Temperature field prediction method and device of core particle system

Abstract

The invention provides a temperature field prediction method and equipment of a core particle system. The method comprises the following steps: determining a target core particle system to be detected; the target core particle system is subjected to simulation processing through a simulation model in the preset hybrid model, a simulation result is obtained, and the simulation result comprises the temperature field variable of the target core particle system; the simulation result is corrected through a network model obtained through pre-training in the post-processing stage in the preset hybrid model, a correction result is obtained, the correction result comprises the corrected temperature field predicted value, and the network model is obtained through fusion training of simulation data of the simulation model and experimental data under the same condition. Through the method, the simulation result of the traditional simulation model can be further processed and corrected through the network model obtained through pre-training, and a temperature field prediction value which is more practical is obtained, so that hot spot positions are eliminated in advance, the design period of a 3D-IC integrated chip is shortened, and the cost is reduced.

Description

Technical Field

[0001] The present invention relates to the field of semiconductor technology, and particularly to a method and device for predicting the temperature field of a chiplet system. Background Art

[0002] Chiplet systems that stack multiple small chips three-dimensionally have become the mainstream direction of advanced packaging due to their high integration, high bandwidth, and high performance. However, the increase in transistor integration has increased the complexity and power density of advanced packaging structures, reduced the heat dissipation specific surface area, and made the thermal management problem particularly prominent. In addition, problems such as vertical thermal coupling, thermal mismatch of heterogeneous materials, and local hot spots caused by multi-layer stacking significantly affect the performance and reliability of chips. At present, there is still a lack of a fast, accurate, and efficient temperature prediction method for three-dimensional stacked chiplet systems, which restricts the design iteration speed of 3D-IC (3Dimensional-Integrated Circuit), and poses challenges to the design of ultra-large-scale 3D-IC packaged integrated chips. Summary of the Invention

[0003] In view of the above problems, the present invention proposes a method and device for predicting the temperature field of a chiplet system that overcome or at least partially solve the above problems.

[0004] One object of the present invention is to utilize the powerful non-linear fitting ability of the network model to capture factors not considered by traditional simulation models, thereby solving the problem of mismatch between simulation and experiment;

[0005] Another further object of the present invention is to save the simulation computing power for complex numerical models.

[0006] Another further object of the present invention is to shorten the simulation cycle of 3D-IC design for chiplet systems.

[0007] Specifically, the present invention provides a method for predicting the temperature field of a chiplet system, including:

[0008] Determine the target chiplet system to be measured;

[0009] The simulation model in the preset hybrid model performs simulation processing on the target chiplet system to obtain simulation results, and the simulation results include the temperature field variables of the target chiplet system;

[0010] In the post-processing stage of the preset hybrid model, the network model obtained through pre-training corrects the simulation results to obtain correction results, and the correction results include the corrected temperature field prediction values. The network model is obtained by fusing and training the simulation data of the simulation model with the experimental data under the same conditions.

[0011] Optionally, the generation steps of the preset hybrid model include:

[0012] Obtain the simulation data and experimental data for training. The simulation data is the simulation result of the simulation model under the set working condition parameters, and the experimental data is the actual measurement result under the set working condition parameters;

[0013] Train a network model based on the simulation data and experimental data;

[0014] Evaluate the network model;

[0015] When the network model passes the model evaluation, construct a preset hybrid model according to the simulation model and the network model.

[0016] Optionally, the steps of obtaining the simulation data and experimental data for training include:

[0017] Determine the parameter space of the simulation model. The parameter space includes the variation ranges of material parameters, heat source density, and boundary conditions;

[0018] Sample in the parameter space to obtain multiple groups of working condition parameters, where the working condition parameters include material parameters, heat source density, and boundary conditions;

[0019] Use the simulation model to perform simulations according to the working condition parameters to obtain the temperature field variables under different working condition parameters as simulation data;

[0020] Replicate the physical scenarios corresponding to different working condition parameters in the experimental environment, measure and record the surface temperature distribution of the die system under different physical scenarios as experimental data;

[0021] Align the simulation data and experimental data in space and time.

[0022] Optionally, the steps of training a network model based on the simulation data and experimental data include:

[0023] Use the temperature field variables in the simulation data as the first input channel, and map the boundary conditions to a three-dimensional grid as the second input channel;

[0024] Concatenate the other parameters in the working condition parameters except the boundary conditions in vector form to obtain a working condition parameter vector;

[0025] Determine the target network architecture;

[0026] Process the target network architecture in combination with the first input channel, the second input channel, and the working condition parameter vector, and finally output the training prediction value of the network model;

[0027] Calculate the loss value of the current round of prediction of the network model by using a preset loss function for the training prediction value and the experimental data under the same working condition parameters;

[0028] Call a preset optimizer to adjust the parameters of the network model according to the loss value;

[0029] Repeat training until a preset training stop condition is met, and finally obtain the network model.

[0030] Optionally, the steps for determining the target network architecture include:

[0031] When the simulation data and experimental data are results in a steady state, the target network architecture selects a three-dimensional convolutional neural network architecture;

[0032] When the simulation data and experimental data are results in a transient state, the target network architecture selects a spatio-temporal hybrid neural network architecture, which includes a three-dimensional convolutional neural network architecture and a long short-term memory network architecture. The three-dimensional convolutional neural network architecture is used to process spatial features, and the long short-term memory network architecture is used to process time series;

[0033] The steps for concatenating other parameters in the working condition parameters except the boundary conditions in vector form include: when the simulation data and experimental data are results in a transient state, obtain the time step parameters and concatenate the time step parameters with the working condition parameters.

[0034] Optionally, the preset loss function is: Loss = MSE(y pred , y exp );

[0035] Where Loss represents the loss value, MSE represents the mean square error, y pred represents the training prediction value, and y exp represents the actual measurement value in the experimental data.

[0036] Optionally, the steps for evaluating the network model include:

[0037] Obtain an evaluation prediction value by predicting with the trained network model under the set working condition parameters;

[0038] Obtain the experimental result under the same set working condition parameters as the evaluation actual value;

[0039] Calculate the error value between the evaluation prediction value and the evaluation actual value;

[0040] If the error value is less than the preset threshold, determine that the network model passes the model evaluation.

[0041] Optionally, the steps for evaluating the network model include:

[0042] Obtain multiple groups of evaluation prediction values by predicting with the trained network model under different set working condition parameters;

[0043] Draw a comparison curve based on the simulation data, the evaluation prediction value, and the experimental data;

[0044] Determine whether the network model passes the model evaluation through the comparison curve.

[0045] Optionally, after the step of constructing a preset hybrid model based on the simulation model and the network model, the following steps are further included:

[0046] Obtain the simulation data and the experimental data under the same set of operating condition parameters;

[0047] Input the simulation data into the network model;

[0048] In the post-processing stage, correct the simulation data through the network model to obtain a corrected prediction value;

[0049] According to the residual between the prediction value and the measured value in the experimental data, reversely optimize the parameters of the network model.

[0050] According to another aspect of the present invention, there is also provided a computer device, including a memory, a processor, and a machine-executable program stored on the memory and running on the processor, and when the processor executes the machine-executable program, the steps of the temperature field prediction method of any one of the above-mentioned chiplet systems are implemented.

[0051] The temperature field prediction method of the chiplet system of the present invention first determines the target chiplet system to be measured; then, the simulation model in the preset hybrid model performs a simulation process on the target chiplet system to obtain a simulation result, and the simulation result includes the temperature field variables of the target chiplet system; then, in the post-processing stage of the preset hybrid model, the simulation result is corrected through a pre-trained network model to obtain a correction result, and the correction result includes a corrected temperature field prediction value, wherein the network model is obtained by fusing and training the simulation data of the simulation model and the experimental data under the same conditions. Through this method, the simulation result of the traditional simulation model can be further processed and corrected by the pre-trained network model to obtain a more practical temperature field prediction value, thereby eliminating the hot spot positions in advance, reducing the design cycle of 3D-IC integrated chips, and reducing costs.

[0052] Based on the following detailed description of the specific embodiments of the present invention in conjunction with the accompanying drawings, those skilled in the art will more clearly understand the above and other objects, advantages, and features of the present invention. Description of the Drawings

[0053] Some specific embodiments of the present invention will be described in detail hereinafter with reference to the accompanying drawings in an exemplary but non-limiting manner. The same reference numerals in the drawings denote the same or similar components or parts. Those skilled in the art should understand that these drawings are not necessarily drawn to scale. In the drawings:

[0054] Figure 1 is a schematic flow chart of a method for predicting the temperature field of a chiplet system according to an embodiment of the present invention;

[0055] Figure 2 is a schematic flow chart of the training of a network model according to an embodiment of the present invention;

[0056] Figure 3 is a schematic diagram of a computer program product according to an embodiment of the present invention;

[0057] Figure 4 is a schematic diagram of a computer-readable storage medium according to an embodiment of the present invention; and

[0058] Figure 5 is a schematic diagram of a computer device according to an embodiment of the present invention. Detailed implementation manners

[0059] Those skilled in the art should understand that the embodiments described below are only a part of the embodiments of the present invention, rather than all the embodiments of the present invention. This part of the embodiments is intended to explain the technical principles of the present invention, rather than to limit the protection scope of the present invention. Based on the embodiments provided by the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts should still fall within the protection scope of the present invention.

[0060] It should be noted that the logic and / or steps represented in the flow chart or described in other ways herein, for example, can be considered as a definite sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch instructions from the instruction execution system, apparatus, or device and execute the instructions), or in combination with these instruction execution systems, apparatus, or devices.

[0061] In the current semiconductor technology field, the following methods are generally used for temperature prediction of three-dimensional stacked chiplet systems:

[0062] 1. Traditional thermal simulation methods (such as finite element analysis, finite difference method). Although this method can provide high-precision temperature field prediction, its calculation cost is high and the time consumption is long, making it difficult to meet the requirements of real-time thermal management.

[0063] 2. Equivalent thermal model method. This method simplifies each structure of the chip into a thermal resistance model. Compared with a complete thermal simulation model, it has the advantage of fast solution speed, but the accuracy of its prediction results highly depends on the quality of constructing the equivalent structure.

[0064] 3. Although the pure data-driven machine learning method has a fast calculation speed, it lacks physical consistency and has a strong dependence on experimental data, making it difficult to generalize to complex and changeable 3D-IC scenarios.

[0065] Based on the above defects and the mismatch between the thermal simulation and experimental test data of the chiplet system, the present invention provides a real-time temperature field prediction method for the chiplet system based on measured values and physical constraints. By using the power map of the chiplet system and the measured temperature on the chip surface, a hybrid model combining a simulation model and a neural network (i.e., a pre-trained network model) is constructed to predict the internal temperature distribution of the chiplet system, eliminate hot spots in advance, reduce the design cycle of 3D-IC integrated chips, and reduce costs. Specifically, Figure 1 is a schematic flowchart of the temperature field prediction method for the chiplet system according to an embodiment of the present invention, as Figure 1 shown, the temperature field prediction method for the chiplet system at least includes the following steps S101 to step S103.

[0066] Step S101, determine the target chiplet system to be measured.

[0067] Step S102, perform simulation processing on the target chiplet system by the simulation model in the preset hybrid model to obtain a simulation result. The simulation result includes the temperature field variables of the target chiplet system. The temperature field variables are generally physical quantities used to describe the temperature distribution and its changes in a physical system. The calculation method of the simulation model can generally be any one of methods such as the Finite Element Method (FEM), the Finite Volume Method (FVM), and the Finite Difference Method (FDM). Those skilled in the art can select the corresponding calculation method of the simulation model according to the actual situation.

[0068] Step S103, perform correction processing on the simulation result through the pre-trained network model in the post-processing stage of the preset hybrid model to obtain a correction result. The correction result includes the corrected temperature field prediction value. The network model is obtained by fusing and training the simulation data of the simulation model and the experimental data under the same conditions. Since the simulation model usually has factors that are not considered, which will lead to an error between the simulation result and the actual temperature of the chiplet system, the method of the present invention selects to fuse and train the network model with the simulation data of the simulation model and the experimental data under the same conditions. The network model trained in this way can capture the factors not considered by the simulation model, thereby correcting the simulation result to make it closer to the real data. In this way, the preset hybrid model obtained by combining the simulation model and the network model can effectively solve the problem of mismatch between simulation and experiment, and can also save a large amount of simulation computing power for complex numerical models.

[0069] Among them, the temperature field is usually used to reflect the internal temperature of the die system, and the temperature values at each spatial point constitute the internal temperature distribution of the die system. The hot spot position is a local position in the temperature field where the temperature is significantly higher than the surrounding area and is an abnormal point in the temperature field. Through the research and analysis of the temperature field, the internal temperature change trend and hot spot distribution of the die system can be clearly understood, providing key information for chip thermal management and design optimization.

[0070] In the hybrid model of the present invention, the simulation model can simulate the heat conduction process of the die system based on material parameters, heat source density, and boundary conditions, and generate internal temperature field data. These data follow the physical laws of heat conduction and reflect the temperature distribution laws under different working conditions. Subsequently, the simulation results obtained from the simulation model are corrected by the trained network model. Since the network model is trained by combining real experimental data and simulation data, such a network model can consider more real factors, thereby correcting the simulation results obtained from the simulation model into a more accurate prediction result closer to the actual situation.

[0071] After the hybrid model accurately predicts the internal temperature distribution of the die system, potential hot spot positions can be identified based on the temperature threshold. Because hot spots are usually areas with extremely high temperatures, when the model predicts that the temperature in a certain area far exceeds the normal range, it can be judged that this area may become a hot spot. For example, in the design stage of 3D-IC chips, through the prediction of the hybrid model of the present invention, it is found that the temperature at the stack of two dies is too high, indicating that a hot spot may form here, thus providing a basis for design optimization to avoid the generation of hot spots. By determining potential hot spot positions in advance, designers can take targeted optimization measures in the design stage of 3D-IC integrated chips, such as adjusting the chip layout and improving the heat dissipation structure, so as to avoid the appearance of hot spots in actual use and improve the performance and reliability of the chips. For example, by changing the heat source distribution or adding heat dissipation materials, the temperature in the potential hot spot area can be reduced to eliminate the hidden danger of hot spots.

[0072] Therefore, through the method of the present invention, the simulation results of the traditional simulation model can be further processed and corrected by the pre-trained network model to obtain a more accurate temperature field prediction value closer to the actual situation, thereby eliminating hot spot positions in advance, reducing the design cycle of 3D-IC integrated chips, and lowering costs.

[0073] In addition, the hybrid model of the present invention can also solve problems such as vertical thermal coupling and heterogeneous material thermal mismatch caused by multi-layer stacking.

[0074] Among them, the vertical thermal coupling problem caused by multi-layer stacking refers to the phenomenon that in 3D-IC, multiple chips or chip layers are stacked to form a multi-layer structure, and heat is transferred vertically between these stacked chip layers and affects each other. Because the chip layers are closely stacked, heat will be transferred from the chip layer with a higher temperature to the adjacent chip layer with a lower temperature. This heat transfer is not isolated, but is mutually coupled. The temperature change of one chip layer will affect the temperature distribution of the adjacent chip layer. For example, when the power components of a certain chip layer generate heat, the generated heat will be vertically transferred to other layers through the medium between the chips (such as interconnect materials, heat dissipation layers, etc.), causing the temperature of other layers to change accordingly. This vertical thermal coupling phenomenon will complicate the thermal management of the chip, and may cause uneven temperature distribution of the entire chip system, affecting chip performance and reliability.

[0075] To address the vertical thermal coupling problem caused by multi-layer stacking, traditional simulation models can model the thermophysical properties of different layers of material. Although this can calculate the vertical heat transfer, including heat diffusion and convergence, vertical thermal coupling involves the interaction of multiple physical factors and has strong nonlinear characteristics, making it difficult for traditional simulation models to accurately describe it. The present invention trains a network model using real experimental data and simulation data, allowing the network model to automatically capture factors not considered in the simulation model and further correct the simulation results to predict the temperature distribution of each layer in the vertical direction that is closer to reality, thereby effectively solving the vertical thermal coupling problem.

[0076] Thermal mismatch of heterogeneous materials is caused by the fact that a variety of different materials are used in 3D-ICs. These materials have different thermophysical properties (such as thermal conductivity, thermal expansion coefficient, etc.). Due to these differences in thermophysical properties, when the temperature changes, the expansion or contraction of different materials is inconsistent, which causes thermal stress and strain at the material interface. For example, the silicon-based material in the chip and the epoxy resin material used for packaging have different thermal expansion coefficients. When the operating temperature of the chip rises, the silicon-based chip and the epoxy resin packaging material expand to different degrees, and thermal stress will be generated at their interface. Long-term thermal stress may lead to a decrease in the bonding strength between materials, resulting in problems such as delamination and cracking, affecting the electrical performance and reliability of the chip. It will also change the heat conduction path inside the chip, thereby affecting the temperature distribution of the chip and causing thermal management problems such as local hot spots.

[0077] Regarding the problem of thermal mismatch in heterogeneous materials, traditional simulation models generally take into account the differences in thermophysical properties of heterogeneous materials, such as thermal conductivity, coefficient of thermal expansion, etc. By modeling the characteristics of different materials, the deformation and thermal stress distribution of heterogeneous materials under thermal action are simulated. However, due to the complexity of the actual situation, there may be certain errors in the simulation model. The method of the present invention can compensate for these errors through a network model trained with experimental data and simulation data. It can capture the subtle characteristics of the temperature distribution changes caused by the thermal mismatch of heterogeneous materials, thereby correcting the simulation results, making the prediction results closer to the actual situation, and further improving the accuracy of temperature prediction and alleviating the problems caused by the thermal mismatch of heterogeneous materials.

[0078] In some optional embodiments, the generation steps of the preset hybrid model generally may include: obtaining the simulation data and experimental data used for training, where the simulation data is the simulation result of the simulation model under the set working condition parameters, and the experimental data is the actual measurement result under the set working condition parameters; training a network model based on the simulation data and experimental data; evaluating the model of the network model; and constructing the preset hybrid model according to the simulation model and the network model when the network model passes the model evaluation.

[0079] Optionally, the steps of obtaining the simulation data and experimental data used for training generally may include: determining the parameter space of the simulation model, where the parameter space generally includes the variation ranges of material parameters, heat source density, and boundary conditions; sampling in the parameter space to obtain multiple sets of working condition parameters, and the working condition parameters include material parameters, heat source density, and boundary conditions; using the simulation model to perform simulations according to the working condition parameters to obtain the temperature field variables under different working condition parameters as the simulation data; reproducing the physical scenarios corresponding to different working condition parameters in the experimental environment, measuring and recording the surface temperature distribution of the die system under different physical scenarios as the experimental data; and aligning the simulation data and experimental data in space and time.

[0080] Among them, the material parameters generally include parameters such as the thermal conductivity of the material. The thermal conductivity determines the efficiency of heat conduction in it. The thermal conductivities of different materials vary significantly and have a significant impact on the temperature field distribution.

[0081] The heat source density reflects the heat generation intensity per unit area of the die system. To determine the variation range of the heat source density, it is necessary to refer to the actual power consumption situation and design requirements of the chip to comprehensively simulate the heat generation conditions under different workloads.

[0082] Boundary conditions generally can include information such as heat transfer coefficient, adiabatic condition, heat radiation, etc. Among them, the heat transfer coefficient is a physical quantity that measures the ability of an object to transfer heat to the surrounding medium. The adiabatic boundary condition means that at this boundary, no heat flows into or out of the system, that is, the heat flux density is 0. Heat radiation is generally a way for an object to transfer heat through electromagnetic waves. In the chiplet system, the chip surface will dissipate a certain amount of heat to the surrounding environment in the form of heat radiation. Heat radiation has a more significant impact on the heat transfer of the chiplet system in a high-temperature environment. When the chip operating temperature is relatively high, the proportion of the heat dissipated by heat radiation in the total dissipated heat will increase. When determining the boundary conditions, it is necessary to consider the factor of heat radiation and reasonably set parameters such as emissivity according to the actual situation to accurately simulate the heat transfer process of the chiplet system under different working conditions.

[0083] To comprehensively and efficiently cover the parameter space, an optional example of the method of the present invention for sampling in the parameter space is to adopt the Latin hypercube sampling method. Latin hypercube sampling can reduce the number of samples while ensuring the uniform distribution of the samples. Those skilled in the art can select different sampling methods according to the actual situation, and finally generate multiple groups of parameter combinations to cover the simulations of multiple working conditions.

[0084] After obtaining multiple groups of working condition parameters by sampling, the simulation model can perform simulations according to the working condition parameters to obtain the temperature field variables under different working condition parameters as simulation data. At the same time, the method of the present invention selects to reproduce the physical scenarios corresponding to different working condition parameters in the experimental environment, that is, to ensure that the ranges of parameters such as material parameters, loads, and boundary conditions are within the working conditions of the simulation model; then measure and record the surface temperature distribution of the chiplet system under different physical scenarios as experimental data. Among them, the method of measuring the surface temperature generally can choose to use an infrared imaging thermometer to record the surface temperature distribution under specific working conditions. Those skilled in the art can select appropriate measurement methods according to the actual situation, such as the scanning thermal probe method, the thermocouple method, the thermal resistance network method, etc.

[0085] After obtaining the simulation data and experimental data, due to the limitation of the spatio-temporal resolution of the measured data, it is necessary to align the simulation data and experimental data spatiotemporally. Specifically, in terms of spatial alignment, if the FEM method is used before, the temperature field variables of the FEM grid nodes can be interpolated to the experimental points; for dynamic problems, resampling can be used to unify the time steps of the simulation and the experiment; finally, standardize the input data and output labels, and an optional method for standardization is, for example, Z-score normalization. Those skilled in the art can select the methods used for spatio-temporal alignment according to the actual situation.

[0086] In some alternative embodiments, the steps of training a network model based on simulation data and experimental data generally may include: using the temperature field variables in the simulation data as the first input channel, and mapping the boundary conditions into a three-dimensional grid as the second input channel; concatenating the other parameters in the operating conditions parameters except the boundary conditions in vector form to obtain an operating conditions parameter vector; determining the target network architecture; processing the target network architecture in combination with the first input channel, the second input channel, and the operating conditions parameter vector, and finally outputting the training prediction value of the network model; calculating the loss value of the current round of prediction of the network model by using a preset loss function for the training prediction value and the experimental data under the same operating conditions parameters; calling a preset optimizer to adjust the parameters of the network model according to the loss value; repeating the training until the preset training stop condition is satisfied, and finally obtaining the network model. Some examples of the preset training stop condition are: reaching the maximum number of iterations or the loss function converges, etc., and those skilled in the art can set it according to actual needs. The first input channel includes the temperature distribution information at different positions inside the system by inputting the temperature field variables in the simulation data, enabling the model to learn the spatial distribution characteristics of the temperature. The second input channel is the data obtained by mapping the boundary conditions into a three-dimensional grid, reflecting the influence of the boundary conditions on the temperature field. These boundary conditions include heat transfer coefficient, adiabatic, thermal radiation, etc. By mapping to the three-dimensional grid, the network model can capture the relationship between the spatial distribution of the boundary conditions and the temperature field. And the other parameters (such as thermal conductivity, load, etc.) in the operating conditions parameters except the boundary conditions are concatenated in vector form to obtain an operating conditions parameter vector. This vector provides more information related to the system working environment and physical characteristics for the model, helping the model comprehensively consider the influence of various factors on the temperature field. The input channel data and the vector data will be integrated at the input end of the network model and used together as the input of the model, providing a basis for subsequent operations.

[0087] After the integrated input data enters the network model, it first undergoes operations in each layer of the target network architecture. For example, in the 3D U-Net architecture, the data first passes through the convolutional layers of the encoder, and the convolutional kernels slide on the input data for convolution operations to extract features; the pooling layer then downsamples the feature maps to reduce the data dimension. After multiple layers of convolution and pooling, the data enters the decoder, and the decoder restores the size of the feature maps step by step through deconvolution and upsampling operations, and integrates the extracted feature information. In this process, the data performs multiplication operations with the weights, adds the biases, and then passes through the activation function for processing, continuously transmitting and transforming, and finally reaches the output layer.

[0088] At the output layer, after the operations and processing of the previous layers, the model will obtain an output result related to the temperature field, which is the training prediction value of the temperature field of the die system. For the temperature field prediction task, the output result may be a three-dimensional array representing the predicted temperature values at different spatial positions, or, when considering the time factor, a four-dimensional array containing the predicted temperature values at different time steps and different spatial positions.

[0089] Subsequently, based on the training prediction value and combined with the preset loss function, the gap between the current network model and the real experimental data can be determined, that is, the loss value of this round of prediction. Then, the preset optimizer can be called to further adjust the parameters of the network model according to the loss value to make its prediction result closer to the actual situation.

[0090] Optionally, the steps to determine the target network architecture generally may include: when the simulation data and the experimental data are the results in the steady state, the target network architecture can select a three-dimensional convolutional neural network architecture; when the simulation data and the experimental data are the results in the transient state, the target network architecture can select a spatio-temporal hybrid neural network architecture, which includes a three-dimensional convolutional neural network architecture and a long short-term memory network architecture. The three-dimensional convolutional neural network architecture is used to process spatial features, and the long short-term memory network architecture is used to process time series. Among them, the steady state is a thermal equilibrium state where the temperature field does not change with time. For example, the chip continuously operates at a constant power and the temperature reaches a stable distribution. The transient state is a non-equilibrium state where the temperature field changes dynamically with time, such as the chip startup / shutdown, power sudden change, ambient temperature fluctuation, etc.

[0091] For the steady state, since the steady-state temperature field does not change with time, the spatial features are mainly concerned. Therefore, the method of the present invention selects to use a three-dimensional convolutional neural network architecture to process such spatial features. An optional example of the three-dimensional convolutional neural network architecture includes 3D U-Net network architecture, 3D CNN convolutional neural network architecture, GNN graph neural network architecture, and PINN physics-informed neural network architecture, etc. Through such a three-dimensional convolutional neural network architecture, accurate prediction of the steady-state temperature field can be achieved.

[0092] For the transient state, optionally, the step of splicing other parameters except the boundary conditions in the working condition parameters in vector form generally may also include: when the simulation data and the experimental data are the results in the transient state, obtaining the time step parameters and splicing the time step parameters with the working condition parameters.

[0093] Due to the dynamic change of the temperature field over time in transient problems and the time-dependent characteristics such as thermal response delay, the method of the present invention selects to add a long short-term memory network architecture on the basis of a three-dimensional convolutional neural network architecture. The three-dimensional convolutional neural network architecture processes spatial features, and the long short-term memory network architecture processes time series. The long short-term memory network architecture generally refers to the LSTM (Long Short-Term Memory) hybrid network model architecture. In this way, LSTM learns the time step parameters and combines the spatial features of each time step extracted by the 3D U-Net to predict the temperature field of future time steps. For example, during the processes of chip startup, shutdown, or sudden power change, the changing trend of temperature over time can be accurately captured.

[0094] In some alternative embodiments, the preset loss function is generally: Loss = MSE(y pred , y exp );

[0095] where Loss represents the loss value, MSE represents the mean square error, y pred represents the training prediction value, and y exp represents the actual measurement value in the experimental data. The loss value reflects the prediction performance of the network model under the current parameter settings. The smaller the loss value, the closer the prediction of the model is to the actual value.

[0096] The steps of calling the preset optimizer to adjust the parameters of the network model according to the loss value generally may include: calculating the gradient of the loss function with respect to each parameter of the model. The gradient represents the rate of change of the loss function with respect to the parameter change. By updating the parameters of the model along the opposite direction of the gradient, the loss function can be gradually reduced. An alternative example of the preset optimizer is: the Adam optimizer. The Adam optimizer is an optimization algorithm with an adaptive learning rate, which combines the advantages of momentum and adaptive learning rate adjustment. In the training of the temperature field prediction model of the chiplet system, due to the large number of model parameters and complex data, an optimization algorithm that can efficiently adjust parameters and quickly converge is required. The Adam optimizer can adaptively adjust the learning rate according to the gradient of each parameter, can quickly decline in the initial stage of training to accelerate the convergence speed, and can avoid the learning rate being too large and missing the optimal solution in the later stage of training, which is suitable for such complex training scenarios.

[0097] The initial learning rate of the Adam optimizer can generally be set to 1.0e-3, and then dynamically adjusted. It should be noted that the setting of the initial learning rate is an empirical value. At the beginning of training, it determines the step size of each update of the model parameters. If the initial learning rate is too large, the model may skip the optimal solution during training, resulting in non-convergence; if the initial learning rate is too small, the training speed will be very slow, requiring more training time and computing resources. Therefore, those skilled in the art can select a specific optimizer according to the actual situation and set its corresponding initial learning rate.

[0098] In addition, to prevent the preset optimizer from overfitting, the method of the present invention can also use early stopping to prevent the preset optimizer from overfitting. Early stopping is a technique used in machine learning and deep learning to prevent model overfitting and improve model generalization ability. During the model training process, as the number of training rounds increases, the loss of the model on the training set usually gradually decreases, while the loss on the validation set first decreases and then increases. This is because the model gradually remembers the characteristics of the training data during training. When the model overfits the training data, its performance on new, unseen data (validation set) will deteriorate. Early stopping monitors the performance metrics (such as loss value, accuracy, etc.) on the validation set, and when the metrics no longer improve or start to deteriorate, training is stopped in advance to avoid model overfitting.

[0099] In some alternative embodiments, the steps of model evaluation for the network model generally may include: obtaining an evaluation prediction value by predicting with the trained network model under set operating condition parameters; obtaining the experimental results under the same set operating condition parameters as the evaluation actual value; calculating the error value between the evaluation prediction value and the evaluation actual value; and determining that the network model passes the model evaluation when the error value is less than a preset threshold. One optional example of the error value is the Mean Absolute Error (MAE) and the Root Mean Squared Error (RMSE). The Mean Absolute Error is the average of the absolute errors between the predicted value and the true value, which can directly reflect the average deviation degree between the predicted value and the true value. The Root Mean Squared Error is the square root of the Mean Squared Error (MSE). The MSE is the average of the squares of the errors between the predicted value and the true value. By squaring the error, the RMSE amplifies larger errors, making the model more sensitive to larger prediction errors.

[0100] Optionally, an example of the preset threshold can be 1K (K is the Kelvin, a temperature unit). If the values of the Mean Absolute Error and the Root Mean Squared Error are less than 1K (K is the Kelvin, a temperature unit), it is generally considered to have good performance. Those skilled in the art can determine the specific value of the preset threshold according to actual requirements.

[0101] In some further alternative embodiments, the step of model evaluation of the network model may further include: obtaining multiple groups of evaluation prediction values by making predictions with the trained network model under different set working condition parameters; plotting a comparison curve based on the simulation data, the evaluation prediction values, and the experimental data; and determining whether the network model passes the model evaluation through the comparison curve. By plotting the comparison curve, it can be visually seen whether the prediction result of the network model is reasonable, thereby helping to determine the performance of the network model.

[0102] If the network model passes the evaluation, it can enter the next stage of model deployment; otherwise, it is necessary to adjust the input feature design (i.e., the selection of input channels and the parameters participating in vector splicing), the selection of network architecture, and the selection of training parameters in the model training stage. Some possible ways include: increasing the simulation data set and expanding the sampling range of working condition parameters; improving the resolution of spatial sampling of the neural network model; controlling the convergence criterion of the loss function, etc. Those skilled in the art can determine the corresponding adjustment methods according to the actual situation.

[0103] In some alternative embodiments, after the step of constructing a preset hybrid model based on the simulation model and the network model, it generally further includes: obtaining the simulation data and the experimental data under the same set working condition parameters; inputting the simulation data into the network model; performing correction processing on the simulation data through the network model in the post-processing stage to obtain the corrected prediction values; and reversely optimizing the parameters of the network model according to the residual between the prediction values and the measured values in the experimental data. An optional example of reverse optimization is: according to the residual between the prediction values and the measured values in the experimental data, updating the thermal conductivity by the gradient descent method, so that the preset hybrid model is closer to the actual situation; and then through iterative processes such as simulation model data acquisition and network model update, obtaining a more accurate hybrid model to form a closed-loop optimization.

[0104] Figure 2 It is a schematic diagram of the training process of the network model according to an embodiment of the present invention. As Figure 2 shown, the training process of the network model at least includes the following steps S210 to step S240.

[0105] Step S210, data preparation. Through this step, simulation data and experimental data to be generated in the data preparation stage are generated, and then these two types of data are aligned and preprocessed.

[0106] Optionally, step S210 generally includes:

[0107] Sub-step S211: Generate simulation data. To generate simulation data, it is first necessary to define the design parameter space of the simulation model, that is, to define the range of variation of material parameters (such as thermal conductivity), heat source density, and boundary conditions (heat transfer coefficient, adiabatic, thermal radiation, etc.). And for the parameter space, an appropriate method usually needs to be selected for parameter space sampling. An optional example is to use the Latin hypercube sampling method to generate multiple groups of parameter combinations to cover simulations of multiple working conditions, so as to obtain multiple groups of working condition parameters. Finally, the temperature field variables are obtained through simulation by the simulation model, and the temperature field variables and working condition parameters obtained from the simulation are saved for subsequent data training.

[0108] The calculation method of the simulation model can generally be any one of methods such as the Finite Element Method (FEM), the Finite Volume Method (FVM), and the Finite Difference Method (FDM). Those skilled in the art can select the corresponding calculation method of the simulation model according to the actual situation.

[0109] Sub-step S212: Collect experimental data. Try to reproduce in the experimental environment according to the simulation conditions, that is, ensure that the ranges of materials, loads, and boundary conditions are within the working conditions of the simulation model. Subsequently, measure and record the surface temperature distribution of the die system under different physical scenarios as experimental data. The method of measuring the surface temperature can generally choose to use an infrared thermal imaging thermometer to record the surface temperature distribution under specific working conditions. Those skilled in the art can select a suitable measurement method according to the actual situation, such as the scanning thermal probe method, the thermocouple method, the thermal resistance network method, etc.

[0110] Sub-step S213: Data alignment and preprocessing. After obtaining the simulation data and experimental data, due to the limitations of the spatio-temporal resolution of the measured data, it is necessary to align the simulation data and experimental data spatio-temporally. Specifically, in terms of spatial alignment, if the FEM method was used before, the temperature field variables of the FEM grid nodes can be interpolated to the experimental points; for dynamic problems, resampling can be used to unify the time steps of the simulation and the experiment; finally, standardize the input data and output labels. An optional method for standardization is, for example, Z-score normalization. Those skilled in the art can select the methods used for spatio-temporal alignment according to the actual situation.

[0111] Step S220: Network model training. In this step, it is first necessary to determine the input feature design, select a specific network architecture, and then conduct the actual network model training and verification.

[0112] Therefore, step S220 generally includes:

[0113] Sub-step S221, input feature design. First, the temperature field variables calculated by the simulation model are used as an input channel, and the boundary conditions can be mapped into the three-dimensional grid as an input channel; while the operating conditions parameters, such as thermal conductivity, load, etc., are concatenated in vector form; for dynamic problems evolving over time, the additional time step parameters are placed together with the operating conditions parameters.

[0114] Sub-step S222, network architecture selection. In this method, for steady-state problems, the network architecture generally can choose a three-dimensional convolutional neural network architecture; for transient problems, generally can choose a three-dimensional convolutional neural network architecture combined with a long short-term memory network architecture, where the three-dimensional convolutional neural network architecture processes spatial features and the long short-term memory network architecture processes time series.

[0115] Among them, the steady state is a thermal equilibrium state where the temperature field does not change with time. For example, when the chip operates continuously at a constant power, the temperature reaches a stable distribution. The transient state is a non-equilibrium state where the temperature field changes dynamically with time, such as when the chip starts / shuts down, the power changes suddenly, or the ambient temperature fluctuates.

[0116] For the steady state, since the steady-state temperature field does not change with time, the spatial features are mainly concerned. Therefore, the method of the present invention selects to use a three-dimensional convolutional neural network architecture to process such spatial features. An optional example of the three-dimensional convolutional neural network architecture is the 3D U-Net network architecture, the 3D CNN convolutional neural network architecture, the GNN graph neural network architecture, and the PINN physics-informed neural network architecture, etc. Through such a three-dimensional convolutional neural network architecture, accurate prediction of the steady-state temperature field can be achieved.

[0117] For the transient state, optionally, the step of concatenating other parameters in the operating conditions parameters except the boundary conditions in vector form generally can also include: when the simulation data and experimental data are the results in the transient state, obtaining the time step parameters and concatenating the time step parameters with the operating conditions parameters.

[0118] Since the temperature field changes dynamically with time in transient problems and there are time-dependent characteristics such as thermal response delay, the method of the present invention selects to add a long short-term memory network architecture on the basis of the three-dimensional convolutional neural network architecture, and the three-dimensional convolutional neural network architecture processes spatial features and the long short-term memory network architecture processes time series, where the long short-term memory network architecture usually refers to the LSTM (Long Short-Term Memory) hybrid network model architecture. In this way, LSTM learns the time step parameters and combines the spatial features of each time step extracted by 3D U-Net to predict the temperature field of future time steps. For example, during the chip startup, shutdown or sudden power change process, accurately capture the change trend of temperature over time.

[0119] Sub-step S223: conduct training and validation. This step generally may include: calculating the loss value predicted by the network model in this round through a preset loss function for the training prediction value and the experimental data under the same operating condition parameters; calling a preset optimizer to adjust the parameters of the network model according to the loss value; repeating the training until the preset training stop condition is met, and finally obtaining the network model.

[0120] In some optional embodiments, the preset loss function is generally: Loss = MSE(y pred , y exp );

[0121] where Loss represents the loss value, MSE represents the mean square error, y pred represents the training prediction value, and y exp represents the actual measurement value in the experimental data. It should be noted that pred (predicted value) and exp (experimental value) are only exemplary symbols and have no special meaning. Those skilled in the art can select specific symbols to represent the training prediction value and the actual measurement value according to the actual situation. The loss value reflects the prediction performance of the network model under the current parameter settings. The smaller the loss value, the closer the prediction of the model is to the actual value.

[0122] An optional example of the preset optimizer is, for example, the Adam optimizer. The Adam optimizer is an optimization algorithm with an adaptive learning rate, which combines the advantages of momentum and adaptive learning rate adjustment. In the training of the dielet system temperature field prediction model, due to the large number of model parameters and complex data, an optimization algorithm that can efficiently adjust parameters and quickly converge is required. The Adam optimizer can adaptively adjust the learning rate according to the gradient of each parameter, and can quickly decline in the initial stage of training to accelerate the convergence speed; in the later stage of training, it can avoid the learning rate being too large and causing the optimal solution to be missed, and is suitable for such complex training scenarios.

[0123] The initial learning rate of the Adam optimizer can generally be set to 1.0e-3, and then dynamically adjusted. It should be noted that the setting of the initial learning rate is an empirical value. At the beginning of training, it determines the step size of each update of the model parameters. If the initial learning rate is too large, the model may skip the optimal solution during training and fail to converge; if the initial learning rate is too small, the training speed will be very slow and more training time and computing resources are required. Therefore, those skilled in the art can select a specific optimizer according to the actual situation and set its corresponding initial learning rate.

[0124] In addition, to prevent the preset optimizer from overfitting, the method of the present invention can also use early stopping to prevent the preset optimizer from overfitting. Early stopping is a technique used in machine learning and deep learning to prevent model overfitting and improve the generalization ability of the model. During the model training process, as the number of training rounds increases, the loss of the model on the training set usually gradually decreases, while the loss on the validation set first decreases and then increases. This is because the model gradually memorizes the features of the training data during training. When the model overfits the training data, its performance on new, unseen data (validation set) will deteriorate. Early stopping monitors performance metrics (such as loss value, accuracy, etc.) on the validation set and stops training early when the metrics no longer improve or start to deteriorate, thus avoiding model overfitting.

[0125] Step S230, model evaluation. After the training of the network model stops, model performance evaluation is required. First, calculate the mean absolute error (MAE) and root mean square error (RMSE). For the temperature field prediction problem, if the values of MAE and RMSE are controlled within 1 K (K is Kelvin, the temperature unit), it is considered that the performance is good. At the same time, the comparison curves of the simulation results, the corrected prediction results of the network model, and the experimental observation results can be plotted to visually compare the temperature distribution of the entire field. If the model evaluation passes, enter the next stage of model deployment; otherwise, the network model needs to be trained and verified again, such as adjusting the input feature design, network architecture, and training parameters in the model training stage. The available methods include: increasing the simulation data set, expanding the sampling range of the operating condition parameters; improving the resolution of the spatial sampling of the U-Net; controlling the convergence criterion of the loss function, etc. Those skilled in the art can determine the corresponding adjustment methods according to the actual situation.

[0126] Step S240, in the case where the judgment in step S230 is yes, perform hybrid model deployment. Optionally, step S240 includes:

[0127] Sub-step S241, embed the network model into the post-processing module of the simulation model. After the network model obtained by training in step S220 is placed after the post-processing module of the simulation model, the network model can correct the simulation data output by the simulation model. That is, when the temperature field result calculated by the simulation model is input into the network model, the corrected prediction value can be output in real time. In this way, a hybrid model can be obtained by combining the simulation model and the network model. The hybrid model can capture factors not considered by the simulation model, thus effectively solving the problem of mismatch between simulation and experiment.

[0128] Sub-step S242, error feedback and closed-loop optimization. Since there are systematic errors between the values calculated by the hybrid model and the measured values, the parameters of the network model can be optimized in reverse by calculating the residuals between the experimental data. For example, the thermal conductivity can be updated by the gradient descent method, so that the hybrid model is closer to the actual situation. Then, through the iterative processes such as simulation model data acquisition and network model update, a more accurate hybrid model can be obtained to form a closed-loop optimization.

[0129] The present invention post-processes the simulation results of the simulation model by training the network model with integrated experimental data, which belongs to the calibration method combining measured data with physical constraints. The advantage of this method is to utilize the powerful non-linear fitting ability of machine learning to capture the factors not considered by the physical model, thus effectively solving the problem of mismatch between simulation and experiment, and can greatly save the simulation computing power for complex numerical models, as well as reduce the cost of experimental testing, and greatly shorten the simulation cycle of the die system 3D-IC design.

[0130] The flowchart provided in this embodiment is not intended to indicate that the operations of the method will be executed in any specific order, or that all operations of the method are included in every case. In addition, the method may include additional operations. Within the scope of the technical concept provided by the method in this embodiment, additional changes can be made to the above method.

[0131] It should be understood that in some embodiments, each part can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system.

[0132] This embodiment also provides a computer program product 10, a computer-readable storage medium 20, and a computer device 30. Figure 3 is a schematic diagram of a computer program product 10 according to an embodiment of the present invention, Figure 4 is a schematic diagram of a computer-readable storage medium 20 according to an embodiment of the present invention, Figure 5 is a schematic diagram of a computer device 30 according to an embodiment of the present invention. The computer program product 10 includes a computer program 11, and when the computer program 11 is executed by a processor 32, it implements the steps of the temperature field prediction method of the die system as described above. The computer-readable storage medium 20 stores the above computer program 11, and when the computer program 11 is executed by a processor 32, it implements the steps of the temperature field prediction method of the die system in any of the above embodiments. The computer device 30 may include a memory 31, a processor 32, and a computer program 11 stored on the memory 31 and running on the processor 32.

[0133] The computer program 11 for performing the operations of the present invention may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, configuration data of an integrated circuit, or source code or object code written in any combination of one or more programming languages and procedural programming languages. The computer program 11 may be executed entirely on the user's computer, partially on the user's computer, executed as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter case, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider via the Internet). In some embodiments, in order to perform aspects of the present invention, an electronic circuit, including, for example, a programmable logic circuit, a field-programmable gate array (FPGA), or a programmable logic array (PLA), may execute computer-readable program instructions by utilizing the state information of the computer-readable program instructions to personalize the electronic circuit.

[0134] For the description of this embodiment, the computer program product 10 is a related product containing the computer program 11. For the description of this embodiment, the computer-readable storage medium 20 is a tangible device capable of retaining and storing the computer program 11, which may be any device that can contain, store, communicate, propagate, or transmit the program 11 for use by or in connection with an instruction execution system, apparatus, or device. More specific examples (non-exhaustive list) of the computer-readable storage medium 20 include the following: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital versatile disc (DVD), memory stick, floppy disk, mechanically encoded device, and any suitable combination of the above.

[0135] The computer device 30 can be, for example, a server, a desktop computer, a laptop computer, a tablet computer, or a smart phone. In some examples, the computer device 30 can be a cloud computing node. The computer device 30 can be described in the general context of computer system executable instructions, such as program modules, executed by a computer system. Generally, program modules can include routines, programs, object programs, components, logic, data structures, etc., which perform specific tasks or implement specific abstract data types. The computer device 30 can be implemented in a distributed cloud computing environment where tasks are executed by remote processing devices linked through a communication network. In a distributed cloud computing environment, program modules can be located on local or remote computing system storage media including storage devices.

[0136] The computer device 30 can include a processor 32 adapted to execute stored instructions and a memory 31 that provides temporary storage space for the operation of the instructions during operation. The processor 32 can be a single-core processor, a multi-core processor, a computing cluster, or any number of other configurations. The memory 31 can include random access memory (RAM), read-only memory, flash memory, or any other suitable storage system.

[0137] The computer device 30 can also include a network adapter / interface and an input / output (I / O) interface. The I / O interface allows data to be input and output with external devices that can be connected to the computer device. The network adapter / interface can provide communication between the computer device and a network, which is generally shown as a communication network.

[0138] At this point, those skilled in the art should recognize that although multiple exemplary embodiments of the present invention have been shown and described in detail herein, many other variations or modifications that conform to the principles of the present invention can still be directly determined or derived from the disclosed content of the present invention without departing from the spirit and scope of the present invention. Therefore, the scope of the present invention should be understood and determined to cover all these other variations or modifications.

Claims

1. A method for predicting the temperature field of a chiplet system, comprising: Determine the target chiplet system to be measured; Perform simulation processing on the target chiplet system by the simulation model in the preset hybrid model to obtain simulation results, where the simulation results include the temperature field variables of the target chiplet system; In the post-processing stage of the preset hybrid model, correct the simulation results through a pre-trained network model to obtain a correction result, where the correction result includes the predicted value of the corrected temperature field, and the network model is obtained by fusing and training the simulation data of the simulation model with the experimental data under the same conditions.

2. The method for predicting the temperature field of a chiplet system according to claim 1, wherein The generation steps of the preset hybrid model include: Obtain the simulation data and experimental data for training. The simulation data is the simulation result of the simulation model under the set working condition parameters, and the experimental data is the actual measurement result under the set working condition parameters; Train the network model according to the simulation data and the experimental data; Evaluate the model of the network model; In the case where the network model passes the model evaluation, construct the preset hybrid model according to the simulation model and the network model.

3. The method for predicting the temperature field of a chiplet system according to claim 2, wherein The step of obtaining the simulation data and experimental data for training includes: Determine the parameter space of the simulation model, where the parameter space includes the variation ranges of material parameters, heat source density, and boundary conditions; Sample in the parameter space to obtain multiple sets of the working condition parameters, where the working condition parameters include material parameters, heat source density, and boundary conditions; The simulation model performs simulations according to the working condition parameters to obtain the temperature field variables under different working condition parameters as the simulation data; Replicate the physical scenarios corresponding to different working condition parameters in the experimental environment, measure and record the surface temperature distribution of the chiplet system under different physical scenarios as the experimental data; Align the simulation data and the experimental data in space and time.

4. The method for predicting the temperature field of a chiplet system according to claim 3, wherein The step of training the network model according to the simulation data and the experimental data includes: Use the temperature field variables as the first input channel, and map the boundary conditions to a three-dimensional grid as the second input channel; Stitch the other parameters in the working condition parameters except the boundary conditions in vector form to obtain a working condition parameter vector; Determine the target network architecture; The target network architecture processes the first input channel, the second input channel, and the working condition parameter vector, and finally outputs the training prediction value of the network model; Calculate the loss value of the current round of prediction of the network model through a preset loss function for the training prediction value and the experimental data under the same working condition parameters; Call a preset optimizer to adjust the parameters of the Network model according to the loss value; Repeat training until the preset training stop condition is met, and finally obtain the network model.

5. The method for predicting the temperature field of a chiplet system according to claim 4, wherein The steps of determining the target network architecture include: When the simulation data and the experimental data are results in a steady state, the target network architecture selects a three-dimensional convolutional neural network architecture; When the simulation data and the experimental data are results in a transient state, the target network architecture selects a spatio-temporal hybrid neural network architecture, which includes the three-dimensional convolutional neural network architecture and a long short-term memory network architecture. The three-dimensional convolutional neural network architecture is used to process spatial features, and the long short-term memory network architecture is used to process time series; The step of splicing other parameters except the boundary conditions in the working condition parameters in vector form includes: when the simulation data and the experimental data are results in a transient state, obtaining time step parameters and splicing the time step parameters with the working condition parameters.

6. The method for predicting the temperature field of the chiplet system according to claim 4, wherein, The preset loss function is: Loss = MSE(y pred , y exp ); wherein the Loss represents the loss value, the MSE represents the mean square error, and the y pred represents the training prediction value, and the y exp represents the actual measured value in the experimental data.

7. The method for predicting the temperature field of the chiplet system according to claim 2, wherein, The step of performing model evaluation on the network model includes: Obtaining an evaluation prediction value through prediction by the trained network model under the set working condition parameters; Obtaining an experimental result under the same set working condition parameters as the evaluation actual value; Calculating an error value between the evaluation prediction value and the evaluation actual value; When the error value is less than a preset threshold, determining that the network model passes the model evaluation.

8. The method for predicting the temperature field of the chiplet system according to claim 2, wherein, The step of performing model evaluation on the network model includes: Obtaining multiple groups of evaluation prediction values through prediction by the trained network model under different set working condition parameters; Drawing a comparison curve according to the simulation data, the evaluation prediction values, and the experimental data; Determining whether the network model passes the model evaluation through the comparison curve.

9. The method for predicting the temperature field of the chiplet system according to claim 2, wherein, After the step of constructing the preset hybrid model according to the simulation model and the network model, the following steps are further included: Obtaining the simulation data and the experimental data under the same set working condition parameters; Inputting the simulation data into the network model; Performing correction processing on the simulation data through the network model in the post-processing stage to obtain a corrected prediction value; Reversely optimizing the parameters of the network model according to the residual between the prediction value and the measured value in the experimental data.

10. A computer device, including a memory, a processor, and a machine-executable program stored on the memory and running on the processor, and when the processor executes the machine-executable program, the steps of the method for predicting the temperature field of the chiplet system according to any one of claims 1 to 9 are implemented.

Images