Chiplet maximum temperature prediction method and system in a co-wos package based on machine learning

Abstract

The application discloses a Chiplet highest temperature prediction method in CoWoS packaging based on machine learning and belongs to the technical field of semiconductor packaging. The method comprises the following steps: S1, establishing an equivalent model with consistent thermal characteristics according to an actual CoWoS packaging structure; S2, setting different characteristic parameter conditions to represent various working conditions; S3, obtaining a thermal resistance matrix corresponding to each parameter through numerical simulation; S4, assembling and preprocessing a data set of the characteristic parameters and the thermal resistance matrix; S5, training the data set by using a machine learning model to establish a mapping relationship; and S6, inputting the actual packaging structure after equivalent processing into the model, predicting the thermal resistance matrix and calculating the highest temperature. The application further discloses a Chiplet highest temperature prediction system in CoWoS packaging based on machine learning. The application reduces the complexity and saves the calculation resources through the equivalent model, and improves the prediction efficiency in combination with the machine learning.

Description

Technical Field

[0001] This invention relates to the field of semiconductor packaging technology, and in particular to a method and system for predicting the highest chiplet temperature in CoWoS packaging based on machine learning. Background Technology

[0002] In recent years, the rapid development of artificial intelligence and high-performance computing has driven the widespread application of Chip-on-Wafer-on-Substrate (CoWoS) technology, which has become a key packaging solution to meet the requirements of high integration and low power consumption. CoWoS forms a chip-on-wafer structure by interconnecting multiple chiplets on a silicon interposer and connecting them to the substrate via solder balls to form an integrated package. It has advantages such as high scalability, low power consumption, and high-speed interconnection, and is now widely used in the packaging design of high-end computing chips.

[0003] However, with increased integration, the heat flux density within the package rises significantly, increasing the difficulty of system thermal management. Excessive temperature can lead to chip performance degradation or even failure; studies show that for every 10°C increase in junction temperature, device lifespan is halved. To ensure the reliability and lifespan of CoWoS packages, precise analysis of their thermal characteristics is crucial, especially accurate prediction and evaluation of the maximum junction temperature of each chipplet.

[0004] Existing methods for predicting junction temperature mainly include experimental measurement and numerical simulation. Experimental methods typically apply power under standard test conditions and establish a thermal resistance model by measuring the chip junction temperature. However, these methods are costly, time-consuming, and difficult to adapt to multi-condition analysis. Numerical simulation methods typically use finite element analysis software to establish a three-dimensional thermal model and calculate the temperature distribution under different boundary conditions. While these methods offer high accuracy, they require extensive mesh generation, consume significant computational resources, and have long analysis cycles, making them unsuitable for the rapid thermal analysis needs of complex multi-chip package structures.

[0005] In summary, existing junction temperature prediction methods still have shortcomings in terms of computational efficiency, model accuracy, and versatility, making it difficult to balance the complexity of multi-chip package structures with prediction accuracy. Particularly in CoWoS packages, the thermal coupling effect between different chiplets is significant, and traditional analytical modeling and finite element simulation methods struggle to achieve efficient and accurate maximum temperature prediction. Therefore, there is an urgent need to propose a machine learning-based method and system for predicting the maximum temperature of chiplets in CoWoS packages that can balance computational efficiency and prediction accuracy, enabling rapid and accurate prediction of the maximum temperature of chiplets in CoWoS packages. Summary of the Invention

[0006] This invention aims to address the problems of high computational cost and large model complexity in existing CoWoS package junction temperature prediction methods by proposing a machine learning-based method and system for predicting the maximum temperature of chiplets in CoWoS packages. This method establishes an equivalent model consistent with the actual package thermal characteristics and combines it with machine learning algorithms to achieve rapid prediction of the thermal resistance matrix. This reduces the computational cost of modeling and simulation while ensuring prediction accuracy, providing an efficient analytical tool for the thermal design and reliability assessment of multi-chiplet packages.

[0007] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0008] A machine learning-based method for predicting the maximum temperature of a Chiplet in a CoWoS package includes the following steps:

[0009] S1. Based on the actual CoWoS packaging structure, establish an equivalent model that matches the actual thermal characteristics of the structure.

[0010] S2. Set different characteristic parameter conditions in the equivalent model to represent various working conditions;

[0011] S3. Through numerical simulation, steady-state thermal simulation is performed on the equivalent model under different parameters to obtain the corresponding thermal resistance matrix.

[0012] S4. Combine the feature parameters and the thermal resistance matrix to form a dataset, and preprocess the dataset.

[0013] S5. Use machine learning models to train datasets and establish the mapping relationship between feature parameters and thermal resistance matrix;

[0014] S6. After performing equivalent processing on the actual packaging structure, input it into the machine learning model to predict the thermal resistance matrix and calculate the maximum temperature of each chiplet.

[0015] Furthermore, S1 includes the following steps:

[0016] Based on the actual CoWoS packaging structure, key geometric features and material parameters are extracted;

[0017] Calculate the equivalent thermal resistance of the package structure to match the thermal resistance of the original structure, thereby ensuring that the maximum temperature of each chiplet in the equivalent structure is approximately the same as that in the original structure.

[0018] Equations were constructed to make the thermal resistance matrices of the original structure and the equivalent structure equal:

[0019]

[0020]

[0021]

[0022] Where R cond,equal R conv,equal and R s,equal These represent the convective thermal resistance, conductive thermal resistance, and diffusion thermal resistance of the equivalent structure, respectively; R conv and R cond R represents the convective thermal resistance and conductive thermal resistance of the original structure. sn This represents the diffusion thermal resistance generated from the nth layer to the (n+1)th layer.

[0023] Furthermore, S2 includes the following steps:

[0024] Based on the actual packaging structure, determine the characteristic parameters and their value range;

[0025] The parameter conditions include the structural parameters, material parameters, and boundary conditions of the equivalent structure.

[0026] Furthermore, S3 includes the following steps:

[0027] The numerical simulation methods include the finite element method, thermal network modeling, and MATLAB simulation.

[0028] For each set of parameter conditions, a unit power is applied to each chipplet individually, and the steady-state temperature rise of each chipplet is obtained through numerical simulation to determine the self-thermal resistance of each chipplet.

[0029] By sequentially selecting any two Chiplet chips and simultaneously applying a unit power input, the corresponding steady-state temperature change is obtained through numerical simulation.

[0030] Based on the principle of linear superposition of heat transfer, the coupling thermal resistance between each chipplet is solved using the above simulation results, and the complete thermal resistance matrix of the system is constructed.

[0031] Furthermore, S4 includes the following steps:

[0032] The feature parameters are combined with the corresponding thermal resistance matrix data to form a sample dataset;

[0033] The dataset is preprocessed, including data normalization, noise reduction, or outlier handling, to improve the training accuracy of subsequent machine learning models.

[0034] The dataset is divided into a training set, a validation set, and a test set for model training and performance evaluation.

[0035] Furthermore, S5 includes the following steps:

[0036] The preprocessed dataset is trained using a machine learning model;

[0037] A nonlinear mapping relationship is established between the characteristic parameters of the packaging structure and the thermal resistance matrix to obtain a training model that can predict the thermal resistance matrix.

[0038] Furthermore, the machine learning model is a BP neural network, the basic idea of ​​which is backpropagation of error. The BP neural network consists of an input layer, a hidden layer, and an output layer.

[0039] The input and output layers are single-layer structures, while the hidden layers can be single-layer or multi-layer, with the number of neurons determined based on the sample complexity and accuracy requirements.

[0040] The input layer is a vector of feature parameters, and the output layer is a vector of elements of a 2×2 thermal resistance matrix.

[0041] Furthermore, the BP neural network training steps include:

[0042] (1) Initialize the connection weights and thresholds, and randomly select values ​​in the interval [-1, 1];

[0043] (2) Set the activation function and input the training samples;

[0044] (3) Perform forward propagation and calculate the error between the predicted thermal resistance matrix and the target matrix;

[0045] (4) If the error exceeds the preset threshold, backpropagation is performed, the weights and thresholds are corrected using the gradient descent method, and the process returns to the third step to continue iterating in order to optimize the model’s fitting performance to the thermal resistance matrix until the error meets the requirements or the maximum number of runs is reached.

[0046] Furthermore, S6 includes the following steps:

[0047] The actual CoWoS packaging structure is equivalentized according to the S1 method;

[0048] The equivalent structural features are input into the trained machine learning model to predict the corresponding thermal resistance matrix.

[0049] Based on the thermal resistance matrix, the maximum junction temperature of each chiplet is calculated in conjunction with the chip power consumption conditions.

[0050] To achieve the above objectives, this invention also discloses a machine learning-based chipplet maximum temperature prediction system for CoWoS packaging, comprising:

[0051] The equivalent model construction module is used to establish an equivalent model that matches the actual thermal characteristics of the CoWoS package structure.

[0052] The parameter setting module is used to set different characteristic parameter conditions in the equivalent model to represent various working conditions;

[0053] The thermal resistance matrix acquisition module is used to perform steady-state thermal simulation on equivalent models under different parameter conditions through numerical simulation to obtain the corresponding thermal resistance matrix.

[0054] The dataset processing module is used to combine the feature parameters and the corresponding thermal resistance matrix to form a dataset, and to preprocess the dataset.

[0055] The model training module is used to train the preprocessed dataset using a machine learning model, establish the mapping relationship between feature parameters and thermal resistance matrix, and obtain the trained prediction model.

[0056] The temperature prediction module is used to input the actual CoWoS package structure into the trained prediction model after it has been equivalentized by the equivalent model building module, predict the corresponding thermal resistance matrix, and calculate the maximum temperature of each chiplet in combination with the chip power consumption conditions.

[0057] Beneficial effects: (1) By establishing an equivalent model consistent with the actual thermal characteristics of the package, the present invention can effectively characterize the thermal characteristics of the CoWoS package while effectively reducing the complexity of the model;

[0058] (2) The present invention uses machine learning algorithms to quickly predict the thermal resistance matrix, which significantly improves prediction efficiency and reduces computational resource consumption while maintaining high accuracy, and is applicable to multi-condition analysis.

[0059] (3) This invention can provide efficient and scalable analysis methods for thermal design and reliability assessment of CoWoS packaging, and provide technical support for thermal optimization of complex multi-chip packaging structures. Attached Figure Description

[0060] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0061] Figure 1 This is the main flowchart of the machine learning-based method for predicting the highest temperature of Chiplet in CoWoS packaging according to an embodiment of the present invention.

[0062] Figure 2 This is a diagram of the CoWoS packaging model structure for the machine learning-based method for predicting the highest temperature of Chiplet in CoWoS packaging, as described in an embodiment of the present invention.

[0063] Figure 3 This is a BP neural network structure diagram of the machine learning-based method for predicting the highest temperature of Chiplet in CoWoS packaging according to an embodiment of the present invention.

[0064] Figure 4 This is a schematic diagram of the structure of the Chiplet maximum temperature prediction system in the CoWoS package based on machine learning, as described in an embodiment of the present invention. Detailed Implementation

[0065] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0066] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0067] Example 1

[0068] See Figure 1-3 A machine learning-based method for predicting the maximum temperature of a Chiplet in a CoWoS package includes the following steps:

[0069] S1. Based on the actual CoWoS packaging structure, establish an equivalent model that matches the actual thermal characteristics of the structure.

[0070] S2. Set different characteristic parameter conditions in the equivalent model to represent various working conditions;

[0071] S3. Through numerical simulation, steady-state thermal simulation is performed on the equivalent model under different parameters to obtain the corresponding thermal resistance matrix.

[0072] S4. Combine the feature parameters and the thermal resistance matrix to form a dataset, and preprocess the dataset.

[0073] S5. Use machine learning models to train datasets and establish the mapping relationship between feature parameters and thermal resistance matrix;

[0074] S6. After performing equivalent processing on the actual packaging structure, input it into the machine learning model to predict the thermal resistance matrix and calculate the maximum temperature of each chiplet.

[0075] This embodiment fully covers the entire process from modeling to temperature prediction through the core technical framework of equivalent model + machine learning. It not only solves the problem of low computational efficiency of traditional numerical simulation, but also reduces the modeling difficulty of complex structures through equivalent model, achieving the dual advantages of high accuracy and high efficiency. It provides a standardized and replicable technical path for thermal analysis of CoWoS packaging.

[0076] In a specific instance, S1 includes the following steps:

[0077] Based on the actual CoWoS packaging structure, key geometric features and material parameters are extracted;

[0078] Calculate the equivalent thermal resistance of the package structure to match the thermal resistance of the original structure, thereby ensuring that the maximum temperature of each chiplet in the equivalent structure is approximately the same as that in the original structure.

[0079] Equations were constructed to make the thermal resistance matrices of the original structure and the equivalent structure equal:

[0080]

[0081]

[0082]

[0083] Where R cond,equal R conv,equal and R s,equal These represent the convective thermal resistance, conductive thermal resistance, and diffusion thermal resistance of the equivalent structure, respectively; R conv and R cond R represents the convective thermal resistance and conductive thermal resistance of the original structure. sn This represents the diffusion thermal resistance generated from the nth layer to the (n+1)th layer.

[0084] This embodiment ensures that the equivalent model matches the actual structural thermal characteristics by extracting key features, matching thermal resistance, and using equations to equalize the thermal resistance matrix. This simplifies the model complexity (reducing mesh generation and computational resource consumption) and avoids the problem of thermal characteristic distortion during the equivalence process, laying the foundation for subsequent efficient simulation and accurate prediction. At the same time, it clarifies the physical meaning of the thermal resistance matrix (self-thermal resistance + coupling thermal resistance), making it suitable for thermal coupling effect scenarios in multi-heat source systems.

[0085] It should be noted that in multi-heat source systems, when there is a significant thermal coupling effect, a single thermal resistance is insufficient to accurately characterize the temperature distribution, and a thermal resistance matrix needs to be introduced for description.

[0086]

[0087] For a system with n heat sources, its thermal resistance matrix is ​​of the form n×n, where R0 ii R represents the self-thermal resistance of the i-th heat source. ij This represents the coupling thermal resistance between the i-th heat source and the j-th heat source;

[0088] In the heat transfer process of the packaging structure, the thermal resistance mainly includes four parts: one-dimensional conductive thermal resistance, convection thermal resistance, diffusion thermal resistance, and radiation thermal resistance.

[0089] One-dimensional thermal resistance is used to characterize the ability of heat to be transferred along the direction of heat flow in a homogeneous structure, and can be expressed as:

[0090]

[0091] Where t represents the structural thickness, A represents the cross-sectional area of ​​the structure, and k is the thermal conductivity of the material;

[0092] When heat flows through multiple materials in sequence, the thermal resistance of these parts is in series, and the equivalent thermal resistance is:

[0093]

[0094] If the heat flow is split into multiple paths, it is a parallel structure, and the equivalent thermal resistance is:

[0095]

[0096] Convective thermal resistance refers to the thermal resistance that arises when heat exchange occurs between a solid surface and the surrounding fluid due to the limited convective heat transfer capacity. It can be expressed as:

[0097]

[0098] Where h is the surface convective heat transfer coefficient;

[0099] Diffusion thermal resistance refers to the additional resistance generated when heat diffuses from a small area to a large area. In this embodiment, perfect interface contact is assumed, and the calculation method is as follows:

[0100]

[0101]

[0102]

[0103]

[0104]

[0105]

[0106]

[0107]

[0108] Where c and d represent the length and width of the heat source, respectively; Xc and Yc are the horizontal and vertical coordinates of the center of the heat source; a and b represent the length and width of the heat diffusion region, respectively; k1 is the thermal conductivity of the heat diffusion region; and h is the convective diffusion coefficient of the heat diffusion region.

[0109] For multilayer heat diffusion structures, the following recursive formula should be used for calculation:

[0110]

[0111] k i and ti denoted as thermal conductivity and thickness of the i-th layer, respectively.

[0112] Radiation thermal resistance refers to the equivalent thermal resistance that occurs during energy exchange through thermal radiation due to the limited surface radiation capacity.

[0113] It reflects the relationship between radiative heat transfer capacity and temperature difference, and its heat transfer intensity is proportional to the fourth power of the temperature;

[0114] At typical packaging operating temperatures, radiation thermal resistance is much smaller than conduction and convection thermal resistance, and its impact on overall heat transfer is relatively small, so it can be ignored in modeling.

[0115] When performing equivalent modeling, the planar dimensions of the equivalent structure are first determined;

[0116] In this embodiment, the equivalent structure is a cuboid, and its length and width are consistent with the length and width of the structure adjacent to the heat source.

[0117] In a specific instance, S2 includes the following steps:

[0118] Based on the actual packaging structure, determine the characteristic parameters and their value range;

[0119] The parameter conditions include the structural parameters, material parameters, and boundary conditions of the equivalent structure.

[0120] Taking a typical dual-Chiplet CoWoS packaging structure as an example, in this embodiment, the characteristic parameters are the equivalent structure thickness t, thermal conductivity k, and convective heat transfer coefficient h, as well as the horizontal position coordinates (x1, y1) and (x2, y2) of Chiplet1 and Chiplet2.

[0121] The value range of the feature parameters is shown in Table 1;

[0122] Table 1

[0123]

[0124] In a specific instance, S3 includes the following steps:

[0125] The numerical simulation methods include the finite element method, thermal network modeling, and MATLAB simulation.

[0126] For each set of parameter conditions, a unit power is applied to each chipplet individually, and the steady-state temperature rise of each chipplet is obtained through numerical simulation to determine the self-thermal resistance of each chipplet.

[0127] By sequentially selecting any two Chiplet chips and simultaneously applying a unit power input, the corresponding steady-state temperature change is obtained through numerical simulation.

[0128] Based on the principle of linear superposition of heat transfer, the coupling thermal resistance between each chipplet is solved using the above simulation results, and the complete thermal resistance matrix of the system is constructed.

[0129] In the specific implementation, COMSOL finite element analysis software is used to apply unit power to Chiplet1 and Chiplet2 separately, and then unit power is applied to Chiplet1 and Chiplet2 simultaneously.

[0130] In this embodiment, Figure 2 Taking the CoWos structure shown as an example, its specific parameters are listed in Table 2;

[0131] Table 2

[0132]

[0133] Finite element simulations were performed under different power conditions, and the simulation results are shown in Table 3.

[0134] Table 3

[0135]

[0136] Based on the principle of linear superposition of heat transfer, the thermal resistance matrix of the original structure is obtained as follows:

[0137]

[0138] The thermal resistance matrix of the equivalent structure obtained by the S1 method is:

[0139]

[0140] For each set of input characteristic parameters, the thermal resistance matrix is ​​calculated using finite element analysis.

[0141] In a specific instance, S4 includes the following steps:

[0142] The feature parameters are combined with the corresponding thermal resistance matrix data to form a sample dataset;

[0143] The dataset is preprocessed, including data normalization, noise reduction, or outlier handling, to improve the training accuracy of subsequent machine learning models.

[0144] The dataset is divided into a training set, a validation set, and a test set for model training and performance evaluation.

[0145] This embodiment improves the training accuracy and generalization ability of the machine learning model through data preprocessing (normalization, denoising, etc.) and dataset partitioning, avoiding the model overfitting problem caused by noisy data or outliers; the standardized dataset processing process ensures the stability of model training, significantly improves the repeatability and reliability of prediction results, and solves the defect of traditional machine learning models being sensitive to data quality.

[0146] In a specific example, S5 includes the following steps:

[0147] The preprocessed dataset is trained using a machine learning model;

[0148] A nonlinear mapping relationship is established between the characteristic parameters of the packaging structure and the thermal resistance matrix to obtain a training model that can predict the thermal resistance matrix.

[0149] In a specific example, the machine learning model is a BP neural network. The basic idea of ​​the BP neural network is backpropagation of error. The BP neural network consists of an input layer, a hidden layer, and an output layer.

[0150] The input and output layers are single-layer structures, while the hidden layers can be single-layer or multi-layer, with the number of neurons determined based on the sample complexity and accuracy requirements.

[0151] like Figure 3 As shown, the input layer is a vector of feature parameters, and the output layer is a vector of elements of a 2×2 thermal resistance matrix.

[0152] It should be noted that in temperature prediction, there is a complex nonlinear relationship between the input variables and the thermal resistance matrix, which the BP neural network can effectively capture.

[0153] In addition, the training process of BP neural networks is relatively simple and suitable for our existing computing resources and data volume.

[0154] In a specific example, the BP neural network training steps include:

[0155] (1) Initialize the connection weights and thresholds, and randomly select values ​​in the interval [-1, 1];

[0156] (2) Set the activation function and input the training samples;

[0157] (3) Perform forward propagation and calculate the error between the predicted thermal resistance matrix and the target matrix;

[0158] (4) If the error exceeds the preset threshold, backpropagation is performed, the weights and thresholds are corrected using the gradient descent method, and the process returns to the third step to continue iterating in order to optimize the model’s fitting performance to the thermal resistance matrix until the error meets the requirements or the maximum number of runs is reached.

[0159] This embodiment, through the training process, can obtain a BP neural network model that can accurately characterize the mapping relationship between input feature parameters and thermal resistance matrix.

[0160] In a specific instance, S6 includes the following steps:

[0161] The actual CoWoS packaging structure is equivalentized according to the S1 method;

[0162] The equivalent structural features are input into the trained machine learning model to predict the corresponding thermal resistance matrix.

[0163] Based on the thermal resistance matrix, the maximum junction temperature of each chiplet is calculated in conjunction with the chip power consumption conditions.

[0164] This embodiment inputs the actual structure into the model after it has been equivalently processed, realizing the transformation from model universality to practical application. It calculates the maximum junction temperature by combining chip power consumption, directly addressing actual engineering needs (packaging thermal design, reliability assessment). The prediction results can directly guide the optimization of the packaging structure (such as adjusting the chiplet position and selecting high thermal conductivity materials), and have extremely strong engineering practical value.

[0165] Example 2

[0166] To achieve the above objectives, see Figure 4 This embodiment also discloses a machine learning-based chiplet maximum temperature prediction system in CoWoS packaging, including:

[0167] The equivalent model construction module is used to establish an equivalent model that matches the actual thermal characteristics of the CoWoS package structure.

[0168] The parameter setting module is used to set different characteristic parameter conditions in the equivalent model to represent various working conditions;

[0169] The thermal resistance matrix acquisition module is used to perform steady-state thermal simulation on equivalent models under different parameter conditions through numerical simulation to obtain the corresponding thermal resistance matrix.

[0170] The dataset processing module is used to combine the feature parameters and the corresponding thermal resistance matrix to form a dataset, and to preprocess the dataset.

[0171] The model training module is used to train the preprocessed dataset using a machine learning model, establish the mapping relationship between feature parameters and thermal resistance matrix, and obtain the trained prediction model.

[0172] The temperature prediction module is used to input the actual CoWoS package structure into the trained prediction model after it has been equivalentized by the equivalent model building module, predict the corresponding thermal resistance matrix, and calculate the maximum temperature of each chiplet in combination with the chip power consumption conditions.

[0173] The machine learning-based chipplet maximum temperature prediction system in CoWoS packaging described in this embodiment has the same advantages over the prior art as the machine learning-based chipplet maximum temperature prediction method in CoWoS packaging, and will not be repeated here.

[0174] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the highest temperature of a chiplet in a CoWoS package based on machine learning, characterized in that, Comprising the following steps: S1. According to the actual CoWoS packaging structure, an equivalent model consistent with the thermal characteristics of the actual structure is established; According to the actual CoWoS packaging structure, the key geometric features and material parameters are extracted; Calculate the equivalent thermal resistance of the packaging structure, so that it matches the thermal resistance of the original structure, thereby ensuring that the maximum temperature of each Chiplet under the equivalent structure and the original structure is approximately consistent; Construct an equation to make the thermal resistance matrices of the original structure and the equivalent structure equal: where R cond,equal , R conv,equal and R s,equal are the convective, conductive and diffusive thermal resistances, respectively; R conv and R cond represent the convective and conductive thermal resistances of the original structure, and R sn represents the diffusive thermal resistance generated from the n to n+1 layer. S2. Set different characteristic parameter conditions in the equivalent model to represent various working conditions; S3. Through numerical simulation, steady-state thermal simulation is performed on the equivalent model under different parameters to obtain the corresponding thermal resistance matrix; The numerical simulation method includes finite element method, thermal network modeling, and MATLAB simulation; For each set of parameter conditions, a unit power is applied to each Chiplet respectively, and the steady-state temperature rise of each Chiplet is obtained through numerical simulation to determine the self-thermal resistance of each Chiplet; Select any two Chiplets in turn and apply a unit power input simultaneously, and obtain the corresponding steady-state temperature change through numerical simulation; According to the linear superposition principle of heat transfer, the coupling thermal resistance between Chiplets is solved based on the simulation results above, and the complete thermal resistance matrix of the system is constructed; S4. Combine the characteristic parameters and the thermal resistance matrix to form a data set, and preprocess the data set; S5. Train the data set using a machine learning model to establish a mapping relationship between the characteristic parameters and the thermal resistance matrix; S6. Input the actual packaging structure into the machine learning model after equivalent processing to predict the thermal resistance matrix and calculate the maximum temperature of each Chiplet.

2. The method of claim 1, wherein, The S2 includes the following steps: According to the actual packaging structure, determine the characteristic parameters and their value ranges; The parameter conditions include the structural parameters, material parameters and boundary conditions of the equivalent structure.

3. The method of claim 1, wherein, The S4 includes the following steps: Combine the characteristic parameters and the corresponding thermal resistance matrix data to form a sample data set; Preprocess the data set, including data normalization, denoising or outlier processing, to improve the training accuracy of the subsequent machine learning model; Divide the data set into training set, validation set and test set for model training and performance evaluation.

4. The method of claim 1, wherein, The S5 includes the following steps: Train the preprocessed data set using a machine learning model; Establish a nonlinear mapping relationship between the packaging structure characteristic parameters and the thermal resistance matrix to obtain a trained model that can predict the thermal resistance matrix.

5. The method of claim 4, wherein, The machine learning model is a BP neural network, and the basic idea of the BP neural network is back propagation of errors. The BP neural network consists of an input layer, a hidden layer and an output layer; The input layer and the output layer are single-layer structures, and the hidden layer can be single-layer or multi-layer. The number of neurons is determined according to the complexity of the sample and the accuracy requirement; The input layer is a characteristic parameter vector, and the output layer is an element vector of a 2x2 thermal resistance matrix.

6. The method of claim 5, wherein, The BP neural network training steps include: (1) Initialize the connection weights and thresholds, and randomly take values in the [-1, 1] interval; (2) Set the activation function and input the training sample; (3) Perform forward propagation to calculate the error between the predicted thermal resistance matrix and the target matrix; (4) If the error exceeds the preset threshold, back propagation is performed, the weight and threshold are corrected by using the gradient descent method, and the third step is returned to continue iteration to optimize the fitting performance of the model to the thermal resistance matrix until the error meets the requirement or the maximum number of operations is reached.

7. The method of claim 1, wherein, The S6 includes the following steps: The actual CoWoS package structure is equivalently processed according to the S1 method; The equivalent structure characteristics are input into the trained machine learning model to predict the corresponding thermal resistance matrix; Based on the thermal resistance matrix, the maximum junction temperature of each Chiplet is calculated under the condition of chip power consumption.

8. A machine learning based Chiplet maximum temperature prediction system in a CoWoS package, the system comprising: It includes: An equivalent model construction module is configured to establish an equivalent model consistent with the thermal characteristics of an actual structure according to an actual CoWoS package structure; According to the actual CoWoS package structure, key geometric features and material parameters are extracted; The equivalent thermal resistance of the package structure is calculated to match the thermal resistance of the original structure, thereby ensuring that the maximum temperatures of each Chiplet under the equivalent structure and the original structure are approximately consistent; An equation is constructed to make the thermal resistance matrices of the original structure and the equivalent structure equal: where R cond,equal , R conv,equal and R s,equal are the convective, conductive and diffusive thermal resistances, respectively; R conv and R cond represent the convective and conductive thermal resistances of the original structure, and R sn represents the diffusive thermal resistance generated from the n th layer to the n+1 th layer. A parameter setting module is configured to set different feature parameter conditions in the equivalent model to represent various working conditions; A thermal resistance matrix acquisition module is configured to perform steady-state thermal simulation on the equivalent model under different parameter conditions through numerical simulation to obtain the corresponding thermal resistance matrix; The numerical simulation method includes finite element method, thermal network modeling, and MATLAB simulation; For each set of parameter conditions, a unit power is applied to each Chiplet respectively, and the steady-state temperature rise of each Chiplet is obtained through numerical simulation to determine the self-thermal resistance of each Chiplet; Any two Chiplets are selected in turn to apply a unit power input simultaneously, and the corresponding steady-state temperature change is obtained through numerical simulation; According to the linear superposition principle of heat transfer, the coupling thermal resistance between each Chiplet is solved through the simulation results above, and the complete thermal resistance matrix of the system is constructed; A data set processing module is configured to form a data set by combining the feature parameters and the corresponding thermal resistance matrix, and to preprocess the data set; A model training module is configured to train the preprocessed data set using a machine learning model to establish a mapping relationship between the feature parameters and the thermal resistance matrix, and to obtain a trained prediction model; A temperature prediction module is configured to input the actual CoWoS package structure equivalently processed by the equivalent model construction module into the trained prediction model to predict the corresponding thermal resistance matrix, and to calculate the maximum temperature of each Chiplet in combination with the chip power consumption condition.

Images