A power device package optimization method based on a BPNN-WOA fusion model

Abstract

This invention discloses a power device packaging optimization method based on a BPNN-WOA fusion model, relating to the field of power device thermal management. It addresses the problems of traditional optimization methods, such as neglecting nonlinear coupling of parameters, difficulty in global optimization, and handling complex discrete variables. This method relies on a database of heat dissipation influencing factors and a high-fidelity 3D thermal simulation model. Parameters are selected and standardized to generate a sample dataset, which is then used to train a BPNN thermal resistance prediction surrogate model. The WOA algorithm's whale position vector is mapped to a combination of packaging parameters, and the BPNN output is mapped to a fitness value. The optimal parameter combination is obtained through global iterative optimization using WOA, and verified by simulation after inverse normalization. If the deviation exceeds a threshold, additional samples are added to retrain the model. This method can capture the synergistic effect between parameters, achieving global optimization in a continuous-discrete hybrid design space, significantly reducing packaging thermal resistance, and improving optimization efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to the field of power device thermal management, and specifically to a power device packaging optimization method based on a BPNN-WOA fusion model. Background Technology

[0002] When power devices operate continuously under high power and high voltage conditions, heat accumulates rapidly in the chip junction region. Package thermal resistance, as a core indicator of heat dissipation efficiency, directly determines the device's junction temperature level and long-term reliability. Therefore, thermal optimization of the package structure has become a crucial step in the power device design process. However, current optimization methods for package thermal resistance face three deep-seated challenges: First, while traditional single-variable methods and orthogonal experimental methods are simple and easy to implement, they are essentially based on the assumption of "independent and decomposable parameters," ignoring the nonlinear coupling and synergistic effects commonly found between package parameters such as chip area, solder layer thickness, and lead frame material. This leads to the "optimal solution" obtained from isolated analysis often deviating from the true optimum during actual system integration. Second, strategies such as orthogonal experiments and response surface methodology, based on discrete level combinations or multivariate fitting, heavily rely on preset level intervals and modulo operations for their optimization results. First, the current method can only select relatively optimal values ​​from a limited number of test points, making it difficult to achieve true global optimization within a continuous design space. It is also prone to overlooking better design points located in unsampled intervals. Second, the widespread discrete variables in packaging design—such as molding materials, solder types, and lead frame materials—make it difficult for methods like response surface methodology and BPNN optimization based on the continuous gradient assumption to model these variables or get trapped in local optima. Forcing continuous processing makes optimization impossible, while purely discrete optimization drastically increases problem complexity and computational overhead. These technical shortcomings are intertwined, making it difficult for existing methods to stably search for the globally optimal combination of packaging parameters from a continuous-discrete hybrid design space while maintaining both efficiency and accuracy. Summary of the Invention

[0003] The purpose of this invention is to provide a power device packaging optimization method based on the BPNN-WOA fusion model, thereby solving the problems mentioned in the background art.

[0004] This invention is achieved through the following technical solution:

[0005] A power device packaging optimization method based on a BPNN-WOA fusion model includes a database storing different factors affecting package heat dissipation and their corresponding structural and material feasible domains, and a three-dimensional thermal simulation model of the power device packaging structure to be optimized, comprising the following steps:

[0006] S1. Retrieve the database through the three-dimensional thermal simulation model to determine and obtain the influencing factors corresponding to the three-dimensional thermal simulation model, as well as the structural feasible domain and material feasible domain corresponding to each influencing factor;

[0007] S2. Select M values ​​from the structural feasible domains corresponding to the influencing factors according to a preset interval threshold to form a selection feasible domain. After standardizing the values ​​in the selection feasible domains, select one element from each selection feasible domain and combine it with one element from each material feasible domain to obtain the encapsulation parameter combination, thereby obtaining a standardized sample dataset.

[0008] S3. After dividing the standardized sample dataset, the BPNN model is trained and validated to obtain the BPNN model for rapid prediction of thermal resistance.

[0009] S4. Map the whale position vector of the WOA algorithm to the combination of encapsulation parameters, and map the thermal resistance prediction value output by the BPNN model to the WOA algorithm fitness value corresponding to the whale position vector. Perform global iterative optimization through the WOA algorithm to obtain the optimal combination of encapsulation parameters.

[0010] S5. After denormalizing the optimal packaging parameter combination, input it into the three-dimensional thermal simulation model for simulation verification to confirm the actual thermal resistance value corresponding to the optimal packaging parameter combination, thus completing the optimization of power device packaging.

[0011] Furthermore,

[0012] The three-dimensional thermal simulation model in S1 is constructed based on the actual geometry, material physical properties, and heat conduction path of the power device package. The model can reflect the complete thermal path from the chip junction to the plastic package. The model includes key components such as the chip, solder layer, lead frame, and plastic package, and sets matching boundary conditions and heat sources for each component.

[0013] Furthermore,

[0014] The influencing factors include chip area, chip thickness, solder layer thickness, lead frame thickness, molding compound, solder layer material, and lead frame material. Each influencing factor has a feasible region. The feasible regions of chip area, chip thickness, solder layer thickness, and lead frame thickness are structural feasible regions, while the feasible regions of molding compound, solder layer material, and lead frame material are material feasible regions.

[0015] Furthermore,

[0016] The standardization process in S2 specifically involves selecting elements within the feasible region and transforming them to the numerical range of [-1, 1]. The transformation formula is as follows: Where X is the value of an element selected from the feasible region, X minX represents the minimum value of the structurally feasible region of the influencing factors corresponding to the element. max y represents the maximum value of the structurally feasible region corresponding to the influencing factor, and y is the standardized value.

[0017] Furthermore,

[0018] The BPNN model in S3 has a three-layer network structure, consisting of an input layer, a single hidden layer, and an output layer. The number of nodes in the hidden layer is determined by combining an empirical formula with a trial-and-error method based on the number of nodes in the input and output layers. The empirical formula is as follows: , where m is the number of input layer nodes, n is the number of output layer nodes, a is a constant from 1 to 10, and h is the number of hidden layer nodes.

[0019] Furthermore,

[0020] In step S3, the standardized sample dataset is divided into a training set, a validation set, and a test set according to a preset ratio. The BPNN model is trained using the training set, the training process is monitored in real time using the validation set to prevent overfitting, and the prediction accuracy of the model is evaluated using the coefficient of determination R² score on the test set. The formula for calculating the coefficient of determination R² score is as follows: , where y i The actual thermal resistance value is obtained by physical simulation calculation after substituting the i-th group of packaging parameters into the three-dimensional thermal simulation model; The standardized encapsulation structure parameter vector of the i-th group is input into the trained BPNN model, and the thermal resistance prediction value output by the BPNN model is ȳ, which is the mean of the actual thermal resistance values ​​of all samples in the test set.

[0021] Furthermore,

[0022] The optimal combination of packaging parameters in S4 specifically aims to minimize the packaging thermal resistance, and its objective function is: Where P is the individual whale position vector corresponding to the combination of encapsulation parameters. P s1 , P s2 , ... P sz For each factor, consider both the structural and material feasible regions, where z represents the type of factor selected, and s represents different elements within the feasible region of factor selection. This is the package thermal resistance value output by the BPNN model.

[0023] Furthermore,

[0024] In the global iterative optimization process of the WOA algorithm in S4, the position update method of bubble net predation or shrinking enclosure is selected according to the preset probability. At the same time, the search strategy of global random search or local optimal approximation is switched according to the absolute value of the coefficient vector in the WOA algorithm. The optimization stops when the preset convergence condition or the maximum number of iterations is met.

[0025] Furthermore,

[0026] The inverse standardization in S5 is the inverse transformation of the standardization process. It restores the standardized parameters to the actual physical parameters based on the maximum and minimum values ​​of the structurally feasible region of the corresponding influencing factors. The inverse standardization formula is as follows: y is the standardized value, and X is the value of an element selected from the feasible region. min X is the minimum value of the structurally feasible region corresponding to the influencing factors. max This represents the maximum value of the structurally feasible region corresponding to the influencing factor.

[0027] Furthermore,

[0028] In step S5, if the actual thermal resistance value verified by simulation deviates from the thermal resistance prediction value of the BPNN model outside the preset range, the corresponding packaging parameter combination and the corresponding actual thermal resistance value are added to the standardized sample dataset. After retraining the BPNN model, global iterative optimization is performed again until the optimal packaging parameter combination with the deviation within the preset deviation range is obtained.

[0029] The beneficial effects of this invention are:

[0030] 1. This invention constructs a nonlinear mapping proxy model between encapsulation parameters and thermal resistance values ​​using a BPNN model. This model can accurately capture the nonlinear coupling and synergistic effect between multiple parameters, solving the problem that traditional optimization methods ignore parameter interactions and cause optimization results to deviate from the true optimum. This significantly improves the accuracy and efficiency of thermal resistance prediction.

[0031] 2. This invention achieves global optimization of the mixed design space of continuous structural parameters and discrete material parameters by deeply integrating the WOA algorithm and the BPNN model. It avoids the shortcomings of traditional methods that are prone to getting trapped in local optima and have difficulty in handling discrete variables. It can stably search for the globally optimal combination of packaging parameters and significantly reduce the thermal resistance of power device packaging.

[0032] 3. This invention constructs a closed-loop optimization process of "simulation modeling - sample construction - model training - global optimization - verification iteration". Through the iterative mechanism of simulation verification and sample supplementation, the prediction error of the surrogate model is continuously corrected, ensuring the engineering feasibility and reliability of the optimization results, and significantly reducing the R&D cycle and computational cost of power device packaging thermal optimization. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the process of the present invention;

[0034] Figure 2 This is a graph showing the average resistance error of the BPNN model of this invention with different numbers of hidden layer nodes;

[0035] Figure 3 A line graph showing the comparison between the actual and predicted values ​​on the test set of the BPNN model;

[0036] Figure 4 A line graph showing the error distribution between the actual and predicted values ​​on the test set of the BPNN model;

[0037] Figure 5 The temperature distribution cloud maps before and after the optimization of the power device packaging structure are compared. Detailed Implementation

[0038] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0039] See the example. Figures 1 to 5 :

[0040] A power device packaging optimization method based on a BPNN-WOA fusion model includes a database storing different factors affecting package heat dissipation and their corresponding structural and material feasible domains, and a three-dimensional thermal simulation model constructed based on the power device packaging structure to be optimized. The method includes the following steps:

[0041] S1. Retrieve the database through the three-dimensional thermal simulation model to determine and obtain the influencing factors corresponding to the three-dimensional thermal simulation model, as well as the structural feasible domain and material feasible domain corresponding to each influencing factor;

[0042] S2. Select M values ​​from the structural feasible domains corresponding to the influencing factors according to a preset interval threshold to form a selection feasible domain. After standardizing the values ​​in the selection feasible domains, select one element from each selection feasible domain and combine it with one element from each material feasible domain to obtain the encapsulation parameter combination, thereby obtaining a standardized sample dataset.

[0043] S3. After dividing the standardized sample dataset, the BPNN model is trained and validated to obtain the BPNN model for rapid prediction of thermal resistance.

[0044] S4. Map the whale position vector of the WOA algorithm to the combination of encapsulation parameters, and map the thermal resistance prediction value output by the BPNN model to the WOA algorithm fitness value corresponding to the whale position vector. Perform global iterative optimization through the WOA algorithm to obtain the optimal combination of encapsulation parameters.

[0045] S5. After denormalizing the optimal packaging parameter combination, input it into the three-dimensional thermal simulation model for simulation verification to confirm the actual thermal resistance value corresponding to the optimal packaging parameter combination, thus completing the optimization of power device packaging.

[0046] Furthermore,

[0047] The three-dimensional thermal simulation model in S1 is constructed based on the actual geometry, material physical properties, and heat conduction path of the power device package. The model can reflect the complete thermal path from the chip junction to the plastic package. The model includes key components such as the chip, solder layer, lead frame, and plastic package, and sets matching boundary conditions and heat sources for each component.

[0048] In this embodiment, the three-dimensional thermal simulation model is built in COMSOL Multiphysics software based on the actual geometry, material physical properties, and heat conduction path of the TO-220 package. The model fully includes key components such as the chip, solder layer, lead frame, and plastic package, completely covering the entire heat conduction path from the chip junction region to the plastic package. The model sets the top of the chip as a constant 2W heat source and the bottom surface of the plastic package as a convection heat transfer boundary condition. Subsequent sample data in this embodiment are all generated based on this high-fidelity benchmark model, ensuring the physical authenticity of the sample data and the engineering credibility of the optimization results.

[0049] Furthermore,

[0050] The influencing factors include chip area, chip thickness, solder layer thickness, lead frame thickness, molding compound, solder layer material, and lead frame material. Each influencing factor has a feasible region. The feasible regions of chip area, chip thickness, solder layer thickness, and lead frame thickness are structural feasible regions, while the feasible regions of molding compound, solder layer material, and lead frame material are material feasible regions.

[0051] The example selected seven factors affecting heat dissipation in the packaging, which were divided into four continuous structural parameters and three discrete material parameters, as detailed below:

[0052] Continuous structural parameters and corresponding feasible regions: chip area [6.65, 36] mm², chip thickness [0.1, 0.3] mm, solder layer thickness [0.03, 0.07] mm, lead frame thickness [0.4, 1] mm;

[0053] Discrete material parameters and corresponding feasible regions: molding compound (CEL1802, CEL9750), solder layer material (SAC305, Sn63Pb37), lead frame material (Cu, Cu-Ni alloy); among them, the upper limit of the feasible region of chip area, 36mm², is the upper limit of the industry standard physical size of TO-220 package, which cannot be expanded due to the limitations of package shape specifications; and through thermal conduction theory analysis and simulation verification, increasing the chip area can effectively reduce the heat flux density on the chip surface, thereby reducing the thermal resistance of the package. Therefore, 36mm² is the optimal value under this package specification. In this embodiment, continuous structural parameters are standardized to the [-1,1] interval before participating in model training, and discrete material parameters participate in parameter combination using an enumeration index method. Through variable classification processing, the fitting accuracy of the BPNN model to the continuous physical trend and discrete material characteristics is ensured.

[0054] Furthermore,

[0055] The standardization process in S2 specifically involves selecting elements within the feasible region and transforming them to the numerical range of [-1, 1]. The transformation formula is as follows: Where X is the value of an element selected from the feasible region, X min X represents the minimum value of the structurally feasible region of the influencing factors corresponding to the element. max y represents the maximum value of the structurally feasible region corresponding to the influencing factor, and y is the standardized value.

[0056] This formula applies only to four continuous structural parameters: chip area, chip thickness, solder layer thickness, and lead frame thickness. This transformation eliminates the differences in dimensions and orders of magnitude between different parameters, preventing the numerical magnitude difference between chip area (6.65~36mm²) and solder layer thickness (0.03~0.07mm) from dominating the BPNN model weight update. This ensures the network's fitting accuracy for all parameters while preserving the reversibility of the parameters, providing a basis for subsequent denormalization to restore engineering parameters.

[0057] Furthermore,

[0058] The BPNN model in S3 has a three-layer network structure, consisting of an input layer, a single hidden layer, and an output layer. The number of nodes in the hidden layer is determined by combining an empirical formula with a trial-and-error method based on the number of nodes in the input and output layers. The empirical formula is as follows: , where m is the number of input layer nodes, n is the number of output layer nodes, a is a constant from 1 to 10, and h is the number of hidden layer nodes.

[0059] The number of hidden layer nodes can be determined using the empirical formula described above. (a=1~10) and determined by trial and error to avoid underfitting due to too few hidden layer nodes and inability to capture nonlinear coupling between parameters, while preventing overfitting and decreased generalization ability due to too many nodes, thus achieving the optimal balance between model complexity and prediction accuracy. In this embodiment, the following is executed: input layer m=7 (corresponding to 7 factors affecting package heat dissipation), output layer n=1 (corresponding to the predicted thermal resistance value). Substituting into the empirical formula, when a takes a constant from 1 to 10, the theoretical reference value of h is obtained. Combining engineering practice and network fitting requirements, the integer range is expanded based on the theoretical value. Experiments show that the average error decreases before the number of hidden nodes is 12, so the part before the number of hidden nodes is removed from the figure, such as Figure 2 As shown, this section displays the number of hidden nodes from 8 to 18, which makes the changes in the average error more obvious. As the number of hidden layer nodes increases from 8 to 12, the average error of the model's thermal resistance prediction continues to decrease, and the average absolute error drops to the lowest point when h=12. When the number of hidden layer nodes exceeds 12, due to the overfitting effect caused by the increase in network complexity, the average absolute error of the model gradually increases with the number of nodes.

[0060] The formula for the forward propagation process of the BPNN model is as follows:

[0061] The output formula of the hidden layer neuron is In the formula, H j x is the output of the j-th neuron in the hidden layer. i w is the i-th input parameter of the input layer. ij b represents the connection weights from the input layer to the hidden layer. j Let f(⋅) be the bias of the j-th neuron in the hidden layer, and f(⋅) be the sigmoid activation function, expressed as follows: This is used to fit the nonlinear coupling relationship between the packaging parameters and thermal resistance of this scheme;

[0062] The formula for calculating the predicted thermal resistance of the output layer is as follows: In the formula, w is the predicted thermal resistance value of the output layer. jk Let b be the connection weight from the j-th neuron in the hidden layer to the k-th neuron in the output layer. k This is the bias of the k-th neuron in the output layer;

[0063] The backpropagation process of the BPNN model uses mean squared error as the loss function to quantify the deviation between the predicted thermal resistance value and the simulated true value, driving the iterative update of network weights and biases. The loss function formula is as follows: In the formula, N is the total number of samples in the training set. This represents the predicted thermal resistance value of the k-th sample output by the output layer. The actual thermal resistance value of the k-th sample obtained from 3D thermal simulation is given. To minimize the above loss function, this scheme uses gradient descent to perform error backpropagation. Based on the partial derivatives of the loss function with respect to each parameter, the weights and bias parameters of the network are iteratively updated layer by layer from the output layer to the input layer. The core update rule is: updated parameter value = original parameter value - preset learning rate × partial derivative of the loss function with respect to that parameter. The corresponding weight and bias update formulas for two layers of the network are as follows:

[0064] The formulas for updating weights and biases from the hidden layer to the output layer are as follows:

[0065] ;

[0066] The formulas for updating the weights and biases from the input layer to the hidden layer are as follows:

[0067] ;

[0068] In the above formulas, η is the preset learning rate, used to control the parameter update step size in a single iteration; t is the current iteration number. Let be the network weight values ​​at the t-th iteration. This represents the network bias value at the t-th iteration; These are the partial derivatives of the loss function with respect to the corresponding weights and bias parameters, respectively. The network weights and biases are iteratively updated using the above formulas to continuously minimize the thermal resistance prediction loss, thus completing the training of the BPNN model.

[0069] Furthermore,

[0070] In step S3, the standardized sample dataset is divided into a training set, a validation set, and a test set according to a preset ratio. The BPNN model is trained using the training set, the training process is monitored in real time using the validation set to prevent overfitting, and the prediction accuracy of the model is evaluated using the coefficient of determination R² score on the test set. The formula for calculating the coefficient of determination R² score is as follows: , where y i The actual thermal resistance value is obtained by physical simulation calculation after substituting the i-th group of packaging parameters into the three-dimensional thermal simulation model; The standardized encapsulation structure parameter vector of the i-th group is input into the trained BPNN model, and the thermal resistance prediction value output by the BPNN model is ȳ, which is the mean of the actual thermal resistance values ​​of all samples in the test set.

[0071] In this embodiment, 648 sets of standardized sample data were randomly divided into a training set of 454 sets, a validation set of 97 sets, and a test set of 97 sets, in a ratio of 7:1.5:1.5. The training set was used for iterative updates of network weights and biases. The validation set was used to calculate the loss value after each training round. Training was terminated immediately when the validation loss did not decrease for 10 consecutive rounds, effectively suppressing model overfitting. The test set was not involved in the training process and was used exclusively for unbiased evaluation of the model's prediction accuracy. Testing showed that the BPNN model trained in this embodiment achieved the following results: [The text abruptly ends here, so the translation stops as well.] Figure 3 As shown, the predicted value almost perfectly matches the actual value, and the deviation between the predicted thermal resistance and the simulated actual thermal resistance is extremely small; the relative error distribution of the predicted thermal resistance is as follows. Figure 4 As shown, the relative errors of all test samples are controlled within the allowable range, demonstrating extremely high fitting accuracy and generalization performance. Figure 3 and Figure 4 Because the number of random test samples is large, they appear densely packed on the graph, leading to... Figure 3 , Figure 4 The line chart could not be displayed more intuitively and clearly, so the portion of the random test sample number 1 to 8 was enlarged for clearer display.

[0072] Furthermore,

[0073] The optimal combination of packaging parameters in S4 specifically aims to minimize the packaging thermal resistance, and its objective function is: Where P is the individual whale position vector corresponding to the combination of encapsulation parameters. P s1 , P s2 , ... P sz These are the parameter values ​​selected from the structural and material feasible regions for each factor, where z represents the type of factor selected, and s represents different elements within the feasible region for the selection of the factor. This is the package thermal resistance value output by the BPNN model.

[0074] In this embodiment, the position vector P of an individual whale is defined as a 7-dimensional real vector, where the first 4 dimensions correspond to standardized continuous structural parameters (range [-1, 1]), and the last 3 dimensions are rounded down and mapped to the enumeration index of discrete material parameters. The position vector P is input into the trained BPNN model, and the thermal resistance prediction value output by the model is used as the fitness value of the WOA algorithm. The population is driven to iteratively evolve by minimizing the fitness value. Through this mapping relationship, the continuous-discrete hybrid encapsulation optimization problem is transformed into a numerical optimization problem that can be efficiently handled by the WOA algorithm, which greatly reduces the computational overhead of the optimization process and solves the problem that directly calling the simulation model for optimization is extremely inefficient and lacks engineering feasibility.

[0075] Furthermore,

[0076] In the global iterative optimization process of the WOA algorithm in S4, the position update method of bubble net predation and shrinking enclosure is selected according to the preset probability. At the same time, the search strategy of global random search or local optimal approximation is switched according to the absolute value of the coefficient vector in the WOA algorithm. The optimization stops when the preset convergence condition or the maximum number of iterations is met.

[0077] In this embodiment, the core optimization process of the WOA algorithm is implemented through the following formula, adapting to the global optimization requirements of continuous-discrete mixed variables in this scheme. In the shrinking encirclement mechanism, the whale individual position update formula is: In the formula, t is the current iteration number, and the maximum number of iterations in this embodiment is set to 100. This represents the globally optimal individual position vector at the current iteration number. This is the current position vector of the individual whale.

[0078] , Let the coefficient vector be the formula for calculation. In the formula, The convergence factor decreases from 2 to 0 with each iteration. , Let be a random vector between [0,1], when | When |<1, the local shrinkage encirclement mechanism is executed to perform a local fine-grained search around the current optimal solution.

[0079] In the bubble network attack mechanism, this embodiment also has the probability of employing a local spiral update method, the formula of which is: In the formula, Let be the distance between the current individual and the global best individual, and b be the logarithmic spiral shape constant; l A random number between [-1, 1];

[0080] Based on the two local update methods described above, the final position update selection formula is: In the formula, p is a random number between [0,1]. This formula balances the two local development modes of shrinkage enclosure and spiral update.

[0081] When | When |≥1, the global search mechanism is executed, and the position update formula is: In the formula, The individual position vector of the current random individual.

[0082] Furthermore,

[0083] The inverse standardization in S5 is the inverse transformation of the standardization process. It restores the standardized parameters to the actual physical parameters based on the maximum and minimum values ​​of the structurally feasible region of the corresponding influencing factors. The inverse standardization formula is as follows: y is the standardized value, and X is the value of an element within the feasible region. min X is the minimum value of the structurally feasible region corresponding to the influencing factors. max This represents the maximum value of the structurally feasible region corresponding to the influencing factor.

[0084] This formula uniquely and accurately restores the optimal position vector obtained by WOA in the [-1,1] normalized space to the actual package physical parameters that can be manufactured and measured, ensuring zero-distortion transmission of optimization results from theoretical optimization to engineering implementation.

[0085] Furthermore,

[0086] In step S5, if the actual thermal resistance value verified by simulation deviates from the thermal resistance prediction value of the BPNN model outside the preset deviation range, the corresponding packaging parameter combination and the corresponding actual thermal resistance value are added to the standardized sample dataset. After retraining the BPNN model, global iterative optimization is performed again until the optimal packaging parameter combination with the deviation within the preset deviation range is obtained.

[0087] This feature, through a closed-loop feedback mechanism, supplements the standardized sample set with the optimal combination and actual thermal resistance value that show significant deviations in simulation verification, and triggers BPNN retraining. It constructs an adaptive optimization closed loop of "surrogate model prediction → WOA optimization → physical simulation confirmation → model iterative evolution," which can continuously correct potential approximation errors of BPNN at the edge of the design space or the boundary of discrete variable combinations, ensuring the engineering robustness of the optimization results. In the example, the optimal combination obtained in the first optimization (chip area 36mm², thickness 0.148mm, solder layer 0.0435mm, lead frame 0.4264mm, CEL9750 molding compound, SAC305 solder, Cu lead frame) was substituted into COMSOL simulation. The measured thermal resistance was 0.261℃ / W, which deviated from the BPNN prediction of 0.259℃ / W by 0.002℃ / W (<0.01℃ / W preset threshold), and directly passed the verification. To verify the effectiveness of this feature, an extreme test was artificially constructed in the example: the combination of "Cu-Ni lead frame + CEL9750 + SAC305" was deleted from the initial sample set, and the model was retrained and optimized. The optimal combination recommended was the Cu-Ni lead frame. After substituting it into the simulation, the actual thermal resistance was 0.312℃ / W, which deviated from the predicted value of 0.283℃ / W by 0.029℃ / W, exceeding the threshold. Then, after adding this deviation group to the sample set and retraining the BPNN, the new model's prediction accuracy for this combination improved to 0.309℃ / W. Further iterative optimization converged back to the Cu lead frame, and the final thermal resistance recovered to 0.259℃ / W. This ensures the continued reliability of the fusion model in practical engineering applications and avoids the key barrier of optimization failure due to sparse initial samples. The optimal packaging parameter combination was obtained and input into a 3D thermal simulation model for verification, and compared with the 3D thermal simulation model before optimization. Figure 5 As shown.

[0088] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A power device packaging optimization method based on a BPNN-WOA fusion model, comprising a database storing different influencing factors affecting package heat dissipation and their corresponding structural and material feasible domains, and a three-dimensional thermal simulation model of the power device packaging structure to be optimized, characterized in that, Includes the following steps: S1. Retrieve the database through the three-dimensional thermal simulation model to determine and obtain the influencing factors corresponding to the three-dimensional thermal simulation model, as well as the structural feasible domain and material feasible domain corresponding to each influencing factor; S2. Select M values ​​from the structural feasible domains corresponding to the influencing factors according to a preset interval threshold to form a selection feasible domain. After standardizing the values ​​in the selection feasible domains, select one element from each selection feasible domain and combine it with one element from each material feasible domain to obtain the encapsulation parameter combination, thereby obtaining a standardized sample dataset. The influencing factors include chip area, chip thickness, solder layer thickness, lead frame thickness, molding compound, solder layer material, and lead frame material. Each influencing factor has a feasible domain. The feasible domains of chip area, chip thickness, solder layer thickness, and lead frame thickness are structural feasible domains, while the feasible domains of molding compound, solder layer material, and lead frame material are material feasible domains. S3. After dividing the standardized sample dataset, the BPNN model is trained and validated to obtain the BPNN model for rapid prediction of thermal resistance. S4. Map the whale position vector of the WOA algorithm to the combination of encapsulation parameters, and map the thermal resistance prediction value output by the BPNN model to the WOA algorithm fitness value corresponding to the whale position vector. Perform global iterative optimization through the WOA algorithm to obtain the optimal combination of encapsulation parameters. S5. After denormalizing the optimal packaging parameter combination, input it into the three-dimensional thermal simulation model for simulation verification to confirm the actual thermal resistance value corresponding to the optimal packaging parameter combination, thus completing the optimization of power device packaging.

2. The power device package optimization method based on the BPNN-WOA fusion model according to claim 1, characterized in that, The three-dimensional thermal simulation model in S1 is specifically constructed based on the actual geometric structure, material physical properties, and heat conduction path of the power device package. The model can reflect the complete thermal path from the chip junction to the plastic package. The model includes key components such as the chip, solder layer, lead frame, and plastic package, and sets matching boundary conditions and heat sources for each component.

3. The power device package optimization method based on the BPNN-WOA fusion model according to claim 1, characterized in that, S2 is normalized, specifically: selecting the element data in the feasible region to transform to [-1, 1] numerical interval, the transformation formula is Wherein X is the value of an element selected in the feasible region, X min is the minimum value of the structural feasible region of the corresponding influence factor of the element, X max is the maximum value of the structural feasible region of the corresponding influence factor, y is the value after normalization.

4. The power device package optimization method based on the BPNN-WOA fusion model according to claim 1, characterized in that, The BPNN model in S3 has a three-layer network structure, consisting of an input layer, a single hidden layer, and an output layer. The number of nodes in the hidden layer is determined by combining an empirical formula with a trial-and-error method based on the number of nodes in the input and output layers. The empirical formula is as follows: , where m is the number of input layer nodes, n is the number of output layer nodes, a is a constant from 1 to 10, and h is the number of hidden layer nodes.

5. The power device package optimization method based on the BPNN-WOA fusion model according to claim 1, characterized in that, The S3 step of dividing the standardized sample dataset specifically involves: dividing the standardized sample dataset into a training set, a validation set, and a test set according to a preset ratio; training the BPNN model using the training set; monitoring the training process in real time using the validation set to prevent overfitting; and evaluating the model's prediction accuracy using the coefficient of determination (R²) score on the test set. The formula for calculating the coefficient of determination (R²) score is as follows: y i The actual thermal resistance value is obtained by physical simulation calculation after substituting the i-th group of packaging parameters into the three-dimensional thermal simulation model; The standardized encapsulation structure parameter vector of the i-th group is input into the trained BPNN model, and the thermal resistance prediction value is output by the BPNN model. ȳ represents the mean of the actual thermal resistance values ​​of all samples in the test set.

6. The power device package optimization method based on the BPNN-WOA fusion model according to claim 1, characterized in that, The optimal combination of packaging parameters in S4 specifically aims to minimize the thermal resistance of the package, and its objective function is: Where P is the individual whale position vector corresponding to the encapsulation parameter combination, P s1 P s2 ...P sz These are the parameter values ​​selected from the structural and material feasible regions for each factor, where z represents the type of factor selected, and s represents different elements within the feasible region for the selection of the factor. This is the package thermal resistance value output by the BPNN model.

7. The power device package optimization method based on the BPNN-WOA fusion model according to claim 1, characterized in that, In the global iterative optimization process of the WOA algorithm in S4, the position update method of bubble net predation or shrinking enclosure is selected according to the preset probability. At the same time, the search strategy of global random search or local optimal approximation is switched according to the absolute value of the coefficient vector in the WOA algorithm. The optimization stops when the preset convergence condition or the maximum number of iterations is met.

8. The power device package optimization method based on the BPNN-WOA fusion model according to claim 1, characterized in that, The inverse standardization in S5 is inverse transformation of standardization, which restores the standardization parameter to the actual physical parameter based on the maximum and minimum of the structure feasible region of the corresponding influencing factor, wherein the inverse standardization formula is: , y is the standardized value, X is the value of an element selected from the feasible region, X min is the minimum of the structure feasible region of the corresponding influencing factor, and X max is the maximum of the structure feasible region of the corresponding influencing factor.

9. The power device package optimization method based on the BPNN-WOA fusion model according to claim 1, characterized in that, In S5, if the actual thermal resistance value verified by simulation deviates from the thermal resistance prediction value of the BPNN model outside the preset range, the corresponding packaging parameter combination and the corresponding actual thermal resistance value are added to the standardized sample dataset. After retraining the BPNN model, global iterative optimization is performed again until the optimal packaging parameter combination with the deviation within the preset deviation range is obtained.

Images