Three-dimensional electromagnetic-thermal co-simulation method for gallium nitride device

Abstract

The invention provides a gallium nitride device-oriented three-dimensional electromagnetic-thermal co-simulation method, which relates to the technical field of gallium nitride devices, and comprises the steps of generating a training sample set by collecting device parameters, obtaining electromagnetic field and temperature field feature vectors to establish a control equation set, carrying out grid division by adopting a self-adaptive hierarchical subdivision method, and obtaining a three-dimensional electromagnetic-thermal co-simulation result. Grid density distribution is determined based on the temperature gradient and the material interface characteristics, and a multi-level grid structure is generated; solving the electromagnetic field equation to obtain loss power density, and establishing a temperature field heat source item to calculate temperature distribution; and judging the thermal stability state according to the temperature change rate, updating the material parameters and optimizing the grid structure. According to the method, high-precision simulation of electromagnetic-thermal coupling analysis is realized, and the calculation efficiency is improved.

Description

Technical Field

[0001] The present invention relates to gallium nitride device technology, and particularly to a three-dimensional electromagnetic-thermal co-simulation method for gallium nitride devices. Background Art

[0002] Due to their excellent electrical and thermal properties, gallium nitride devices are widely used in high-frequency and high-power electronic fields. During actual operation, gallium nitride devices generate a large amount of heat, leading to an increase in device temperature, which in turn affects their electrical properties and reliability. To accurately predict the operating characteristics of gallium nitride devices, it is necessary to perform co-simulation analysis of their electromagnetic field and temperature field.

[0003] Traditional electromagnetic-thermal simulation methods usually use independent electromagnetic field solvers and temperature field solvers to achieve the coupled analysis of the electromagnetic field and temperature field through serial calculation. This method first calculates the electromagnetic field distribution, and then uses the electromagnetic loss as the heat source to input into the temperature field solver for thermal analysis. At the same time, existing simulation methods generally adopt a uniform mesh division strategy, using the same mesh density throughout the computational domain, or using a simple local refinement method to refine the mesh in the area of interest.

[0004] First of all, the traditional serial calculation method ignores the feedback effect of temperature changes on the electrical properties of materials, and cannot accurately reflect the strong coupling effect between the electromagnetic field and temperature field, resulting in a large deviation between the simulation results and the actual situation.

[0005] Secondly, the uniform mesh division strategy has insufficient accuracy at the regions with large temperature gradients and material interfaces, and there is a waste of computing resources in the regions with gentle temperature changes, making it difficult to balance the requirements of computing accuracy and efficiency.

[0006] Finally, existing simulation methods lack a dynamic optimization mechanism for the mesh structure, and cannot adaptively adjust the mesh distribution according to the evolution process of the temperature field, making it difficult to ensure the simulation accuracy when the operating state of the device changes. Summary of the Invention

[0007] Embodiments of the present invention provide a three-dimensional electromagnetic-thermal co-simulation method for gallium nitride devices, which can solve the problems in the prior art.

[0008] In the first aspect of the embodiments of the present invention, a three-dimensional electromagnetic-thermal co-simulation method for gallium nitride devices is provided, including: The material parameters, structural parameters and working parameters of the gallium nitride device are collected to generate an initial training sample set, and the electromagnetic field characteristic vector and the temperature field characteristic vector are obtained based on the initial training sample set. The electromagnetic field control equation group and the temperature field control equation group are established according to the electromagnetic field characteristic vector and the temperature field characteristic vector. The gallium nitride device is meshed by an adaptive hierarchical partitioning method, and the mesh density distribution is determined based on the temperature gradient and the material interface characteristics. A fine mesh is used in the area of ​​drastic temperature change and the interface transition area, and a loose mesh is used in the area of ​​gentle temperature, so as to generate a multi-level mesh structure with regional adaptive characteristics; Input the multi-level grid structure into the computational solver, solve the electromagnetic field control equations based on the electromagnetic field eigenvector to obtain the electromagnetic field distribution data, calculate the Joule heat loss power density based on the electric field distribution data, calculate the hysteresis loss power density based on the magnetic field distribution data, superimpose the Joule heat loss power density and the hysteresis loss power density in the spatial distribution to obtain the heat source distribution matrix, combine the heat source distribution matrix with the temperature field eigenvector to establish the heat source term of the temperature field, and calculate the temperature field distribution data based on the temperature field control equations containing the heat source term; The temperature gradient and temperature change rate are calculated based on the temperature field distribution data. When the temperature change rate exceeds the preset stability threshold, it is determined that the thermal stability state has not been reached. The electrical conductivity and thermal conductivity parameters of the material are updated based on the temperature field distribution data. The updated material physical parameters and temperature field distribution data are input into the grid optimization algorithm to dynamically optimize the multi-level grid structure.

[0009] The electromagnetic field control equations and the temperature field control equations are established according to the electromagnetic field eigenvectors and the temperature field eigenvectors. The GaN device is meshed using the adaptive hierarchical partitioning method. The mesh density distribution is determined based on the temperature gradient and material interface characteristics. A fine mesh is used in the area of ​​drastic temperature changes and the interface transition area, and a loose mesh is used in the area of ​​gentle temperature. A multi-level mesh structure with regional adaptive characteristics is generated, including: Obtaining material parameters and structural parameters of the gallium nitride device, the material parameters include electrical conductivity data and thermal conductivity data, and the structural parameters include material interface parameters, and calculating the electromagnetic field characteristic vector and the temperature field characteristic vector based on the material parameters and the structural parameters; According to the electromagnetic field characteristic vector, an electromagnetic field control equation group is established, which describes the electric field distribution law and the magnetic field distribution law. According to the temperature field characteristic vector, a temperature field control equation group is established, which describes the temperature field distribution law and the heat transfer law. Calculate temperature gradient distribution data based on the temperature field feature vector, divide the temperature change area according to the temperature gradient distribution data, mark the area with a temperature gradient greater than a first preset threshold as a temperature change drastic area, and mark the area with a temperature gradient less than a second preset threshold as a temperature gentle area; Determine the interface transition area based on the material interface parameters, the interface transition area is the area within a preset range around the material interface, and the range of the interface transition area is determined according to the distance to the material interface; Construct a mesh density distribution function, input the temperature gradient distribution data and the data of the interface transition area into the mesh density distribution function, and calculate the target mesh size of different areas. The mesh density distribution function ensures that the target mesh size of the area with a larger temperature gradient value is smaller, and the target mesh size of the area closer to the material interface is smaller; The temperature-changing area and interface transition area are divided into fine grids, and the size of the fine grid is calculated by the grid density distribution function. The temperature-smooth area is divided into loose grids, and the size of the loose grid is calculated by the grid density distribution function. A transition grid is constructed between the fine grid and the loose grid. The grid size of the transition grid gradually changes between adjacent grid units according to a preset proportional coefficient to ensure a continuous and smooth transition of the grid size. The fine grid, loose grid and transition grid are organized into a multi-level grid structure through node connection relationships.

[0010] Based on the electromagnetic field eigenvector, the electromagnetic field control equations are solved to obtain the electromagnetic field distribution data. The Joule heat loss power density is calculated based on the electric field distribution data. The hysteresis loss power density is calculated based on the magnetic field distribution data. The Joule heat loss power density and the hysteresis loss power density are superimposed on the spatial distribution to obtain the heat source distribution matrix. The heat source distribution matrix is ​​combined with the temperature field eigenvector to establish the heat source term of the temperature field. The temperature field distribution data obtained based on the temperature field control equations containing the heat source term include: Input the electromagnetic field eigenvector into the electromagnetic field control equation group to establish the curl field equation and the divergence field equation, use the vector finite element method to divide the curl field equation and the divergence field equation into space domain grid units, construct the electromagnetic field node basis function in each grid unit, assemble the global stiffness matrix and the mass matrix based on the electromagnetic field node basis function, solve the global stiffness matrix and the mass matrix through the conjugate gradient iterative algorithm to obtain the electromagnetic field distribution data, and the electromagnetic field distribution data includes the electric field distribution data and the magnetic field distribution data; The Joule heat loss power density is calculated based on the electric field distribution data, and the hysteresis loss power density is calculated based on the magnetic field distribution data. The Joule heat loss power density and the hysteresis loss power density are superimposed to generate a heat source distribution matrix. The heat source distribution matrix and the temperature field feature vector are input into a deep neural network for feature extraction and feature fusion to obtain a fused feature vector. A dynamic source term model of the temperature field is constructed based on the fused feature vector. For the time term of the temperature field control equation based on the dynamic source term model of the temperature field, the backward implicit difference scheme is adopted, and for the spatial term, the central difference scheme is adopted to establish the temperature field difference equations. The temperature field difference equations are transformed into linear algebraic equations, and the successive over-relaxation iteration algorithm is used to solve the linear algebraic equations to obtain the temperature field distribution data.

[0011] Using the vector finite element method, the curl field equation and the divergence field equation are divided into spatial domain grid cells. The electromagnetic field nodal basis functions are constructed within each grid cell, and the global stiffness matrix and mass matrix are assembled based on the electromagnetic field nodal basis functions. The global stiffness matrix and mass matrix are solved by the conjugate gradient iteration algorithm to obtain the electromagnetic field distribution data including: The temperature field distribution data is divided into multiple calculation units according to the spatial grid. The temperature value difference equations are constructed for each calculation unit, and the least squares method is used to solve the temperature value difference equations to obtain the temperature gradient values of each grid node. All the temperature gradient values are combined to generate the temperature gradient matrix. The mapping relationship function between the material parameters and the temperature gradient is constructed. The temperature gradient matrix is substituted into the mapping relationship function to calculate the parameter sensitivity values of each grid node. The parameter sensitivity function is constructed based on the parameter sensitivity values and the initial material parameters. The dynamic mapping equation between the temperature field and the material parameters is established. The dynamic mapping equation includes a parameter update term and a stability adjustment term. The parameter update term controls the parameter update rate, and the stability adjustment term suppresses the parameter oscillation. The gradient descent method is used to iteratively solve the dynamic mapping equation. In each iteration, the gradient value of the parameter sensitivity function is calculated, and the parameter update direction and step size are determined according to the gradient value to obtain the material parameter correction amount for the current iteration. The material parameter correction amounts obtained from each iteration are weighted and averaged to obtain the final material parameter correction amount.

[0012] Based on the electric field distribution data, the Joule heat loss power density is calculated. Based on the magnetic field distribution data, the hysteresis loss power density is calculated. The Joule heat loss power density and the hysteresis loss power density are superimposed to generate the heat source distribution matrix including: The electric field distribution data is discretized into an electric field strength vector matrix within the calculation domain, and the magnetic field distribution data is discretized into a magnetic field strength vector matrix within the calculation domain. The current density distribution is calculated based on the electric field strength vector matrix. The current density distribution matrix is established through the product relationship between the current density and the electric field strength. The Joule heat loss power density is calculated according to the current density distribution matrix. The Joule heat loss power density matrix is obtained by dividing the square of the value of each unit in the current density distribution matrix by the conductivity value at the corresponding position. Calculate the magnetic induction intensity distribution based on the magnetic field intensity vector matrix, establish the magnetic induction intensity distribution matrix through the product relationship between the magnetic induction intensity and the magnetic field intensity, determine the peak magnetic induction intensity distribution according to the magnetic induction intensity distribution matrix, and obtain the hysteresis loss coefficient distribution matrix by multiplying the peak magnetic induction intensity distribution by the magnetic field change frequency; Calculate the hysteresis loss power density based on the hysteresis loss coefficient distribution matrix, and obtain the hysteresis loss power density matrix by multiplying the value of each unit in the hysteresis loss coefficient distribution matrix by the sum of the squares of the peak magnetic induction intensity and the magnetic field change frequency at the corresponding position; Overlay the numerical values of the corresponding position units of the Joule heat loss power density matrix and the hysteresis loss power density matrix within the calculation domain to obtain the total loss power density matrix, and the value of each matrix unit represents the total heat source intensity at the corresponding position; Reconstruct the total loss power density matrix according to the spatial grid division, perform volume integration on the total loss power density within each grid unit and divide by the grid unit volume to obtain the average loss power density value of the grid unit; Stack the average loss power density values of the grid units according to the spatial position relationship of the grid units in the calculation domain to form the heat source distribution matrix.

[0013] Calculate the temperature gradient and the temperature change rate based on the temperature field distribution data. When the temperature change rate exceeds the preset stability threshold, it is determined that the thermal stability state has not been reached. Update the conductivity and thermal conductivity parameters of the material based on the temperature field distribution data, and input the updated material physical parameters and temperature field distribution data into the grid optimization algorithm to dynamically optimize the multi-level grid structure, including: Obtain the temperature field distribution data at the initial moment, calculate the temperature gradient values of each grid node based on the temperature field distribution data, and the temperature gradient value is obtained through the spatial derivative of the temperature; Obtain the temperature field distribution data at two adjacent moments, calculate the temperature change amount per unit time to obtain the temperature change rate, and the temperature change rate characterizes the transient change characteristics of the temperature field; Compare the temperature change rate with the preset stability threshold. When the temperature change rate is greater than the preset stability threshold, it is determined that the temperature field is in an unstable state. Update the conductivity parameter according to the temperature-conductivity characteristic curve based on the temperature field distribution data at the current moment, and the conductivity parameter changes according to a preset law as the temperature rises; When it is determined that the temperature field is in a stable state, update the thermal conductivity parameter according to the temperature-thermal conductivity characteristic curve based on the temperature field distribution data at the current moment, and the thermal conductivity parameter changes according to a preset law as the temperature rises; Construct a grid optimization algorithm, taking the temperature field distribution data, updated conductivity parameters, and updated thermal conductivity parameters as input parameters. Using the calculation results of the grid optimization algorithm, construct the first layer of grids in the area where the temperature gradient value is higher than the preset gradient threshold. The first layer of grids has a grid density higher than the preset density threshold to accurately capture the characteristics of drastic temperature changes; using the calculation results of the grid optimization algorithm, construct the second layer of grids in the area where the temperature gradient value is lower than the preset gradient threshold. The second layer of grids has a grid density lower than the preset density threshold to improve the calculation efficiency; Construct a transition layer of grids between the first layer of grids and the second layer of grids. The density of the transition layer of grids changes continuously according to the preset gradient coefficient to ensure a smooth transition of the grid structure; connect the first layer of grids, the transition layer of grids, and the second layer of grids to form a multi-level grid structure, and establish the topological connection relationship of the grid nodes; continue to execute the temperature field analysis, material parameter update, and grid optimization processes until the temperature change rate is less than the preset stability threshold, to achieve the dynamic optimization of the multi-level grid structure In the second aspect of the embodiments of the present invention, Provide an electronic device, including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the foregoing method.

[0014] In the third aspect of the embodiments of the present invention, Provide a computer-readable storage medium, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the foregoing method is implemented.

[0015] The beneficial effects of this application are as follows: 1. The present invention establishes an initial training sample set by collecting the material parameters, structural parameters, and working parameters of gallium nitride devices, and based on this, obtains the electromagnetic field and temperature field feature vectors, establishes the corresponding control equations, and uses the adaptive hierarchical dissection method for grid division, realizing the accurate modeling of the electromagnetic-thermal coupling characteristics of gallium nitride devices and improving the simulation accuracy.

[0016] 2. The present invention calculates the power density of joule heat loss and hysteresis loss from the electromagnetic field distribution data, superimposes them to form a heat source distribution matrix, and combines it with the temperature field feature vector to establish a heat source term, realizing the accurate coupling between the electromagnetic field and the temperature field, and making the simulation results more in line with the actual working conditions.

[0017] 3. The present invention calculates the temperature gradient and temperature change rate based on the temperature field distribution data, and judges the thermal stability state accordingly. At the same time, it dynamically updates the material parameters and grid structure, realizing the adaptive optimization of the simulation process, improving the calculation efficiency while ensuring the accuracy of the results. Description of the Drawings

[0018] Figure 1 It is a schematic flow chart of the three-dimensional electromagnetic-thermal co-simulation method for gallium nitride devices according to an embodiment of the present invention; Figure 2 It is a schematic diagram for comparing the device temperature distributions according to an embodiment of the present invention; Figure 3 It is an interface diagram of the vector finite element electromagnetic field analysis software based on temperature field coupling according to an embodiment of the present invention; Figure 4 It is a schematic diagram for comparing the joule heat loss power density distributions according to an embodiment of the present invention; Figure 5 It is a schematic diagram of the temperature field stability analysis data according to an embodiment of the present invention. Detailed Embodiments

[0019] 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. Apparently, the described embodiments are only a part rather than all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0020] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments may be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0021] Figure 1 It is a schematic flow chart of the three-dimensional electromagnetic-thermal co-simulation method for gallium nitride devices according to an embodiment of the present invention, as Figure 1 shown. The method includes: Collect the material parameters, structural parameters, and working parameters of the gallium nitride device to generate an initial training sample set. Obtain the electromagnetic field feature vector and the temperature field feature vector based on the initial training sample set. Establish the electromagnetic field control equations and the temperature field control equations according to the electromagnetic field feature vector and the temperature field feature vector. Use the adaptive hierarchical dissection method to perform mesh division on the gallium nitride device. Determine the mesh density distribution based on the temperature gradient and the material interface characteristics. Use fine meshes in the regions with drastic temperature changes and the interface transition regions, and use sparse meshes in the regions with gentle temperature changes to generate a multi-level mesh structure with regional adaptive characteristics; Input the multi-level grid structure into the computational solver, solve the electromagnetic field control equations based on the electromagnetic field eigenvector to obtain the electromagnetic field distribution data, calculate the Joule heat loss power density based on the electric field distribution data, calculate the hysteresis loss power density based on the magnetic field distribution data, superimpose the Joule heat loss power density and the hysteresis loss power density in the spatial distribution to obtain the heat source distribution matrix, combine the heat source distribution matrix with the temperature field eigenvector to establish the heat source term of the temperature field, and calculate the temperature field distribution data based on the temperature field control equations containing the heat source term; The temperature gradient and temperature change rate are calculated based on the temperature field distribution data. When the temperature change rate exceeds the preset stability threshold, it is determined that the thermal stability state has not been reached. The electrical conductivity and thermal conductivity parameters of the material are updated based on the temperature field distribution data. The updated material physical parameters and temperature field distribution data are input into the grid optimization algorithm to dynamically optimize the multi-level grid structure.

[0022] In an optional implementation, an electromagnetic field control equation group and a temperature field control equation group are established according to the electromagnetic field eigenvector and the temperature field eigenvector, and a grid is divided for the gallium nitride device using an adaptive hierarchical partitioning method. The grid density distribution is determined based on the temperature gradient and the material interface characteristics. A fine grid is used in the temperature change area and the interface transition area, and a loose grid is used in the temperature gentle area. The generation of a multi-level grid structure with regional adaptive characteristics includes: Obtaining material parameters and structural parameters of the gallium nitride device, the material parameters include electrical conductivity data and thermal conductivity data, and the structural parameters include material interface parameters, and calculating the electromagnetic field characteristic vector and the temperature field characteristic vector based on the material parameters and the structural parameters; According to the electromagnetic field characteristic vector, an electromagnetic field control equation group is established, which describes the electric field distribution law and the magnetic field distribution law. According to the temperature field characteristic vector, a temperature field control equation group is established, which describes the temperature field distribution law and the heat transfer law. Calculate temperature gradient distribution data based on the temperature field feature vector, divide the temperature change area according to the temperature gradient distribution data, mark the area with a temperature gradient greater than a first preset threshold as a temperature change drastic area, and mark the area with a temperature gradient less than a second preset threshold as a temperature gentle area; Determine the interface transition area based on the material interface parameters, the interface transition area is the area within a preset range around the material interface, and the range of the interface transition area is determined according to the distance to the material interface; Construct a mesh density distribution function, input the temperature gradient distribution data and the data of the interface transition area into the mesh density distribution function, and calculate the target mesh size of different areas. The mesh density distribution function ensures that the target mesh size of the area with a larger temperature gradient value is smaller, and the target mesh size of the area closer to the material interface is smaller; The areas with drastic temperature changes and the interface transition areas are meshed to obtain fine grids, and the sizes of the fine grids are calculated by the grid density distribution function. The areas with gentle temperature changes are meshed to obtain sparse grids, and the sizes of the sparse grids are calculated by the grid density distribution function; Transition grids are constructed between the fine grids and the sparse grids, and the grid sizes of the transition grids gradually change according to a preset proportionality coefficient between adjacent grid cells to ensure a continuous and smooth transition of the grid sizes; the fine grids, the sparse grids, and the transition grids are organized into a multi-level grid structure through node connection relationships.

[0023] The original conductivity data is recorded at 5°C intervals, from 0.8 S / m at 25°C to 0.3 S / m at 250°C, with a total of 46 data points. The change rate is calculated for adjacent points: for example, in the range from 25°C to 30°C, the conductivity drops from 0.8 S / m to 0.79 S / m, and the change rate is -0.002 S / (m·°C). The thermal conductivity data is processed in the same way: 160 W / (m·K) at 25°C and 159.1 W / (m·K) at 30°C, resulting in a change rate of -0.18 W / (m·K·°C). This calculation process is repeated for all temperature ranges to obtain a complete dataset of change rates.

[0024] The conductivity values at 46 temperature points are paired with the corresponding temperatures, and feature extraction is performed every 10 data points. For example, in the range from 25°C to 75°C, 10 points with conductivity values of 0.8, 0.76, 0.72 S / m, etc. are taken, and the average change rate of -0.016 S / (m·°C) is calculated. This process is repeated to obtain 5 sets of interval feature values. The thermal conductivity features are obtained using the same method to get the change characteristics in the range from 160 to 150 W / (m·K).

[0025] Take the initial temperature field data. For example, if the temperature at a certain point is 200°C, and the temperature at a point 0.1 mm away in the X direction from its adjacent point is 170°C, the temperature gradient of 300 K / mm is calculated. The temperature at a point 0.1 mm away in the Y direction is 160°C, resulting in a gradient of 400 K / mm. The temperature at a point 0.1 mm away in the Z direction is 180°C, resulting in a gradient of 200 K / mm. The gradients in the three directions are synthesized to obtain the total gradient value at this point. This calculation is repeated for all grid points in the entire computational domain.

[0026] Scan the temperature gradient values of all grid points. When the gradient is greater than 300 K / mm, for example, if the gradient at a point in the active layer is 350 K / mm, mark this point and the surrounding area within a range of 0.05 mm as the area with drastic changes. When the gradient is less than 50 K / mm, for example, if the gradient at a point in the substrate layer is 30 K / mm, mark this point and the surrounding area within a range of 0.15 mm as the gentle area. This process is repeated for all grid points to complete the area division.

[0027] Taking the active layer - substrate layer interface (located at 1 mm) as an example, expand 0.1 mm upward and downward. In the range of 0.9 - 1.1 mm, a calculation point is taken every 0.01 mm. The influence coefficient is 0.9 at 0.01 mm from the interface, 0.8 at 0.02 mm, and so on decreasingly. Pair the influence coefficients of each point with their coordinates to generate a complete dataset of the interface influence function.

[0028] Input the temperature gradient value and the interface distance, and calculate the target grid size. For example, if the temperature gradient at a certain point is 350 K / mm and the interface distance is 0.02 mm, calculate the influence of the two factors respectively: the gradient influence term gives a basic grid size of 0.01 mm, and the interface influence term gives a correction coefficient of 0.8. Finally, determine the grid size at this point to be 0.008 mm. Repeat this calculation process for all regions.

[0029] At the junction of the fine grid (0.01 mm) and the coarse grid (0.1 mm), set the grid size ratio to 1.2. The size of the first transition grid is 0.012 mm, the second is 0.0144 mm, the third is 0.0173 mm, and so on until the coarse grid size is reached. The specific size of each transition layer is obtained by multiplying the previous layer by 1.2.

[0030] The solution of this application can: Adaptive determine the grid density distribution according to the temperature gradient and material interface characteristics, adopt fine grids in areas with drastic temperature changes and interface transition areas, and adopt coarse grids in areas with gentle temperature, improving the pertinence and calculation efficiency of grid division. Adopt a multi - level grid structure and a smoothly - transitioning grid size distribution, avoiding numerical calculation instability caused by sudden changes in grid size, and improving the calculation accuracy and convergence of finite - element analysis. Establish a control equation set based on the eigenvectors of the electromagnetic field and temperature field, accurately describing the distribution law of field quantities inside the device, and providing a reliable theoretical basis for subsequent thermal - electrical coupling analysis.

[0031] In an optional implementation manner, solve the electromagnetic - field control equation set based on the eigenvectors of the electromagnetic field to obtain electromagnetic - field distribution data, calculate the joule heat - loss power density according to the electric - field distribution data, calculate the hysteresis - loss power density according to the magnetic - field distribution data, superimpose the joule heat - loss power density and the hysteresis - loss power density in the spatial distribution to obtain a heat - source distribution matrix, combine the heat - source distribution matrix with the eigenvectors of the temperature field to establish the heat - source term of the temperature field, and calculate the temperature - field distribution data based on the temperature - field control equation set including the heat - source term, including: Input the electromagnetic field eigenvector into the electromagnetic field control equations to establish the curl field equation and the divergence field equation. Use the vector finite element method to divide the curl field equation and the divergence field equation into spatial domain grid cells. Construct the electromagnetic field nodal basis functions within each grid cell. Assemble the global stiffness matrix and the mass matrix based on the electromagnetic field nodal basis functions. Solve the global stiffness matrix and the mass matrix through the conjugate gradient iteration algorithm to obtain the electromagnetic field distribution data, which includes the electric field distribution data and the magnetic field distribution data. Calculate the Joule heat loss power density based on the electric field distribution data, calculate the hysteresis loss power density based on the magnetic field distribution data. Superimpose the Joule heat loss power density and the hysteresis loss power density to generate the heat source distribution matrix. Input the heat source distribution matrix and the temperature field eigenvector into the deep neural network for feature extraction and feature fusion to obtain the fused eigenvector. Construct the temperature field dynamic source term model based on the fused eigenvector. Adopt the backward implicit difference scheme for the time term of the temperature field control equation and the central difference scheme for the spatial term based on the temperature field dynamic source term model to establish the temperature field difference equations. Convert the temperature field difference equations into linear algebraic equations. Solve the linear algebraic equations using the successive over-relaxation iteration algorithm to obtain the temperature field distribution data.

[0032] Initialize the electromagnetic field eigenvector, including the temperature coefficient of conductivity -0.002 S / (m·°C) and the temperature coefficient of permeability -0.15 H / m·°C. Divide the computational domain into 10,000 hexahedral grid cells, with each cell side length of 0.1 mm. Set 8 nodes within each cell and construct the nodal basis functions. Each nodal basis function takes a value of 1 at that node, 0 at other nodes, and its value within the cell is determined by linear interpolation.

[0033] When assembling the global matrix, scan all grid cells. Taking a certain cell as an example, the nodal coordinates are 8 points such as (0,0,0), (0.1,0,0), (0,0.1,0), etc. Calculate the contribution of this cell to the global stiffness matrix: Substitute the 8 nodal basis functions within the cell into the curl field equation and the divergence field equation to obtain an 8×8 cell stiffness matrix. Similarly, calculate the cell mass matrix. Repeat this process to complete the global matrix assembly.

[0034] Solve using the conjugate gradient method, setting the convergence threshold to 1e-6. Update the solution vector in each iteration and stop the iteration when the residual is less than the threshold. Obtain the electric field distribution data: For example, the electric field strength at a certain point is 1000 V / m, and the direction is the positive x-axis. Magnetic field distribution data: The magnetic induction intensity at a certain point is 0.1 T, and the direction is the positive z-axis.

[0035] Calculating Joule heat based on the electric field distribution: The electric field strength at a certain point is 1000 V / m, and the conductivity is 0.8 S / m. The calculated Joule heat power density is 800 W / m³. The magnetic field distribution is used to calculate the hysteresis loss: The magnetic induction intensity is 0.1 T, and the frequency is 1 MHz. The obtained hysteresis loss power density is 200 W / m³. Adding the two losses at the corresponding positions in space, the total heat source density at this point is 1000 W / m³. Repeat this calculation for all grid points.

[0036] Construct a deep neural network. The number of nodes in the input layer is the dimension of the heat source distribution matrix, which is 10000, and the dimension of the temperature field feature vector is 100. The first hidden layer has 1000 nodes and uses the ReLU activation function. The second hidden layer has 500 nodes and uses the tanh activation function. The output layer has 100 nodes to obtain the fused feature vector.

[0037] Taking a certain grid point as an example: Input the heat source density of this point, which is 1000 W / m³, and the temperature characteristic value (temperature 300 K, thermal conductivity 150 W / m·K). After network processing, the fused feature value of this point is 0.85. Repeat this process for all grid points to generate a complete fused feature field.

[0038] The time discretization uses a step size of 1 ms, and the space discretization uses a grid spacing of 0.1 mm. For each time step, construct a difference equation: Substitute the temperature at the current moment, the temperature at the previous moment, and the temperatures of adjacent nodes into the equation, and combine with the heat source term to obtain a linear equation system.

[0039] Set the over-relaxation factor to 1.5 and the iteration convergence threshold to 0.01 K. Update the temperature field for each iteration. Stop when the temperature difference between two adjacent iterations is less than the threshold. For example, the initial temperature of a certain point is 300 K, and it converges to the steady-state temperature of 320 K after 100 iterations. Repeat the solution process for all time steps.

[0040] Figure 2 Schematic diagram for comparing the device temperature distributions in the embodiments of the present invention: This figure shows the predicted results of the device temperature distributions by different methods. The circular markers and solid lines represent the present technical solution (the electromagnetic-thermal coupling calculation model based on a deep neural network), the square markers and dashed lines represent the traditional finite element method, and the triangular markers and dotted lines represent the conventional difference method. At the heat source position (0 mm), the predicted temperature by the coupling calculation model based on a deep neural network is 375 K, which is closer to the measured value than 365 K of the traditional finite element method and 360 K of the conventional difference method. As the distance increases, at 5 mm, the three methods predict 305 K, 303 K, and 300 K respectively. The present technical solution shows better prediction accuracy throughout the measurement range, especially in the near-source region (0 - 2 mm) with a large temperature gradient, and the predicted temperature curve is more in line with the actual temperature distribution characteristics.

[0041] The solution of the present application can: This technical solution aims to solve the problems existing in the calculation of the existing temperature field distribution. The existing technologies mainly use the traditional finite element method and the conventional difference method to calculate the temperature field. These methods usually solve the electromagnetic field and the temperature field as independent systems respectively, and then establish the correlation through simple data transfer. This processing method ignores the coupling effect between the electromagnetic field and the temperature field, resulting in a deviation between the calculation result and the actual physical phenomenon. Especially in the area with a large temperature gradient, it is difficult to ensure the calculation accuracy. To solve the above problems, this technical solution proposes an electromagnetic-thermal coupling calculation model based on a deep neural network. This model first extracts the key information of the electromagnetic field distribution through the electromagnetic field eigenvector, and then uses the deep neural network to establish a non-linear mapping relationship between the electromagnetic field and the temperature field, realizing the deep coupling of the two physical fields. This method not only considers the influence of the electromagnetic field on the temperature field, but also can capture the feedback effect of the temperature field change on the electromagnetic field. By adopting the above technical means, this technical solution has achieved remarkable technical effects in the calculation of the temperature field distribution: First, in the high temperature gradient area near the heat source, the prediction result of this technical solution is closer to the measured value, reflecting a higher calculation accuracy; Second, within the entire calculation domain, the temperature distribution curve is smooth and continuous, which is more in line with the characteristics of the actual physical process; Finally, compared with the existing technologies, this technical solution can more accurately describe the interaction between the electromagnetic field and the temperature field, so as to provide a more reliable temperature field distribution prediction result. The realization of these improvement effects benefits from the innovation at the algorithm level of this technical solution. By introducing a deep neural network to process the complex physical field coupling relationship, the limitations of traditional numerical methods in dealing with multi-physical field coupling problems are overcome. At the same time, the information extraction method based on the eigenvector effectively reduces the calculation complexity and improves the solution efficiency. These technical innovations have significantly improved the accuracy, reliability and practicality of this technical solution in the calculation of the temperature field distribution.

[0042] In an optional implementation manner, the curl field equation and the divergence field equation are divided into spatial domain grid cells by using the vector finite element method. The electromagnetic field node basis functions are constructed within each grid cell, and the global stiffness matrix and the mass matrix are assembled based on the electromagnetic field node basis functions. The electromagnetic field distribution data is obtained by solving the global stiffness matrix and the mass matrix through the conjugate gradient iteration algorithm, including: The temperature field distribution data is divided into multiple calculation cells according to the spatial grid. The temperature value difference equations are constructed for each calculation cell, and the temperature gradient values of each grid node are obtained by solving the temperature value difference equations using the least squares method. All the temperature gradient values are combined to generate a temperature gradient matrix; Construct a mapping relationship function between the construction material parameters and the temperature gradient. Substitute the temperature gradient matrix into the mapping relationship function to calculate the parameter sensitivity values of each grid node, and construct a parameter sensitivity function based on the parameter sensitivity values and the initial material parameters. Establish a dynamic mapping equation between the temperature field and the material parameters. The dynamic mapping equation includes a parameter update term and a stability adjustment term. The parameter update term controls the parameter update rate, and the stability adjustment term suppresses parameter oscillation. Use the gradient descent method to iteratively solve the dynamic mapping equation. Calculate the gradient value of the parameter sensitivity function in each iteration, determine the parameter update direction and step size according to the gradient value, and obtain the material parameter correction amount for the current iteration. Perform weighted averaging on the material parameter correction amounts obtained in each iteration to obtain the final material parameter correction amount.

[0043] First, perform a vector finite element analysis on the electromagnetic field, and divide the calculation domain into tetrahedral mesh elements. Construct zero-order vector basis functions within each mesh element. The basis functions are composed of the normal component and the tangential component of the unit boundary surface. Based on the constructed basis functions, assemble the global stiffness matrix and the mass matrix. The stiffness matrix reflects the coupling relationship between the mesh elements, and the mass matrix characterizes the distribution characteristics of the field quantities within the elements. Use the conjugate gradient iteration algorithm to solve the matrix equation and obtain the electromagnetic field distribution data.

[0044] Next, discretize the temperature field according to the same mesh division scheme. For each tetrahedral element, construct a temperature value difference equation set based on the central difference format. The equation set includes the temperature values of the internal nodes and the adjacent node temperatures of the element. Use the least squares method to solve the difference equation set to obtain the temperature gradient values at each node. Combine the temperature gradient values of all nodes into a temperature gradient matrix.

[0045] Then establish the mapping relationship between the material parameters and the temperature gradient. For isotropic materials, the mapping relationship can be expressed as a linear change of the material parameters with the temperature gradient. Substitute the temperature gradient matrix into the mapping relationship to calculate the parameter sensitivity values of each node. Construct a parameter sensitivity function based on the sensitivity values and the initial parameter values. This function reflects the response characteristics of the material parameters to temperature changes.

[0046] Establish a dynamic mapping equation between the temperature field and the material parameters. The parameter update term adopts an adaptive step size strategy, and the step size is proportional to the gradient value of the sensitivity function. The stability adjustment term introduces a damping factor to increase the damping when the parameters change violently to suppress oscillation.

[0047] Finally, the gradient descent method is used to solve the dynamic mapping equation. In each iteration, the gradient value of the parameter sensitivity function with respect to the current parameter is calculated to determine the search direction. The Armijo criterion is adopted for the step size selection to ensure sufficient descent of the objective function. The exponentially weighted average is performed on the parameter correction amount obtained by iteration, and the weight factor decays with the number of iterations. The iteration stops when the change in the correction amount between two consecutive iterations is less than the set threshold.

[0048] For a cubic computational domain with a side length of 0.1 meters, it is divided into 10,000 tetrahedral elements. The initial value of the temperature field is linearly distributed between 25 and 100 degrees Celsius. The initial relative permittivity of the material is 4.0, and the relative permeability is 1.0. The initial value of the parameter update step size is taken as 0.1, and the damping factor is taken as 0.5. The iteration convergence threshold is set to 0.001. The final material parameter distribution is obtained after 50 iterations.

[0049] Figure 3 The following is the software interface diagram of the vector finite element electromagnetic field analysis based on temperature field coupling according to the embodiment of the present invention: The calculation system interface shows the whole process numerical simulation of a cube calculation domain with a side length of 0.1 meters. The calculation domain is divided into 10,000 tetrahedral units. From the grid division diagram, it can be seen that the grid structure is uniform and regular, the minimum dihedral angle is greater than 30 degrees, and the maximum aspect ratio is controlled within 3:1, which meets the grid quality requirements of finite element calculation. Based on this grid structure, the temperature field shows a linear distribution feature from 25℃ to 100℃, and the color gradient of the thermal map is uniform, reflecting the continuity of the temperature field. The current calculation has completed 35 iterations (the total number of iterations is 50 times), and the convergence error has reached 0.00085, which is better than the set convergence threshold of 0.001, indicating that the temperature field calculation has good convergence. In terms of material parameter settings, the initial relative dielectric constant is 4.0 and the relative magnetic permeability is 1.0. Through the regulation of parameter update step size 0.1 and damping factor 0.5, the material parameter correction amount reaches 0.0324. The density change of the stripe pattern in the material parameter distribution diagram intuitively shows the spatial distribution characteristics of the parameters. Its uniformity and continuity confirm the stability of the solution of the dynamic mapping equation, and there is no obvious numerical oscillation. The electromagnetic field distribution diagram uses a radial gradient effect to show the field quantity distribution calculated based on the vector finite element method, which meets the requirements of the curl field equation and the divergence field equation. The radial gradient from the center to the outside accurately reflects the spatial attenuation characteristics of the electromagnetic field. The continuity and smoothness of the field quantity indicate that the zero-order vector basis function can effectively describe the field quantity distribution. The entire calculation process is progressing smoothly, and the current progress has reached 75%, and the remaining time is expected to be 2 minutes. The real-time calculation monitoring area on the right side of the interface clearly shows the complete process from meshing, temperature field solution, parameter correction to electromagnetic field analysis through four sub-areas. The calculation results show that within the given calculation domain size, the coupled calculation of the temperature field and the electromagnetic field meets the expected accuracy requirements, the dynamic correction process of the material parameters is stable and controllable, and the overall calculation effect is good. The control panel on the left provides a complete parameter adjustment function, including the number of grid cells, material parameters, algorithm control parameters, etc., so that the calculation process can be flexibly adjusted as needed. This interface not only displays the calculation results, but also realizes real-time monitoring of the calculation process through various parameters and progress indicators, reflecting the professionalism and practicality of the system in electromagnetic field analysis. This interface design embodies the organic combination of complex numerical calculations and intuitive visualization, which not only meets professional calculation needs, but also ensures ease of use. Through multi-dimensional data display and real-time monitoring, users can intuitively grasp every link of the calculation process to ensure the reliability and accuracy of the calculation results. Overall, this interface successfully realizes the complete display and effective control of the vector finite element electromagnetic field analysis based on temperature field coupling and the dynamic correction calculation process of material parameters.

[0050] The solution of this application can: By constructing appropriate electromagnetic field nodal basis functions and using the conjugate gradient iteration algorithm, the accuracy and efficiency of electromagnetic field calculation are improved, ensuring the accuracy of electromagnetic field distribution data. By adopting the dynamic mapping relationship between temperature gradient and material parameters, combining parameter sensitivity analysis and adaptive step size strategy, the accurate correction of material parameters is realized, and the reliability of parameter inversion is improved. By introducing a stability adjustment mechanism and a weighted average strategy for parameter correction amounts, the numerical oscillation in the iteration process is effectively suppressed, the convergence stability of the algorithm is enhanced, and the accuracy of the final result is ensured.

[0051] In an alternative embodiment, the joule heat loss power density is calculated based on the electric field distribution data, the hysteresis loss power density is calculated based on the magnetic field distribution data, and the superposition of the joule heat loss power density and the hysteresis loss power density to generate a heat source distribution matrix includes: The electric field distribution data is discretized into an electric field strength vector matrix within the calculation domain, and the magnetic field distribution data is discretized into a magnetic field strength vector matrix within the calculation domain; the current density distribution is calculated based on the electric field strength vector matrix, a current density distribution matrix is established through the product relationship between the current density and the electric field strength, the joule heat loss power density is calculated according to the current density distribution matrix, and the joule heat loss power density matrix is obtained by dividing the square of the value of each unit in the current density distribution matrix by the conductivity value at the corresponding position; The magnetic induction intensity distribution is calculated based on the magnetic field strength vector matrix, a magnetic induction intensity distribution matrix is established through the product relationship between the magnetic induction intensity and the magnetic field strength, the magnetic induction intensity peak distribution is determined according to the magnetic induction intensity distribution matrix, and the hysteresis loss coefficient distribution matrix is obtained by multiplying the magnetic induction intensity peak distribution by the magnetic field change frequency; The hysteresis loss power density is calculated based on the hysteresis loss coefficient distribution matrix, and the hysteresis loss power density matrix is obtained by multiplying the value of each unit in the hysteresis loss coefficient distribution matrix by the product of the square of the magnetic induction intensity peak and the magnetic field change frequency at the corresponding position; The joule heat loss power density matrix and the hysteresis loss power density matrix are numerically superimposed on the corresponding position units within the calculation domain to obtain a total loss power density matrix, and the value of each matrix unit represents the total heat source intensity at the corresponding position; The total loss power density matrix is reconstructed according to the spatial grid division, the volume integral of the total loss power density within each grid unit is performed and divided by the grid unit volume to obtain the average loss power density value of the grid unit; the average loss power density values of the grid units are superimposed according to the spatial position relationship of the grid units in the calculation domain to form a heat source distribution matrix.

[0052] A three-dimensional space coordinate system is established within the computational domain, and the electric field distribution data is discretized into an electric field intensity vector matrix within the computational domain. Taking a cubic computational domain of 1 m × 1 m × 1 m as an example, it is uniformly divided into 100 × 100 × 100 grid cells, and the size of each grid cell is 0.01 m × 0.01 m × 0.01 m. For each grid cell, the electric field intensity vector value at that position is obtained according to the electric field distribution data, including the x, y, and z direction components. For example, for the grid cell at the coordinate (0.5, 0.5, 0.5), the electric field intensity vector is (100, 200, 300) volts / meter. Similarly, the magnetic field distribution data is discretized into a magnetic field intensity vector matrix.

[0053] Based on the electric field intensity vector matrix, the current density distribution is calculated. For each grid cell, the current density vector is calculated according to the material conductivity and the electric field intensity vector at that position. For example, for a grid cell with a conductivity of 5.8 × 10^7 Siemens / meter and an electric field intensity of 100 volts / meter, the current density is 5.8 × 10^9 amperes per square meter. The current density values of all grid cells are combined to form a current density distribution matrix.

[0054] The Joule heat loss power density is calculated. For each cell in the current density distribution matrix, the square of the current density value is divided by the conductivity value at the corresponding position. For example, if the current density is 5.8 × 10^9 amperes per square meter and the conductivity is 5.8 × 10^7 Siemens / meter, then the Joule heat loss power density at that position is 5.8 × 10^11 watts per cubic meter.

[0055] Based on the magnetic field intensity vector matrix, the magnetic induction intensity distribution is calculated. For each grid cell, the magnetic induction intensity vector is calculated according to the material permeability and the magnetic field intensity vector at that position. For example, for a grid cell with a permeability of 4π × 10^-7 henries / meter and a magnetic field intensity of 1000 amperes / meter, the magnetic induction intensity is 1.256 × 10^-3 tesla. The peak magnetic induction intensity of each grid cell is determined.

[0056] The magnetic hysteresis loss power density is calculated. For each grid cell, the peak magnetic induction intensity, the magnetic field change frequency, and the material magnetic hysteresis coefficient are multiplied. For example, if the peak magnetic induction intensity is 1 tesla, the frequency is 50 hertz, and the magnetic hysteresis coefficient is 100, then the magnetic hysteresis loss power density at that position is 5000 watts per cubic meter.

[0057] The numerical values of the corresponding positions of the Joule heat and magnetic hysteresis loss power density matrices are added together to obtain the total loss power density matrix. The total loss power density within each grid cell is volume-integrated and averaged to generate the final heat source distribution matrix.

[0058] Figure 4 This is a schematic diagram for comparing the Joule heat loss power density distribution of the embodiments of the present invention: The figure shows the distribution results of the joule heat loss power density calculated by different methods, where the circular markers and solid lines represent the technical solution of the present invention (joule heat loss calculation model based on conductivity temperature correction and vector finite element), which considers the influence of temperature on material conductivity by updating the conductivity parameter matrix in real time. The square markers and dashed lines represent the traditional constant conductivity model, which uses a fixed conductivity value for calculation. At the center of the calculation domain (0 mm), the loss power density predicted by the technical solution of the present invention is 850 W / m³, which is closer to the measured value than 800 W / m³ of the traditional model. As the distance increases to 6 mm, the predicted values of the two methods are 660 W / m³ and 650 W / m³ respectively, and at 10 mm they are 580 W / m³ and 570 W / m³ respectively. The technical solution of the present invention shows more accurate prediction ability throughout the calculation domain, especially in the central region with a large temperature gradient, where the prediction accuracy is improved more significantly.

[0059] The solution of this application can: Regarding the calculation problem of the Joule heat loss power density, the existing technologies usually adopt a constant conductivity model for calculation. This method treats the conductivity of the material as a fixed value, ignoring the influence of temperature on conductivity, resulting in a large deviation between the calculation result and the actual physical phenomenon in the region with non-uniform temperature distribution, especially in the region near the heat source with a large temperature gradient. At the same time, the traditional method often uses a scalar calculation method when performing electromagnetic field calculations, which cannot accurately reflect the vector characteristics of the electromagnetic field, further affecting the calculation accuracy of the Joule heat loss. To address the above problems, the present technical solution proposes a Joule heat loss calculation model based on conductivity temperature correction and vector finite element. This model first establishes a dynamic correlation between the temperature field and conductivity by updating the conductivity parameter matrix in real time, accurately reflecting the influence of temperature changes on the conductivity of the material. Secondly, the vector finite element method is used for electromagnetic field calculation, fully considering the directional characteristics of the electromagnetic field and improving the accuracy of field quantity calculation. Finally, through the product relationship between the current density and the electric field strength, combined with the corrected conductivity value, the accurate calculation of the Joule heat loss is realized. Through the above improvements, the present technical solution has achieved remarkable technical effects: First, by establishing a dynamic correlation between temperature and conductivity, the calculation accuracy of material parameters is improved; second, the accuracy of electromagnetic field calculation is improved by using the vector finite element method; third, in the region with a large temperature gradient, the calculation result of the present technical solution is closer to the actual physical process; finally, a significant improvement in the loss prediction accuracy is achieved throughout the calculation domain. These improvements enable the present technical solution to more accurately describe the spatial distribution characteristics of the Joule heat loss and provide more reliable heat source data for subsequent temperature field analysis. The innovation of the present technical solution lies in introducing the dynamic correlation mechanism between the temperature field and conductivity into the calculation of the Joule heat loss and realizing high-precision field quantity calculation by combining the vector finite element method. This method breaks through the limitations of the traditional constant parameter model and establishes a calculation framework that is more in line with physical reality, providing an effective technical means to improve the accuracy of electromagnetic-thermal coupling analysis.

[0060] In an alternative embodiment, the temperature gradient and the temperature change rate are calculated according to the temperature field distribution data. When the temperature change rate exceeds a preset stability threshold, it is determined that the thermal stability state has not been reached. The conductivity and thermal conductivity parameters of the material are updated based on the temperature field distribution data, and the updated material physical parameters and temperature field distribution data are input into the grid optimization algorithm to dynamically optimize the multi-level grid structure, including: Obtain the temperature field distribution data at the initial moment, calculate the temperature gradient values of each grid node based on the temperature field distribution data, and the temperature gradient values are obtained by calculating the spatial derivative of temperature; obtain the temperature field distribution data at two adjacent moments, calculate the temperature change amount per unit time to obtain the temperature change rate, and the temperature change rate characterizes the transient change characteristics of the temperature field; Compare the temperature change rate with a preset stability threshold. When the temperature change rate is greater than the preset stability threshold, it is determined that the temperature field is in an unstable state. According to the temperature field distribution data at the current moment, update the conductivity parameter using the temperature-conductivity characteristic curve. The conductivity parameter changes according to a preset rule as the temperature increases; when it is determined that the temperature field is in a stable state, update the thermal conductivity parameter using the temperature-thermal conductivity characteristic curve according to the temperature field distribution data at the current moment. The thermal conductivity parameter changes according to a preset rule as the temperature increases; Construct a grid optimization algorithm. Use the temperature field distribution data, the updated conductivity parameter, and the updated thermal conductivity parameter as input parameters. Utilize the calculation result of the grid optimization algorithm to construct the first layer of grids in the area where the temperature gradient value is higher than the preset gradient threshold. The first layer of grids has a grid density higher than the preset density threshold to accurately capture the characteristics of drastic temperature changes; utilize the calculation result of the grid optimization algorithm to construct the second layer of grids in the area where the temperature gradient value is lower than the preset gradient threshold. The second layer of grids has a grid density lower than the preset density threshold to improve the calculation efficiency; Construct a transition layer of grids between the first layer of grids and the second layer of grids. The density of the transition layer of grids changes continuously according to a preset gradient coefficient to ensure a smooth transition of the grid structure; connect the first layer of grids, the transition layer of grids, and the second layer of grids to form a multi-level grid structure, and establish the topological connection relationship of the grid nodes; continue to execute the temperature field analysis, material parameter update, and grid optimization processes until the temperature change rate is less than the preset stability threshold to achieve the dynamic optimization of the multi-level grid structure.

[0061] Select an arbitrary point P(x, y, z) in the computational domain and collect the temperatures of its six adjacent points: the temperature of the point to the right of P (x + 0.1mm, y, z) is 325K, and the temperature of the point to the left of P (x - 0.1mm, y, z) is 305K. Divide the temperature difference of 200mK between two adjacent points by the distance of 0.2mm to obtain the temperature gradient in the x direction of 1000K / mm. Similarly, the temperature of the point in front of P (x, y + 0.1mm, z) is 315K, and the temperature of the point behind P (x, y - 0.1mm, z) is 295K, calculate the gradient in the y direction of 1000K / mm. The temperature of the point above P (x, y, z + 0.1mm) is 320K, and the temperature of the point below P (x, y, z - 0.1mm) is 290K, calculate the gradient in the z direction of 1500K / mm.

[0062] Synthesize the total temperature gradient: Substitute the gradient values in the three directions into the three-dimensional space vector synthesis formula to obtain the total temperature gradient value of 2062K / mm at this point. Repeat this calculation process for all nodes in the computational domain to generate the complete temperature gradient distribution data.

[0063] Record the temperature data at two adjacent moments: At time t1, the temperature at point P is 315K, and at time t2 (with an interval of 1ms), the temperature rises to 320K. Calculate the temperature change per unit time: Divide the temperature difference of 5K by the time interval of 1ms to obtain a temperature change rate of 5K / ms. Repeat the calculation of the temperature change rate for all nodes.

[0064] Taking point P as an example, when the current temperature is 320K, query the temperature-conductivity curve: Calculate the conductivity value through linear interpolation. 320K is between 315K (corresponding to 0.75S / m) and 325K (corresponding to 0.73S / m). Interpolate according to the temperature ratio to obtain the conductivity at 320K as 0.74S / m. Similarly, calculate the thermal conductivity: 320K is between 315K (corresponding to 155W / m·K) and 325K (corresponding to 152W / m·K), and interpolate to obtain the thermal conductivity at 320K as 153.5W / m·K.

[0065] Take the temperature gradient at point P as 2062K / mm, which exceeds the preset threshold of 300K / mm. Calculate the required grid density: Multiply the reference density of 100 per mm³ by the temperature gradient ratio (2062 / 300) to obtain the optimized grid density of 687 per mm³. This means that hexahedral grid cells with a side length of approximately 0.05mm need to be constructed around point P.

[0066] Transition from a fine grid (density of 687 per mm³) to a sparse grid (density of 100 per mm³), and set the gradient coefficient to 1.2. The density of the first transition layer is 687 / 1.2 = 572.5 per mm³, corresponding to a grid side length of 0.061mm. The density of the second transition layer is 477 per mm³, and the side length is 0.073mm. The density of the third transition layer is 397.5 per mm³, and the side length is 0.088mm. And so on until the sparse grid density is reached.

[0067] Fine grid area: The volume of a single grid cell is 1 / 687 = 0.001456mm³. Take the cube root to obtain a side length of 0.0496mm. The first layer of the transition layer: The volume is 1 / 572.5 = 0.001747mm³, and the side length is 0.0561mm. Calculate layer by layer until the side length of the sparse grid reaches 0.1mm.

[0068] Taking two adjacent grid cells as an example, the side lengths are 0.0496mm and 0.0561mm respectively. The vertex coordinates of the smaller cell are (0,0,0), (0.0496,0,0), (0,0.0496,0), (0.0496,0.0496,0). The corresponding vertices of the larger cell need to be aligned, and interpolation is used to determine the exact coordinates of the shared nodes.

[0069] After completing one round of optimization, calculate the new temperature field. The temperature change rate at point P decreases from 5 K / ms to 4.5 K / ms, and the total temperature gradient decreases from 2062 K / mm to 1850 K / mm. Recalculate the grid density according to the new gradient value: 1850 / 300×100 = 617 per mm³. Adjust the grid size and transition layer parameters accordingly until the temperature field reaches a stable state.

[0070] Figure 5 Schematic diagram of the temperature field stability analysis data for the embodiments of the present invention: Analysis of the temperature field data at three typical test points shows that the present technology has achieved remarkable results in temperature field stability control. The temperature gradient at point P1 (0, 0, 0) is optimized from 2062 K / mm to 1850 K / mm, with an improvement rate of 10.28%; the temperature at point P2 (5, 5, 5) decreases from 1895 K / mm to 1677 K / mm, with an improvement rate of 11.50%; the temperature at point P3 (-5, -5, -5) decreases from 1978 K / mm to 1723 K / mm, with an improvement rate as high as 12.89%. In response to the relatively low improvement rate at the central point P1, improvement measures such as increasing the grid density stratification and introducing an adaptive time step algorithm are proposed. To address the problem of asymmetric optimization effects at points P2 and P3, plans are made to implement progressive boundary constraints, introduce boundary layer grid optimization technology, and establish a dynamic boundary compensation mechanism. The implementation of these improvement measures is expected to increase the overall stability of the temperature field by at least 20%, providing better basic conditions for subsequent optimization.

[0071] The solution of the present application can: Determine the thermal stability through the temperature gradient and change rate, realize the real-time dynamic update of material parameters, and improve the accuracy of temperature field calculation. Based on the temperature gradient threshold, a multi-level grid structure is constructed, with high-density grids in areas of rapid temperature change and low-density grids in areas of gentle temperature change, realizing the reasonable allocation of computing resources. A gradual transition layer is used to achieve a smooth transition of grid density, avoiding numerical instability caused by grid mutations, and ensuring the reliability and convergence of calculation results.

[0072] In the second aspect of the embodiments of the present invention, Provide an electronic device, including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the foregoing method.

[0073] In the third aspect of the embodiments of the present invention, Provide a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the foregoing method is implemented.

[0074] The present invention may be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having thereon computer-readable program instructions for performing various aspects of the present invention.

[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A three-dimensional electromagnetic-thermal collaborative simulation method for GaN devices, characterized in that: include: The material parameters, structural parameters and working parameters of the gallium nitride device are collected to generate an initial training sample set, and the electromagnetic field characteristic vector and the temperature field characteristic vector are obtained based on the initial training sample set. The electromagnetic field control equation group and the temperature field control equation group are established according to the electromagnetic field characteristic vector and the temperature field characteristic vector. The gallium nitride device is meshed by an adaptive hierarchical partitioning method, and the mesh density distribution is determined based on the temperature gradient and the material interface characteristics. A fine mesh is used in the area of ​​drastic temperature change and the interface transition area, and a loose mesh is used in the area of ​​gentle temperature, so as to generate a multi-level mesh structure with regional adaptive characteristics; Input the multi-level grid structure into the computational solver, solve the electromagnetic field control equations based on the electromagnetic field eigenvector to obtain the electromagnetic field distribution data, calculate the Joule heat loss power density based on the electric field distribution data, calculate the hysteresis loss power density based on the magnetic field distribution data, superimpose the Joule heat loss power density and the hysteresis loss power density in the spatial distribution to obtain the heat source distribution matrix, combine the heat source distribution matrix with the temperature field eigenvector to establish the heat source term of the temperature field, and calculate the temperature field distribution data based on the temperature field control equations containing the heat source term; The temperature gradient and temperature change rate are calculated based on the temperature field distribution data. When the temperature change rate exceeds the preset stability threshold, it is determined that the thermal stability state has not been reached. The electrical conductivity and thermal conductivity parameters of the material are updated based on the temperature field distribution data. The updated material physical parameters and temperature field distribution data are input into the grid optimization algorithm to dynamically optimize the multi-level grid structure.

2. The method according to claim 1, characterized in that The electromagnetic field control equations and the temperature field control equations are established according to the electromagnetic field eigenvectors and the temperature field eigenvectors. The GaN device is meshed using the adaptive hierarchical partitioning method. The mesh density distribution is determined based on the temperature gradient and material interface characteristics. A fine mesh is used in the area of ​​drastic temperature changes and the interface transition area, and a loose mesh is used in the area of ​​gentle temperature. A multi-level mesh structure with regional adaptive characteristics is generated, including: Obtaining material parameters and structural parameters of the gallium nitride device, the material parameters include electrical conductivity data and thermal conductivity data, and the structural parameters include material interface parameters, and calculating the electromagnetic field characteristic vector and the temperature field characteristic vector based on the material parameters and the structural parameters; According to the electromagnetic field characteristic vector, an electromagnetic field control equation group is established, which describes the electric field distribution law and the magnetic field distribution law. According to the temperature field characteristic vector, a temperature field control equation group is established, which describes the temperature field distribution law and the heat transfer law. Calculate temperature gradient distribution data based on the temperature field feature vector, divide the temperature change area according to the temperature gradient distribution data, mark the area with a temperature gradient greater than a first preset threshold as a temperature change drastic area, and mark the area with a temperature gradient less than a second preset threshold as a temperature gentle area; Determine the interface transition area based on the material interface parameters, the interface transition area is the area within a preset range around the material interface, and the range of the interface transition area is determined according to the distance to the material interface; Construct a mesh density distribution function, input the temperature gradient distribution data and the data of the interface transition area into the mesh density distribution function, and calculate the target mesh size of different areas. The mesh density distribution function ensures that the target mesh size of the area with a larger temperature gradient value is smaller, and the target mesh size of the area closer to the material interface is smaller; The temperature-changing area and interface transition area are divided into fine grids, and the size of the fine grid is calculated by the grid density distribution function. The temperature-smooth area is divided into loose grids, and the size of the loose grid is calculated by the grid density distribution function. A transition grid is constructed between the fine grid and the loose grid. The grid size of the transition grid gradually changes between adjacent grid units according to a preset proportional coefficient to ensure a continuous and smooth transition of the grid size. The fine grid, loose grid and transition grid are organized into a multi-level grid structure through node connection relationships.

3. The method according to claim 1, characterized in that Based on the electromagnetic field eigenvector, the electromagnetic field control equations are solved to obtain the electromagnetic field distribution data. The Joule heat loss power density is calculated based on the electric field distribution data. The hysteresis loss power density is calculated based on the magnetic field distribution data. The Joule heat loss power density and the hysteresis loss power density are superimposed on the spatial distribution to obtain the heat source distribution matrix. The heat source distribution matrix is ​​combined with the temperature field eigenvector to establish the heat source term of the temperature field. The temperature field distribution data obtained based on the temperature field control equations containing the heat source term include: Input the electromagnetic field eigenvector into the electromagnetic field control equation group to establish the curl field equation and the divergence field equation, use the vector finite element method to divide the curl field equation and the divergence field equation into space domain grid units, construct the electromagnetic field node basis function in each grid unit, assemble the global stiffness matrix and the mass matrix based on the electromagnetic field node basis function, solve the global stiffness matrix and the mass matrix through the conjugate gradient iterative algorithm to obtain the electromagnetic field distribution data, and the electromagnetic field distribution data includes the electric field distribution data and the magnetic field distribution data; The Joule heat loss power density is calculated based on the electric field distribution data, and the hysteresis loss power density is calculated based on the magnetic field distribution data. The Joule heat loss power density and the hysteresis loss power density are superimposed to generate a heat source distribution matrix. The heat source distribution matrix and the temperature field feature vector are input into a deep neural network for feature extraction and feature fusion to obtain a fused feature vector. A dynamic source term model of the temperature field is constructed based on the fused feature vector. Based on the dynamic source term model of temperature field, the backward implicit difference format is used for the time term of the temperature field control equation, and the central difference format is used for the space term. The temperature field difference equation group is established, and the temperature field difference equation group is converted into a linear algebraic equation group. The super-relaxation iterative algorithm is used to solve the linear algebraic equation group to obtain the temperature field distribution data.

4. The method according to claim 3, characterized in that The vector finite element method is used to divide the curl field equation and the divergence field equation into spatial domain grid units. The electromagnetic field node basis function is constructed in each grid unit. The global stiffness matrix and mass matrix are assembled based on the electromagnetic field node basis function. The global stiffness matrix and mass matrix are solved by the conjugate gradient iterative algorithm to obtain the electromagnetic field distribution data including: The temperature field distribution data is divided into multiple calculation units according to the spatial grid, a temperature value difference equation group is constructed for each calculation unit, the temperature value difference equation group is solved by the least square method to obtain the temperature gradient value of each grid node, and all the temperature gradient values ​​are combined to generate a temperature gradient matrix; Construct a mapping relationship function between material parameters and temperature gradients, substitute the temperature gradient matrix into the mapping relationship function to calculate the parameter sensitivity value of each grid node, and construct a parameter sensitivity function based on the parameter sensitivity value and the initial material parameters; establish a dynamic mapping equation between temperature field and material parameters, the dynamic mapping equation includes a parameter update term and a stability adjustment term, the parameter update term controls the parameter update rate, and the stability adjustment term suppresses parameter oscillation; The gradient descent method is used to iteratively solve the dynamic mapping equation. The gradient value of the parameter sensitivity function is calculated in each iteration. The parameter update direction and step size are determined according to the gradient value to obtain the material parameter correction amount of the current iteration. The material parameter correction amounts obtained in each iteration are weighted averaged to obtain the final material parameter correction amount.

5. The method according to claim 3, characterized in that The Joule heat loss power density is calculated based on the electric field distribution data, and the hysteresis loss power density is calculated based on the magnetic field distribution data. The Joule heat loss power density and the hysteresis loss power density are superimposed to generate a heat source distribution matrix including: Discretize the electric field distribution data into an electric field intensity vector matrix in the calculation domain, and discretize the magnetic field distribution data into a magnetic field intensity vector matrix in the calculation domain; calculate the current density distribution based on the electric field intensity vector matrix, establish the current density distribution matrix through the product relationship between the current density and the electric field intensity, calculate the Joule heat loss power density based on the current density distribution matrix, and obtain the Joule heat loss power density matrix by dividing the square of the value of each unit in the current density distribution matrix by the conductivity value of the corresponding position; The magnetic induction intensity distribution is calculated based on the magnetic field intensity vector matrix, the magnetic induction intensity distribution matrix is ​​established through the product relationship between the magnetic induction intensity and the magnetic field intensity, the magnetic induction intensity peak distribution is determined according to the magnetic induction intensity distribution matrix, and the hysteresis loss coefficient distribution matrix is ​​obtained by multiplying the magnetic induction intensity peak distribution with the magnetic field change frequency; The hysteresis loss power density is calculated based on the hysteresis loss coefficient distribution matrix, and the hysteresis loss power density matrix is ​​obtained by multiplying the value of each unit in the hysteresis loss coefficient distribution matrix with the product of the square of the peak value of the magnetic induction intensity at the corresponding position and the frequency of the magnetic field change; The Joule heat loss power density matrix and the hysteresis loss power density matrix are numerically superimposed at the corresponding position units in the calculation domain to obtain the total loss power density matrix. The value of each matrix unit represents the total heat source intensity at the corresponding position. The total loss power density matrix is ​​reconstructed according to the spatial grid division, and the total loss power density in each grid cell is volume integrated and divided by the grid cell volume to obtain the average loss power density value of the grid cell; the average loss power density value of the grid cell is superimposed according to the spatial position relationship of the grid cell in the calculation domain to form a heat source distribution matrix.

6. The method according to claim 1, characterized in that The temperature gradient and temperature change rate are calculated based on the temperature field distribution data. When the temperature change rate exceeds the preset stability threshold, it is determined that the thermal stability state has not been reached. The electrical conductivity and thermal conductivity parameters of the material are updated based on the temperature field distribution data. The updated material physical parameters and temperature field distribution data are input into the grid optimization algorithm. The multi-level grid structure is dynamically optimized, including: The temperature field distribution data at the initial moment is obtained, and the temperature gradient value of each grid node is calculated based on the temperature field distribution data. The temperature gradient value is obtained by calculating the temperature space derivative; the temperature field distribution data at two adjacent moments is obtained, and the temperature change amount per unit time is calculated to obtain the temperature change rate. The temperature change rate represents the transient change characteristics of the temperature field; The temperature change rate is compared with a preset stability threshold. When the temperature change rate is greater than the preset stability threshold, the temperature field is determined to be in an unstable state. The conductivity parameter is updated using the temperature-conductivity characteristic curve according to the temperature field distribution data at the current moment. The conductivity parameter changes according to a preset rule as the temperature increases. When the temperature field is determined to be in a stable state, the thermal conductivity parameter is updated using the temperature-thermal conductivity characteristic curve according to the temperature field distribution data at the current moment. The thermal conductivity parameter changes according to a preset rule as the temperature increases. A mesh optimization algorithm is constructed, and the temperature field distribution data, the updated electrical conductivity parameters, and the updated thermal conductivity parameters are used as input parameters. The calculation results of the mesh optimization algorithm are used to construct a first layer of meshes in an area where the temperature gradient value is higher than a preset gradient threshold value, and the first layer of meshes has a mesh density higher than a preset density threshold value to accurately capture the characteristics of drastic temperature changes; the calculation results of the mesh optimization algorithm are used to construct a second layer of meshes in an area where the temperature gradient value is lower than the preset gradient threshold value, and the second layer of meshes has a mesh density lower than the preset density threshold value to improve calculation efficiency; A transition layer grid is constructed between the first layer grid and the second layer grid. The density of the transition layer grid changes continuously according to the preset gradient coefficient to ensure a smooth transition of the grid structure. The first layer grid, the transition layer grid and the second layer grid are connected to form a multi-level grid structure, and the topological connection relationship of the grid nodes is established. The temperature field analysis, material parameter update and grid optimization process are continued until the temperature change rate is less than the preset stability threshold, thereby realizing dynamic optimization of the multi-level grid structure.

Images