Silicon carbide power module thermal management method and system based on double-sided heat dissipation heat pipe

Abstract

The invention discloses a silicon carbide power module thermal management method and system based on a double-sided heat dissipation heat pipe, and relates to the technical field of thermal management of power electronic devices. The method comprises the steps that S1, working condition parameters of a silicon carbide power module are collected in real time, and a material attribute database is called to obtain preset chip layout parameters; s2, inputting the working condition parameters and the chip layout parameters into a pre-trained heat flow distribution prediction model, generating a heat flow density distribution diagram, and calculating a double-sided heat dissipation distribution scheme through a thermal resistance balance algorithm; s3, generating a configuration scheme containing a heat pipe layout path and a thermal interface material based on the heat flow density distribution diagram and the double-sided heat dissipation distribution scheme, and completing the assembly of the heat dissipation system; and S4, during operation of the system, monitoring temperature data of the heat dissipation system in real time, and generating a fan rotating speed control instruction. Heat flow is accurately predicted through data driving, double-face heat dissipation is optimized, and efficient temperature equalization and low-noise operation are achieved in combination with intelligent temperature control.

Description

Technical Field

[0001] This invention relates to the field of thermal management technology for power electronic devices, specifically to a thermal management method and system for silicon carbide power modules based on double-sided heat pipes. Background Technology

[0002] Silicon carbide power devices are increasingly widely used in modern power electronic devices due to their high-frequency, high-voltage, and high-temperature operating characteristics. As power density continues to increase, thermal management becomes particularly important, directly affecting the reliability, lifespan, and compactness of the devices. Traditional single-sided heat dissipation solutions struggle to efficiently dissipate high-density heat flow, easily leading to localized hot spots and limiting the full potential of the devices.

[0003] To enhance heat dissipation capabilities, the industry has introduced technologies such as heat pipes and bi-directional cooling. However, existing methods still lack systematic solutions for optimizing heat pipe network layouts based on actual heat source distribution, dynamically allocating heat loads across bi-directional cooling paths, and achieving quiet and efficient operation of cooling systems. Typical thermal design relies on experience and simplified simulations, making it difficult to accurately match complex operating conditions, resulting in redundant or inefficient cooling structures. Furthermore, operational heat dissipation control often employs simple feedback mechanisms, making it difficult to achieve an optimal balance between heat dissipation efficiency, noise levels, and dynamic response.

[0004] Therefore, there is an urgent need for a more intelligent, precise, and systematic thermal management method to address the severe heat dissipation challenges of high-power silicon carbide modules and meet the comprehensive requirements of high reliability, high power density, and low noise operation. Summary of the Invention

[0005] In view of the shortcomings of the prior art described above, the purpose of this invention is to provide a thermal management method and system for silicon carbide power modules based on double-sided heat pipes, so as to solve the above-mentioned technical problems.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a thermal management method for silicon carbide power modules based on double-sided heat pipes, comprising:

[0007] S1: Real-time acquisition of operating parameters of silicon carbide power modules, and retrieval of preset chip layout parameters from the material property database;

[0008] S2: Input the operating parameters and chip layout parameters into the pre-trained heat flow distribution prediction model to generate a heat flow density distribution map, and calculate the double-sided heat dissipation distribution scheme through the thermal resistance balance algorithm.

[0009] S3: Based on the heat flux density distribution map and the double-sided heat dissipation distribution scheme, generate a configuration scheme that includes heat pipe layout paths and thermal interface materials, and complete the assembly of the heat dissipation system accordingly.

[0010] S4: During system operation, monitor the temperature data of the heat dissipation system in real time and generate fan speed control commands.

[0011] The present invention is further configured such that S1 includes:

[0012] The operating parameters are constructed by acquiring current, voltage, and switching frequency signals in real time through current sensors, voltage sensors, and controller registers.

[0013] The chip layout parameters are constructed by retrieving data from a pre-defined material property database, including chip geometric coordinates, dimensions, package layer thickness, and material thermal properties.

[0014] The present invention is further configured such that the pre-trained heat flow distribution prediction model is obtained through the following steps:

[0015] Collect operating data of silicon carbide power modules under different operating conditions and their corresponding infrared thermal imaging temperature field data to construct a training sample set;

[0016] A hybrid neural network structure is constructed using convolutional networks and long short-term memory networks. The convolutional network is used to process the spatial features of chip layout, and the long short-term memory network is used to process the temporal series features of running data.

[0017] The hybrid neural network structure is trained under supervision using a training sample set, with running data as input and temperature field data as supervision labels. The network parameters are then iteratively optimized to obtain a heat flow distribution prediction model.

[0018] The present invention is further configured such that S2 includes: a heat flow map generation step and an allocation calculation step.

[0019] The present invention is further configured such that the heat flow map generation step includes:

[0020] The time-series signal in the operating condition parameters is time-space aligned with the chip geometric coordinates in the chip layout parameters, and all parameters are normalized.

[0021] The normalized operating parameters and chip layout parameters are concatenated and merged into a unified multi-dimensional feature vector.

[0022] The multidimensional feature vector is input into the pre-trained heat flux distribution prediction model, and the linear transformation, nonlinear activation and feature transfer calculation of each network layer in the hybrid neural network structure are performed in sequence to output the heat flux density distribution map.

[0023] The present invention is further configured such that the allocation calculation step includes:

[0024] Based on the heat flux density distribution map and chip layout parameters, an equivalent thermal resistance network model is constructed.

[0025] Based on the heat flux density distribution map and the equivalent thermal resistance network model, the predicted temperature of each chip node is calculated.

[0026] The optimization problem is established with minimizing the maximum temperature difference inside the silicon carbide power module as the optimization objective and the constraint that the predicted temperature of all chip nodes does not exceed a preset safety threshold.

[0027] The optimal thermal resistance configuration parameters are obtained by iteratively solving the optimization problem using a numerical optimization algorithm.

[0028] Based on the optimal thermal resistance configuration parameters, a dual-sided heat dissipation distribution scheme is directly output; wherein, the dual-sided heat dissipation distribution scheme defines the distribution ratio of heat in each local area of ​​the module surface in the upper and lower heat dissipation paths in the form of a matrix or mapping table.

[0029] The present invention is further configured such that S3 includes:

[0030] Based on the heat flux density distribution map, high heat flux density areas are identified, a set of evaporation end positioning coordinates is constructed, and the type and thickness configuration of thermal interface materials in each area are determined according to the double-sided heat dissipation distribution scheme and chip layout parameters.

[0031] Based on the evaporator end positioning coordinate set, the optimal heat pipe layout path connecting each evaporator end to the preset condenser end is generated through geometric path planning.

[0032] The present invention is further configured such that S3 further includes:

[0033] The optimal heat pipe layout path and the type and thickness configuration of the thermal interface material are converted into structured manufacturing instructions;

[0034] The heat dissipation system is assembled by driving the equipment according to structured manufacturing instructions.

[0035] The present invention is further configured such that S4 includes:

[0036] During system operation, the temperature at multiple preset locations in the heat dissipation system is monitored in real time to obtain real-time temperature data;

[0037] Based on real-time temperature data and combined with preset temperature control targets, the fan adjustment amount is calculated through model predictive control algorithms.

[0038] Convert the fan adjustment amount into a specific fan speed control command.

[0039] This invention also provides a thermal management system for silicon carbide power modules based on double-sided heat pipes, the system comprising:

[0040] Parameter acquisition and scheduling module: Real-time acquisition of operating parameters of silicon carbide power modules, and access to the material property database to obtain preset chip layout parameters;

[0041] Simulation and allocation decision module: Input the operating condition parameters and chip layout parameters into the pre-trained heat flow distribution prediction model, generate a heat flow density distribution map, and calculate the double-sided heat dissipation allocation scheme through the thermal resistance balance algorithm;

[0042] Heat dissipation structure configuration generation module: Based on the heat flux density distribution map and the double-sided heat dissipation distribution scheme, it generates a configuration scheme including heat pipe layout path and thermal interface material, and completes the assembly of the heat dissipation system accordingly.

[0043] Intelligent temperature control module: During system operation, it monitors the temperature data of the heat dissipation system in real time and generates fan speed control commands.

[0044] This invention provides a thermal management method and system for silicon carbide power modules based on double-sided heat pipes. The method comprises: S1: Real-time acquisition of operating parameters of the silicon carbide power module and retrieval of preset chip layout parameters from a material property database; S2: Inputting the operating parameters and chip layout parameters into a pre-trained heat flux distribution prediction model to generate a heat flux density distribution map, and calculating a double-sided heat dissipation allocation scheme using a thermal resistance balance algorithm; S3: Based on the heat flux density distribution map and the double-sided heat dissipation allocation scheme, generating a configuration scheme including heat pipe layout paths and thermal interface materials, thereby completing the assembly of the heat dissipation system; S4: During system operation, real-time monitoring of the temperature data of the heat dissipation system and generation of fan speed control commands. The resulting benefits include:

[0045] It achieves predictability and accuracy in thermal management: the heat flow distribution is accurately predicted by a pre-trained hybrid neural network model, and the double-sided heat dissipation path is optimized based on the thermal resistance balance algorithm. This enables the heat dissipation structure design to shift from relying on experience to data-driven design, significantly improving the matching accuracy with the actual heat source distribution and avoiding local overheating from the source.

[0046] The efficiency and temperature uniformity of the heat dissipation system were optimized: by constructing an equivalent thermal resistance network model and optimizing the solution, the heat load of the upper and lower heat dissipation paths was dynamically allocated, which effectively reduced the maximum temperature and temperature difference inside the module, improved heat dissipation efficiency and reliability, and created conditions for increasing power density.

[0047] It ensures quiet and efficient operation: During operation, a model predictive control algorithm is used to dynamically adjust the fan speed based on real-time temperature data, achieving an intelligent balance between heat dissipation capacity and noise level, so that the system can maintain lower operating noise while meeting strict temperature control requirements.

[0048] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0050] Figure 1 A flowchart illustrating a thermal management method for a silicon carbide power module based on a double-sided heat pipe, as an exemplary embodiment of the present invention;

[0051] Figure 2 This is a schematic diagram illustrating the structure of a silicon carbide power module thermal management system based on a double-sided heat pipe, as an exemplary embodiment of the present invention. Detailed Implementation

[0052] The embodiments of the present invention will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention and not for limiting the scope of protection of the present invention.

[0053] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0054] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0055] Example 1:

[0056] Thermal management methods for silicon carbide power modules based on double-sided heat pipes, such as Figure 1 As shown, it includes:

[0057] S1: Real-time acquisition of operating parameters of silicon carbide power modules, and retrieval of preset chip layout parameters from the material property database;

[0058] S2: Input the operating parameters and chip layout parameters into the pre-trained heat flow distribution prediction model to generate a heat flow density distribution map, and calculate the double-sided heat dissipation distribution scheme through the thermal resistance balance algorithm.

[0059] S3: Based on the heat flux density distribution map and the double-sided heat dissipation distribution scheme, generate a configuration scheme that includes heat pipe layout paths and thermal interface materials, and complete the assembly of the heat dissipation system accordingly.

[0060] S4: During system operation, monitor the temperature data of the heat dissipation system in real time and generate fan speed control commands.

[0061] The present invention is further configured such that S1 includes:

[0062] The operating parameters are constructed by acquiring current, voltage, and switching frequency signals in real time through current sensors, voltage sensors, and controller registers.

[0063] The system retrieves data from a pre-defined material property database, including chip geometric coordinates, dimensions, package layer thickness, and material thermophysical parameters, to construct chip layout parameters. Specifically, operating parameters include current reflecting real-time load, voltage characterizing voltage stress, and switching frequency determining switching loss frequency. These parameters are obtained through a combination of physical measurement and direct reading from the controller: a closed-loop Hall effect current sensor converts the bus current into a proportional analog voltage signal; a high-impedance differential voltage probe performs a safe step-down measurement of the DC bus voltage; and the system controller reads the frequency setting value of the pulse width modulation generator via a digital communication interface. These analog signals are synchronously sampled and converted into digital codes by a high-precision analog-to-digital converter, then processed by a digital filtering algorithm in the microcontroller firmware to suppress noise, and converted into values ​​with actual physical units according to a pre-defined calibration coefficient. Simultaneously, based on the power module model identifier, the system accesses a pre-defined material property relational database, retrieves and loads the corresponding chip layout parameters by executing structured query language commands. These static parameters include the two-dimensional geometric coordinates of each chip on the substrate, the chip's length, width, and thickness dimensions, the thickness of each package material layer, and thermophysical properties such as the thermal conductivity of the chip and each layer at different temperatures. The complete implementation process is as follows: when the calculation cycle is triggered, the system first queries the database and loads all chip layout data of the corresponding model into memory; then, the analog signals sensed by the current and voltage sensors are digitized by the analog-to-digital converter, and after filtering and scaling transformation, real-time physical values ​​are obtained. At the same time, the real-time switching frequency value is obtained through the controller local area network communication protocol; finally, the microcontroller encapsulates the current value, voltage value, switching frequency value processed in this cycle, together with the chip layout parameter references already associated in memory, into a structured input data object: operating condition parameters and chip layout parameters.

[0064] The present invention is further configured such that the pre-trained heat flow distribution prediction model is obtained through the following steps:

[0065] Collect operating data of silicon carbide power modules under different operating conditions and their corresponding infrared thermal imaging temperature field data to construct a training sample set;

[0066] A hybrid neural network structure is constructed using convolutional networks and long short-term memory networks. The convolutional network is used to process the spatial features of chip layout, and the long short-term memory network is used to process the temporal series features of running data.

[0067] Supervised training of the hybrid neural network structure is performed using a training sample set. Operating data is used as input, and temperature field data is used as the supervision label. The network parameters are iteratively optimized to obtain a heat flux distribution prediction model. Specifically, the heat flux distribution prediction model establishes a high-precision nonlinear mapping relationship between the real-time operating state of the silicon carbide power module and its detailed surface heat flux distribution. The implementation of the heat flux distribution prediction model consists of three stages: data acquisition and sample construction, network architecture design and implementation, and model training and optimization.

[0068] The first phase involves data acquisition and training sample set construction. This phase is completed on a controlled laboratory testing platform. The testing platform includes a programmable DC power supply, electronic load, gate driver, and an open heat sink fixture with the target silicon carbide power module mounted on it. Key measurement equipment includes a high-precision power analyzer and data acquisition card for acquiring real-time operating data, as well as a high-resolution infrared thermal imaging camera. When constructing the sample set, a series of test conditions covering the expected operating range of the module are first planned, including different combinations of DC bus voltage, load current waveforms, and switching frequencies. Under each steady-state or transient test condition, the system synchronously records the time-series operating data output by the power analyzer, including the precise waveforms of current, voltage, and switching frequency of each bridge arm. Simultaneously, the infrared thermal imaging camera captures images of the temperature field distribution on the surface of the power module at a fixed frame rate, and converts the pixel grayscale values ​​in the infrared images into absolute temperature values ​​through temperature calibration, generating a two-dimensional temperature field matrix. To ensure spatiotemporal alignment of the data, all data acquisition devices are triggered and synchronized by the same clock source, and each set of time-corresponding operating data sequences is associated with the infrared temperature field images to form an "input-output" paired sample. Finally, thousands of such paired samples were organized into a standardized dataset and randomly divided into training, validation, and test sets according to proportions.

[0069] The second stage involves the design and construction of a hybrid neural network structure. To address the simultaneous spatial layout dependency and temporal dynamics in power module thermal prediction, the network employs a parallel fusion architecture of convolutional neural networks and long short-term memory networks. The convolutional network branch handles the spatial topological features of the chip layout. Its input is encoded chip layout parameters, such as chip coordinates and dimensions, which are converted into two-dimensional feature maps corresponding to the module surface grid. The network consists of multiple layers of two-dimensional convolutional and pooling layers. The convolutional layers use small filters sliding across the feature maps to progressively extract the geometric and spatial correlation features of heat sources from local to global perspectives. The long short-term memory network branch specifically handles the temporal series features of operating parameters. Its input is sequential data such as current and voltage arranged in time steps. Internally, the network learns and memorizes long-term dependencies in the time series through its unique gating mechanism, effectively capturing the dynamic process of switching losses changing with frequency and duty cycle. The two branches perform feature fusion in the middle layer of the network, concatenating the extracted high-level spatial feature vector with the temporal feature vector, and then performing joint abstraction through several fully connected layers, finally outputting a two-dimensional heat flux density distribution map with the same resolution as the input temperature field image.

[0070] The third stage is supervised training and optimization of the model. The training process is performed on a computing server equipped with a graphics processing unit (GPU). During training, a batch of samples is taken from the training set, and the running data is input into the network to perform forward propagation calculations to obtain the predicted heat flow map. Subsequently, the difference between the predicted heat flow map and the actual infrared temperature field data is calculated, using mean squared error as the loss function to quantify this difference. The core of the training is the collaborative work of the backpropagation algorithm and the optimizer. The system automatically calculates the gradient of the loss relative to millions of network parameters, and then uses an adaptive moment estimation optimization algorithm to dynamically adjust the learning step size of each parameter based on the gradient direction and historical momentum, updating the network parameters to reduce the loss. This process is iterated for tens to hundreds of cycles on the training set. To prevent the model from overfitting the training data, after each training cycle, the model performance is evaluated using an independent validation set, and training is stopped early when the validation loss no longer decreases. Finally, the model parameters that perform best on the test set and whose prediction error is consistently below a preset threshold are selected as the final heat flow distribution prediction model. This training process comprehensively utilizes supervised learning, backpropagation, adaptive optimization algorithms, and regularization techniques to ensure that the obtained heat flow distribution prediction model has both prediction accuracy and generalization ability.

[0071] The present invention is further configured such that S2 includes a heat flow map generation step and an allocation calculation step. Specifically, the heat flow map generation step converts abstract operating parameters such as current and voltage into an intuitive and detailed "heat map" of the module surface, i.e., a heat flow density distribution map, by calling a pre-trained intelligent model, which is used to accurately reveal where it is hot and where it is even hotter.

[0072] The allocation calculation steps are based on the "heat map" generated by the above steps. Through optimized calculation and intelligent decision-making, the total heat is dynamically allocated to the upper and lower heat dissipation surfaces in the best proportion to generate a dual-sided heat dissipation allocation scheme. It is like planning the most efficient "dual lane" for heat dissipation to achieve uniform heat dissipation and prevent local overheating.

[0073] The present invention is further configured such that the heat flow map generation step includes:

[0074] The time-series signal in the operating condition parameters is time-space aligned with the chip geometric coordinates in the chip layout parameters, and all parameters are normalized.

[0075] The normalized operating parameters and chip layout parameters are concatenated and merged into a unified multi-dimensional feature vector.

[0076] The multidimensional feature vector is input into a pre-trained heat flux distribution prediction model, and linear transformations, nonlinear activations, and feature transfer calculations are sequentially performed on each network layer of the hybrid neural network structure to output a heat flux density distribution map. Specifically, the implementation process includes three main stages: data alignment and standardization, feature fusion, and neural network inference.

[0077] The first stage involves temporal and spatial alignment and normalization. Since operating parameters are time-varying sequential signals, while chip layout is a static spatial coordinate system, they need to be unified onto a computable common basis. The system assigns a unique spatial identifier to each chip and extracts features from the current and voltage waveforms within a complete switching cycle, calculating their effective values ​​and representative scalars such as average switching power loss. Normalization is then performed to eliminate the impact of differences in the dimensions of different physical quantities on model convergence. Typically, all input parameters are linearly scaled to the range of 0 to 1. The global maximum and minimum values ​​required for normalization are derived from the statistical results of the entire training sample during model training. For example, current might be normalized based on the maximum value of the sensor range. Using training set statistics instead of real-time extreme values ​​ensures consistency in data distribution between online inference and offline training, preventing performance degradation due to input data distribution drift.

[0078] The second stage involves feature concatenation and multi-dimensional feature vector construction. The aligned and normalized scalar parameters are organized into a one-dimensional feature array. Simultaneously, the chip's geometric coordinates and dimensional parameters are converted into a spatial code, for example, discretized and mapped onto a low-resolution two-dimensional grid. Each grid point contains information such as whether a chip exists at that location and its basic attributes. Subsequently, the one-dimensional feature array representing operating conditions and the two-dimensional encoding matrix representing spatial layout are concatenated and flattened along the feature dimension, fusing them into a unified, high-dimensional feature vector. This vector simultaneously encapsulates both "when and at what power" and "where heat is generated."

[0079] The third stage involves performing forward propagation calculations in the neural network and outputting a heat flux map. The multidimensional feature vectors obtained in the previous step are input into the pre-trained heat flux distribution prediction model loaded into memory. The hybrid neural network structure within this model performs calculations sequentially. First, the feature vectors are linearly transformed and weighted by fully connected layers, where the connection weights are the key parameters learned by the model during training. The convolutional network branches process the spatial feature part, and the filter weights of their convolutional layers are typically initialized using the Kemming uniform distribution method. This method can maintain the stability of the variance of the output data of each layer in the early stage of training, which is beneficial to model convergence. Subsequently, the result of the linear transformation is passed through a non-linear activation function, such as the modified linear unit function, whose default threshold is 0, that is, retaining all positive inputs and setting negative inputs to 0. This introduces non-linear mapping capability into the model. The features are passed layer by layer between network layers. After multiple convolution, activation, and pooling operations, the deep feature map is restored to the same spatial resolution as the target heat flux map at the output layer through deconvolution or upsampling operations. Finally, the model outputs a two-dimensional matrix, where each element of the matrix represents the predicted heat flux density of the corresponding small region on the module surface. This matrix is ​​denormalized and converted back to real physical units, which is then output as the final heat flux density distribution map.

[0080] The present invention is further configured such that the allocation calculation step includes:

[0081] Based on the heat flux density distribution map and chip layout parameters, an equivalent thermal resistance network model is constructed.

[0082] Based on the heat flux density distribution map and the equivalent thermal resistance network model, the predicted temperature of each chip node is calculated.

[0083] The optimization problem is established with minimizing the maximum temperature difference inside the silicon carbide power module as the optimization objective and the constraint that the predicted temperature of all chip nodes does not exceed a preset safety threshold.

[0084] The optimal thermal resistance configuration parameters are obtained by iteratively solving the optimization problem using a numerical optimization algorithm.

[0085] Based on the optimal thermal resistance configuration parameters, a bi-sided heat dissipation scheme is directly output. This scheme, in the form of a matrix or mapping table, defines the proportion of heat distribution in the upper and lower heat dissipation paths for each local area on the module surface. Specifically, the implementation process includes five stages: thermal resistance network modeling, node temperature calculation, optimization problem construction, numerical solution, and allocation scheme generation.

[0086] The first stage involves constructing an equivalent thermal resistance network model. This model discretizes the continuous physical structure into a computational network consisting of nodes and thermal resistance branches. The modeling process uses a heat flux density distribution map as the heat source input, where the heat flux density of each high-resolution grid cell is converted into the thermal power applied to the corresponding spatial node. Simultaneously, the thermal conductivity and thickness information of each material layer extracted from the chip layout parameters are used to calculate the thermal resistance value. For example, the vertical thermal resistance connecting adjacent nodes is obtained by dividing the thickness of the corresponding material layer by its thermal conductivity and the horizontal area represented by the node. The lateral thermal resistance connecting adjacent horizontal nodes reflects the ease with which heat diffuses through lateral thermally conductive layers such as copper layers. Ultimately, the model forms a detailed thermal resistance network containing hundreds to thousands of nodes, characterizing all major heat conduction paths from the chip junction to the upper and lower heat sinks.

[0087] The second stage involves calculating the predicted temperature of each chip node. The system expresses the thermal resistance network model obtained in the first stage as a large system of linear equations. The coefficient matrix of the equations consists of the reciprocals of each thermal resistance, and the vector on the right-hand side of the equations contains the heat source power at all nodes. By solving this system of linear equations with node temperatures as unknowns, the steady-state predicted temperature of each chip node and other locations within the module can be obtained under a given heat flux distribution and current thermal resistance configuration. Efficient sparse matrix solving algorithms, such as the preprocessed conjugate gradient method, are typically used to adapt to large-scale network computations.

[0088] The third stage involves establishing a mathematical optimization problem. The decision variables for optimization are the adjustable thermal resistance parameters in the model, primarily specifying the proportion of equivalent thermal conductance allocated to the upper and lower heat dissipation paths. The primary optimization objective is to directly minimize the maximum predicted temperature of all chip nodes. The core constraint is that the predicted temperature of all chip nodes must be below a preset safety threshold, typically set 1 to 15 degrees Celsius lower than the maximum allowable junction temperature of silicon carbide devices, for example, 160 degrees Celsius, to retain sufficient design and operational margins. Simultaneously, the optimization variables themselves also have upper and lower bound constraints to ensure that the allocation proportions are physically achievable.

[0089] The fourth stage involves iterative solving using numerical optimization algorithms. For the constrained nonlinear optimization problems mentioned above, algorithms such as sequential quadratic programming are typically employed. In each iteration, this algorithm constructs and solves an approximate quadratic programming subproblem at the current solution, gradually approaching the optimal solution of the original problem. The algorithm requires setting a convergence tolerance as the iteration stopping condition; the default value is typically set to 1*10^-6, meaning that when the change in the objective function value or solution is less than this threshold, the optimal solution that meets the accuracy requirements is considered to have been found. This threshold balances computational accuracy with solution time.

[0090] The fifth stage involves generating and outputting a dual-sided heat dissipation distribution scheme. The optimal thermal resistance configuration parameters obtained from the solution essentially define the ratio of thermal conductivity from each local region to the upper and lower heat dissipation surfaces. The system converts this ratio into a specific distribution scheme. This scheme is output as a two-dimensional matrix or a key-value pair mapping table. The row and column indices of the matrix correspond to the region grid coordinates on the module surface, and each cell stores a pair of values ​​representing the proportion of heat allocated to the upper and lower heat dissipation paths, respectively, with the sum of the two always being 100%. This scheme serves as the direct input basis for subsequent heat dissipation structure design and optimization.

[0091] The present invention is further configured such that S3 includes:

[0092] Based on the heat flux density distribution map, high heat flux density areas are identified, a set of evaporation end positioning coordinates is constructed, and the type and thickness configuration of thermal interface materials in each area are determined according to the double-sided heat dissipation distribution scheme and chip layout parameters.

[0093] Based on the evaporator end location coordinate set, the optimal heat pipe layout path connecting each evaporator end to the preset condenser end is generated through geometric path planning. Specifically, the implementation process includes two main stages in sequence: hotspot identification and material configuration design, and heat pipe path geometric planning.

[0094] The first stage involves identifying high-heat areas and determining the location of the evaporation end and the configuration of the interface material. The system first performs a region scan and feature analysis on the input heat flux density distribution map to identify local hotspots requiring focused heat dissipation. Identification is typically based on a dynamic threshold, which is set by default to 2.5 times the average heat flux density of the entire module. This multiple is chosen based on engineering experience, effectively filtering out key heat-generating areas that decisively influence the overall temperature rise of the module, avoiding over-design of low-temperature areas. All consecutive pixel regions with heat flux densities exceeding this threshold are marked, and their geometric center coordinates are used as recommended positioning points for the flat heat pipe evaporation section, forming the evaporation end positioning coordinate set. Next, the system determines the configuration of the thermal interface material for each local area according to the double-sided heat dissipation allocation scheme. The upper and lower surfaces of the chip are designed independently. The allocation scheme specifies the proportion of heat flow to the upper and lower paths for each area, and the system translates this proportion into specific requirements for the thermal resistance of the heat conduction paths on both the upper and lower sides. The material selection logic is based on a predefined database of material properties. For example, when a region's upper surface needs to bear the main heat dissipation and has an extremely high heat flux density, a metal-based composite phase change material is automatically selected as the preferred material type, with a typical default thickness of 0.15 mm. The rationale for choosing this default thickness is to provide sufficient material volume to withstand long-term thermal cycling stress while ensuring good surface adhesion to reduce contact thermal resistance. For areas with lower heat dissipation loads, lower-cost silicone grease or thermal pads may be selected, and the thickness may be adjusted to 0.2 to 0.3 mm to accommodate larger installation tolerances. All configuration decisions are guided by the theoretical target value calculated to achieve a thermal resistance balance scheme.

[0095] The second stage generates the optimal heat pipe layout path based on a coordinate set using a geometric algorithm. The system inputs the coordinates of the heat pipe evaporator ends, module boundaries, and geometric information of inaccessible areas such as chips and pads into a path planning algorithm. This algorithm abstracts the layout problem into a graph theory optimization problem, treating each evaporator end coordinate and the preset condenser end position located at the module edge as nodes in the graph, and potential connection paths as edges, with the edge weight determined by the path length and turning angle. The algorithm first uses the Delaunay triangulation method to establish an initial triangular mesh connection between the evaporator end nodes. Then, it applies a minimum spanning tree algorithm, such as Prim's algorithm, to select a set of edges with the shortest total length that connects all evaporator and condenser ends, forming the backbone path of the heat pipe network. During path generation, the system enforces a series of geometric constraints, such as a minimum spacing between paths of greater than 0.5 mm to prevent manufacturing interference, and a turning radius of no less than three times the heat pipe diameter to ensure unimpaired capillary performance.

[0096] The present invention is further configured such that S3 further includes:

[0097] The optimal heat pipe layout path and the type and thickness configuration of the thermal interface material are converted into structured manufacturing instructions;

[0098] The heat dissipation system is assembled using equipment driven by structured manufacturing instructions. Specifically, the process includes two main stages: conversion of structured manufacturing instructions and automated equipment-driven assembly.

[0099] The first stage involves converting design parameters into structured manufacturing instructions. The optimal heat pipe layout path received by the system is data in vector coordinate sequence format, while the thermal interface material type and thickness configuration are data in region mapping table format. The conversion process begins with the heat pipe layout path. The system calls its built-in computer-aided design kernel engine to automatically generate corresponding 2D engineering drawing files based on the path coordinate sequence, typically outputting in DXF format. Simultaneously, each path segment is labeled with its corresponding processing technology code. For example, for embedded grooves to be processed on copper-clad ceramic substrates, the path is converted into a sequence of G-code instructions to control a laser etching machine or precision milling machine. Key processing parameters such as laser power or cutting feed rate are set to default values. For instance, the default laser power is set to 120% of the minimum effective power required for complete vaporization of the corresponding copper layer thickness. This default value is set to ensure one-time processing while avoiding damage to the underlying ceramic insulating layer due to excessive power. For the thermal interface material configuration information, the system converts it into control instructions suitable for automatic dispensing machines or precision coating equipment. Based on the material type and target thickness, the system matches the corresponding dispensing valve model, pressure curve, and movement speed from the process database. For example, for a phase change material with a default thickness of 0.15 mm, the system will instruct the selection of a high-precision spiral valve and adopt a dispensing path strategy of first drawing the outline at high speed and then filling the interior at low speed. The default value of the extrusion pressure is set to 0.3 MPa, which is determined based on experimental data of the minimum pressure required to form a continuous and uniform film layer at a set thickness for a material of a specific viscosity.

[0100] The second stage involves system assembly driven by structured manufacturing instructions. All generated manufacturing instructions are integrated into a unified production execution system (MES) work order. The system first sends an instruction package containing the G-code for heat pipe path processing to the designated CNC machining center via an open production network connection protocol. The machining center controller parses the instructions and drives the spindle and worktable to complete the precision machining of the substrate. After machining, the system triggers a vision inspection unit to scan the machining features and compares the measurement results with the design coordinates. The default positional tolerance is no more than ±0.05 mm; this strict tolerance is set to ensure that the heat pipe can be accurately embedded subsequently. After confirming the machining is qualified, the system then schedules the automatic assembly unit. The robotic arm picks up the corresponding heat pipe from the material library according to the instructions, precisely places it in the machining groove of the substrate, and then transfers it to the vacuum brazing furnace. The furnace temperature profile is automatically set by the system according to the solder type and sent to the furnace controller. Simultaneously or after brazing, another high-precision dispensing device, based on the received coating instructions, completes the coating of the thermal interface material on the designated areas of the upper and lower heat dissipation surfaces of the module. All critical process parameters, such as the peak temperature and holding time of the brazing furnace or the ambient temperature of the dispensing machine, are clearly specified in the instructions. For example, the peak temperature for vacuum brazing is set by default to 30 degrees Celsius above the liquidus temperature of the solder used. The reason for setting this temperature offset is to provide the necessary process window to achieve optimal solder flowability and interfacial bonding strength. Finally, under the coordination of the central production execution system, each subsystem operates sequentially, assembling all components such as the processed substrate, chip, and casing into a complete heat dissipation system module. The completion status of each assembly step is confirmed through sensor feedback, achieving automated closed-loop production without human intervention.

[0101] The present invention is further configured such that S4 includes:

[0102] During system operation, the temperature at multiple preset locations in the heat dissipation system is monitored in real time to obtain real-time temperature data;

[0103] Based on real-time temperature data and combined with preset temperature control targets, the fan adjustment amount is calculated through model predictive control algorithms.

[0104] The fan adjustment parameters are converted into specific fan speed control commands. Specifically, the implementation process includes three main stages: multi-point, multi-frequency temperature monitoring, online optimization calculation by the model predictive controller, and control command issuance.

[0105] The first stage involves the highly reliable acquisition and preprocessing of real-time temperature data. Digital temperature sensors are installed at several pre-defined key locations in the heat dissipation system, such as near the center points of the upper and lower surfaces of the silicon carbide chip casing, the base of the heat sink fins, and the airflow outlet. These locations are chosen based on coverage of critical nodes in the heat flow path. The sensors synchronously report digital temperature readings via a single-bus protocol or integrated circuit bus at a fixed sampling period, e.g., ten times per second. After receiving the raw temperature data stream, the system first performs validity verification, eliminating outliers caused by communication errors. Subsequently, a first-order low-pass digital filter algorithm is applied to smooth the data. The filter time constant is set to two seconds by default, which is sufficient to suppress high-frequency measurement noise without significantly delaying the actual temperature change trend, thus providing a stable and timely feedback signal for the control algorithm. The processed temperature data is updated to a shared memory circular buffer, forming a real-time temperature dataset that the controller can access at any time.

[0106] The second stage is the online calculation of fan adjustment based on a model predictive control algorithm. The system's built-in model predictive controller is activated in each control cycle. At its core is a simplified discrete state-space model reflecting the dynamic characteristics of the cooling system. Its state variables include the chip junction temperature and the heat sink's thermal capacity temperature; the input variable is the fan speed; and the output variable is the controlled temperature. At each control moment, the controller first uses algorithms such as recursive least squares, combined with the latest input and output data, to fine-tune the model parameters to improve prediction accuracy. Then, using the current temperature as the initial state, the controller predicts the evolution of the system temperature under different fan speed sequences in a fixed-step rolling prediction over a future prediction time domain, such as the next ten seconds. The optimized objective function aims to minimize the deviation between the predicted temperature and the preset target temperature, while penalizing drastic changes in fan speed to achieve quiet operation. The weight of the output error term in the objective function is set to 1.0 by default, and the weight of the control increment term is set to 0.01 by default. The reason for this default weighting is to focus the primary control objective on temperature tracking accuracy, while imposing appropriate constraints on fan speed changes to smooth the speed curve. The solution also requires strict constraints, namely that all predicted chip junction temperatures must be below the safety threshold of 155 degrees Celsius. The controller obtains a series of optimal fan speed sequences in the future control time domain by solving this constrained quadratic programming problem, and outputs only the first value, i.e., the fan adjustment that should be executed immediately at the current moment.

[0107] The third stage involves converting the theoretical adjustment value into specific executable control commands. The calculated fan adjustment value is a dimensionless value or target speed representing the desired heat dissipation intensity. The system converts this value into specific hardware control signals based on the actual fan's drive interface protocol. For four-wire fans supporting pulse width modulation (PWM) speed control, the system calculates the corresponding duty cycle; for example, a target speed of 2000 RPM corresponds to a 67% duty cycle, and generates a digital waveform with the corresponding pulse width through the microcontroller's timer channel. For fans controlled via a bus protocol, the system encapsulates the target speed value into a data frame conforming to the protocol format and sends it to the fan driver via the controller area network (CLAN) or serial communication interface. After the command is issued, the system verifies the control action by reading the fan's speed feedback signal, ensuring accurate execution and thus closing the entire real-time control loop of "perception-decision-execution".

[0108] Example 2:

[0109] Please see Figure 2 The exemplary silicon carbide power module thermal management system based on double-sided heat pipes includes:

[0110] Parameter acquisition and scheduling module: Real-time acquisition of operating parameters of silicon carbide power modules, and access to the material property database to obtain preset chip layout parameters;

[0111] Simulation and allocation decision module: Input the operating condition parameters and chip layout parameters into the pre-trained heat flow distribution prediction model, generate a heat flow density distribution map, and calculate the double-sided heat dissipation allocation scheme through the thermal resistance balance algorithm;

[0112] Heat dissipation structure configuration generation module: Based on the heat flux density distribution map and the double-sided heat dissipation distribution scheme, it generates a configuration scheme including heat pipe layout path and thermal interface material, and completes the assembly of the heat dissipation system accordingly.

[0113] Intelligent temperature control module: During system operation, it monitors the temperature data of the heat dissipation system in real time and generates fan speed control commands.

[0114] It should be noted that the silicon carbide power module thermal management system based on double-sided heat pipes provided in the above embodiments and the silicon carbide power module thermal management method based on double-sided heat pipes provided in the above embodiments belong to the same concept. The specific operation methods of each module and unit have been described in detail in the method embodiments and will not be repeated here. In practical applications, the silicon carbide power module thermal management system based on double-sided heat pipes provided in the above embodiments can be assigned to different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above. This is not a limitation here.

[0115] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A thermal management method for silicon carbide power modules based on double-sided heat pipes, characterized in that, include: S1: Real-time acquisition of operating parameters of silicon carbide power modules, and retrieval of preset chip layout parameters from the material property database; S2: Input the operating parameters and chip layout parameters into the pre-trained heat flow distribution prediction model to generate a heat flow density distribution map, and calculate the double-sided heat dissipation distribution scheme through the thermal resistance balance algorithm. S3: Based on the heat flux density distribution map and the double-sided heat dissipation distribution scheme, generate a configuration scheme that includes heat pipe layout paths and thermal interface materials, and complete the assembly of the heat dissipation system accordingly. S4: During system operation, monitor the temperature data of the heat dissipation system in real time and generate fan speed control commands.

2. The thermal management method for silicon carbide power modules based on double-sided heat pipes according to claim 1, characterized in that, S1 includes: The operating parameters are constructed by acquiring current, voltage, and switching frequency signals in real time through current sensors, voltage sensors, and controller registers. The chip layout parameters are constructed by retrieving data from a pre-defined material property database, including chip geometric coordinates, dimensions, package layer thickness, and material thermal properties.

3. The thermal management method for silicon carbide power modules based on double-sided heat pipes according to claim 1, characterized in that, The pre-trained heat flow distribution prediction model is obtained through the following steps: Collect operating data of silicon carbide power modules under different operating conditions and their corresponding infrared thermal imaging temperature field data to construct a training sample set; A hybrid neural network structure is constructed using convolutional networks and long short-term memory networks. The convolutional network is used to process the spatial features of chip layout, and the long short-term memory network is used to process the temporal series features of the running data. The hybrid neural network structure is trained under supervision using a training sample set, with running data as input and temperature field data as supervision labels. The network parameters are then iteratively optimized to obtain a heat flow distribution prediction model.

4. The thermal management method for silicon carbide power modules based on double-sided heat pipes according to claim 1, characterized in that, The S2 includes: a heat flow map generation step and an allocation calculation step.

5. The thermal management method for silicon carbide power modules based on double-sided heat pipes according to claim 4, characterized in that, The heat flow map generation step includes: The time-series signal in the operating condition parameters is time-space aligned with the chip geometric coordinates in the chip layout parameters, and all parameters are normalized. The normalized operating parameters and chip layout parameters are concatenated and merged into a unified multi-dimensional feature vector. The multidimensional feature vector is input into the pre-trained heat flux distribution prediction model, and the linear transformation, nonlinear activation and feature transfer calculation of each network layer in the hybrid neural network structure are performed in sequence to output the heat flux density distribution map.

6. The thermal management method for silicon carbide power modules based on double-sided heat pipes according to claim 5, characterized in that, The allocation calculation steps include: Based on the heat flux density distribution map and chip layout parameters, an equivalent thermal resistance network model is constructed. Based on the heat flux density distribution map and the equivalent thermal resistance network model, the predicted temperature of each chip node is calculated. The optimization problem is established with minimizing the maximum temperature difference inside the silicon carbide power module as the optimization objective and the constraint that the predicted temperature of all chip nodes does not exceed a preset safety threshold. The optimal thermal resistance configuration parameters are obtained by iteratively solving the optimization problem using a numerical optimization algorithm. Based on the optimal thermal resistance configuration parameters, a dual-sided heat dissipation distribution scheme is directly output; wherein, the dual-sided heat dissipation distribution scheme defines the distribution ratio of heat in each local area of ​​the module surface in the upper and lower heat dissipation paths in the form of a matrix or mapping table.

7. The thermal management method for silicon carbide power modules based on double-sided heat pipes according to claim 1, characterized in that, S3 includes: Based on the heat flux density distribution map, high heat flux density areas are identified, a set of evaporation end positioning coordinates is constructed, and the type and thickness configuration of thermal interface materials in each area are determined according to the double-sided heat dissipation distribution scheme and chip layout parameters. Based on the evaporator end positioning coordinate set, the optimal heat pipe layout path connecting each evaporator end to the preset condenser end is generated through geometric path planning.

8. The thermal management method for silicon carbide power modules based on double-sided heat pipes according to claim 7, characterized in that, S3 further includes: The optimal heat pipe layout path and the type and thickness configuration of the thermal interface material are converted into structured manufacturing instructions; The heat dissipation system is assembled by driving the equipment according to structured manufacturing instructions.

9. The thermal management method for silicon carbide power modules based on double-sided heat pipes according to claim 1, characterized in that, S4 includes: During system operation, the temperature at multiple preset locations in the heat dissipation system is monitored in real time to obtain real-time temperature data; Based on real-time temperature data and combined with preset temperature control targets, the fan adjustment amount is calculated through model predictive control algorithms. Convert the fan adjustment amount into a specific fan speed control command.

10. A thermal management system for silicon carbide power modules based on double-sided heat pipes, used to implement the thermal management method for silicon carbide power modules based on double-sided heat pipes as described in any one of claims 1-9, characterized in that, include: Parameter acquisition and scheduling module: Real-time acquisition of operating parameters of silicon carbide power modules, and access to the material property database to obtain preset chip layout parameters; Simulation and allocation decision module: Input the operating condition parameters and chip layout parameters into the pre-trained heat flow distribution prediction model, generate a heat flow density distribution map, and calculate the double-sided heat dissipation allocation scheme through the thermal resistance balance algorithm; Heat dissipation structure configuration generation module: Based on the heat flux density distribution map and the double-sided heat dissipation distribution scheme, it generates a configuration scheme including heat pipe layout path and thermal interface material, and completes the assembly of the heat dissipation system accordingly. Intelligent temperature control module: During system operation, it monitors the temperature data of the heat dissipation system in real time and generates fan speed control commands.

Images