Integrated core particle thermal performance prediction method driven by physical information neural network

Abstract

The invention discloses an integrated core particle thermal performance prediction method driven by a physical information neural network, and the method comprises the steps: constructing a parameterized geometric model of a to-be-predicted integrated core particle system, and the parameterized geometric model comprises a solid domain geometric model and a fluid domain geometric model; based on the parameterized geometric model, design parameters of the to-be-predicted integrated core particle system are obtained, and the design parameters comprise geometric parameters of a solid domain and geometric parameters of a fluid domain; acquiring a plurality of space coordinates from the parameterized geometric model to obtain a target space coordinate; and inputting the design parameters and the target space coordinates into a trained thermal field physical information neural network model, so that the model outputs temperature field distribution of the to-be-predicted integrated core particle system under the design parameters through forward reasoning, the model is obtained by training based on a trained flow field physical information neural network model, design parameters and a preset thermal field loss function. According to the method, the prediction time can be shortened while the thermal performance prediction accuracy of the integrated core particles is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of electronic chip thermal performance prediction, and particularly relates to a physical information neural network driven integrated chiplet thermal performance prediction method. BACKGROUND

[0002] With the rapid development of integrated circuit technology, chiplet integration technology has become an important technical path in the fields of high-performance computing, artificial intelligence, autonomous driving, etc. due to its advantages in improving chip performance, reducing manufacturing costs and enhancing design flexibility. However, the continuous improvement of chiplet integration degree leads to a significant increase in power density per unit area of the chip, and the heat dissipation problem is increasingly prominent. If the thermal performance of different heat dissipation schemes cannot be quickly and accurately evaluated at the design stage, it will seriously affect the reliability, life and even system safety of the chip. Therefore, developing a method capable of quickly predicting the thermal performance of integrated chiplet systems is of great significance for optimizing heat dissipation design and improving product competitiveness.

[0003] Currently, the thermal performance evaluation of integrated chiplet systems mainly relies on traditional numerical simulation software based on the finite element method or the finite volume method, such as ANSYS Icepak, COMSOL Multiphysics, Simcenter Flotherm, etc. Although these methods have high simulation accuracy, they have obvious limitations when facing multi-parameter and high-degree of freedom heat sink design space: each design change requires re-geometric modeling, meshing and numerical solving, which is time-consuming and labor-intensive, and it is difficult to cover continuous design space; in addition, traditional methods rely heavily on the experience of designers and computing resources, making it difficult to achieve rapid iteration and optimization at the early design stage. In recent years, although data-driven machine learning methods have improved the prediction speed to some extent, they rely on a large amount of labeled data for training, and the prediction results lack physical consistency, resulting in insufficient generalization ability under unknown design conditions.

[0004] Therefore, how to improve the accuracy of integrated chiplet thermal performance prediction while reducing data dependency and ensuring that the prediction results meet the physical laws is a technical problem that needs to be solved at present. SUMMARY

[0005] In order to solve the problem of how to improve the accuracy of integrated chiplet thermal performance prediction while reducing data dependency and ensuring that the prediction results meet the physical laws, the application provides a physical information neural network driven integrated chiplet thermal performance prediction method. The technical problems to be solved by the application are realized by the following technical solutions: The application provides a physical information neural network driven integrated core particle thermal performance prediction method, comprising the following steps: step 1, constructing a parameterized geometric model of an integrated core particle system to be predicted, wherein the parameterized geometric model comprises geometric models of a solid domain and a fluid domain; Step 2, obtaining design parameters of the integrated core particle system to be predicted based on the parameterized geometric model, wherein the design parameters comprise geometric parameters of the solid domain and geometric parameters of the fluid domain; Step 3, collecting a plurality of spatial coordinates from the parameterized geometric model to obtain target spatial coordinates; Step 4, inputting the design parameters and the target spatial coordinates into a trained thermal field physical information neural network model, so that the trained thermal field physical information neural network model outputs a temperature field distribution of the integrated core particle system to be predicted under the design parameters through forward reasoning; The trained thermal field physical information neural network model is obtained based on a trained flow field physical information neural network model, the design parameters and a preset thermal field loss function, the trained flow field physical information neural network model is obtained based on the design parameters and a preset flow field loss function, and the trained flow field physical information neural network model is used to output predicted velocity components and pressure of the fluid domain.

[0006] In an embodiment of the application, the geometric parameters of the solid domain comprise heat dissipation design parameters of the integrated core particle system to be predicted and geometric parameters of a heat sink, and the geometric parameters of the fluid domain comprise flow channel geometric parameters.

[0007] In an embodiment of the application, the training process of the flow field physical information neural network model comprises the following steps: Step 401A, constructing a parameterized geometric model and collecting a plurality of first spatial coordinates from a geometric model of a fluid domain to obtain first spatial coordinate samples; Step 402A, constructing a first fluid domain residual based on a fluid motion control equation and constructing a first boundary residual based on a fluid boundary condition, and determining a weighted sum of the fluid domain residual and the first boundary residual as a flow field loss function; Step 403A, inputting the first spatial coordinate samples and the design parameters into an initial flow field physical information neural network model, so that the initial flow field physical information neural network model outputs predicted velocity components and pressure of the fluid domain through forward propagation; Step 404A, calculating first derivatives of the predicted velocity components and pressure of the fluid domain with respect to the first spatial coordinate samples by using an automatic differentiation technique, and substituting the first derivatives into the flow field loss function to obtain a loss value of the flow field loss function; At step 405A, the network parameters of the initial flow field physical information neural network model are updated by optimizing the flow field loss function until the flow field loss function converges, so as to obtain the trained flow field physical information neural network model.

[0008] In an embodiment of the present application, the expression of the flow field loss function is:

[0009] wherein, , represents the first loss weight in the fluid domain, represents the first residual in the fluid domain, represents the weight matrix and the bias vector of the flow field physical information neural network model, represents the design parameter, represents the number of first spatial coordinate samples collected in the fluid domain, , represents the second loss weight of the fluid domain boundary, represents the first boundary residual, represents the number of first spatial coordinate samples collected on the fluid domain boundary, and respectively represent the first spatial coordinate samples collected from the fluid domain and the first spatial coordinate samples collected from the fluid domain boundary, is the first nonlinear partial differential operator in the fluid domain, is the L2 norm, is the second nonlinear partial differential operator of the fluid domain boundary, is the predicted velocity component and pressure in the fluid domain, is the predicted velocity component and pressure on the fluid domain boundary.

[0010] In an embodiment of the present application, the thermal field physical information neural network model comprises a first thermal field physical information neural network model for predicting the temperature of the solid domain and a second thermal field physical information neural network model for predicting the temperature of the fluid domain; The training process of the thermal field physical information neural network model comprises: At step 401B, a parameterized geometric model is constructed, and a plurality of second spatial coordinates and a plurality of third spatial coordinates are collected from the geometric models of the solid domain and the fluid domain, respectively, to obtain second spatial coordinate samples and third spatial coordinate samples; ​At step 402B, a residual in the solid domain is constructed based on a solid heat conduction equation, a residual in the second fluid domain is constructed based on a fluid energy control equation, a second interface residual is constructed based on an interface temperature continuity condition and a heat flux continuity condition of fluid-structure coupling, a third boundary residual is constructed based on a heat source boundary, a fourth boundary residual is constructed based on a fluid inlet temperature boundary and an adiabatic boundary, and a weighted sum of the residual in the solid domain, the second interface residual and the third boundary residual is determined as a first thermal field physical information neural network model first thermal field loss function, and a weighted sum of the residual in the second fluid domain, the second interface residual and the fourth boundary residual is determined as a second thermal field physical information neural network model second thermal field loss function, wherein the fluid velocity in the fluid energy equation is provided by the trained flow field physical information neural network model output velocity component; At step 403B, the second spatial coordinate sample and the design parameter are input to the initial first thermal field physical information neural network model, so that the initial first thermal field physical information neural network model outputs a predicted temperature field distribution of the solid domain by forward propagation, and the third spatial coordinate sample and the design parameter are input to the initial second thermal field physical information neural network model, so that the initial second thermal field physical information neural network model outputs a predicted temperature field distribution of the fluid domain by forward propagation; At step 404B, the second derivative of the predicted temperature field distribution of the solid domain with respect to the second spatial coordinate sample is calculated using automatic differentiation technology, and the second derivative is substituted into the first thermal field loss function to obtain a loss value of the first thermal field loss function, and the third derivative of the predicted temperature field distribution of the fluid domain with respect to the third spatial coordinate sample is calculated, and the third derivative is substituted into the second thermal field loss function to obtain a loss value of the second thermal field loss function; At step 405B, the network parameters of the initial first thermal field physical information neural network model and the initial second thermal field physical information neural network model are updated by optimizing the first thermal field loss function and the second thermal field loss function to minimize the first thermal field loss function and the second thermal field loss function until the first thermal field loss function and the second thermal field loss function converge, to obtain a trained flow field physical information neural network model.

[0011] In an embodiment of the present application, the method further comprises: In constructing the first thermal field loss function and the second thermal field loss function, a thermal boundary condition in the form of Monte Carlo integration is used as an additional integral constraint, and the integral constraint is used to constrain the heat flux integral of the heat sink bottom surface to be consistent with the heat source power consumption; The expression of the integral constraint is:

[0012] Wherein, is the integral domain, is a constant heat flux generated by the heat source at the bottom of the heat sink, Power consumption of the heat source.

[0013] In one embodiment of the present application, the expression of the fluid energy equation is:

[0014] wherein, , and are the temperature, the constant-pressure heat capacity and the thermal conductivity of the fluid, is the fluid velocity, is the Laplace operator, is the density of the fluid.

[0015] In one embodiment of the present application, the expression of the first thermal field loss function is:

[0016] wherein, , represents the loss weight in the solid domain, represents the residual in the solid domain, represents the weight matrix and the bias vector of the first thermal field physical information neural network model, represents the design parameter, represents the number of second spatial coordinate samples collected in the solid domain, , represents the loss weight of the solid domain boundary, represents the weighted sum of the second interface residual and the third boundary residual, represents the number of second spatial coordinate samples collected on the solid domain boundary, and respectively represent the second spatial coordinate samples collected from the solid domain and the second spatial coordinate samples collected from the solid domain boundary, is a nonlinear partial differential operator in the solid domain, is the L2 norm, is a nonlinear partial differential operator of the solid domain boundary, is the predicted temperature field distribution in the solid domain, is the predicted temperature field distribution on the solid domain boundary.

[0017] In one embodiment of the present application, the expression of the second thermal field loss function is:

[0018]

[0019] wherein, , This represents the third loss weight within the fluid domain. This represents the residual within the second fluid domain. The weight matrix and bias vector represent the neural network model of the second thermal field physical information. Indicates design parameters, This indicates the number of third-space coordinate samples collected within the fluid domain. , This represents the fourth loss weight within the fluid domain. This represents the weighted sum of the residuals at the second interface and the fourth boundary. This indicates the number of spatial coordinate samples collected at the boundary of the fluid domain. and These represent third-space coordinate samples collected from within the fluid domain and third-space coordinate samples collected from the boundary of the fluid domain, respectively. It is the third nonlinear partial differential operator in the fluid domain. It is the L2 norm. For the fourth nonlinear partial differential operator at the boundary of the fluid domain, To predict the temperature field distribution within the fluid domain, This represents the predicted temperature field distribution at the boundary of the fluid domain.

[0020] Another aspect of the present invention provides a storage medium storing a computer program for performing the steps of the physical information neural network-driven integrated core thermal performance prediction method described in any of the above embodiments.

[0021] Another aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor, when calling the computer program in the memory, implements the steps of the physical information neural network-driven integrated core thermal performance prediction method as described in any of the above embodiments.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention provides a method for predicting the thermal performance of integrated core particles driven by physical information neural networks. The physical information neural network is applied to the temperature prediction of integrated core particle systems. Compared with traditional numerical simulation software, it greatly reduces human intervention and shortens the time-consuming mesh division during preprocessing. Compared with data-driven neural networks, the parameterized geometric model established by this invention does not require a labeled dataset. It can output the predicted temperature field distribution by relying only on physical knowledge after training the thermal field physical information neural network model. This can avoid the results deviating from physical knowledge, improve the prediction accuracy, and make the model prediction consistent with the actual results in the real world.

[0023] (2) The present application only needs to sample multiple space coordinates on the parametric geometric model, and takes the design parameters and the collected multiple space coordinates as inputs of the trained thermal field physical information neural network model, so that the network model outputs the temperature field distribution of the integrated core particle system under the design parameters through forward inference, the physical information neural network model is a kind of meshless method, compared with the traditional numerical solver, the time required for pre-processing is greatly reduced, thereby the prediction time is reduced, and the prediction efficiency is improved.

[0024] (3) The present application uses the physical information neural network for integrated core particle system thermal design, the heat dissipation design parameters of the integrated core particle system, the geometric parameters of the heat sink (the geometric parameters of the solid domain) and the flow channel geometric parameters (the geometric parameters of the fluid domain) are input into the physical information neural network for training, after training of the physical information neural network of integrated design parameters, the solving result can cover the entire design space, and the research and development personnel only need to input the design variable values concerned as the input of the physical neural network, so as to predict the corresponding core particle system thermal performance under the design variable / design variable group, that is, the system temperature distribution under the design, so as to improve the research and development efficiency of the research and development personnel.

[0025] (4) The present application embeds physical laws into the physical information neural network architecture, fully utilizes the automatic derivation characteristics in the back propagation algorithm of the physical information neural network, constructs the physical equation residual as a loss function to guide the training of the physical information neural network, without label data, and only by introducing complete physical knowledge to constrain the training of the physical information neural network, high-precision prediction conforming to physical laws can be obtained, and the dependence of the traditional neural network architecture on label data is greatly reduced.

[0026] The present application will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 is a 2.5D integrated core particle system packaging structure schematic diagram provided by an embodiment of the present application; Figure 2 is a flowchart of an integrated core particle thermal performance prediction method driven by a physical information neural network provided by an embodiment of the present application; Figure 3 is an architecture schematic diagram of a flow field physical information neural network provided by an embodiment of the present application; Figure 4 is a fluid domain sampling schematic diagram provided by an embodiment of the present application; Figure 5 is a solid domain sampling schematic diagram provided by an embodiment of the present application; Figure 6 is a whole sampling schematic diagram of a solid domain and a fluid domain provided by an embodiment of the present application; Figure 7 is a simulation diagram of running time required for training a flow field physical information neural network model provided by an embodiment of the present application; Figure 8 is a simulation diagram of running time required for training a thermal field physical information neural network model provided by an embodiment of the present application; Figure 9 is a simulation diagram of running time required for a single COMSOL simulation provided by an embodiment of the present application; Figure 10 is a comparison diagram of peak temperature predicted by a single running of a thermal field physical information neural network model provided by the present method and COMSOL simulation software. DETAILED DESCRIPTION

[0028] In order to further illustrate the technical means and effects adopted by the present application to achieve the predetermined object, a physical information neural network driven integrated chip particle thermal performance prediction method according to the present application is described in detail below in combination with the drawings and specific embodiments.

[0029] The foregoing and other technical contents, features and effects of the present application can be clearly presented in the detailed description of the specific embodiments below in combination with the drawings. Through the description of the specific embodiments, the technical means and effects adopted by the present application to achieve the predetermined object can be more deeply and specifically understood. However, the accompanying drawings are provided for reference and illustration only, and are not intended to limit the technical solutions of the present application.

[0030] It should be noted that, in this document, relational terms such as first and second and the like can only be used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that these entities or operations exist in any such actual relationship or order. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus that includes a list of elements does not necessarily include only those elements in the list, but can include other elements not expressly listed or included. Without more limitations, the element defined by the statement "comprising a" does not exclude the presence of additional identical elements in the process, method, article or apparatus including the element.

[0031] The present application aims to solve the problem of how to improve the accuracy of integrated chip particle thermal performance prediction while reducing data dependency and ensuring that the prediction results comply with physical laws. A physical information neural network driven integrated chip particle thermal performance prediction method is proposed. In order to better understand the method provided by the embodiments of the present application, first, refer to Figure 1 , Figure 1 is a schematic diagram of a 2.5D integrated chip particle system packaging structure provided by an embodiment of the present application, asFigure 1 As shown, the system includes, from bottom to top, in sequence: An organic substrate (101) for supporting and electrically connecting the entire integrated chip system.

[0032] C4 bumps and underfill (102) for electrically interconnecting and mechanically reinforcing the chip (105) and the organic substrate (101).

[0033] A board (103) for high-density interconnection between multiple chips.

[0034] Micro bumps and underfill (104) for connecting the chip (105) and the board (103).

[0035] The chip (105) as the main heat source.

[0036] Thermal interface material (106) for enhancing the heat conduction between the chip (105) and the upper heat dissipation structure (108).

[0037] Thermal conduction plate (107) for spreading heat.

[0038] Heat sink (108) for contacting with an external cooling fluid to take away the heat generated inside the integrated chip system through convective heat transfer.

[0039] In which, Figure 1 The red line in the above figure represents the convective heat dissipation boundary of the integrated chip system, located on the outer surface of the heat sink (108); the yellow line represents the heat flux boundary, located at the bottom of the chip (105), used to simulate the chip power consumption as the heat source boundary of the integrated chip system, and specifies a constant heat flux input; the black line represents the adiabatic boundary, distributed on the outer surface of the integrated chip system except the above-mentioned boundaries, including the flow channel wall surface, etc., indicating that these boundaries have no heat exchange with the outside world.

[0040] It should be noted that the solid domain in the embodiment of the present application is Figure 1 The chip (105), the thermal interface material (106), the thermal conduction plate (107) and the heat sink (108) in the above figure, and the fluid domain is the cooling fluid space flowing through the outside of the heat sink (108). Further, the present application also includes a fluid inlet boundary and a fluid outlet boundary, the fluid inlet boundary is located at the starting end of the fluid domain, used to specify the fluid inlet velocity and the fluid inlet temperature boundary condition, and the fluid outlet boundary is located at the end of the fluid domain, used to specify the static pressure outlet condition.

[0041] It can be understood that the target of the present application is to quickly predict the temperature field distribution under the fluid-structure coupling model through the physical information neural network. Next, please refer to Figure 2 The method comprises the following steps: Step 1: Construct a parametric geometric model of the integrated core system to be predicted. The parametric geometric model includes geometric models of the solid domain and the fluid domain.

[0042] Step 2: Based on the parametric geometric model, obtain the design parameters of the integrated core system to be predicted. The design parameters include the geometric parameters of the solid domain and the geometric parameters of the fluid domain.

[0043] The geometric parameters of the solid domain include the heat dissipation design parameters of the integrated core system to be predicted and the geometric parameters of the heat sink, while the geometric parameters of the fluid domain include the flow channel geometric parameters.

[0044] In embodiments of the present invention, the heat dissipation design parameters include at least one of the following: the number of core particles (also known as heat sources), the layout of the core particles, the power consumption of the core particles, the number of fins of the heat sink, the fin length, the fin thickness, the fin height, and the fin spacing; the geometric parameters of the heat sink include at least one of the following: the material of the heat sink, the process parameters, the parameters of the substrate structure and the fin structure; the flow channel geometric parameters include at least one of the following: the flow channel length, the flow channel width, the flow channel height, and the relative position of the heat sink.

[0045] Step 3: Collect multiple spatial coordinates from the parametric geometric model to obtain the target spatial coordinates.

[0046] In embodiments of the present invention, multiple spatial coordinates can be randomly collected from the parametric geometric model (i.e., the geometric model of the solid domain and the fluid domain) using the Monte Carlo method to obtain the target spatial coordinates.

[0047] Step 4: Input the design parameters and target spatial coordinates into the trained thermal field physical information neural network model so that the trained thermal field physical information neural network model can output the temperature field distribution of the integrated core system to be predicted under the design parameters through forward inference.

[0048] The trained thermal field physical information neural network model is obtained by training the flow field physical information neural network model, design parameters, and preset thermal field loss function. The trained flow field physical information neural network model is obtained by training the design parameters and preset flow field loss function. The trained flow field physical information neural network model is used to output the predicted velocity components and pressure of the fluid domain.

[0049] It should be noted that, in order to predict the heat dissipation performance of the integrated core system, it is necessary to study the convective heat transfer between the integrated core system and the surrounding fluid, as well as the heat conduction within the fluid domain and the solid domain of the integrated core system. The differential equations involved in this conjugate heat transfer model can be expressed in the following general form:

[0050] in, It is a nonlinear partial differential operator. This is the antiderivative, i.e., the solution to the differential equation. and Domains Spatial coordinates and design parameters on (solid or fluid domain) in this invention .

[0051] Furthermore, solving the above differential equation requires specifying boundary conditions, the general expression of which is:

[0052] in, To cover constraint operators of the first, second, and mixed boundary conditions, Representative domain The boundary, For the boundary Spatial coordinates on the [space].

[0053] In the embodiments of this application, both the thermal field physical information neural network model and the flow field physical information neural network model are fully connected neural network architectures, which include an input layer, a hidden layer and an output layer.

[0054] Next, taking the flow field physical information neural network as an example, we will introduce the structure of a fully connected neural network. See [link to relevant documentation]. Figure 3 , Figure 3 This is a schematic diagram of the architecture of a flow field physical information neural network provided in an embodiment of the present invention. The network has six hidden layers, each with 512 neurons. The input layer receives spatial coordinates and design parameters, the hidden layers integrate and represent physical field features, and the output layer outputs the flow field velocity components and pressure; that is, it outputs the corresponding flow field velocity components for each spatial coordinate. and pressure and using vectors To indicate, among which, fluid velocities Velocity components in the x, y, and z directions in three-dimensional space.

[0055] Specifically, in a fully connected neural network, the output of each neuron in one layer serves as the input to the neuron in the next layer. The mathematical expression for this forward propagation process is:

[0056] Among them, superscript Subscript representing the layer number of the neural network. Represents a neuron index, such as and The first The first in the layer The first neuron to the second The first in the layer The weights and biases of each neuron, and These represent the spatial coordinates and design parameters in the fluid domain, respectively. As a non-linear activation function, the embodiment of this invention selects the tanh function, with the following specific form:

[0057] Next, we will introduce the training process of the neural network model for flow field physical information, which includes the following five steps: Step 401A: Construct a parametric geometric model and collect multiple first spatial coordinates from the geometric model of the fluid domain to obtain first spatial coordinate samples.

[0058] The parameterized geometric model described here is the same as that in the previous embodiment and will not be repeated here. Specifically, during training, multiple first spatial coordinates are randomly sampled on the geometric model of the fluid domain. For example, see [link to example]. Figure 4 , Figure 4 This is a schematic diagram of fluid domain sampling provided in an embodiment of the present invention, wherein, Figure 4 The red dot in the diagram represents the first spatial coordinates that can be collected.

[0059] Step 402A: Construct the residual in the first fluid domain based on the fluid motion control equation and construct the residual in the first boundary based on the fluid boundary conditions, and determine the weighted sum of the residual in the fluid domain and the residual in the first boundary as the flow field loss function.

[0060] Specifically, the expressions for the fluid motion control equations (continuity equation and Navier-Stokes equations) are as follows:

[0061]

[0062] in, The velocity of a fluid (which can be understood as either a gas or a liquid). The components in the x, y, and z directions in three-dimensional space are respectively , , ,Right now =[ , , ], For the density of the fluid, For fluid pressure, The viscosity of the fluid.

[0063] For the fluid boundary conditions, the fluid velocity at the channel inlet is known ( Pressure at the channel outlet There is no slippage between the channel and the solid wall. .

[0064] Furthermore, the sum of the residuals of the above physical equations is used as a soft constraint. To measure the difference between the flow field physical information neural network model and the constraint, the loss function of the flow field physical information neural network model is defined as the weighted sum of the squares of the L2 norm of the PDE residuals and the boundary residuals. That is, the residuals in the first fluid domain are constructed based on the above fluid motion control equations and the residuals in the first boundary are constructed based on the fluid boundary conditions. The weighted sum of the residuals in the fluid domain and the residuals in the first boundary is determined as the flow field loss function.

[0065] Specifically, the expression for the flow field loss function is:

[0066] in, , This represents the first loss weight within the fluid domain. This represents the residual within the first fluid domain. The weight matrix and bias vector represent the neural network model of the flow field physical information. This indicates the number of the first spatial coordinate samples collected within the fluid domain. = , The second loss weight represents the boundary of the fluid domain. Represents the first boundary residual. This indicates the number of the first spatial coordinate samples collected at the boundary of the fluid domain. and These represent the first spatial coordinate sample collected from within the fluid domain and the first spatial coordinate sample collected from the boundary of the fluid domain, respectively. This is the first nonlinear partial differential operator in the fluid domain. It is the L2 norm. For the second nonlinear partial differential operator at the boundary of the fluid domain, For the predicted velocity components and pressure within the fluid domain, The velocity components and pressure at the predicted fluid domain boundary are given.

[0067] Step 403A: Input the first spatial coordinate sample and design parameters into the initial flow field physical information neural network model so that the initial flow field physical information neural network model outputs the predicted velocity components and pressure of the fluid domain through forward propagation.

[0068] Specifically, the first spatial coordinate sample and design parameters are concatenated, and then the concatenated vector is input into the initial flow field physical information neural network model. After forward propagation, the initial flow field physical information neural network model can obtain the fluid velocity components and pressure corresponding to each first spatial coordinate position. Then, through the backpropagation algorithm, the residual of the physical equation at that point is output to determine the accurate dynamic characteristics of the flow field. The specific solution process is as follows: First, we propose to construct a neural network model of the initial flow field physical information. The fluid motion control equations satisfy the above-mentioned boundary conditions at the channel inlet and outlet (i.e., at the inlet). Exit There is no slippage in the channel or on the solid wall. Analytical solution The proxy model can generate a proxy model that is compatible with... Same dimension (e.g., dimension 4, i.e., output) The output vector of ) is shown below:

[0069] Among them, subscript This indicates that the model is a neural network model of flow field physical information, indicated by the superscript. L The output layer of the neural network model representing the physical information of the flow field.

[0070] Step 404A: Using automatic differentiation technology, calculate the first derivative of the predicted velocity component and pressure of the fluid domain with respect to the first spatial coordinate sample, and substitute the first derivative into the flow field loss function to obtain the loss value of the flow field loss function.

[0071] Specifically, the first derivative of the predicted velocity components and pressure in the fluid domain with respect to the first spatial coordinate sample is calculated using automatic differentiation techniques. Relative to input The gradient information.

[0072] Step 405A: Minimize the flow field loss function using an optimization algorithm, update the network parameters of the initial flow field physical information neural network model until the flow field loss function converges, and obtain the trained flow field physical information neural network model.

[0073] Furthermore, gradient-based optimizers (such as SGD, Adam, and L-BFGS) are used to minimize the flow field loss function to train the weights and biases of the neural network model of the initial flow field physical information. Specifically, through this formula, The weights and biases are used to train the initial flow field physical information neural network model, where, The weights and biases represent the weights and biases of the trained flow field physical information neural network model. This occurs when the flow field loss function converges, i.e., when the initial flow field physical information neural network model converges to a preset accuracy level. It can then be considered an analytical solution. The proxy model, that is, the well-trained neural network model of flow field physical information can be obtained at this time.

[0074] It should be noted that in this embodiment, the flow field physical information neural network model is trained first, followed by the thermal field physical information neural network model. When training the thermal field physical information neural network model, the network parameters of the flow field physical information neural network model need to be fixed, and the fluid velocity information in the flow field physical information neural network model is used to calculate the convection term. At this time, the already trained flow field physical information neural network model is integrated as a fixed and precise module. Specifically, when training the thermal field physical information neural network model, if the sampled spatial coordinate point is located in the fluid domain, the system will call the already trained flow field physical information neural network model to obtain the fluid velocity at that point and substitute it as a known quantity into the fluid energy equation, thereby ensuring that the thermal prediction conforms to the actual physical convection heat transfer process.

[0075] It should be noted that, in order to enable the network to accurately capture the heat flux jump caused by the difference in thermal conductivity between the chip packaging system and the fluid, in the embodiments of this application, a first thermal field physical information neural network model for predicting the solid domain temperature and a second thermal field physical information neural network model for predicting the fluid domain temperature were established respectively.

[0076] The training process of the thermal field physical information neural network model is described below, which includes the following five steps: Step 401B: Construct a parametric geometric model and collect multiple second spatial coordinates and multiple third spatial coordinates from the geometric models of the solid domain and the fluid domain, respectively, to obtain second spatial coordinate samples and third spatial coordinate samples.

[0077] Specifically, during training, multiple second-space coordinates and multiple third-space coordinates are randomly sampled on the geometric models of the solid and fluid domains. For example, see [link to example]. Figure 4 and Figure 5 , Figure 5 This is a schematic diagram of solid-state domain sampling provided in an embodiment of the present invention, wherein, Figure 5 The red dots in the diagram represent the collectable third-space coordinates. Further details can be found in [link to relevant documentation]. Figure 6 , Figure 6 This is a schematic diagram of an overall sampling of a solid domain and a fluid domain provided in an embodiment of the present invention, wherein, Figure 6 In the diagram, red dots represent channel inlets, green dots represent channel outlets, blue dots represent channels (fluid domains), and yellow dots represent chip packaging systems (solid domains).

[0078] Step 402B: Construct residuals within the solid domain based on the solid heat conduction equation; construct residuals within the second fluid domain based on the fluid energy control equation; construct residuals at the second interface based on the temperature continuity condition and heat flux continuity condition at the fluid-structure interaction interface; construct residuals at the third boundary based on the heat source boundary; and construct residuals at the fourth boundary based on the fluid inlet temperature boundary and the adiabatic boundary. The weighted sum of the residuals within the solid domain, the second interface residual, and the third boundary residual is determined as the first thermal field loss function of the first thermal field physical information neural network model. The weighted sum of the residuals within the second fluid domain, the second interface residual, and the fourth boundary residual is determined as the second thermal field loss function of the second thermal field physical information neural network model. The fluid velocity in the fluid energy equation is provided by the velocity component output by the trained flow field physical information neural network model.

[0079] Within the chip package, only heat conduction occurs. For steady-state heat conduction without an internal heat source, the temperature field... It satisfies the Laplace equation; specifically, the expression for the solid heat conduction equation is:

[0080] in, For solid temperature; , and These represent the thermal conductivity, density, and constant-pressure heat capacity of a solid, respectively. For the Laplace operator.

[0081] The expression for the fluid energy control equation is:

[0082] in, , and These are the fluid's temperature, constant-pressure heat capacity, and thermal conductivity, respectively. The density of the fluid.

[0083] At the interface between the fluid and solid domains, the continuity of temperature and heat flux transfer is controlled by the following boundary condition constraints, namely, the expression for the temperature continuity condition at the fluid-structure interaction interface is:

[0084]

[0085] Subscript and These represent the solid domain wall and the fluid domain wall at the fluid-solid interface, respectively. This represents the temperature of the solid domain wall. This represents the temperature of the fluid domain wall. This represents the heat flux through the walls of a solid domain. This represents the heat flux through the walls of the fluid domain.

[0086] Among them, heat flux can be directly calculated from the temperature gradient, that is, the expression for the heat flux continuity condition is:

[0087]

[0088] in, Let be the unit normal vector pointing outwards from the wall. and These are the temperature gradients of the solid domain wall and the fluid domain wall, respectively. This represents the thermal conductivity of the wall surface of a fluid domain. This represents the thermal conductivity of the solid domain wall.

[0089] It should be noted that, in the embodiments of this application, it is assumed that the fluid temperature remains constant at the channel inlet, i.e., the fluid inlet temperature boundary. The channel walls and outlet are considered as adiabatic, meaning the expression for the adiabatic boundary is: Finally, assuming the chip generates a uniform heat flux at the bottom of the heatsink, a constant heat flux is specified for the input system. The thermal boundary condition, i.e., the setting method for the heat source boundary, is as follows:

[0090] in, This indicates the temperature of the heat source.

[0091] Furthermore, the sum of the residuals of the above heat transfer physical equations is used as a soft constraint. Similarly, to measure the difference between the thermal field physical information neural network model and the constraints, the loss function of the thermal field physical information neural network model is defined as the weighted sum of the squares of the L2 norm of the PDE residuals and the boundary residuals. That is, the residuals in the solid domain are constructed based on the solid heat conduction equation, the residuals in the second fluid domain are constructed based on the fluid energy control equation, the residuals in the second interface are constructed based on the temperature continuity condition and heat flux continuity condition of the fluid-structure interaction interface, the residuals in the third boundary are constructed based on the heat source boundary, and the residuals in the fourth boundary are constructed based on the fluid inlet temperature boundary and the adiabatic boundary. The weighted sum of the residuals in the solid domain, the residuals in the second interface, and the residuals in the third boundary are determined as the first thermal field loss function of the first thermal field physical information neural network model, and the weighted sum of the residuals in the second fluid domain, the residuals in the second interface, and the residuals in the fourth boundary are determined as the second thermal field loss function of the second thermal field physical information neural network model.

[0092] The expression for the first thermal field loss function is:

[0093]

[0094] in, , Represents the loss weight within the solid domain. Represents the residual within the solid domain. This represents the weight matrix and bias vector of the neural network model for the first thermal field physical information. This indicates the number of second spatial coordinate samples collected within the solid domain. , The loss weights represent the boundaries of the solid domain. This represents the weighted sum of the residuals at the second interface and the third boundary. This indicates the number of second spatial coordinate samples collected at the boundary of the solid domain. and These represent the second spatial coordinate samples collected from within the solid domain and the second spatial coordinate samples collected from the boundary of the solid domain, respectively. For nonlinear partial differential operators in the solid domain, It is the L2 norm. For the nonlinear partial differential operator of the solid domain boundary, To predict the temperature field distribution within the solid domain, This represents the predicted temperature field distribution at the boundary of the solid domain.

[0095] The expression for the second thermal field loss function is:

[0096]

[0097] in, , This represents the third loss weight within the fluid domain. This represents the residual within the second fluid domain. The weight matrix and bias vector represent the neural network model of the second thermal field physical information. This indicates the number of third-space coordinate samples collected within the fluid domain. , This represents the fourth loss weight within the fluid domain. This represents the weighted sum of the residuals at the second interface and the fourth boundary. This indicates the number of spatial coordinate samples collected at the boundary of the fluid domain. and These represent third-space coordinate samples collected from within the fluid domain and third-space coordinate samples collected from the boundary of the fluid domain, respectively. It is the third nonlinear partial differential operator in the fluid domain. It is the L2 norm. For the fourth nonlinear partial differential operator at the boundary of the fluid domain, To predict the temperature field distribution within the fluid domain, This represents the predicted temperature field distribution at the boundary of the fluid domain.

[0098] To further improve the convergence performance of the neural network model for solving the physical model of the thermal field physical information, in the embodiments of this application, when constructing the first thermal field loss function and the second thermal field loss function, a Monte Carlo integral form of thermal boundary condition can also be used as an additional integral constraint. This integral constraint is used to ensure that the integral of the heat flux on the bottom surface of the heat sink is consistent with the power consumption of the heat source. Specifically, the expression of the integral constraint is:

[0099] in, For the integration domain, The constant heat flux generated by the heat source at the bottom of the radiator. This represents the power consumption of the heat source.

[0100] It should be noted that adopting Monte Carlo integral form of thermal boundary conditions as an additional integral constraint increases the constraint capability of the chip's heat source. Compared with discrete point constraints, the added integral constraints can form an overall constraint on the spatial coordinate points of the heat source sampling, thereby improving the convergence performance of the thermal field physical information neural network model and further reducing the number of spatial coordinate points that need to be sampled.

[0101] Step 403B: Input the second spatial coordinate sample and design parameters into the initial first thermal field physical information neural network model so that the initial first thermal field physical information neural network model outputs the predicted temperature field distribution of the solid domain through forward propagation; and input the third spatial coordinate sample and design parameters into the initial second thermal field physical information neural network model so that the initial second thermal field physical information neural network model outputs the predicted temperature field distribution of the fluid domain through forward propagation.

[0102] Specifically, the second spatial coordinate samples and design parameters are concatenated, and then the concatenated vector is input into the initial first thermal field physical information neural network model. After forward propagation, the initial first thermal field physical information neural network model can obtain the temperature at each second spatial coordinate position, thus obtaining the temperature field distribution of the entire solid domain. Similarly, the third spatial coordinate samples and design parameters are concatenated, and then the concatenated vector is input into the initial second thermal field physical information neural network model. After forward propagation, the initial second thermal field physical information neural network model can obtain the temperature at each third spatial coordinate position, thus obtaining the temperature field distribution of the entire fluid domain.

[0103] Step 404B: Using automatic differentiation technology, calculate the second derivative of the predicted temperature field distribution of the solid domain with respect to the second spatial coordinate sample, substitute the second derivative into the first thermal field loss function to obtain the loss value of the first thermal field loss function, and calculate the third derivative of the predicted temperature field distribution of the fluid domain with respect to the third spatial coordinate sample, substitute the third derivative into the second thermal field loss function to obtain the loss value of the second thermal field loss function.

[0104] This step is similar in principle to step 404A in the above embodiment. For details, please refer to the above embodiment, which will not be repeated here.

[0105] Step 405B: Minimize the first thermal field loss function and the second thermal field loss function using an optimization algorithm, and update the network parameters of the initial first thermal field physical information neural network model and the initial second thermal field physical information neural network model until the first thermal field loss function and the second thermal field loss function converge to obtain the trained flow field physical information neural network model.

[0106] Furthermore, gradient-based optimizers (such as SGD, Adam, and L-BFGS) are used to minimize the flow field loss function to train the weights and biases of the initial first thermal field physical information neural network model and the initial second thermal field physical information neural network model. , Specifically, through the following two formulas, , The weights and biases of the initial first thermal field physical information neural network model and the initial second thermal field physical information neural network model are used to train the model. This represents the weights and biases of the trained neural network model containing the physical information of the first thermal field. This represents the weights and biases of the trained second thermal field physical information neural network model. The first and second thermal field loss functions converge when they converge, i.e., when the initial first and second thermal field physical information neural network models converge to a preset accuracy level. and It can then be considered an analytical solution. and The proxy model, that is, the well-trained neural network model of thermal field physical information can be obtained at this time.

[0107] It should be noted that the trained thermal field physics information neural network model will have forward inference capability for any "spatial coordinates and design parameters." At this point, it is no longer necessary to call the flow field physics information neural network model because, during the training phase, the thermal field physics information neural network model has already learned and internalized the influence of the flow field. Through backpropagation training, the weights and biases of the thermal field physics information neural network model have been adjusted to their optimal state, enabling it to directly output a temperature solution that satisfies both solid heat conduction and convective heat transfer coupled with a specific flow field, given design parameters and spatial coordinates.

[0108] In summary, the embodiment of the present invention provides a method for predicting the thermal performance of integrated core particles driven by physical information neural networks. (1) Applying physical information neural networks to the temperature prediction of integrated core particle systems significantly reduces human intervention compared with traditional numerical simulation software and shortens the time-consuming mesh division during preprocessing. Compared with data-driven neural networks, the parameterized geometric model established by the present invention does not require a labeled dataset and can output the predicted temperature field distribution by relying solely on physical knowledge. This avoids the results deviating from physical knowledge, improves the accuracy of prediction, and makes the model prediction consistent with the actual results in the real world.

[0109] (2) In this invention, only multiple spatial coordinates need to be sampled on the parametric geometric model, and the design parameters and the collected multiple spatial coordinates are used as inputs to the trained thermal field physical information neural network model so that the network model can output the temperature field distribution of the integrated core system under the design parameters through forward inference. As a meshless method, the physical information neural network model greatly reduces the time required for preprocessing compared with the traditional numerical solver, thereby reducing the prediction time and improving the prediction efficiency.

[0110] (3) In this invention, a physical information neural network is used for the thermal design of an integrated core system. The heat dissipation design parameters of the integrated core system, the geometric parameters of the radiator (geometric parameters of the solid domain), and the geometric parameters of the flow channel (geometric parameters of the fluid domain) are fed into the physical information neural network as input for training. After the physical information neural network with integrated design parameters is trained, its solution results can cover the entire design space. Researchers only need to use the design variable values ​​of interest as input to the physical neural network to predict the thermal performance of the core system under the design variable / set of design variables, that is, the system temperature distribution under the design, thereby improving the research and development efficiency of researchers.

[0111] (4) This invention embeds physical laws into the physical information neural network architecture, making full use of the automatic differentiation feature in the backpropagation algorithm of the physical information neural network, constructing the physical equation residual as a loss function to guide the training of the physical information neural network. No label data is required. The physical information neural network can be trained by introducing complete physical knowledge constraints to obtain high-precision predictions that conform to physical laws, which greatly reduces the dependence of traditional neural network architecture on label data.

[0112] The technical effects of the integrated core thermal performance prediction method driven by physical information neural network provided in this invention are verified through simulation experiments below.

[0113] Specifically, the experimental example used a core-particle system heat sink with a 4-fin structure. Fin height, length, thickness, and heat source power consumption were integrated as parametric design variables into the neural network input. The height, length, and thickness ranged from [0.3, 0.5], [0.5, 1], and [0.05, 0.1] cm, respectively, and the heat source power consumption ranged from [50~150] W. The experiment revealed that the flow field required 400,000 training iterations (28 hours) (details as follows). Figure 7 As shown), at this point, the loss function can reach the order of e⁻³, and the network inference has good accuracy; with the above flow field training completed, the thermal field needs to be trained 220,000 times (25 hours) (specifically as shown). Figure 8 As shown in the figure, the loss function can reach the order of e-5, and the network has good accuracy. Therefore, the training of the model using this method takes a total of 25+28=53 hours.

[0114] Correspondingly, a parametric geometric model is established in a traditional solver such as COMSOL, and finite element analysis is performed using the same parameter range. Each design variable considers 10 values, resulting in a total of 10^4 = 10,000 working conditions to be solved. Experiments show that the average total time for mesh generation and finite element analysis for each working condition is 63 seconds, meaning the total time for finite element analysis is 175 hours (details omitted). Figure 9 (As shown).

[0115] As shown above, this method can achieve the same predictive ability at a lower computational cost of approximately 3 times (53 hours). Furthermore, Figure 10 This paper presents a comparison of the peak temperatures predicted during a single run of the experimental example using COMSOL simulation software and the model built using this method. Figure 10It can be seen that the temperature difference between the highest temperature predicted by the physical information neural network model and the highest temperature predicted by COMSOL is only 351.93-351.79=0.14, and the temperature difference between the lowest temperature predicted by the physical information neural network model and the lowest temperature predicted by COMSOL is only 293.22-292.66=0.56. Therefore, it can be seen that the simulation results of this method are close to those of COMSOL with a lower computational cost, that is, it has good accuracy and has advantages over COMSOL.

[0116] In the several embodiments provided by this invention, it should be understood that the methods disclosed in this invention can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of modules is merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0117] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or in the form of hardware plus software functional modules.

[0118] Another embodiment of the present invention provides a storage medium storing a computer program for executing the steps of the integrated core thermal performance prediction method driven by the physical information neural network described in the above embodiments.

[0119] Another aspect of the present invention provides an electronic device, including a memory and a processor. The memory stores a computer program, and when the processor invokes the computer program in the memory, it implements the steps of the integrated core thermal performance prediction method driven by a physical information neural network as described in the above embodiments. Specifically, the integrated modules implemented in the form of software functional modules can be stored in a computer-readable storage medium. The software functional modules stored in the storage medium include several instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0120] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the thermal performance of integrated core particles driven by a physical information neural network, characterized in that, include: Step 1: Construct a parameterized geometric model of the integrated core system to be predicted. The parameterized geometric model includes geometric models of the solid domain and the fluid domain. Step 2: Based on the parametric geometric model, obtain the design parameters of the integrated core system to be predicted. The design parameters include the geometric parameters of the solid domain and the geometric parameters of the fluid domain. Step 3: Collect multiple spatial coordinates from the parametric geometric model to obtain the target spatial coordinates; Step 4: Input the design parameters and the target spatial coordinates into the trained thermal field physical information neural network model, so that the trained thermal field physical information neural network model can output the temperature field distribution of the integrated core system to be predicted under the design parameters through forward inference; The trained thermal field physical information neural network model is obtained by training a trained flow field physical information neural network model, the design parameters, and a preset thermal field loss function. The trained flow field physical information neural network model is obtained by training a trained flow field physical information neural network model, the design parameters, and a preset flow field loss function. The trained flow field physical information neural network model is used to output the predicted fluid velocity components and pressure of the fluid domain.

2. The integrated core thermal performance prediction method driven by physical information neural network according to claim 1, characterized in that, The geometric parameters of the solid domain include the heat dissipation design parameters of the integrated chip system to be predicted and the geometric parameters of the heat sink, and the geometric parameters of the fluid domain include the flow channel geometric parameters.

3. The integrated core thermal performance prediction method driven by physical information neural network according to claim 1, characterized in that, The training process of the flow field physical information neural network model includes: Step 401A: Construct the parameterized geometric model and collect multiple first spatial coordinates from the geometric model of the fluid domain to obtain first spatial coordinate samples; Step 402A: Construct the residual in the first fluid domain based on the fluid motion control equation and construct the residual in the first boundary based on the fluid boundary conditions, and determine the weighted sum of the residual in the fluid domain and the residual in the first boundary as the flow field loss function; Step 403A: Input the first spatial coordinate sample and the design parameters into the initial flow field physical information neural network model, so that the initial flow field physical information neural network model outputs the predicted fluid velocity components and pressure of the fluid domain through forward propagation; Step 404A: Using automatic differentiation technology, calculate the first derivative of the fluid velocity component and pressure of the predicted fluid domain with respect to the first spatial coordinate sample, and substitute the first derivative into the flow field loss function to obtain the loss value of the flow field loss function; Step 405A: Minimize the flow field loss function using an optimization algorithm, update the network parameters of the initial flow field physical information neural network model, until the flow field loss function converges, to obtain the trained flow field physical information neural network model.

4. The integrated core thermal performance prediction method driven by physical information neural network according to claim 3, characterized in that, The expression for the flow field loss function is: in, , This represents the first loss weight within the fluid domain. This represents the residual within the first fluid domain. This represents the weight matrix and bias vector of the neural network model representing the physical information of the flow field. Indicates the design parameters, This indicates the number of the first spatial coordinate samples collected within the fluid domain. = , The second loss weight represents the boundary of the fluid domain. This represents the first boundary residual. This indicates the number of the first spatial coordinate samples collected at the boundary of the fluid domain. and These represent the first spatial coordinate sample collected from within the fluid domain and the first spatial coordinate sample collected from the boundary of the fluid domain, respectively. This is the first nonlinear partial differential operator in the fluid domain. It is the L2 norm. The second nonlinear partial differential operator is the boundary of the fluid domain. For the predicted fluid velocity components and pressure within the fluid domain, The predicted fluid velocity components and pressures at the boundary of the fluid domain.

5. The method for predicting the thermal performance of integrated core particles driven by a physical information neural network according to claim 1, characterized in that, The thermal field physical information neural network model includes a first thermal field physical information neural network model for predicting the temperature of the solid domain and a second thermal field physical information neural network model for predicting the temperature of the fluid domain. The training process of the thermal field physical information neural network model includes: Step 401B: Construct the parameterized geometric model and collect multiple second spatial coordinates and multiple third spatial coordinates from the geometric models of the solid domain and the fluid domain, respectively, to obtain second spatial coordinate samples and third spatial coordinate samples; Step 402B: Construct residuals within the solid domain based on the solid heat conduction equation; construct residuals within the second fluid domain based on the fluid energy control equation; construct residuals at the second interface based on the temperature continuity condition and heat flux continuity condition at the fluid-structure interaction interface; construct residuals at the third boundary based on the heat source boundary; and construct residuals at the fourth boundary based on the fluid inlet temperature boundary and the adiabatic boundary. The weighted sum of the residuals within the solid domain, the second interface residual, and the third boundary residual is determined as the first thermal field loss function of the first thermal field physical information neural network model. The weighted sum of the residuals within the second fluid domain, the second interface residual, and the fourth boundary residual is determined as the second thermal field loss function of the second thermal field physical information neural network model. The fluid velocity in the fluid energy equation is provided by the velocity component output by the trained flow field physical information neural network model. Step 403B: Input the second spatial coordinate sample and the design parameters into the initial first thermal field physical information neural network model, so that the initial first thermal field physical information neural network model outputs the predicted temperature field distribution of the solid domain through forward propagation; and input the third spatial coordinate sample and the design parameters into the initial second thermal field physical information neural network model, so that the initial second thermal field physical information neural network model outputs the predicted temperature field distribution of the fluid domain through forward propagation. Step 404B: Using automatic differentiation technology, calculate the second derivative of the predicted temperature field distribution of the solid domain with respect to the second spatial coordinate sample, substitute the second derivative into the first thermal field loss function to obtain the loss value of the first thermal field loss function, and calculate the third derivative of the predicted temperature field distribution of the fluid domain with respect to the third spatial coordinate sample, substitute the third derivative into the second thermal field loss function to obtain the loss value of the second thermal field loss function; Step 405B: Minimize the first thermal field loss function and the second thermal field loss function using an optimization algorithm, and update the network parameters of the initial first thermal field physical information neural network model and the initial second thermal field physical information neural network model until the first thermal field loss function and the second thermal field loss function converge to obtain the trained flow field physical information neural network model.

6. The integrated core thermal performance prediction method driven by a physical information neural network according to claim 5, characterized in that, The method further includes: When constructing the first thermal field loss function and the second thermal field loss function, a Monte Carlo integral form of thermal boundary condition is used as an additional integral constraint. The integral constraint is used to ensure that the heat flux integral on the bottom surface of the heat sink is consistent with the heat source power consumption. The expression for the integral constraint is: in, For the integration domain, The constant heat flux generated by the heat source at the bottom of the radiator. This represents the power consumption of the heat source.

7. The method for predicting the thermal performance of integrated core particles driven by a physical information neural network according to claim 5, characterized in that, The expression for the fluid energy equation is: in, , and These are the fluid's temperature, constant-pressure heat capacity, and thermal conductivity, respectively. The fluid velocity, For the Laplace operator, The density of the fluid.

8. The integrated core thermal performance prediction method driven by physical information neural network according to claim 5, characterized in that, The expression for the first thermal field loss function is: in, , This represents the loss weight within the solid domain. This represents the residual within the solid domain. This represents the weight matrix and bias vector of the first thermal field physical information neural network model. Indicates the design parameters, This indicates the number of second spatial coordinate samples collected within the solid domain. , This represents the loss weight at the boundary of the solid domain. This represents the weighted sum of the second interface residual and the third boundary residual. This indicates the number of second spatial coordinate samples collected on the boundary of the solid domain. and These represent the second spatial coordinate samples collected from within the solid domain and the second spatial coordinate samples collected from the boundary of the solid domain, respectively. For the nonlinear partial differential operator in the solid domain, It is the L2 norm. For the nonlinear partial differential operator of the boundary of the solid domain, To predict the temperature field distribution within the solid domain, This refers to the predicted temperature field distribution at the boundary of the solid domain.

9. The method for predicting the thermal performance of integrated core particles driven by a physical information neural network according to claim 5, characterized in that, The expression for the second thermal field loss function is: in, , This represents the third loss weight within the fluid domain. This represents the residual within the second fluid domain. This represents the weight matrix and bias vector of the second thermal field physical information neural network model. Indicates the design parameters, This indicates the number of third spatial coordinate samples collected within the fluid domain. , This represents the fourth loss weight within the fluid domain. This represents the weighted sum of the second interface residual and the fourth boundary residual. This indicates the number of spatial coordinate samples collected at the boundary of the fluid domain. and These represent the third spatial coordinate samples collected from within the fluid domain and the third spatial coordinate samples collected from the boundary of the fluid domain, respectively. This is the third nonlinear partial differential operator within the fluid domain. It is the L2 norm. This is the fourth nonlinear partial differential operator for the boundary of the fluid domain. To predict the temperature field distribution within the fluid domain, This refers to the predicted temperature field distribution at the boundary of the fluid domain.

10. A storage medium storing a computer program, characterized in that, The computer program is used to perform the steps of the integrated core thermal performance prediction method driven by the physical information neural network according to any one of claims 1 to 9.

Images