1 .

[0081] This characteristic curve describes the variation of thermal resistance with temperature and position, taking into account the non-linear effect of temperature and the linear influence of position. The coefficients are obtained by performing multiple non-linear regression on the measured data.

[0082] The gradient thermal resistance vector is specifically expressed as follows:

[0083] R0·

[0084] (1 + δ T ·(T - T0) + η T ·(T - T0) 2 )·ξ y , R0·(1 + δ T ·(T - T0) + η T ·(T - T0) 2 )·ξ z );

[0085] In the formula, is the gradient thermal resistance vector, with the unit of K / (W·m); are the partial derivatives of the thermal resistance in the x, y, and z directions respectively.

[0086] This vector points in the direction where the thermal resistance increases fastest and is used to guide the optimization of heat flow. The calculation method is: taking the partial derivative of the spatial coordinates according to the thermal resistance characteristic curve.

[0087] The heat dissipation temperature distribution matrix constructed in step S03 is specifically expressed as follows:

[0088]

[0089] In the formula, T 分布 is the heat dissipation temperature distribution matrix, with the unit of K; T ij is the temperature of the point in the i-th row and j-th column, with the unit of K; m and n are the matrix dimensions, usually m = 107 and n = 78.

[0090] This matrix characterizes the temperature distribution of each point on the module surface. The construction method is: first measure the temperature at the key points using thermocouples, and then calculate the temperature at other positions using the bicubic spline interpolation algorithm. The interpolation function is where the coefficient a ij is obtained by solving the following system of equations: S(x k , y k ) = T k ,

[0091] The temperature gradient field is specifically expressed as follows:

[0092]

[0093] In the formula, is the temperature gradient field, with the unit of K / m; are the partial derivatives of temperature in the x, y, and z directions respectively.

[0094] This gradient field describes the spatial rate of change of temperature. The calculation method is as follows: Use the central difference method to calculate the partial derivatives for the heat dissipation temperature distribution matrix, and also use the central difference method to calculate after measuring the temperatures at different depths.

[0095] The heat diffusion equation in step S04 is specifically expressed as follows:

[0096]

[0097] In the formula, is the partial derivative of temperature with respect to time, with the unit of K / s; α is the thermal diffusivity, with the unit of m 2 / s; is the Laplacian operator of temperature, with the unit of K / m 2 ; are the second-order partial derivatives of temperature in the x, y, and z directions respectively.

[0098] This equation describes the unsteady heat conduction process. The calculation method of the thermal diffusivity α is as follows: By the pulse heat source response method, record the temperature change curve T(t) over time, and then calculate according to where L is the characteristic length, and t 1 / 2 is the time required for the temperature to decay to half of the initial value.

[0099] The heat diffusion distribution matrix is specifically expressed as follows:

[0100]

[0101] In the formula, D is the heat diffusion distribution matrix, with the unit of m 2 / s; D ij is the thermal diffusivity of the point in the i-th row and j-th column, with the unit of m 2 / s; m and n are the matrix dimensions, which are the same as those of the temperature distribution matrix.

[0102] This matrix characterizes the spatial distribution of the thermal conductivity of the material. The construction method is as follows: Use an infrared thermal imager to obtain the data of temperature change over time, and then apply the heat diffusion equation to each point to solve for the α value.

[0103] The principle of heat flux conservation in step S05 is specifically expressed as follows:

[0104]

[0105] In the formula, is the divergence of the heat flux density vector, with the unit of W / m 3 ; are the partial derivatives of the heat flux density components in the x, y, and z directions, respectively.

[0106] This equation is derived from the law of conservation of energy and describes the distribution characteristics of heat flux under steady-state conditions. The calculation method is as follows: The module is discretized into grids, and the heat balance equation is applied to each grid cell where Q ij is the heat flux from node i to adjacent node j.

[0107] The calculation of the thermal resistance between nodes is specifically expressed as follows:

[0108]

[0109] In the formula, R ij is the thermal resistance between node i and node j, with the unit of K / W; T i , T j are the temperatures of node i and node j, respectively, with the unit of K; Q ij is the heat flux from node i to node j, with the unit of W; k ij is the thermal conductivity of the material between node i and node j, with the unit of W / (m·K); A ij is the cross-sectional area through which the heat flux passes, with the unit of m 2 ; l ij is the distance between node i and node j, with the unit of m.

[0110] This formula is based on the definition of thermal resistance and Fourier's law of heat conduction and is used to construct an equivalent thermal network model. The calculation method is as follows: Calculate the distance l ij according to the node coordinates, determine the cross-sectional area A ij according to the grid size, and obtain the thermal conductivity k ij from the thermal diffusion distribution matrix.

[0111] The thermal resistance sparse matrix is specifically expressed as follows:

[0112]

[0113] In the formula, R 稀疏 is the thermal resistance sparse matrix, with the unit of K / W; R ijk is the thermal resistance value at the corresponding position, i, j, and k are spatial indices; R 计算值 is the thermal resistance value calculated by the above formula.

[0114] This matrix only retains the thermal resistance values on the key heat conduction paths. The threshold condition is that the thermal resistance value is less than 0.1 K / W and the heat flux passing through is greater than 10% of the total heat flux.

[0115] The singular value decomposition in step S06 is specifically expressed as follows:

[0116]

[0117] In the formula, R 稀疏 is the thermal resistance sparse matrix; U is the left singular matrix, and its column vector u i is the eigenvector of; ∑ is the singular value diagonal matrix, and the diagonal element σ i is the singular value of R 稀疏 ; V is the right singular matrix, and its column vector v i is the eigenvector of; r is the rank of the matrix.

[0118] This decomposition method is used for dimensionality reduction and extraction of main features. The calculation steps are as follows: (1) Calculate and (2) Solve the eigenvalue problem to obtain eigenvalues and eigenvectors; (3) Take the square root of the eigenvalues to obtain singular values, and the eigenvectors form the singular matrix.

[0119] The low-dimensional thermal resistance matrix is specifically represented as follows:

[0120]

[0121] In the formula, R 低维 is the low-dimensional thermal resistance matrix; U r is the truncated left singular matrix, only retaining the first k columns; ∑ r is the truncated singular value diagonal matrix, only retaining the first k singular values; V r is the truncated right singular matrix, only retaining the first k columns; k is the retained dimension, usually 5 to 8.

[0122] This matrix retains the main information of the original matrix. The selection criterion for the dimension k is that the cumulative contribution rate of the retained singular values exceeds 95%, and the calculation formula is

[0123] The optimization objective function in step S07 is specifically represented as follows:

[0124]

[0125] In the formula, f(R) is the optimization objective function, dimensionless; max(T(R)) is the highest temperature, with the unit of K; min(T(R)) is the lowest temperature, with the unit of K; σ T (R) is the temperature standard deviation, with the unit of K; is the total thermal resistance, with the unit of K / W; w1, w2, w3, w4 are weight coefficients, taking the values of 0.4, 0.1, 0.3, and 0.2 respectively, dimensionless.

[0126] This function comprehensively considers the temperature extreme value, temperature uniformity, and total thermal resistance. The temperature T(R) is obtained by solving the heat conduction equation, and the standard deviation where is the average temperature.

[0127] The constraint conditions of the thermal resistance gradient optimization function are specifically expressed as follows:

[0128]

[0129] Q ≤ 150 W / cm 2 ;

[0130] 120 W / (m·K) ≤ k ≤ 500 W / (m·K);

[0131] R_contact ≤ 0.1 K / W;

[0132] In the formula, is the amplitude of the temperature gradient, with the unit of K / cm; Q is the heat flux density, with the unit of W / cm 2 ; k is the thermal conductivity of the material, with the unit of W / (m·K); R 接触 is the interfacial contact thermal resistance, with the unit of K / W.

[0133] These constraint conditions consider the limitations of actual processes and materials. The temperature gradient constraint comes from the material's resistance to thermal stress, the upper limit of the heat flux density comes from the heat dissipation ability, the range of the thermal conductivity comes from the material properties, and the contact thermal resistance constraint comes from the interfacial connection process.

[0134] The attention calculation formula in the thermal resistance prediction network model is the general attention calculation formula in the prior art, specifically expressed as follows:

[0135]

[0136] In the formula, Attention is the attention function; Q is the query matrix, with the dimension of (n, d k ); K is the key matrix, with the dimension of (m, d k ); V is the value matrix, with the dimension of (m, d v ); d k is the dimension of the key vector, with the value of 64; softmax is the normalization function

[0137] This formula implements the attention mechanism, enabling the network to focus on important features. The calculation steps are as follows: (1) Calculate the similarity QK between the query and the key T ; (2) Divide by for scaling; (3) Apply softmax normalization; (4) Multiply by the value matrix to obtain the weighted result.

[0138] The difference from the prior art lies in that the self-attention sparsity parameter is specifically expressed as follows:

[0139]

[0140] In the formula, S is the self-attention sparsity parameter, dimensionless; γ is the proportionality coefficient, with a value of 0.8, dimensionless; is the magnitude of the gradient thermal resistance vector, with the unit of K / (W·m); β is the attenuation coefficient, with a value of 0.03, with the unit of K / (W·m).

[0141] This parameter controls the focusing degree of the attention mechanism. Regions with large gradient thermal resistance obtain smaller sparsity and thus receive more attention. The calculation method is: substituting the magnitude of the gradient thermal resistance vector into the formula to generate the corresponding sparsity parameter matrix.

[0142] The loss function of the network model is specifically expressed as follows:

[0143]

[0144] In the formula, L is the loss function, dimensionless; MSE is the mean square error, with the unit of (K / W) 2 ; λ is the balance coefficient, with a value of 0.25, dimensionless; U Q is the heat flux non-uniformity penalty term, dimensionless; R 预测，i is the predicted thermal resistance value, with the unit of K / W; R 真实，i is the true thermal resistance value, with the unit of K / W; σ Q is the standard deviation of the heat flux density, with the unit of W / m 2 ; μ Q is the average value of the heat flux density, with the unit of W / m 2 .

[0145] This loss function simultaneously optimizes the thermal resistance prediction accuracy and the heat flux uniformity. The calculation steps of the mean square error are: (1) calculating the square of the difference between the predicted value and the true value; (2) finding the average value. The calculation steps of the heat flux non-uniformity are: (1) finding the mean and standard deviation of the heat flux density; (2) calculating the coefficient of variation as the non-uniformity measure.

[0146] Specifically, the principle of the present invention is: Based on the gradient thermal resistance theory and deep learning technology, through multi-level thermal resistance analysis and optimization strategies, the present invention realizes the global thermal resistance optimization of the silicon carbide double-sided heat dissipation module packaging. Its core principle lies in decomposing the heat conduction behavior into two levels of internal conduction and interface conduction, finding the global optimal solution through a game model, and at the same time introducing the gradient thermal resistance vector to guide the heat flux optimization direction, establishing a quantitative relationship between the heat flux distribution and the thermal resistance optimization.

[0147] First, the present invention constructs upper and lower heat conduction functions to describe the relationship between the internal heat flux density distribution and the temperature gradient in the module, as well as the interface heat conduction and contact thermal resistance characteristics, forming a game model of double-layer heat conduction functions. This model breaks through the limitations of traditional single-layer heat flux analysis, can take into account both the internal heat network structure and the interface thermal resistance characteristics at the same time, and provides a more complete theoretical framework for thermal resistance optimization. Secondly, by calculating the gradient thermal resistance vector, the main path and secondary path of the heat flux are determined, realizing the quantitative analysis of the heat conduction path and overcoming the shortcoming that the heat flux path is difficult to accurately characterize in traditional methods.

[0148] At the implementation level, the present invention obtains the surface heat diffusion data of the module through thermal infrared imaging technology, constructs a thermal resistance sparse matrix in combination with the temperature distribution matrix, and uses singular value decomposition technology to reduce the dimension of the matrix, extracting the key features of heat dissipation. This method effectively reduces the computational complexity while retaining the key information of the thermal resistance distribution. More importantly, the present invention introduces a deep attention network model for thermal resistance prediction, and accurately predicts and optimizes the thermal resistance distribution through the multi-head self-attention mechanism, realizing the dynamic adjustment and global optimization of the thermal resistance distribution. Based on the optimization results, a sandwich structure is designed, and a silver paste sintering process is used to achieve low thermal resistance connection, thereby forming an efficient double-sided heat dissipation packaging structure.

[0149] The logic of the method of the present invention is reflected in that it follows the basic physical laws of heat conduction, and at the same time introduces advanced mathematical tools and deep learning technologies to establish a complete technical chain from heat flux analysis, thermal resistance calculation to structural design. In particular, the establishment of the game model of double-layer heat conduction functions provides a theoretical basis for solving the thermal resistance optimization problem in the double-sided heat dissipation structure, and the introduction of the deep attention network for thermal resistance prediction provides an efficient implementation tool for thermal resistance optimization. The two together ensure the scientificity and effectiveness of the packaging method.

[0150] A specific embodiment 1 of the present invention is provided below, and the specific implementation manners of each step in this embodiment 1 are described in detail as follows.

[0151] The specific implementation manner of step S01 is to construct a heat conduction function system for the silicon carbide double-sided heat dissipation module. First, based on Fourier's law of heat conduction, the upper-layer function of heat conduction is established where k(T) is the thermal conductivity of the silicon carbide material, which is a function of temperature, with the unit of W / (m·K). The thermal conductivity of the silicon carbide material at different temperatures is measured through experiments and then fitted as a function of temperature k(T) = k0(1 + α T ·(T - T0) + β T ·(T - T0) 2 ), where k0 is the thermal conductivity at the reference temperature T0, α T is the first-order temperature coefficient, and β T is the second-order temperature coefficient; is the temperature gradient, with the unit of K / m; x, y, and z are spatial coordinates, with the unit of m; at the same time, establish the lower-layer function Q of heat conduction 下 (x, y, T) = h(T)·(T 模块 (x, y) - T 基板 (x, y)), where h(T) is the interfacial heat transfer coefficient, which is a function of temperature, with the unit of W / (m 2 ·K), and is measured at different temperatures by an interfacial thermal resistance tester and fitted as h(T) = h0·(1 + γ T ·(T - T0)), where h0 is the interfacial heat transfer coefficient at the reference temperature T0, and γ T is the temperature coefficient; T 模块 is the module temperature, and T 基板 is the substrate temperature, with the unit of K. Then combine the upper and lower layer functions to form a two-layer heat conduction function game model, and this model is solved by minimizing the objective function J = ∫ A (T max (x, y) - T min (x, y)) 2 dxdy + α∫ A |Q(x, y)| 2 dxdy, where the first term is used to minimize the temperature difference, and the second term is used to optimize the heat flux distribution. α is the weight coefficient, which is obtained through parameter scanning, and the value range is 0.01 - 0.05. The purpose of this step is to establish a mathematical description of heat conduction and lay a theoretical foundation for subsequent thermal resistance optimization.

[0152] The specific implementation of step S02 is to establish a thermal resistance characteristic curve based on the two-layer heat conduction function game model. First, solve the model equation by the finite element method to obtain the temperature distribution T(x, y, z) and the heat flux density distribution Q(x, y, z); then calculate the thermal resistance where ΔT(x, y, z) is the temperature difference, with the unit of K; Q(x, y, z) is the heat flux, with the unit of W; T ref is the reference temperature, usually taken as the ambient temperature or the heat sink temperature, with the unit of K; then establish the characteristic curve of the thermal resistance varying with temperature and position R = f(T, x, y, z) = R0(x, y, z)·(1 + δ T ·(T - T0) + η T ·(T - T0) 2 )·(1 + ξ x ·x + ξ y ·y + ξ z ·z), where R0(x, y, z) is the thermal resistance reference value at the position (x, y, z) under the reference temperature T0, δ T is the first-order temperature coefficient of the thermal resistance, η T is the second-order temperature coefficient of the thermal resistance, and ξx , ξ y , ξ z are the position-dependent coefficients of the thermal resistance in the x, y, and z directions respectively; then calculate the gradient thermal resistance vector

[0153] R0·(1 + δ T ·(T - T0) + η T ·(T - T0) 2 )·ξ y , h0·(1 + δ T ·(T - T0) + η T ·(T - T0) 2 )·ξ z ), and this vector points to the direction where the thermal resistance increases fastest; finally, according to the heat flow line density and the direction of the gradient thermal resistance vector, determine the main heat flow path (the path where the magnitude of the gradient thermal resistance vector is less than 0.02 K / (W·m)) and the secondary path (the path where the magnitude of the gradient thermal resistance vector is between 0.02 and 0.05 K / (W·m)). The function of this step is to quantify the thermal resistance distribution characteristics and provide a basis for subsequent hot spot identification.

[0154] The specific implementation of step S03 is to measure and construct the heat dissipation temperature distribution matrix. First, select key measurement points on the main path and the secondary path, and the spacing between key points is usually 3 - 5 mm; then measure the temperature of each point using a high-precision thermocouple (accuracy better than 0.1 °C) under different power loads; then use the bicubic spline interpolation algorithm to generate a complete temperature distribution matrix where T ij is the temperature of the point in the i-th row and j-th column, m and n are the matrix dimensions, usually m = 107, n = 78, covering the entire module surface, and the interpolation function is The coefficient a ij is obtained by solving the system of equations; finally, calculate the temperature gradient field is calculated by the central difference method, is calculated by the central difference method after measuring the temperature at different depths, and is used to quantify the temperature change rate. The purpose of this step is to obtain the actual temperature distribution data of the module and provide a basis for heat diffusion analysis.

[0155] The specific implementation of step S04 is to obtain heat diffusion data using thermal infrared imaging technology. First, use an infrared thermal imager with a resolution of not less than 640×480 pixels and a temperature measurement accuracy better than 0.05 °C to scan the module surface under the condition of a constant ambient temperature (25 ± 1 °C); then apply the pulsed heat source response method to excite the module through an excitation signal (10 - 50 ms square wave pulse) and record the temperature transient response; then use the heat diffusion equation to analyze the temperature change rate, where is the partial derivative of temperature with respect to time, α is the thermal diffusion coefficient, is the Laplacian operator of temperature, and the calculation method of the thermal diffusion coefficient α is where L is the characteristic length, and t 1 / 2 is the time required for the temperature to decay to half of the initial value; finally, a thermal diffusion distribution matrix is constructed to characterize the thermal diffusion coefficient at each point. This step aims to obtain the spatial distribution of the material's thermal conductivity and provide parameters for thermal resistance calculation.

[0156] The specific implementation of step S05 is to construct a thermal resistance sparse matrix. First, based on the thermal diffusion distribution matrix and the temperature distribution matrix, using the principle of heat flux conservation calculate the heat flux between each point, where is the divergence of the heat flux density vector, are the partial derivatives of the heat flux density components in the x, y, and z directions respectively; then establish an equivalent thermal network model, discretize the module into a three-dimensional grid, and calculate the thermal resistance between nodes as where R ij is the thermal resistance between node i and node j, T i , T j are the temperatures of node i and node j respectively, Q ij is the heat flux from node i to node j, k ij is the thermal conductivity of the material between node i and node j, A ij is the cross-sectional area through which the heat flux passes, and l ij is the distance between node i and node j; then construct a thermal resistance sparse matrix Only the thermal resistance values between the nodes on the key heat conduction paths are recorded in the matrix, and the non-key paths are set to zero; finally, apply the threshold screening algorithm to select the nodes with a thermal resistance value less than 0.1 K / W and a heat flux passing through greater than 10% of the total heat flux as the key heat dissipation nodes. The role of this step is to identify and quantify the main heat dissipation paths and reduce the computational complexity of subsequent optimization.

[0157] The specific implementation of step S06 is to reduce the dimension of the thermal resistance sparse matrix to a low-dimensional thermal resistance matrix. First, perform singular value decomposition on the thermal resistance sparse matrix where U is the left singular matrix, and its column vector u i is the eigenvector of; ∑ is the singular value diagonal matrix, and the diagonal element σ i is the singular value of R 稀疏 ; V is the right singular matrix, and its column vector v i is the eigenvector of; r is the rank of the matrix; then, according to the singular value size, retain the main singular values with a cumulative contribution rate exceeding 95%, usually the first 5 - 8, and the calculation formula is Then reconstruct to obtain the low-dimensional thermal resistance matrix Among them, U r , ∑ r , V r are respectively the matrices after interception; finally, the key heat dissipation features are extracted, including the distribution of the main heat flow channels, the eigenvector of the thermal resistance gradient, and the main mode of the thermal resistance spatial distribution. The purpose of this step is to reduce the complexity of the problem, extract the key features, and improve the efficiency for the optimization calculation.

[0158] The specific implementation of step S07 is to calculate the optimal heat dissipation path by applying the thermal resistance gradient optimization function. First, construct the optimization objective function where f(R) is the optimization objective function, max(T(R)) is the highest temperature, min(T(R)) is the lowest temperature, σ T (R) is the standard deviation of temperature, is the average temperature, is the total thermal resistance, w1 to w4 are weight coefficients, and their usual values are 0.4, 0.1, 0.3, and 0.2 respectively; then set the constraint conditions: the temperature gradient does not exceed 20 °C / cm, the upper limit of the heat flux density does not exceed 150 W / cm 2 , Q ≤ 150 W / cm 2 ; the range of the material thermal conductivity vector is [120, 500] W / (m·K), 120 W / (m·K) ≤ k ≤ 500 W / (m·K); the interfacial contact thermal resistance does not exceed 0.1 K / W, R 接触 ≤ 0.1 K / W; then use the conjugate gradient method to solve the optimization problem, set the iteration accuracy to 10 -5 , and the maximum number of iterations is 500 times; finally, generate the optimal heat dissipation path parameter set, including the optimized thermal resistance distribution scheme and the corresponding temperature uniformity index (the standard deviation of temperature is less than 3 °C). This step aims to find the theoretical optimal thermal resistance distribution through mathematical optimization.

[0159] The specific implementation of step S08 is to use the deep attention network model for thermal resistance prediction to verify and fine-tune. The core of this model is the multi-head self-attention mechanism, and the attention calculation formula is

[0160] where Q is the query matrix, K is the key matrix, V is the value matrix, d kis the key vector dimension, with a value of 64, and softmax is the normalization function. The model consists of an encoder and a decoder. The encoder has three layers of multi-head self-attention layers and two layers of feed-forward neural networks. Each layer of multi-head self-attention layers contains 8 attention heads, and the query, key, and value vector dimensions of each head are 64, with a total dimension of 512; the decoder has two layers of cross-attention layers and one layer of feed-forward neural networks. The self-attention sparsity parameter is determined by the amplitude distribution function of the gradient thermal resistance vector, and the calculation formula is

[0161] where γ takes a value of 0.8 and β takes a value of 0.03. The convolutional layer uses a 3×3 convolutional kernel, with a total of 32 feature channels, a stride of 1, and zero padding at the edges to maintain the feature map size. The loss function is where MSE is the mean squared error, λ is the balance coefficient, with a value of 0.25, and U Q is the heat flux non-uniformity penalty term, and σ Q is the standard deviation of the heat flux density, and μ Q is the average value of the heat flux density. First, input the optimal heat dissipation path parameter set into this network; then calculate the predicted thermal resistance distribution through forward propagation; then compare the prediction result with the theoretical optimal solution to calculate the deviation; finally, based on the deviation, apply backpropagation to fine-tune the network parameters to obtain the final thermal resistance distribution scheme. The purpose of this step is to verify the theoretical optimization result and fine-tune it in combination with actual data to improve the reliability of the scheme.

[0162] The specific implementation of step S09 is to design the structural dimensions and material ratios of the silicon carbide double-sided heat dissipation module. First, according to the final thermal resistance distribution scheme, determine the overall size of the module, usually with a thickness of 0.3 - 0.5 mm, and the area is determined according to the power demand; then optimize the material ratio, with the pure silicon carbide matrix accounting for more than 95%, and adding a small amount (<5%) of high thermal conductivity fillers such as carbon nanotubes and graphene to enhance the thermal conductivity; then design the structural details, including surface microstructures (etching 10 - 30 μm deep V-shaped grooves) to increase the heat dissipation area, and edge chamfers (chamfer radius 0.2 - 0.3 mm) to reduce stress concentration; finally, pre-assemble the silicon carbide double-sided heat dissipation module with the upper and lower heat dissipation substrates respectively to form a sandwich structure, and the pre-assembly pressure is controlled at 0.5 - 1 MPa. This step aims to transform the theoretical optimization result into the actual product design and realize the physical structure implementation.

[0163] The specific implementation of step S10 is to achieve low thermal resistance connection by using silver paste sintering process. First, select silver nanoparticle paste with a particle size of 15 - 25 nm, silver content of 85% - 92%, and organic carrier content of 8% - 15%; then evenly coat the silver paste on the surface of the heat dissipation substrate through screen printing process, with the thickness controlled at 50 - 80 μm; then perform pre-drying at a temperature of 75 - 85 °C for 10 - 15 minutes to remove part of the organic solvent; then place the silicon carbide module on the silver paste layer and apply a pressure of 0.8 - 1.2 MPa; finally, perform sintering at a temperature controlled at 240 - 260 °C for 20 - 30 minutes to form a dense silver layer to achieve low thermal resistance connection (interface thermal resistance less than 0.03 K / W). The function of this step is to achieve efficient thermal coupling between the module and the heat dissipation substrate through highly thermally conductive connection materials, and finally complete the encapsulation.

[0164] The specific implementation of establishing the training dataset for the deep attention network model for thermal resistance prediction includes systematically collecting data of various different structures of silicon carbide heat dissipation modules. First, select 10 - 15 samples of silicon carbide heat dissipation modules with different sizes, thicknesses, and surface treatment processes; then in a standard experimental environment (temperature 25 ± 1 °C, humidity 45% - 55%), use a heat flow meter with an accuracy of 0.01 W / cm 2 and a high-precision thermal imager with an accuracy of 0.05 °C (resolution not less than 1024 × 768 pixels) to measure the heat flux density distribution and the corresponding temperature field distribution of each sample under different power loads (20 - 200 W / cm 2 , with a step of 10 W / cm 2 ); then inversely calculate the actual thermal resistance distribution according to Fourier's law of heat conduction as the label data; then normalize all the data, and use the Z-score normalization method Z = (X - μ) / σ for processing, where μ is the mean value and σ is the standard deviation; then randomly divide the dataset according to the ratio of 70% training set, 15% validation set, and 15% test set; finally, perform data augmentation processing, including rotation transformation (0°, 90°, 180°, 270°), reflection transformation (horizontal, vertical), and scale transformation (scaling by 0.9 - 1.1 times), expand the number of samples to 4 times the original dataset, and convert all samples into tensor format. The purpose of this process is to establish a comprehensive and representative training dataset to provide a reliable learning basis for the network model.

[0165] The specific implementation of training the deep attention network model for thermal resistance prediction is to adopt a step-by-step training strategy. First, initialize the network weight parameters. The weights of the convolutional layer are randomly initialized using the standard normal distribution N(0, 0.02), and the weights of the self-attention layer are initialized using the uniform distribution where k = 1 / d in , d inis the input dimension; then set the batch size to 32, and use the Adam optimizer with a learning rate of 0.001, where β1 = 0.9, β2 = 0.999, ∈ = 10 -8 ; then adopt a learning rate warm-up strategy in the first 10 training epochs, with the learning rate linearly increasing from 0.0001 to 0.001; then implement a cosine annealing learning rate scheduling strategy, and the learning rate change formula is η t = η min + 0.5(η max - η min )(1 + cos(πt / T)), where T is the total number of epochs 200, η min is 0.00001, η max is 0.001; then evaluate the model performance metrics on the validation set every 5 epochs, including the mean squared error of thermal resistance prediction (threshold 0.005 K / W) and temperature uniformity (threshold 3 °C); when the performance on the validation set no longer improves for 20 consecutive epochs, apply an early stopping strategy to avoid overfitting; finally, evaluate all saved checkpoints on the test set, and select the model with the smallest thermal resistance prediction error and the highest temperature uniformity as the final model. Generally, the relative error of thermal resistance prediction should be less than 3%, and the temperature uniformity index (the difference between the highest temperature and the lowest temperature) should be less than 5 °C. This process aims to obtain a high-precision thermal resistance prediction model through scientific training methods, providing a reliable tool for module structure optimization.

[0166] This embodiment combines heat conduction theory, gradient thermal resistance optimization, and deep learning technology to establish a complete encapsulation method for silicon carbide double-sided heat dissipation modules. The construction of the heat conduction function reflects the physical essence of heat conduction; the introduction of the gradient thermal resistance vector provides a clear direction for heat flow optimization; the construction and dimensionality reduction of the thermal resistance sparse matrix greatly reduce the computational complexity; the application of the deep attention network model improves the accuracy of thermal resistance prediction; finally, a low thermal resistance connection is achieved through the silver paste sintering process. The whole method forms a complete technical route from theoretical analysis, data acquisition, model optimization to actual encapsulation, which can effectively solve the thermal resistance optimization and temperature uniformity problems of silicon carbide heat dissipation modules and meet the heat dissipation requirements of high-power density electronic devices. Compared with traditional methods, this method has significant improvements in temperature uniformity, thermal resistance reduction, and heat dissipation efficiency, comprehensively improving the performance of silicon carbide double-sided heat dissipation modules.

[0167] To better understand and implement the present invention, the following provides Embodiment 2 of a specific application scenario of the present invention: Researchers carried out research on the encapsulation of all-silicon carbide double-sided heat dissipation modules based on gradient thermal resistance optimization for the heat dissipation problem of silicon carbide power modules in a certain high-power density power electronic device. The working power density of this power electronic device reaches 120 W / cm 2, there are extremely high requirements for heat dissipation performance and temperature uniformity. The traditional heat dissipation solution uses a single-sided aluminum substrate for heat dissipation. Under high-power operating conditions, there is a large temperature difference on the surface of the module, and the maximum temperature can reach 175°C, seriously affecting the performance and reliability of the equipment. The encapsulation method of a fully silicon carbide double-sided heat dissipation module based on gradient thermal resistance optimization is applied as follows.

[0168] First, the upper and lower layer functions of heat transfer conduction for the silicon carbide double-sided heat dissipation module are established. The upper layer function of heat transfer conduction describes the heat transfer conduction behavior inside the module. By measuring, the thermal conductivity of the silicon carbide material is 370 W / (m·K) at a reference temperature of 25°C, and the first-order temperature coefficient is -3.2×10 -4 K -1 , and the second-order temperature coefficient is -8.5×10 -7 K -2 . The lower layer function of heat transfer conduction describes the heat transfer conduction behavior at the interface between the module and the heat dissipation substrate. The interface heat transfer coefficient is 2.8×10 4 W / (m 2 ·K) at a reference temperature of 25°C, and the temperature coefficient is -2.1×10 -3 K -1 . Then, a game model of the double-layer heat conduction function is constructed, and the weight coefficient α is taken as 0.025.

[0169] The researchers solved the model equation by the finite element method, obtained the temperature distribution and heat flux density distribution, and then calculated the thermal resistance distribution to establish a thermal resistance characteristic curve. On this basis, the gradient thermal resistance vector is calculated, and the main heat flow path (the amplitude of the gradient thermal resistance vector < 0.015 K / (W·m)) and the secondary path (the amplitude of the gradient thermal resistance vector is between 0.015 and 0.04 K / (W·m)) are determined. The main path covers 75% of the area under the power chip, and the secondary path covers the edge area.

[0170] After determining the primary and secondary paths, the researchers measured the key point temperatures under typical load conditions (power density 120 W / cm 2 ), and the spacing between key points was set to 4 mm. The temperatures of 64 key points were measured using a thermocouple with an accuracy of 0.08°C, and a complete temperature distribution matrix of 128×96 was generated through the bicubic spline interpolation algorithm. The calculated temperature gradient field shows that the temperature gradient in the central area of the chip is 16.8°C / cm, and in the edge area is 12.3°C / cm.

[0171] An infrared thermal imager with a resolution of 768×576 pixels and a temperature measurement accuracy of 0.04°C is used to obtain the surface heat diffusion data of the module. The module is excited with a 30 ms square wave pulse, and the temperature transient response is recorded. The calculated heat diffusion coefficient distribution matrix is shown in Table 1:

[0172] Table 1 Heat diffusion coefficients in different regions of the silicon carbide heat dissipation module (unit: mm2 / s)

[0173] Region Central region Intermediate transition region Edge region Upper surface 92.4 87.6 81.3 Middle layer 98.6 95.2 89.7 Lower surface 93.5 88.9 82.1

[0174] Based on the heat diffusion distribution matrix and the temperature distribution matrix, the heat flux between points is calculated using the principle of heat flux conservation, and an equivalent heat network model is established. The module is discretized into a three-dimensional grid of 32×24×8, the thermal resistance between calculation nodes is calculated, and a thermal resistance sparse matrix is constructed. 73 key heat dissipation nodes with thermal resistance values less than 0.085 K / W and heat flux passing through greater than 12% of the total heat flux are selected.

[0175] Perform singular value decomposition on the thermal resistance sparse matrix. Based on the criterion of a cumulative contribution rate of 95%, the first 6 singular values are retained, and a low-dimensional thermal resistance matrix is reconstructed. The extracted key heat dissipation features show that the main heat flux channels are concentrated in the area slightly below the center of the module, and the eigenvector of the thermal resistance gradient points to the edge area.

[0176] In the optimization stage, an optimization objective function is constructed, and the weight coefficients w1, w2, w3, and w4 take the values of 0.38, 0.12, 0.32, and 0.18 respectively. The constraint conditions are set as follows: the temperature gradient does not exceed 18 °C / cm, the upper limit of the heat flux density does not exceed 135 W / cm 2 , the thermal conductivity range of the material is between [150, 480] W / (m·K), and the interfacial contact thermal resistance does not exceed 0.08 K / W. The conjugate gradient method is used for optimization calculation, and the iteration accuracy is set to 5×10 -6 . After 387 iterations, an optimal heat dissipation path parameter set is obtained. The comparison of the temperature distributions before and after optimization is shown in Table 2:

[0177] Table 2 Comparison of temperature distributions before and after optimization (unit: °C)

[0178] Index Before optimization After optimization Improvement rate (%) Highest temperature 175.3 152.6 12.9 Lowest temperature 138.7 135.8 2.1 Temperature difference 36.6 16.8 54.1 Temperature standard deviation 12.3 4.7 61.8

[0179] The optimal heat dissipation path parameter set is input into a pre-trained deep attention network model for thermal resistance prediction for verification and fine-tuning. The model uses test data of 12 different silicon carbide heat dissipation module samples in the training stage, covering a power load range of 30 - 180 W / cm 2 . The model uses 8 attention heads, the convolution kernel size is 3×3, and the number of feature channels is 32. After verification and fine-tuning, the relative error of thermal resistance prediction is reduced to 2.4%, and the temperature uniformity index (the difference between the highest temperature and the lowest temperature) is 15.6 °C.

[0180] Based on the final thermal resistance distribution scheme, the structural dimensions and material ratios of the silicon carbide double-sided heat dissipation module were designed. The module thickness is 0.42 mm, the area is 48 mm × 36 mm, and the material ratio is 97.5% silicon carbide matrix, with 2.5% graphene added as a high thermal conductivity filler. V-shaped micro-grooves with a depth of 25 μm are designed on the surface, and the edge chamfer radius is 0.25 mm. The module is pre-assembled with the upper and lower heat dissipation substrates to form a sandwich structure, and the pre-assembly pressure is set at 0.85 MPa.

[0181] Finally, a silver paste sintering process is used to achieve low thermal resistance connection. Silver nanoparticle paste with a particle size of 18 nm is selected, with a silver content of 90% and an organic carrier content of 10%. The silver paste is evenly coated on the surface of the heat dissipation substrate through a screen printing process, and the thickness is controlled at 65 μm. The pre-drying temperature is 80 °C and the time is 12 minutes. A pressure of 1.0 MPa is applied, the sintering temperature is 250 °C, and the time is 25 minutes to form a dense silver layer, and the interface thermal resistance is reduced to 0.026 K / W.

[0182] The thermal performance of the packaged module was tested, and the test results are shown in Table 3:

[0183] Table 3 Comparison between the double-sided heat dissipation module and the traditional single-sided heat dissipation scheme

[0184]

[0185] The traditional heat dissipation scheme mainly relies on the combination of a single-sided aluminum substrate and a radiator, using an empirical design method, without considering the system optimization of thermal resistance distribution and temperature uniformity. The interface connection mostly uses welding or thermal conductive gel, resulting in a large thermal resistance. In contrast, the method of packaging the all-silicon carbide double-sided heat dissipation module based on gradient thermal resistance optimization in the present invention has significant advantages: First, a gradient thermal resistance vector is introduced to guide the optimization of heat flow, and the physical characteristics of heat flow are used to actively guide the heat dissipation path; second, singular value decomposition is used to reduce the computational complexity and extract key heat dissipation characteristics, greatly improving the optimization efficiency; third, a deep attention network is combined to achieve accurate prediction of thermal resistance, making the optimization results more reliable; fourth, a double-sided heat dissipation structure and a silver paste sintering connection process are adopted to fundamentally improve the heat dissipation path and interface thermal resistance. The measured results show that the total thermal resistance of this method is reduced by 56.5% compared with the traditional scheme, the temperature uniformity is improved by 57.4%, the power density carrying capacity is increased by 48.0%, and the reliability is increased by more than twice, providing an effective solution for the heat dissipation problem of high-power density power electronic devices.

[0186] It should be noted that the detailed explanations of the variables involved in the present invention are shown in Tables 4, 5, and 6 below.

[0187] Table 4 Variable Explanation Table (Part 1)

[0188]

[0189]

[0190] Table 5 Variable Explanation Table (Second Part)

[0191]

[0192] Table 6 Variable Explanation Table (Third Part)

[0193]

[0194]

[0195] As described above, it is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention.

Claims

1. A packaging method for a fully silicon carbide double-sided heat dissipation module based on gradient thermal resistance optimization, characterized in that It includes: Construct the upper-layer function and the lower-layer function of heat transfer conduction of the silicon carbide double-sided heat dissipation module to form a double-layer heat conduction function game model; Based on this model, establish a thermal resistance characteristic curve, calculate the gradient thermal resistance vector, and determine the main heat transfer conduction path and the secondary path; measure the key point temperatures according to the heat flow path, and construct a heat dissipation temperature distribution matrix; Use thermal infrared imaging to obtain surface heat diffusion data and construct a heat diffusion distribution matrix; combine the two matrices to calculate the thermal resistance of the heat dissipation path, establish a thermal resistance sparse matrix, and screen key heat dissipation nodes; reduce the thermal resistance sparse matrix to a low-dimensional thermal resistance matrix by singular value decomposition, and extract the key heat dissipation characteristics; Apply the thermal resistance gradient optimization function to calculate the optimal heat dissipation path parameter set; input the parameter set into the deep attention network model of thermal resistance prediction for verification and fine-tuning to generate the final thermal resistance distribution scheme; design the module structure and material ratio according to the scheme to form a sandwich structure; use the silver paste sintering process to achieve low thermal resistance connection and complete the encapsulation.

2. The packaging method of the all-silicon carbide double-sided heat dissipation module based on gradient thermal resistance optimization according to claim 1, wherein The upper-layer function of heat transfer conduction specifically refers to the mathematical function that describes the heat transfer conduction behavior inside the silicon carbide double-sided heat dissipation module, including the expressions of heat flux density distribution and temperature gradient relationship; the lower-layer function of heat transfer conduction specifically refers to the mathematical function that describes the heat transfer conduction behavior at the interface between the silicon carbide double-sided heat dissipation module and the heat dissipation substrate, reflecting the relationship between interface thermal resistance and heat flux density.

3. The method for packaging a fully silicon carbide double-sided heat dissipation module optimized based on gradient thermal resistance according to claim 2, wherein, The double-layer heat conduction function game model specifically refers to the optimization model formed by combining the upper-layer function and the lower-layer function of heat transfer conduction, and is the mathematical expression for finding the global optimal heat dissipation path through iterative calculation.

4. The method for packaging a fully silicon carbide double-sided heat dissipation module optimized based on gradient thermal resistance according to claim 3, wherein, The gradient thermal resistance vector specifically refers to the vector that describes the gradient of the thermal resistance in the spatial distribution, and its direction points to the direction where the thermal resistance increases fastest, and is used to guide the heat flow optimization path.

5. The method for packaging a fully silicon carbide double-sided heat dissipation module optimized based on gradient thermal resistance according to claim 4, wherein The heat dissipation temperature distribution matrix specifically refers to the two-dimensional data matrix that characterizes the temperature distribution of each point on the surface of the silicon carbide double-sided heat dissipation module, and is used for hot spot identification and temperature field analysis.

6. The method for packaging a fully silicon carbide double-sided heat dissipation module optimized based on gradient thermal resistance according to claim 5, wherein The heat diffusion distribution matrix specifically refers to the two-dimensional data matrix that characterizes the spatial distribution of the heat diffusion coefficient on the surface of the silicon carbide double-sided heat dissipation module, reflecting the spatial variation of the material's thermal conductivity; the thermal resistance sparse matrix specifically refers to the matrix that describes the internal thermal network structure of the silicon carbide double-sided heat dissipation module, and most of its elements are zero, and only the thermal resistance values between the nodes of the key heat conduction paths are recorded.

7. The packaging method of the all-silicon carbide double-sided heat dissipation module optimized based on gradient thermal resistance according to claim 6, characterized in that, The thermal resistance gradient optimization function is used to perform non-linear optimization calculations on the low-dimensional thermal resistance matrix to find the best thermal resistance distribution scheme. The inputs include the low-dimensional thermal resistance matrix, the temperature gradient constraint parameters obtained from the heat dissipation temperature distribution matrix, the upper limit parameter of heat flux density obtained from the upper-layer function of heat transfer conduction, the material thermal conductivity vector of the silicon carbide double-sided heat dissipation module, and the interface contact thermal resistance parameter obtained from the lower-layer function of heat transfer conduction. The output is the optimal heat dissipation path parameter set, and the optimal heat dissipation path parameter set includes the optimized thermal resistance distribution scheme and the corresponding temperature uniformity index.

8. The encapsulation method of the all-silicon carbide double-sided heat dissipation module optimized based on gradient thermal resistance according to claim 7, characterized in that, The specific structure of the thermal resistance prediction deep attention network model is a deep convolutional neural network based on the multi-head self-attention mechanism, including an encoder-decoder architecture. The encoder consists of three layers of multi-head self-attention layers and two layers of feed-forward neural networks, which are used to extract the thermal resistance distribution features. The weight dimension of each self-attention head is determined by the eigenvalues of the heat dissipation temperature distribution matrix. The decoder consists of two layers of cross-attention layers and one layer of feed-forward neural network, which is used to generate an optimized thermal resistance distribution scheme.

9. The method for packaging a fully silicon carbide double-sided heat dissipation module optimized based on gradient thermal resistance according to claim 8, characterized in that, The thermal resistance low-dimensional matrix specifically refers to the key feature matrix extracted from the thermal resistance sparse matrix through the singular value decomposition technique, which reduces the computational complexity while retaining the main heat conduction information.

10. The packaging method of the all-silicon carbide double-sided heat dissipation module optimized based on gradient thermal resistance according to claim 9, characterized in that, The sandwich structure specifically refers to a stacked heat dissipation structure formed by placing a silicon carbide double-sided heat dissipation module between the upper and lower heat dissipation substrates, realizing the two-way heat dissipation function.

Images