A mobile phone thermal simulation method and system based on multi-physical field coupling

Abstract

The application discloses a mobile phone thermal simulation method and system based on multi-physical field coupling, and relates to the field of mobile phone thermal simulation. A hybrid simulation coupling machine learning and physical solving is constructed. The machine learning agent model and the throttling logic script are innovatively introduced. The machine learning agent model is used to quickly predict the temperature according to the current power consumption, so as to instantaneously respond to the power consumption change; the throttling logic script is used to simulate the real temperature control frequency reduction strategy, and dynamically adjusts the target power consumption at the next moment according to the predicted temperature. In this way, the originally black-box temperature control logic is explicit, and a closed-loop feedback path of power consumption-temperature is constructed. Finally, the adjusted power consumption is taken as a heat source, and a physical field is solved by a traditional CAE solver. In this way, the time scale difference problem of heat conduction and electrical change is effectively solved, and efficient and accurate simulation of real performance and temperature performance of a mobile phone under a long-time high-load scene is realized.

Description

Technical Field

[0001] This application relates to the field of mobile phone thermal simulation, and more specifically, to a mobile phone thermal simulation method and system based on multiphysics coupling. Background Technology

[0002] As smartphone performance improves, chip power density continues to rise, making heat dissipation an increasingly prominent issue. High temperatures not only affect component lifespan and stability and reduce grip comfort, but may also trigger thermal protection, limiting performance release. Therefore, thermal simulation has become crucial for optimizing heat dissipation and ensuring performance and user experience during the design phase. However, existing thermal simulation methods have significant limitations. Typically, engineers preset static or simply scripted power consumption values ​​as heat sources, input them into CAE software for steady-state or transient analysis, and obtain temperature distribution. This process is essentially a one-way, open-loop simulation, ignoring the dynamic closed-loop feedback of power consumption, temperature, and performance during actual phone operation. In reality, when the temperature reaches a threshold, the system dynamically reduces frequency and temperature through mechanisms such as DVFS. Existing methods cannot simulate this dynamic throttling process, resulting in significant deviations between simulated temperature and performance predictions and real-world conditions under prolonged high-load scenarios, severely weakening their guiding significance.

[0003] The reasons for this core pain point are multifaceted. First, the CAE software used for thermal simulation and the toolchain used for power consumption performance evaluation are functionally separate. The temperature results calculated by the CAE software lack a standardized interface to automatically feed back to the power consumption model for dynamic power adjustment. Second, the actual temperature control strategy of a mobile phone is usually a complex software logic implemented in the operating system kernel or firmware. It makes decisions based on the readings and trends of multiple sensors, which is a black box that is difficult for simulation engineers to accurately reproduce in simulation software. Furthermore, there is a huge difference between the physical timescale (seconds) of heat conduction and the electrical timescale (milliseconds) of chip power consumption changes. Simulating these two scales accurately in a single simulation would result in a transient simulation computation cost that is unacceptable from an engineering perspective.

[0004] Therefore, there is an urgent need for a simulation method that can efficiently and accurately simulate the dynamic thermal management behavior of mobile phones to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing open-loop simulations that cannot simulate the dynamic closed-loop relationship between power consumption and temperature, this application provides a mobile phone thermal simulation method based on multiphysics coupling, comprising:

[0006] S1: The preprocessor module of the CAE software analyzes the 3D model of the mobile phone that can be subjected to finite element analysis to obtain the basic simulation environment.

[0007] S2: Load the trained machine learning agent model and throttling logic script, define the initial power consumption profile and input it into the basic simulation environment, wherein the throttling logic script defines the mobile phone temperature control strategy;

[0008] S3: Input the power consumption vector at the current simulation moment into the trained machine learning proxy model to obtain the predicted temperature vector;

[0009] S4: Input the predicted temperature vector into the throttling logic script to obtain a new target power consumption vector;

[0010] S5: Using the new target power consumption vector as the heat source and the temperature field of the previous simulation moment as the initial condition, input it into the CAE solver to obtain the temperature field of the current simulation moment.

[0011] S6: Determine whether the current simulation time is less than the total simulation time. If the current simulation time is less than the total simulation time, return to steps S3 to S5. If the current simulation time is greater than or equal to the total simulation time, end the transient simulation.

[0012] According to another aspect of this application, a mobile phone thermal simulation system based on multiphysics coupling is provided, comprising:

[0013] The model analysis module is used to analyze the 3D mobile phone model that can be subjected to finite element analysis through the preprocessor module of CAE software to obtain the basic simulation environment;

[0014] The model script loading module is used to load the trained machine learning agent model and throttling logic script, define the initial power consumption profile and input it into the basic simulation environment, wherein the throttling logic script defines the mobile phone temperature control strategy;

[0015] The temperature prediction module is used to input the power consumption vector at the current simulation moment into the trained machine learning proxy model to obtain the predicted temperature vector.

[0016] A new target power consumption generation module is used to input the predicted temperature vector into the throttling logic script to obtain a new target power consumption vector;

[0017] The simulation temperature field generation module is used to take the new target power consumption vector as the heat source and the temperature field of the previous simulation moment as the initial condition, and input it into the CAE solver to obtain the temperature field of the current simulation moment.

[0018] The simulation duration judgment module is used to determine whether the current simulation time is less than the total simulation time. If the current simulation time is less than the total simulation time, it returns to the temperature prediction module and then to the simulation temperature field generation module. If the current simulation time is greater than or equal to the total simulation time, the transient simulation ends.

[0019] Compared with existing technologies, this application provides a mobile phone thermal simulation method and system based on multi-physics coupling. To address the shortcomings of existing open-loop simulations and their inability to simulate the dynamic closed-loop relationship between power consumption and temperature, this application constructs a hybrid simulation that couples machine learning and physical solution. This innovatively introduces a machine learning proxy model and a throttling logic script. The machine learning proxy model is used to quickly predict the temperature based on the current power consumption, responding instantaneously to power consumption changes; the throttling logic script is used to simulate a realistic temperature control and frequency reduction strategy, dynamically adjusting the target power consumption for the next moment based on the predicted temperature. This makes the originally black-box temperature control logic explicit and constructs a closed-loop feedback path between power consumption and temperature. Finally, the adjusted power consumption is used as a heat source, and a traditional CAE solver performs physical field solutions. By replacing some of the time-consuming physical solutions with fast inference from machine learning, the problem of time-scale differences between heat conduction and electrical changes is effectively solved, achieving efficient and accurate simulation of the real performance and temperature behavior of mobile phones under long-term high-load scenarios. Attached Figure Description

[0020] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0021] Figure 1 This is a flowchart of a mobile phone thermal simulation method based on multiphysics coupling according to an embodiment of this application.

[0022] Figure 2 This is a schematic diagram of data flow in a mobile phone thermal simulation method based on multiphysics coupling according to an embodiment of this application.

[0023] Figure 3 This is a flowchart of step S4 in the mobile phone thermal simulation method based on multiphysics coupling according to an embodiment of this application.

[0024] Figure 4 This is a block diagram of a mobile phone thermal simulation system based on multiphysics coupling according to an embodiment of this application. Detailed Implementation

[0025] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0026] Based on the technical bottlenecks mentioned above, this application proposes a mobile phone thermal simulation method based on multi-physics coupling. Figure 1 This is a flowchart of a mobile phone thermal simulation method based on multiphysics coupling according to an embodiment of this application. Figure 2 This is a schematic diagram of data flow in a mobile phone thermal simulation method based on multiphysics coupling according to an embodiment of this application. Figure 1 and Figure 2 As shown, the mobile phone thermal simulation method based on multiphysics coupling according to an embodiment of this application includes: S1: The preprocessor module of CAE software analyzes the three-dimensional model of the mobile phone that can be analyzed by finite element analysis to obtain the basic simulation environment; S2: Load the trained machine learning agent model and throttling logic script, define the initial power consumption profile and input it into the basic simulation environment, wherein the throttling logic script defines the mobile phone temperature control strategy; S3: Input the power consumption vector at the current simulation moment into the trained machine learning agent model to obtain the predicted temperature vector; S4: Input the predicted temperature vector into the throttling logic script to obtain the new target power consumption vector; S5: Use the new target power consumption vector as the heat source and the temperature field at the previous simulation moment as the initial condition, input it into the CAE solver to obtain the temperature field at the current simulation moment; S6: Determine whether the current simulation moment is less than the total simulation time. If the current simulation moment is less than the total simulation time, return to steps S3 to S5; if the current simulation moment is greater than or equal to the total simulation time, end the transient simulation. This application constructs a dynamic closed-loop system driven by both electricity and heat. On the one hand, the power consumption profile output by the temperature control throttling logic is used as a heat source to drive a machine learning surrogate model or finite element solver to perform thermodynamic simulations to calculate the temperature field; on the other hand, this temperature field is used as a feedback signal input to the temperature control throttling logic to dynamically adjust the power consumption at the next moment. This process of closely linking and interacting with electrical behavior (power consumption) and thermodynamic state (temperature) through control logic constitutes multiphysics coupling.

[0027] In step S1, the preprocessor module of the CAE software analyzes the 3D model of the mobile phone suitable for finite element analysis to obtain the basic simulation environment. It should be understood that before performing any physics-based numerical simulation, the continuous geometric model describing the physical structure of the mobile phone must be transformed into a discretized mathematical model that can be recognized and processed by the numerical solver. This is a necessary prerequisite for the entire simulation process. Although the original 3D model of the mobile phone accurately expresses the shape and assembly relationship of each component, its essence is geometric information and cannot be directly used to solve physical partial differential equations such as heat conduction. Computer solvers can only perform algebraic operations on a finite number of discrete points or elements. Therefore, the analysis of the 3D model of the mobile phone in step S1 is to construct a digital foundation containing all the information required for subsequent physics calculations. That is, to transform the idealized geometric model into a computable finite element model containing information such as mesh, material properties, boundary conditions, and physical contact, providing an accurate and efficient solution domain for subsequent transient thermal simulations.

[0028] In one exemplary operation, step S1 is implemented as follows: First, a 3D model of the mobile phone suitable for finite element analysis needs to be prepared. This model is not the original product design drawing, but a simplified geometric model specifically for simulation. For example, a 3D model of the mobile phone stored in a standard exchange format (such as STEP or IGES) has had minor features that have little impact on simulation accuracy, such as screw holes, chamfers, and silkscreen patterns, removed or filled, and all components are ensured to be closed, undamaged solids (i.e., watertight models) to avoid errors in subsequent mesh generation. This model precisely defines the geometry and spatial location of key heat-generating and heat-dissipating components inside the mobile phone, such as a 12mm x 12mm x 0.5mm system-on-a-chip (SoC), a 70mm x 50mm x 0.8mm multilayer printed circuit board (PCB), battery, metal frame, graphite heat dissipation film, vapor chamber (VC), screen, and back cover.

[0029] Next, launch the CAE software, such as the preprocessor module in commercial software like ANSYS Workbench, Abaqus / CAE, or Simcenter 3D. This preprocessor module is a collection of functions within the CAE software suite specifically responsible for simulation preprocessing. It provides a graphical interface and a suite of tools for importing geometry, assigning materials, defining physical interactions, applying boundary conditions, and generating finite element meshes.

[0030] The first step in the parsing process is geometry import and component identification. The preprocessor module reads the aforementioned 3D mobile phone model file and loads the geometric entities into the workspace. At this point, the module will automatically or with user assistance identify each individual component, such as the SoC, PCB, and battery, based on the model's hierarchical structure or naming conventions.

[0031] The second step in the analysis process is to assign precise material physical properties to each component. This is the foundation for achieving multiphysics coupled thermal simulation. An editable material library is built into the preprocessor module. For different components in the mobile phone model, their thermophysical parameters need to be defined. For example, for the silicon chip portion of the SoC, its thermal conductivity is set to 150 W / (m·K), density to 2330 kg / m³, and specific heat capacity to 710 J / (kg·K). For the PCB, due to its anisotropic nature caused by the stacking of copper foil and FR-4 substrate, its in-plane (X / Y direction) equivalent thermal conductivity needs to be defined as 35 W / (m·K), while its equivalent thermal conductivity perpendicular to the board surface (Z direction) is 0.8 W / (m·K). For the aluminum alloy frame, an isotropic thermal conductivity of 180 W / (m·K) is assigned. For high thermal conductivity graphite films, the in-plane thermal conductivity can reach as high as 1500 W / (m·K), while the normal thermal conductivity is only 5 W / (m·K). This strong anisotropy must be precisely reflected in the material definition. These parameters are all stored and associated with the corresponding geometric components.

[0032] The third step in the analysis process is to define the thermal contact relationships between components. Inside a mobile phone, heat is transferred through the contact surfaces between components. The preprocessor module provides the functionality to define these contact behaviors. For example, a thermally conductive gel (TIM) is typically coated between the SoC and the graphite heat dissipation film above it. In the preprocessor, a contact pair needs to be created between the upper surface of the SoC and the lower surface of the graphite film, and an equivalent thermal contact resistance needs to be defined for it. This value is calculated based on the thickness of the TIM (e.g., 0.1 mm) and its thermal conductivity (e.g., 5 W / (m·K)), and its value is 0.00002 square kJ / W. For components that are tightly fixed by screws or clips, such as the mid-frame and back cover, a bonded contact can be defined, where the contact thermal resistance is zero. For non-tight contacts with tiny air gaps, an equivalent contact thermal resistance that takes into account air conduction and radiation needs to be set.

[0033] The fourth step in the analysis process is to set the boundary conditions for interaction with the external environment. The phone's heat ultimately needs to be dissipated into the surrounding environment. In the preprocessor, all external surfaces of the phone (screen, back cover, side bezels) need to be selected, and mixed boundary conditions need to be applied to these surfaces. Specifically, a convective heat transfer boundary condition is applied, setting the ambient temperature to 25 degrees Celsius, and an average convective heat transfer coefficient is set based on empirical values ​​for natural convection, for example, 8 watts per square kelvin (W / (m²·K)). Simultaneously, a surface radiation boundary condition is applied to these external surfaces, also setting the ambient temperature to 25 degrees Celsius, and adjusting the emissivity based on the material, for example, 0.92 for a glass screen and 0.95 for a frosted back cover.

[0034] The fifth step in the analytical process is finite element mesh generation. The preprocessor module discretizes continuous geometric entities into a collection of millions of tiny elements (such as tetrahedrons or hexahedrons), i.e., the finite element mesh. To balance computational accuracy and efficiency, a non-uniform mesh generation strategy is employed. In critical regions with high heat flux density and large temperature gradients, such as the SoC and its adjacent PCB area, very small mesh elements are used, for example, with element sizes controlled below 0.2 mm, to accurately capture temperature changes. In regions with gentler temperature changes, such as the battery or screen edges far from the chip, larger mesh elements, such as 2 mm, can be used. For thin-layer structures like PCBs or graphite films, swept meshes or shell elements with fewer layers along the thickness direction but appropriate in-plane dimensions can be used to ensure computational accuracy and stability. After mesh generation, the preprocessor also checks the mesh quality, evaluating parameters such as element skewness and aspect ratio to ensure that the generated mesh quality meets the solver's requirements.

[0035] After the above five steps, the original 3D model of the mobile phone, suitable for finite element analysis, is successfully parsed, ultimately generating the basic simulation environment. This is a structured dataset containing all the aforementioned information. Specifically, it includes: complete finite element mesh data (coordinate information of all nodes and node connections of all elements), material property pointers for each element, boundary conditions (convection and radiation) defined on specific surfaces, and thermal contact models defined on the contact surfaces of the components. This dataset is packaged by a preprocessor into a solver input file, which formattably records all the discretization information required to solve the partial differential equations of heat conduction, forming the digital foundation for all subsequent transient simulation calculations.

[0036] In step S2, the trained machine learning agent model and throttling logic script are loaded, and an initial power consumption profile is defined and input into the basic simulation environment. The throttling logic script defines the phone's temperature control strategy. Correspondingly, after constructing a static basic simulation environment containing discretized physical information, the entire simulation process still lacks the core element for simulating the phone's real dynamic behavior: a closed-loop feedback mechanism that allows power consumption and temperature to influence and constrain each other, and a solution to the time-scale differences in these changes. Traditional transient simulations, if directly coupled for solution, would face the problems of enormous computational load and excessive time consumption, making them unsuitable for engineering practice. Therefore, to inject a dynamic and intelligent soul into this static physical environment, this application constructs an efficient simulation framework capable of simulating closed-loop feedback by loading a machine learning agent model that can quickly predict temperature and a throttling logic script that can simulate a real temperature control strategy, and sets a clear simulation starting state for it.

[0037] Firstly, it relies on a pre-trained machine learning proxy model. The training process of this model is the cornerstone of achieving the technical effect of this invention. It should be understood that the core challenge in constructing a simulation method capable of simulating the dynamic closed-loop power consumption-temperature relationship of a mobile phone lies in how to efficiently establish the instantaneous correlation between power consumption changes and temperature response. While traditional finite element physical simulation is accurate, its solution process involves iterative calculations of large-scale linear equations, which can take minutes or even hours. If a complete physical solution is performed in every tiny time step of the transient simulation to predict the temperature for throttling logic decisions, the computational cost of the entire simulation process will be too high, and the time will be too long to be practically applicable in engineering. Therefore, a high-fidelity and computationally inexpensive digital surrogate is created by training a machine learning proxy model. This surrogate learns from a large amount of pre-calculated physical simulation data, mastering the complex nonlinear mapping law from power consumption to temperature, thus enabling it to replace the traditional CAE solver to complete the temperature prediction task at millisecond speeds, providing the technical prerequisite for achieving efficient and fast closed-loop feedback simulation.

[0038] In an exemplary operation, the training process of the machine learning agent model includes: S11: inputting a power consumption vector formed by power consumption combination into the three-dimensional model of the mobile phone capable of finite element analysis to obtain a temperature vector composed of the temperatures of key monitoring points; S12: encapsulating the power consumption vector and the temperature vector into data pairs as training samples; S13: repeatedly executing steps S11 and S12 to obtain a training sample set; S14: training the machine learning agent model based on the training sample set.

[0039] Specifically, in step S11, the process first needs to identify the main heat sources inside the phone that exhibit significant power consumption variations. In this embodiment, four key heat sources are selected: the large core cluster of the central processing unit (CPU) of the system-on-a-chip (SoC), the graphics processing unit (GPU), the 5G radio frequency front-end module (RF), and the OLED display. Next, the power consumption variation range that these heat sources may experience in actual use is defined. These ranges are set based on power consumption model data provided by chip manufacturers, laboratory test data, or industry experience. For example, the power consumption range of the CPU is set to [0.5 watts, 8.0 watts], the power consumption range of the GPU is [0.5 watts, 10.0 watts], the power consumption range of the 5G RF module is [0.3 watts, 2.5 watts], and the power consumption range of the display is [0.5 watts, 2.0 watts]. Then, sampling needs to be performed from this four-dimensional power consumption space to generate a series of power consumption combinations. To ensure that the samples are evenly distributed and representative throughout the entire power consumption space, and to avoid data clustering in certain local areas, this embodiment uses the Latin hypercube sampling method. This method divides the value range of each power source into equally probable intervals and ensures that only one sample point is extracted in each interval, thereby achieving efficient coverage of the entire parameter space with a small number of samples. This method generates a list containing, for example, 5000 different power combinations. Each power combination is a power vector. For example, one power vector can be represented as: [P_cpu=6.5W, P_gpu=8.2W, P_rf=2.1W, P_display=1.2W]. For each generated power vector, it needs to be input as a heat source into a pre-built 3D mobile phone model capable of finite element analysis, i.e., the basic simulation environment. The input process here is a precise and automated physical load application process. In the preprocessing stage of step S1, the geometric entities corresponding to heat sources such as the CPU, GPU, RF, and display (e.g., the silicon chip portion of the SoC) have been explicitly identified and assigned specific names or labels. The automated script here reads each component in the power vector and matches it with the corresponding geometric entity. For example, if the script detects a CPU power consumption of 6.5 watts, it will query the volume of the geometry labeled CPU_die in the base simulation environment, such as its volume being 72 cubic millimeters. The script then calculates the volumetric heat generation rate, which is 6.5 watts divided by 72 cubic millimeters, yielding a specific numerical unit of watts per cubic meter. This volumetric heat generation rate is then applied as a physical load to all finite element mesh elements constituting the CPU_die geometry. Similarly, the GPU power consumption of 8.2 watts, the RF power consumption of 2.1 watts, and the display power consumption of 1.2 watts will also be calculated as their respective volumetric heat generation rates and applied to all mesh elements of their corresponding geometries.Subsequently, after correctly applying the loads defined by the power consumption vector to all heat source components of the model, a steady-state thermal simulation is performed using the CAE solver. The steady-state simulation calculates the global temperature field distribution when the phone reaches thermal equilibrium under this specific power consumption combination. After the simulation, the temperature values ​​of pre-defined key monitoring points are extracted from this temperature field. These monitoring points are the locations that the phone's temperature control strategy actually focuses on, such as: the center point temperature T_soc of the SoC package surface, the highest temperature T_battery of the battery cell surface, the average temperature T_frame of the metal frame area that the user's hand mainly contacts, and the highest temperature T_screen of the screen display area. These extracted temperature values ​​constitute a temperature vector corresponding to the input power consumption vector. For example, corresponding to the above power consumption vector, the obtained temperature vector might be: [T_soc=89.1℃, T_battery=43.5℃, T_frame=42.0℃, T_screen=45.5℃].

[0040] Next, in step S12, each power consumption vector obtained in step S11 and its unique corresponding temperature vector are encapsulated into a data pair. This data pair constitutes a complete training sample, where the power consumption vector is the input feature of the model, and the temperature vector is the target label that the model needs to learn and predict.

[0041] Next, in step S13, steps S11 and S12 are repeated. Through automated script control, all 5000 power consumption vectors generated by Latin hypercube sampling are subjected to steady-state simulation and data encapsulation one by one. After all calculations are completed, a dataset containing 5000 independent training samples is obtained. This training sample set constitutes the entire knowledge source for the machine learning agent model's learning.

[0042] Finally, in step S14. In this embodiment, the machine learning proxy model used is a deep neural network, specifically a multilayer perceptron (MLP) structure with multiple hidden layers. The specific architecture of this MLP model is as follows: The input layer of the model contains 4 neurons, the number of which strictly corresponds to the dimension of the power consumption vector, corresponding to the four input features in the power consumption vector: CPU power consumption, GPU power consumption, 5G RF power consumption, and display power consumption. The signal enters three fully connected hidden layers from the input layer. The first layer has 128 neurons, the second layer expands to 256 neurons, and the third layer shrinks back to 128 neurons. This bottleneck structure, from narrow to wide and then back to narrow, helps the model learn more representative features in the data. That is, the wider layers in the middle can provide the model with sufficient capacity to learn and combine the highly complex nonlinear interactions between input features, while the subsequent shrinking layers help the model refine and integrate these learned features to form a more generalized and robust internal representation of the temperature response, while suppressing overfitting to a certain extent. Each neuron in the hidden layer uses the modified linear unit ReLU as the activation function. The ReLU function, f(x) = max(0,x), introduces non-linearity, making it suitable for fitting the complex relationship between power consumption and temperature arising from various physical phenomena such as heat conduction, convection, and radiation. It also helps alleviate the vanishing gradient problem and accelerates model convergence. The model's output layer contains four neurons, corresponding to the dimension of the target temperature vector, representing the four target values ​​to be predicted: SoC temperature, battery temperature, mid-frame temperature, and screen temperature. The neurons in the output layer use linear activation functions because temperature is a continuous and unbounded physical quantity, and linear output directly matches its numerical range. The training process unfolds as follows: First, the connection weights between all neurons in the model and the biases of each neuron are randomly initialized. Then, the training sample set containing 5000 samples is pre-divided into a training set (4000 samples) and a validation set (1000 samples) in an 8:2 ratio. The training process is iterative, with each complete iteration traversing the entire training set. In practice, to improve training stability and computational efficiency, the training set is not input into the model all at once, but is divided into several small batches, for example, each batch containing 64 samples. In each forward propagation, the power vector of one batch is fed into the model, and matrix multiplication and activation function calculations are performed layer by layer according to the network structure, ultimately resulting in a batch of 64 predicted temperature vectors at the output layer. Then, a loss function is used to accurately quantify the difference between the predicted and true values ​​of this batch. This embodiment uses the mean squared error (MSE) as the loss function, which calculates the average of the sum of squares of the differences between the predicted and true temperature vectors of all samples within a batch. This MSE value intuitively reflects the accuracy of the model's current predictions; the smaller the value, the better the model performance.To minimize this loss function, an adaptive moment estimator (Adam) is employed. The Adam optimizer maintains an independent, adaptive learning rate for each parameter (weight and bias) in the model based on the gradient calculated from the loss function, and incorporates momentum (the exponentially decaying average of past gradients) to smooth the update process, enabling it to converge quickly and stably to the global or local minimum of the loss function. The core mechanism for this parameter update is implemented through the backpropagation algorithm. Starting from the output layer, this algorithm uses the chain rule to calculate the partial derivatives (gradients) of the loss function with respect to the parameters of each layer, indicating the direction of parameter adjustment. This forward propagation—loss calculation—backpropagation—parameter update process is repeated across all batches of the training set. After each iteration, the model performs a performance evaluation on an independent validation set, calculating the validation loss. The training process is repeated thousands or even tens of thousands of times until the loss value on the validation set no longer decreases significantly over several consecutive iterations, or the preset maximum number of training iterations is reached. Monitoring the validation loss to determine when to stop training is an effective strategy to prevent the model from overfitting to the training data. After training, all weights and bias parameters within the model are optimized from initial random values ​​to a set of highly optimized, specific values ​​that accurately reflect the power consumption-temperature mapping relationship. These fixed parameters, together with the model's network structure, constitute a trained machine learning proxy model. Finally, the model, along with its complete parameter set, is saved in a standardized, portable file format (such as HDF5 or ONNX) for rapid loading and recall in subsequent transient simulation steps, enabling near-instantaneous temperature prediction.

[0043] Specifically, when performing mobile phone thermal simulation, if a traditional multilayer perceptron (MLP) is used as a surrogate model, it essentially treats the input power consumption vector as a flat, unstructured list of features and directly maps it to the output temperature vector. This approach ignores a crucial physical fact within the mobile phone: heat does not instantly affect all monitoring points from the power consumption source, but rather follows specific physical laws, being conducted along a three-dimensional path composed of different thermally conductive materials. For example, the heat generated by the chip must pass sequentially through thermally conductive gel, a vapor chamber, and a graphite sheet before finally reaching the casing. Therefore, the power consumption of one component has a much greater impact on the temperature of its neighboring components than on components further away. Multilayer perceptrons lack this structural inductive bias, making it difficult to efficiently learn and express this complex spatial dependency determined by the physical layout. Therefore, this application introduces a graph neural network as a machine learning surrogate model, explicitly encoding the physical heat dissipation structure of the mobile phone into the model architecture. This allows the model to naturally follow the physical topology of heat conduction when learning the mapping relationship from power consumption to temperature, thereby significantly improving the accuracy and physical realism of predictions.

[0044] Based on this, in a preferred exemplary operation, the machine learning agent model is trained to obtain a machine learning agent model trained based on a graph structure, wherein the machine learning agent model is a graph neural network, comprising:

[0045] Power consumption and temperature vectors are obtained as training samples. It's understandable that machine learning proxy models need to learn patterns from a large number of question-answer pairs. Here, the question is the power consumption distribution of the phone under specific operating conditions, and the answer is the stable temperature field ultimately caused by that power consumption distribution. To generate these high-quality training samples, high-precision CAE physical simulation software is required. Specifically, the power consumption variation range of all major heat sources in the phone (such as CPU, GPU, RF, display, etc.) is first defined. Then, advanced experimental design methods such as Latin hypercube sampling are used to generate thousands of widely covered and evenly distributed power consumption vector combinations within these ranges. Each power consumption vector is used as input to run a complete CAE steady-state thermal simulation until the temperature field no longer changes. After the simulation, the temperature values ​​of key monitoring points (such as SoC, battery, specific points on the casing) are extracted to form temperature vectors that correspond one-to-one with the input power consumption vectors, for example. This pair of power consumption and temperature vectors constitutes a complete training sample. The effect of this step is to build a large and diverse training dataset, which provides sufficient and reliable data for subsequent model training, ensuring that the model can learn the thermal response behavior under various complex operating conditions. In other words, the acquisition process here is the same as the exemplary acquisition process described above.

[0046] A directed weighted graph is constructed to represent the heat dissipation structure of a mobile phone. This graph consists of nodes, edges, and edge weights. Nodes represent the various thermodynamic components of the phone, and their characteristics are static attributes and the corresponding power values ​​in the power consumption vector. Edge weights are calculated from the three-dimensional distance and contact thermal resistance between nodes. Accordingly, the heat dissipation performance of a mobile phone is not solely determined by power consumption, but more importantly by the physical layout of the components and the heat conduction paths between them. In practice, the heat dissipation structure of the phone is first abstracted as a graph structure. Key thermodynamic components in the phone, such as the CPU core, GPU, PMIC, battery, heat sink, and casing, are defined as nodes in the graph. Then, each node is assigned two types of features: static attributes, such as component volume and material specific heat capacity extracted from the CAE model (these attributes remain constant in the simulation); and dynamic attributes, namely the power values ​​corresponding to that node in the power consumption vector of the current training sample. Next, the heat conduction topology between nodes needs to be characterized, i.e., the edges and edge weights of the graph are constructed. Two core physical factors need to be considered: one is the three-dimensional distance between nodes. That is, the Euclidean distance from the geometric center of the component. ,in, , and These are the coordinate differences along the X, Y, and Z axes, respectively. It is a node and nodes The distance between them is a three-dimensional spatial distance, i.e., the Euclidean distance; the other part is based on the thermal conductivity of the contact, which is determined by the contact thermal resistance. The edge weight is measured by its reciprocal. Therefore, the edge weight can be expressed as... ,in, It is a node and nodes Contact thermal resistance occurs when physical contact takes place. Thermal resistance is a physical quantity that measures the ease with which heat is transferred; a higher value indicates that heat is more difficult to transfer. For example, the contact thermal resistance between a CPU and a tightly fitted heat spreader. The value is very small, and it is different from the value of a battery that is not in contact with the distance. The value approaches infinity. It is a node and nodes The edge weights between the components are scalar values ​​representing the strength of their thermal association. This constructs a static, directed, weighted graph rich in physical information, acting like a physical map. It provides structured path guidance for information transmission in the subsequent graph neural network, ensuring the physical consistency of the model's inference.

[0047] A graph neural network is trained using power consumption and temperature vectors to obtain a machine learning proxy model trained on the graph structure. In other words, after preparing the training data and graph structure, an effective learning mechanism is needed to adjust the model's internal parameters. Specifically, an iterative training loop is used. In each iteration, a sample is taken from the training dataset. The power consumption value of this sample is used as a dynamic feature, combined with static attributes to form a complete node feature vector, and input into the graph neural network along with the static directed weighted graph constructed in the previous step. Inside the network, through a multi-layer message passing mechanism, each node aggregates information from its neighbors, with the importance of neighbor information adjusted by the edge weights connecting them. After several layers of propagation, each node obtains a rich representation that integrates its neighborhood structure and features. Finally, a fully connected layer maps the final node representation to the predicted temperature value. The predicted temperature vector output by the model is compared with the true temperature vector of the sample, and the difference between the two is calculated using loss functions such as mean squared error. Then, the gradient of the loss with respect to all learnable parameters of the model is calculated using the backpropagation algorithm, and these parameters are updated using optimizers such as Adam to minimize the loss in the next prediction. This process is repeated until the model's total loss across all training samples converges to a sufficiently small value. The result is a well-trained machine learning proxy model that deeply understands the physical topology of heat conduction in a mobile phone. This model can quickly and accurately predict the temperature field, conforming to physical laws, upon receiving new power consumption inputs. The machine learning proxy model trained on a graph structure hard-coded the phone's internal physical heat dissipation topology into the network architecture. This allows the model to naturally follow the physical laws of heat conduction when predicting power consumption to temperature, avoiding false predictions caused by the lack of structural information in traditional models. Therefore, it not only significantly improves prediction accuracy but also ensures the physical authenticity of the simulation results, providing a highly reliable and efficient temperature field input for subsequent temperature control throttling closed-loop simulations, ensuring the stability and accuracy of the entire dynamic simulation process.

[0048] It's worth noting that while static graph structures based on physical layout can depict the basic paths of heat conduction, they cannot reflect the dynamic changes in the intensity and direction of heat flux caused by instantaneous power consumption variations. For example, during the operation of demanding games, the high heat flux state of a chip and its low heat flux state during standby should physically correspond to drastically different heat conduction weight distributions. Static graphs cannot capture this difference, potentially leading to deviations in the model's prediction accuracy under different operating conditions. Therefore, enhancing and refining the static edge weights aims to dynamically incorporate the real-time physical state of heat sources, i.e., the difference in volumetric heat source generation rates, into the edge weight calculation. This allows the graph's topology to simulate real-time heat flux trends, thereby improving the consistency between the model and the underlying CAE solver in terms of physical mechanisms and ensuring the physical self-consistency of the prediction results.

[0049] Based on this, in another preferred exemplary operation, the machine learning agent model is trained to obtain a machine learning agent model trained based on a graph structure, wherein the machine learning agent model is a graph neural network, including:

[0050] The power consumption vector and temperature vector are obtained as training samples; the heat dissipation structure of the mobile phone is constructed to obtain a directed weighted graph, which consists of nodes, edges and edge weights. The nodes are the various thermodynamic components of the mobile phone, and the node features are the static attributes and the corresponding power values ​​in the power consumption vector. The edge weights are calculated from the difference between the three-dimensional distance between nodes, the contact thermal resistance and the volume heat source generation rate. The implementation and description of the nodes and node features here are the same as those in the above-mentioned preferred method, so they will not be described again.

[0051] The volumetric heat source generation rate difference between any two nodes is calculated to obtain the volumetric heat source generation rate difference for each node. It should be understood that, according to the laws of thermodynamics, heat always flows from regions of high heat source intensity to regions of low intensity; this intensity difference directly determines the trend and magnitude of heat flow. In specific implementation, for each heat source node in the current input power consumption vector, its power value is divided by its corresponding component volume, based on its known static volume, to obtain the volumetric heat source generation rate for that node. Then, for any two nodes in the graph... and nodes Calculate the difference in their volumetric heat source generation rates. This difference is a dynamic quantity that changes in real time with the power consumption vector, representing the potential intensity and direction of heat flow between the two points. In this way, the abstract power consumption input is transformed into a physical quantity that can directly characterize the instantaneous heat flow intensity and direction, providing a crucial quantitative basis for the next step of constructing dynamic edge weights that are highly consistent with the physical process.

[0052] The edge weights of each edge are obtained based on the differences in heat source generation rates, three-dimensional distances, and contact thermal resistance for each volume, i.e.:

[0053]

[0054] in, and It is a node and nodes Volumetric heat source generation rate, It is a node and nodes The ω-coefficient is a fixed response coefficient representing the physical properties of heat conduction between two points. It is a parameter calibrated empirically or experimentally to characterize the sensitivity of the temperature difference between two points relative to the difference in their volumetric heat source generation rates. For example, 0.2 indicates a relatively sensitive thermal response. For instance, the ω-coefficient between a CPU and a heat spreader might be set to 0.2. It is a node and nodes The three-dimensional spatial distance between them, i.e., the Euclidean distance. It is a node and nodes Contact thermal resistance when physical contact occurs It is a node and nodes The edge weights between nodes are then calculated. Furthermore, static physical structure information is fused with dynamic heat flow driving force information to generate the final edge weights that accurately characterize the strength of the heat conduction path at the current moment. In practice, the above formula is used to calculate the edge weights between any two nodes. and nodes The edge weights between these edges create a dynamic graph topology that changes in real time with the current power consumption. The edge weights of this structure deeply embody the physical layout and real-time heat conduction information, enabling the graph neural network to maintain a high degree of consistency between its path and intensity and the actual physical heat flow during information transmission.

[0055] A graph neural network is trained using power consumption and temperature vectors to obtain a machine learning surrogate model trained on a graph structure. The implementation and explanation here are the same as the preferred method described above, and therefore will not be repeated. The surrogate model trained with these dynamic weights can deeply integrate real-time heat flow physical driving forces into the network structure. This allows the model to not only perceive the static physical layout but also understand how power consumption changes dynamically reshape the strength of the heat conduction path. Therefore, when making temperature predictions, it can achieve a higher level of consistency with the physical mechanism of the CAE solver, greatly improving prediction accuracy and physical consistency. In subsequent closed-loop simulations, this high-fidelity temperature field input ensures that the decision-making basis for the temperature control throttling strategy is more accurate and reliable, thereby significantly enhancing the stability of the entire dynamic simulation system and the effectiveness of the final results.

[0056] After the machine learning agent model is trained, step S2 is executed to integrate the trained agent model with the explicit throttling logic and the initial power consumption profile, thereby building a complete and ready-to-run framework for the subsequent closed-loop transient solution.

[0057] In one exemplary operation, step S2 is implemented as follows: First, the machine learning agent model that has been trained is loaded. Specifically, a master simulation script (e.g., a Python script) reads the standard format file (e.g., an HDF5 or ONNX file) storing the model. This file contains the complete network architecture of the model (number of neurons and connections in the input, hidden, and output layers) and all trained weights and biases. The master script reads this information into program memory, thereby instantiating a directly callable model object or function.

[0058] Next, the throttling logic script is loaded and parsed. This script is another key component for implementing dynamic closed-loop feedback. It defines the originally black-box temperature control strategy in the phone's internal firmware or operating system kernel in an explicit and editable form. In this embodiment, the throttling logic script is a separate text file written in Python. Its core content is one or more temperature control decision functions. For example, the script defines a function called thermal_throttling_policy. The input parameters of this function are designed to be a temperature vector that is completely consistent with the output of the machine learning agent model, i.e., a list containing key temperature values ​​such as [T_soc, T_battery, T_frame, T_screen]. The output of this function is a new target power consumption vector with the same dimension as the input power consumption vector. The function internally implements the specific temperature control and frequency reduction logic of the phone, which is precisely set and verified by the phone manufacturer or chip designer through a large number of experimental tests and engineering debugging based on the hardware's thermal tolerance, performance goals, and user experience standards. For example, its internal logic might include the following conditional checks: setting the SoC performance throttling threshold to 88 degrees Celsius and the emergency throttling threshold to 95 degrees Celsius; setting the battery's safe temperature threshold to 45 degrees Celsius. When the function receives a predicted temperature vector, it performs the following checks: if the predicted SoC temperature T_soc exceeds 95 degrees Celsius, the target CPU power consumption is forcibly reduced to its base power consumption, such as 2.0 watts, and the GPU power consumption is reduced by 60%; if T_soc is between 88 and 95 degrees Celsius, the target power consumption of the CPU and GPU is proportionally reduced from their current values ​​according to a linear or piecewise function; if the battery temperature T_battery exceeds 45 degrees Celsius, the charging power (if a simulation variable) is limited to a low level regardless of the SoC temperature, and the transmit power of the 5G RF module may also be limited. It checks whether the screen display area temperature T_screen46.5°C exceeds the threshold of 47 degrees Celsius. If any surface temperature exceeds its threshold, even if the internal component temperatures are within safe limits, the function should execute the predetermined power reduction strategy. The process of loading this script involves the main control simulation script using import commands to import it, making the thermal_throttling_policy function a callable part of the main control script.

[0059] Next, an initial power consumption profile needs to be defined. This is the starting point for the entire transient simulation, representing the system load state at time t=0. This initial power consumption profile is a specific power consumption vector, whose value represents the initial peak load of the specific user scenario to be simulated. For example, to simulate a scenario running a large 3D game, the initial power consumption profile could be defined as [P_cpu=8.0W, P_gpu=10.0W, P_rf=1.5W, P_display=1.8W]. This vector represents the instantaneous maximum power consumption reached by the phone to provide ultimate performance before the temperature control mechanism intervenes. To simulate a high-definition video call scenario, the initial power consumption profile might be set to [P_cpu=3.5W, P_gpu=2.0W, P_rf=2.2W, P_display=1.5W]. This initial power consumption profile will serve as the first input to the simulation loop.

[0060] Finally, the loaded machine learning agent model, throttling logic script, and defined initial power consumption profile are input into the basic simulation environment. This input is a logical integration and preparation step, marking the transformation of the simulation environment from a static physical model into a complete framework with dynamic decision-making capabilities. Specifically, the master simulation script has now completed all preparations: it holds a reference or path to the basic simulation environment (i.e., the solver input file) generated in step S1; it instantiates the machine learning agent model in memory and prepares the throttling logic function for immediate use; simultaneously, it stores the defined initial power consumption profile as a variable, serving as the first input for the upcoming transient simulation loop. Thus, a dynamic closed-loop thermal simulation environment has been constructed.

[0061] In step S3, the power consumption vector at the current simulation moment is input into the trained machine learning proxy model to obtain the predicted temperature vector. It is understood that after constructing the simulation framework containing intelligent decision-making components and setting the initial power consumption, the entire simulation process enters the core dynamic iterative loop. In this loop, the primary task is to establish an instantaneous correlation between power consumption and temperature to provide a basis for subsequent temperature control decisions. However, if a traditional CAE solver is called at each time step to calculate the temperature field, the minute-level time consumption would make simulating a complete dynamic scenario impractical, completely violating the original intention of efficient simulation. Therefore, this application utilizes a trained machine learning proxy model to predict the temperature corresponding to the current power consumption state, replacing the expensive physical solution with millisecond-level computational cost. This efficiently provides decision input for the throttling logic and is a key step in achieving efficient operation of the entire closed-loop feedback simulation.

[0062] In an exemplary operation, step S3 is implemented as follows: the power consumption vector at the current simulation moment is the initial power consumption profile defined in step S2 at the beginning of the entire transient simulation, i.e., t=0. For example, taking the running of a large 3D game scene as an example, at t=0, the input power consumption vector is [P_cpu=8.0W, P_gpu=10.0W, P_rf=1.5W, P_display=1.8W]. In subsequent simulation moments (t>0), this power consumption vector is the new target power consumption vector determined by steps S4 and S5 in the previous simulation cycle.

[0063] The trained machine learning agent model, namely the Multilayer Perceptron (MLP) mentioned earlier, has been instantiated in memory. Its network structure consists of four input neurons, three hidden layers with 128, 256, and 128 neurons respectively (using the ReLU activation function), and four output neurons (using the linear activation function). The millions of weights and bias parameters stored internally in the model are the optimization results obtained after learning from 5000 sets of power consumption-temperature data pairs; these parameters contain a highly condensed mathematical expression of the phone's thermophysical characteristics.

[0064] When the power consumption vector at the current simulation moment is input into the model, the model performs a complete forward propagation calculation to obtain the predicted temperature vector. The calculation process is as follows: First, the original input power consumption vector, for example [8.0, 10.0, 1.5, 1.8], is fed into the input layer of the network. Then, this vector is multiplied by the 4x128 weight matrix connecting the input layer and the first hidden layer, and then the 128-dimensional bias vector of that layer is added. The result of this linear transformation is processed by the ReLU activation function, where all negative values ​​are set to zero and positive values ​​remain unchanged, thus obtaining the output of the first hidden layer. Then, the output of the first hidden layer is used as input to continue forward propagation to the second hidden layer. It is multiplied by the 128x256 weight matrix connecting the first and second hidden layers, added to the 256-dimensional bias vector of the second hidden layer, and then processed by the ReLU activation function. This process is repeated in the third hidden layer, where the 256-dimensional vector is transformed and activated using a 256x128 weight matrix and a 128-dimensional bias vector, resulting in the output of the third hidden layer. Finally, the 128-dimensional output vector of the third hidden layer is fed into the output layer and multiplied by the 128x4 weight matrix connecting the third hidden layer and the output layer, and then the 4-dimensional bias vector of the output layer is added. Since the output layer uses a linear activation function, the result of this linear transformation is the final output of the model, a 4-dimensional prediction vector. After the above algebraic operations, the model finally outputs a predicted temperature vector. Taking the aforementioned input power consumption vector as an example, the model may calculate and output the predicted temperature vector [T_soc=93.5℃, T_battery=44.1℃, T_frame=43.8℃, T_screen=46.5℃] within milliseconds. This output predicted temperature vector precisely corresponds to the steady-state or quasi-steady-state temperature that the key monitoring points of the mobile phone will reach under the current power consumption state in the simulation.

[0065] In step S4, the predicted temperature vector is input into the throttling logic script to obtain a new target power consumption vector. It should be understood that this predicted temperature vector is merely a state description and cannot directly drive the simulation's evolution. In the actual operation of a mobile phone, it is based on the judgment of the current and future temperatures that the temperature control subsystem takes a series of proactive intervention actions such as frequency reduction and voltage reduction to ensure device safety and user experience. Without this decision-making step, the simulation will remain in an open-loop state, unable to reflect the core dynamic closed-loop feedback between power consumption and temperature. Therefore, in order to simulate the decision-making brain of the mobile phone's internal temperature control strategy, this application transforms the temperature prediction information into specific, executable power consumption adjustment instructions.

[0066] In one exemplary operation Figure 3 This is a flowchart of step S4 in the mobile phone thermal simulation method based on multiphysics coupling according to an embodiment of this application. Figure 3 As shown, step S4 includes: S41, loading the temperature control decision function from the throttling logic script; S42, inputting the predicted temperature vector into the temperature control decision function to obtain the new target power consumption vector.

[0067] In the above example, step S4 is implemented as follows: First, in step S41, this step logically prepares for executing the temperature control decision. In step S2, a throttling logic script defining the complete temperature control strategy of the mobile phone has been loaded into the runtime environment of the simulation main control program. This script is a text file written in a high-level language such as Python, which can be directly read and modified by engineers, and contains one or more core temperature control decision functions. Looping into loading the temperature control decision function at each time step of the transient simulation means that the main control simulation process calls the specific function representing the temperature control strategy that is already prepared in memory at this moment. In this embodiment, this function is named thermal_policy_decision. The design of this function follows strict input-output specifications: its input parameter is a temperature vector, the structure of which is completely consistent with the predicted temperature vector output by the machine learning proxy model; at the same time, in order to perform differential or proportional adjustments, it also needs to receive the power consumption vector at the current moment as another input. The return value of this function is a new power consumption vector, i.e., the new target power consumption vector. The `thermal_policy_decision` function is a precise coded representation of the phone's temperature control logic. Its internal logic and parameters, such as temperature thresholds, frequency throttling, and response latency, are all set by thermal design engineers based on product specifications, firmware design documents, or observation data from stress tests on real devices. By making this logic an independent, editable script function, this invention achieves transparency and configurability of the temperature control strategy, greatly enhancing the flexibility and realism of the simulation.

[0068] Next, in step S42, the main control simulation program takes the predicted temperature vector generated in step S3 and the power consumption vector at the current moment as input and calls the previously loaded thermal_throttling_policy temperature control decision function. For example, at the current simulation time t, the input power consumption vector in step S3 is [P_cpu=8.0W, P_gpu=10.0W, P_rf=1.5W, P_display=1.8W], and its output predicted temperature vector is [T_soc=93.5℃, T_battery=44.1℃, T_frame=43.8℃, T_screen=46.5℃]. After receiving these two vectors, the function's internal execution logic, which was set in step S2, begins to run. First, the function initializes a new power consumption vector, whose initial value is equal to the input power consumption vector. Subsequent logic modifies this. Then, the function judges a series of temperature control rules according to a preset, high-to-low priority order. Next, the function checks each temperature control rule one by one according to a preset priority order. The first step is to evaluate the most critical SoC temperature. The set emergency throttling threshold is 95 degrees Celsius. Its predicted value, T_soc93.5℃, is compared with the set 95℃ emergency throttling threshold. In this example, it is not exceeded, so this rule is skipped. If it is exceeded, the function will execute the most severe frequency reduction strategy, such as directly setting the CPU and GPU power consumption in the new power vector to the predefined minimum safe power consumption values, such as P_cpu=2.0W and P_gpu=1.5W, to prevent hardware damage. Since emergency throttling was not triggered, the function then checks the regular performance throttling rules, with a threshold set at 88 degrees Celsius. The function determines whether T_soc93.5℃ is within the range of 88℃ to 95℃. In this example, the condition is met. Therefore, the function calculates the power reduction based on a predefined frequency reduction curve. For example, a linear interpolation strategy is used: (T_soc-88) / (95-88)≈0.786, where this factor represents the relative position of the current temperature within the throttling range. Then, the function calculates the new power consumption based on the maximum allowable reduction ratio for different components. For instance, within this range, the CPU can reduce its current performance power consumption by a maximum of 70%, and the GPU by a maximum of 80%. The new target CPU power consumption is calculated as 8.0*(1-0.786*0.70)≈3.60W. The new target GPU power consumption is calculated as 10.0*(1-0.786*0.80)3.71W. The function updates the CPU and GPU components in the new power consumption vector with these two newly calculated values. The second step involves checking the battery safety rules. The set safety threshold is 45 degrees Celsius. The function determines whether T_battery44.1℃ exceeds 45℃. In this example, it does not exceed 45℃, and this rule is not triggered.If the temperature exceeds the limit, even if the SoC temperature is normal, the function may force a reduction in the power consumption of heat sources that significantly affect battery temperature. For example, it might limit the power level of the 5G RF module, or significantly reduce charging power consumption if the model includes it. The third step checks surface temperature comfort rules. For example, the comfort threshold for the mid-frame grip area is set to 44 degrees Celsius. The function checks if T_frame43.8℃ exceeds 44℃. In this example, it does not. If it does, even if the internal component temperatures are all within safe limits, the function may perform a mild, global power reduction, such as uniformly reducing the power consumption of all major heat sources by 5% to improve the user's grip experience. The fourth step checks the set screen comfort rules. The screen comfort threshold is set to 47 degrees Celsius, and the function checks if T_screen46.5℃ does not exceed this value. If any surface temperature exceeds its threshold, even if the internal component temperatures are all within safe limits, the function should execute a predetermined power reduction strategy, such as uniformly reducing the power consumption of all major heat sources by 5% to improve the user experience. After all rules have been checked, the thermal_throttling_policy function returns the final modified vector, which is the new target power vector, such as [3.60W, 3.71W, 1.5W, 1.8W].

[0069] In step S5, the new target power consumption vector is used as the heat source, and the temperature field at the previous simulation moment is used as the initial condition. This is then input into the CAE solver to obtain the temperature field at the current simulation moment. Correspondingly, as a physical entity, the temperature of a mobile phone does not jump instantaneously due to changes in power consumption commands. Instead, it gradually evolves over time according to the physical laws of heat conduction, convection, and radiation, a process exhibiting significant thermal inertia. Therefore, this application further uses the new target power consumption vector as the heat source and the temperature field at the previous simulation moment as the initial condition, inputting it into the CAE solver to perform a precise transient thermophysical solution. This calculates the true and complete temperature field distribution inside the mobile phone under the new power load after a small time step, thus advancing the simulation state from the current moment to the next moment. This is the physical basis and core computational component for simulating the entire dynamic process.

[0070] In an exemplary operation, step S5 includes: S51, extracting finite element mesh data from the 3D mobile phone model capable of finite element analysis; S52, extracting the power values ​​of each heat source term from the new target power consumption vector; S53, locating all mesh cells associated with each heat source term in the finite element mesh data; S54, dividing the power value of each heat source term by the total volume of all mesh cells associated with each heat source term to obtain the volumetric heat source generation rate; S55, the CAE solver calculates the discretized form of the time derivative term in the heat conduction equation based on the temperature field at the previous simulation time and the current simulation time; S56, constructing physical control partial differential equations based on the volumetric heat source generation rate and the discretized form of the time derivative term in the heat conduction equation; S57, discretizing the physical control partial differential equations spatially and temporally on the finite element mesh data to obtain a linear equation system; S58, the CAE solver performs physical field iterative solutions on the linear equation system to obtain the temperature field at the current simulation time.

[0071] In the above example, step S5 is implemented as follows: First, in step S51, the main simulation program locates the mobile phone 3D model file suitable for finite element analysis according to the preset path configuration. Then, the CAE solver or its embedded file parsing module starts reading and parsing the file's contents. This file is in text format, with a strict grammatical structure, organizing different types of data through specific keywords. The parser scans the file line by line, and when it encounters a keyword representing a node definition, such as *NODE, it knows that the following data block describes the information of all mesh nodes. For example, a line in the file might be: 1001,15.23,30.51,2.15. Reading this line, the parser knows that there is a node with ID 1001, whose coordinates in 3D space are (15.23,30.51,2.15). The parser continues to read all lines under this data block, reading in all the IDs and coordinate information of millions of nodes. The parser then continues scanning until it encounters a keyword representing a cell (or element) definition, such as *ELEMENT, TYPE=C3D4. This indicates that the following data block defines tetrahedral cells that make up the 3D solid. For example, a line in the file might be: 5001,1001,1002,1150,1428. The parser reads this line and knows that there exists a tetrahedral cell with ID 5001, which is composed of four nodes with IDs 1001, 1002, 1150, and 1428. The parser traverses the data block, reading the IDs of all cells and their corresponding node connectivity information. During the reading and parsing process, this information is dynamically constructed into efficient data structures within the solver. For example, the coordinates of all nodes are stored in a large array or hash table, allowing for quick indexing of their coordinates by node ID. The information for all cells is stored in another data structure, where each cell object contains its type (e.g., tetrahedral, hexahedral) and a list of pointers or indices to the node objects that constitute it. Finally, after all the mesh-related information in the file has been parsed, a complete finite element mesh data structure that corresponds exactly to the contents of the disk file is built in the solver's memory. This data structure in memory is the finite element mesh data, which is a digital object that contains the precise coordinates of all nodes and the topological connections of all elements, and can be directly accessed and manipulated by the program.

[0072] Next, in step S52, upon receiving a new target power consumption vector, for example [3.60, 3.71, 1.5, 1.8] (in watts), the process of extracting the power values ​​of each heat source item begins. This is a structured data parsing operation based on predefined mapping rules. These mapping rules are established during simulation method initialization, creating a fixed correspondence between the index position of each value in the power consumption vector and a specific physical heat source identifier. For example, this mapping relationship is defined as follows: the first element of the input vector is the CPU heat source, the second element is the GPU heat source, the third element is the RF heat source, and the fourth element is the Display heat source. Based on this rule, the parsing process deconstructs the input numerical vector to obtain the power values ​​of each heat source item, i.e., {"CPU":3.60,"GPU":3.71,"RF":1.5,"Display": 1.8}.

[0073] Next, in step S53, the key to achieving location and association lies in the naming operation of the geometric model during the CAE preprocessing stage (such as step S1). In this stage, geometries representing different physical components (e.g., a cube representing a CPU chip) are assigned names completely consistent with the heat source item identifier, such as CPU, GPU, etc. When these named geometries are divided into finite element meshes, all mesh elements generated within these geometries automatically inherit the name label or are grouped into an element set named with that name. This naming and classification process is a standard function of CAE software, recording this association information in a structured manner in the exported model file. This process iterates through the heat source item set output in step S52. For the first heat source item, CPU, in the set, the main program queries the finite element mesh data using CPU as the keyword. This query retrieves all mesh elements labeled CPU or directly extracts a predefined element set named CPU. The query result is a list containing all mesh element IDs associated with the CPU heat source item, for example, an array containing thousands of element IDs. This process is repeated for each heat source item. For example, the program will then search using "GPU" as the keyword to obtain a list of all grid cells that make up the GPU chip, then search using "RF" as the keyword to find the set of cells representing the 5G radio frequency module, and finally search using "Display" as the keyword to locate the set of cells representing the display's light-emitting layer. The result is a mapping or dictionary structure where the key is the heat source item identifier and the value is a list of corresponding grid cell IDs. For example: { "CPU": [cell ID_7001, cell ID_7002, ..., cell ID_8500], "GPU": [cell ID_9101, cell ID_9102, ..., cell ID_11200], ...}. This output data structure precisely establishes a correlation between macroscopic power values ​​and microscopic grid cells.

[0074] Next, in step S54, taking the CPU heat source item as an example, the specific implementation is as follows: First, the program obtains the CPU power value from the input, which is 3.60 watts. Simultaneously, it obtains a list of mesh cell IDs associated with the CPU, such as [cell ID_7001, cell ID_7002, ..., cell ID_8500]. Then, it calculates the total volume of this cell set. The program initializes an accumulator variable Total_Volume_CPU to zero. Then, it iterates through the CPU's cell ID list. For each cell ID in the list, such as cell ID_7001, the program queries the complete mesh data loaded in S51 for detailed information about that cell. This information includes the cell type (e.g., tetrahedral, hexahedral) and the coordinates of all the nodes constituting the cell. Using these node coordinates, the program calculates the volume of this individual cell using standard geometric formulas. For example, for a tetrahedral cell consisting of four nodes, its volume can be accurately calculated using the vector mixture product formula. The calculated volume of each individual element is accumulated into Total_Volume_CPU. After iterating through all elements associated with the CPU, the value of Total_Volume_CPU is the precise total volume of the CPU component in the finite element model. For example, the calculated result is 1.2e-8 cubic meters, or 12 cubic millimeters. Finally, the program performs a division operation to obtain the volumetric heat source generation rate of the CPU: CPU power value / Total_Volume_CPU = 3.60W / 1.2e-8m³ = 3.0e8 W / m³. This calculation process is repeated sequentially for all heat source items output in step S52. For example, the program will then calculate the total volume of the GPU component and divide its power value of 3.71 watts by that volume to obtain the volumetric heat source generation rate of the GPU, and so on. After calculating all heat source items, a new data structure is obtained, which maps the identifier of each heat source item to its calculated volumetric heat source generation rate. For example: {"CPU": 3.0e8, "GPU": 2.8e8, "RF": 5.0e7, "Display": 1.5e6}, where the unit is watts per cubic meter.

[0075] Next, in step S55, the temperature field at the previous simulation moment is the initial temperature condition of the entire transient simulation at the first time step (e.g., from t=0 to t=Δt). This temperature field is set before the simulation begins, typically as a uniform ambient temperature, for example, setting the initial temperature of all nodes to 25 degrees Celsius. In all subsequent time steps, this temperature field is the result of the temperature field at the current simulation moment calculated in step S58 of the previous simulation cycle, thus forming an iterative relationship over time. This discretization process is based on approximating the time partial derivative using the finite difference method. The time term in the transient heat conduction equation is ρc(∂T / ∂t), where ρ is the material density, c is the specific heat capacity, and ∂T / ∂t is the partial derivative of temperature with respect to time. The CAE solver uses an implicit time integration scheme, such as the backward Euler method, to discretize this partial derivative. Specifically, let the current simulation time be t_n, the previous simulation time be t_{n-1}, and the time step Δt = t_n - t_{n-1}. The time step Δt is a key parameter that needs to be preset, and its value needs to balance computational accuracy and efficiency; for example, it can be set to 0.5 seconds. For any point in the model, the partial derivative of its temperature with respect to time at time t_n can be approximated as: (∂T / ∂t) ≈ (T(t_n) - T(t_{n-1})) / Δt, where T(t_n) is the unknown temperature at the current simulation time, and T(t_{n-1}) is the known temperature at the previous simulation time. Therefore, the discretized form of the entire time derivative term is ρc*(T(t_n) - T(t_{n-1})) / Δt.

[0076] Subsequently, in step S56, the process of constructing the physical governing partial differential equation is based on Fourier's law of transient heat conduction and the law of conservation of energy. The general mathematical form of this physical governing partial differential equation is: ρc(∂T / ∂t) = ∇·(k∇T) + Q, where ρc(∂T / ∂t) is the transient term, representing the energy absorbed or released per unit volume of material as the temperature changes over time; ∇·(k∇T) is the conduction term, representing the net heat flux flowing into or out of a unit volume due to the temperature gradient via heat conduction; and Q is the source term, representing the heat generated within a unit volume. This step involves replacing the transient and source terms in this general equation with the specific forms calculated in the previous steps. Specifically, the CAE solver logically constructs this equation for each point in the model (or more precisely, each material domain). For regions with internal heat sources: taking the region where the CPU component is located as an example. The solver identifies the material properties of the region (density ρ_cpu, specific heat capacity c_cpu, thermal conductivity k_cpu) and obtains its volumetric heat source generation rate Q_cpu = 3.0e8 W / m³ from the input. Simultaneously, it uses the discretized time term obtained from S55. Therefore, at any point within the CPU region, the constructed physical control partial differential equation is: ρ_cpu*c_cpu*(T(t_n)-T(t_{n-1})) / Δt=∇·(k_cpu∇T)+3.0e8. In this equation, T(t_{n-1}) is the known temperature of that point at the previous moment, while T in T(t_n) and ∇·(k_cpu∇T) both refer to the unknown temperature field at the current moment. For regions without internal heat sources: taking the mid-frame or outer casing of a mobile phone as an example. These regions do not have active heating elements, so their volumetric heat source generation rate Q is zero. The solver identifies the material properties of this region (e.g., ρ_frame, c_frame, k_frame of aluminum alloy) and constructs the corresponding equation: ρ_frame*c_frame*(T(t_n)-T(t_{n-1})) / Δt=∇·(k_frame∇T)+0.

[0077] Next, in step S57, the time-part discretization of this process, which has already been completed in S55 and S56, focuses on spatial discretization. This process is achieved using a weighted residual method (such as the Galerkin method). The solver performs the following operations independently for each mesh element in the model (e.g., a tetrahedral element): First, within that element, the continuous unknown temperature field T(t_n) can be approximated by the unknown temperature values ​​of all nodes (e.g., four nodes) of that element, using a set of predefined polynomial functions (called shape functions or interpolation functions) that are only related to the spatial coordinates. Then, this approximate temperature field expression is substituted into the physical governing partial differential equations. Since it is an approximate solution, the left and right sides of the equations are not strictly equal, resulting in a residual. The idea behind the Galerkin method is to require that the integral of this residual multiplied by each shape function within the element domain be zero. The physical meaning of this mathematical operation is that the error of the approximate solution is minimized over the entire element in the space defined by the shape functions. By performing the weighted integration operation on each node (i.e., each shape function) of the element, the differential operator ∇·(k∇T) is transformed into an algebraic expression that depends only on the nodal temperature, ultimately generating a small system of linear equations for this element, in the form [k_e]{T_e}={f_e}. Here, [k_e] is a combination of the element's conduction matrix (derived from the integral of the conduction term) and mass matrix (derived from the integral of the time derivative term), containing information such as material thermal conductivity, density, specific heat capacity, element geometry, and time step. {T_e} is a column vector containing the unknown temperatures of all nodes in the element, and {f_e} is the element's load vector (containing the internal heat source Q, the temperature field T(t_{n-1}) at the previous time step, and contributions from boundary conditions such as convection and radiation). After calculating the element matrix [k_e] and element vector {f_e} for all millions of elements in the model, the solver performs a global assembly operation. It accumulates the corresponding elements of these small unit matrices and vectors into the corresponding positions of two global, initially empty large matrices [A] and vector {B} based on the node ID of each unit. The final output is this fully constructed large sparse linear equation system, whose standard form is [A]{T}={B}. Here, {T} is a global column vector containing the unknown temperature values ​​of all nodes in the model at the current time, which is the final solution target; [A] is the global coefficient matrix, reflecting the heat transfer characteristics and transient effects of the entire model; and {B} is the global load vector, reflecting the combined effects of all heat sources, initial conditions, and boundary conditions.

[0078] Finally, in step S58. In an exemplary operation, step S58, the CAE solver performs a physical field iterative solution on the linear equations to obtain the temperature field at the current simulation moment, including: the CAE solver performs a physical field iterative solution on the linear equations based on the conjugate gradient method and combined with an algebraic multiplex network to obtain the temperature field at the current simulation moment. Specifically, first, the CAE solver initializes the iteration process. It sets an initial guess value for the temperature vector {T} to be solved, such as using the temperature field at the previous moment as the initial guess, because the temperature field changes little between two adjacent time steps, which helps to accelerate convergence. Next, it enters the iterative loop of the conjugate gradient method. In each iteration, the conjugate gradient method does not directly solve the original equation, but gradually approximates the true solution through a series of carefully constructed calculation steps. It first calculates the residual vector corresponding to the current guessed solution, which precisely quantifies the degree of deviation of the current solution from the right-hand side {B} after being substituted into the original equation [A]{T}={B}. Then, it updates the solution vector based on the residual and a new search direction conjugate to all previous search directions. However, for large ill-conditioned equation systems generated by heat conduction problems, the condition number of the coefficient matrix [A] is very large, resulting in extremely slow convergence of the simple conjugate gradient method. Therefore, this embodiment introduces Algebraic Multigrid (AMG) as a preprocessor. Within each iteration of the conjugate gradient method, when calculations related to matrix [A] are needed, [A] is not used directly, but rather transformed by the AMG preprocessor. The AMG preprocessor automatically constructs a series of coarser grid problems with decreasing levels from the original fine-grid problem using pure algebra. It quickly solves a small problem at the coarsest level and then passes the solution information back to the finer levels for correction. This multi-level process efficiently eliminates errors at different frequencies in the solution, greatly improving the condition number of the original equation system and making the transformed equation system easier to solve. Therefore, the actual iterative process is the preprocessed conjugate gradient method (PCG). In each main iteration, when the conjugate gradient method needs to solve a preprocessed equation of the form Mz=r (where r is the current residual), it calls the AMG solver. The AMG approximates the equation by performing an efficient multigrid loop and returns the result z. This approximate solution z provided by the AMG is then used by the conjugate gradient method as a highly efficient search direction to update the solution vector and residuals of its main iteration. This master-slave combined iterative process continues until a preset convergence criterion is met. This criterion determines whether the norm (i.e., the magnitude) of the residual vector in the main iteration loop is less than a sufficiently small threshold. This threshold is determined empirically or experimentally based on the required accuracy and computational cost; for example, a convergence tolerance of 1e-6 can be set.When the solver calculates that the residual norm after a certain iteration is less than this value, it considers the current solution vector {T} to be sufficiently close to the true solution, and the iteration process terminates. The final output is the solution vector {T} obtained after the iteration converges. This vector is the temperature field at the current simulation moment, a complete dataset containing the precise temperature values ​​of each node in the model at the current moment. This temperature field is not only the final result of this round of physical solution, but will also serve as the temperature field of the previous simulation moment in the next simulation cycle, providing initial conditions for the next iteration calculation.

[0079] In step S6, it is determined whether the current simulation time is less than the total simulation duration. If the current simulation time is less than the total simulation duration, the process returns to steps S3 to S5; if the current simulation time is greater than or equal to the total simulation duration, the transient simulation ends. That is, in the process of steps S3 to S5, the method of this invention completes a full closed-loop calculation of a single time step from power input to temperature field output. However, the core objective of mobile phone thermal simulation is to reveal the dynamic process of temperature and power consumption evolving over time under continuous workload, rather than an isolated snapshot at a specific point in time. Therefore, executing step S6 establishes a top-level loop control and termination mechanism for the entire transient simulation process. It periodically checks the current simulation progress to determine whether to continue advancing the time step and repeating the core calculation process, or to orderly end the simulation after reaching the preset total duration, thereby ensuring that the entire dynamic process is completely simulated.

[0080] In an exemplary operation, step S6 is implemented as follows: Before the entire simulation task begins, a specific total simulation duration needs to be set. For example, to simulate a complete game scenario, the total simulation duration can be set to 1800 seconds (i.e., 30 minutes), determined based on the typical duration of the target application scenario (such as a high-load game). Simultaneously, a variable used to track simulation progress, the current simulation time, is initialized to 0. After completing one round of calculations from steps S3 to S5, for example, after completing the calculation from t=0 to t=0.5 seconds, the main control program enters the judgment logic of step S6.

[0081] First, the program reads the value of the current simulation moment. After the first loop, this value is 0.5. The program compares this with the preset total simulation duration of 1800 seconds. If the result is that the current simulation moment is less than the total simulation duration, for example, 0.5 < 1800, then a return operation is performed. This operation first updates the time variable, increasing the current simulation moment by a time step Δt, which is consistent with the value used in step S55, for example, 0.5 seconds, making the current simulation moment 1 second. Subsequently, the main control program redirects the execution flow to step S3. At this time, the new target power consumption vector generated in the previous step S4, and the temperature field calculated in step S5, will be seamlessly used as the initial conditions for the next loop. Specifically, the temperature field obtained in the previous step S5 will be used as the temperature field of the previous simulation moment in the next step S5 calculation, and the new target power consumption vector obtained in the previous step S4 will be used as the new input of the surrogate model in the next step S3, thereby driving the simulation to enter the calculation of the next time step.

[0082] This process of judging, updating, and returning will continue to repeat. For example, after the 100th loop, the current simulation time will be updated to 50 seconds. Since 50 seconds is still less than 1800 seconds, the simulation loop will continue.

[0083] If the result indicates that the current simulation time exceeds or equals the total simulation duration, the termination operation is executed. For example, after the 3600th loop, the current simulation time will be updated to 1800 seconds. At this point, the condition 1800 < 1800 is no longer true. The main control program will exit the loop and will not return to step S3. The simulation process then terminates. At this point, all key data such as the temperature field and power consumption vector for all time steps have been calculated and stored, and the entire transient simulation is declared complete.

[0084] In summary, a mobile phone thermal simulation method based on multiphysics coupling, as described in this application, has been elucidated. To address the shortcomings of existing open-loop simulations that cannot simulate the dynamic closed-loop relationship between power consumption and temperature, this application constructs a hybrid simulation that couples machine learning and physical solution. This innovatively introduces a machine learning proxy model and a throttling logic script. The machine learning proxy model is used to quickly predict the temperature based on the current power consumption, responding instantaneously to power consumption changes; the throttling logic script is used to simulate a realistic temperature control and frequency reduction strategy, dynamically adjusting the target power consumption for the next moment based on the predicted temperature. This makes the originally black-box temperature control logic explicit and constructs a closed-loop feedback path between power consumption and temperature. Finally, the adjusted power consumption is used as a heat source, and a traditional CAE solver performs physical field solutions. By replacing some of the time-consuming physical solutions with fast inference from machine learning, the problem of time-scale differences between heat conduction and electrical changes is effectively solved, achieving efficient and accurate simulation of the real performance and temperature behavior of mobile phones under long-term high-load scenarios.

[0085] Figure 4 This is a block diagram of a mobile phone thermal simulation system based on multiphysics coupling according to an embodiment of this application. Figure 4 As shown, the mobile phone thermal simulation system 100 based on multiphysics coupling according to an embodiment of this application includes: a model parsing module 110, used to parse a finite element analysis-compatible 3D mobile phone model using a preprocessor module of CAE software to obtain a basic simulation environment; a model script loading module 120, used to load a trained machine learning proxy model and a throttling logic script, define an initial power consumption profile, and input it into the basic simulation environment, wherein the throttling logic script defines the mobile phone temperature control strategy; and a temperature prediction module 130, used to input the power consumption vector at the current simulation moment into the trained machine learning proxy model to obtain a predicted temperature. The simulation includes a temperature vector; a new target power consumption generation module 140, which inputs the predicted temperature vector into the throttling logic script to obtain a new target power consumption vector; a simulation temperature field generation module 150, which uses the new target power consumption vector as a heat source and the temperature field of the previous simulation moment as an initial condition, and inputs it into the CAE solver to obtain the temperature field of the current simulation moment; and a simulation duration judgment module 160, which judges whether the current simulation moment is less than the total simulation duration. If the current simulation moment is less than the total simulation duration, the simulation is returned to the temperature prediction module and then to the simulation temperature field generation module. If the current simulation moment is greater than or equal to the total simulation duration, the transient simulation is terminated.

[0086] Here, those skilled in the art will understand that the specific operations of each step in the above-described mobile phone thermal simulation system based on multiphysics coupling have been referenced above. Figures 1 to 3 The description of the mobile phone thermal simulation method based on multiphysics coupling has been detailed, and therefore, its repeated description will be omitted.

Claims

1. A method for thermal simulation of a mobile phone based on multi-physics coupling, characterized in that, Comprise: S1: the preprocessor module of the CAE software analyzes the mobile phone three-dimensional model that can be subjected to finite element analysis to obtain a basic simulation environment, the basic simulation environment includes a structured data set of the following information, specifically including: complete finite element grid data, material attribute pointer corresponding to each unit, boundary condition defined on the surface, and thermal contact model defined on the component contact surface; S2: load the trained machine learning agent model and the throttling logic script, and define and input the initial power consumption profile into the basic simulation environment, wherein the throttling logic script defines the mobile phone temperature control strategy; S3: input the power consumption vector of the current simulation time into the trained machine learning agent model to obtain a predicted temperature vector; S4: input the predicted temperature vector into the throttling logic script to obtain a new target power consumption vector; S5: input the new target power consumption vector as a heat source and the temperature field of the last simulation time as an initial condition into the CAE solver to obtain the temperature field of the current simulation time; S6: determine whether the current simulation time is less than the total simulation time, if the current simulation time is less than the total simulation time, return to steps S3 to S5; if the current simulation time is greater than or equal to the total simulation time, end the transient simulation.

2. The multi-physics coupling based thermal simulation method for mobile phone according to claim 1, wherein, The training process of the machine learning agent model comprises: S11: input the power consumption vector formed by the power consumption combination into the mobile phone three-dimensional model that can be subjected to finite element analysis to obtain a temperature vector composed of key monitoring point temperatures; S12: encapsulate the power consumption vector and the temperature vector into a data pair as a training sample; S13: repeatedly execute steps S11 and S12 to obtain a training sample set; S14: train the machine learning agent model based on the training sample set.

3. The multi-physics coupling based thermal simulation method for mobile phone according to claim 1, wherein, Step S4 comprises: Load the temperature control decision function from the throttling logic script; Input the predicted temperature vector into the temperature control decision function to obtain the new target power consumption vector.

4. The multi-physics coupling based thermal simulation method for mobile phone according to claim 1, wherein, Step S5 comprises: Extract the finite element grid data from the mobile phone three-dimensional model that can be subjected to finite element analysis; Extract the power value of each heat source term from the new target power consumption vector; Locate all grid cells associated with the each heat source term in the finite element grid data; Divide the power value of the each heat source term by the total volume of all grid cells associated with the each heat source term to obtain the volumetric heat source generation rate; The CAE solver calculates the discretized form of the time derivative term in the heat conduction equation based on the temperature field of the last simulation time and the current simulation time; Based on the volumetric heat source generation rate and the discretized form of the time derivative term in the heat conduction equation, construct a physical control partial differential equation; Discretize the physical control partial differential equation in space and time on the finite element grid data to obtain a linear equation system; The CAE solver performs physical field iterative solving on the linear equation system to obtain the temperature field of the current simulation time.

5. The multi-physics coupling based thermal simulation method for mobile phone according to claim 4, wherein, The CAE solver iteratively solves the linear equations to obtain the temperature field at the current simulation time, including: the CAE solver iteratively solves the linear equations to obtain the temperature field at the current simulation time based on the conjugate gradient method and in combination with algebraic multi-grid.

6. A mobile phone thermal simulation system based on multi-physical field coupling, characterized in that, Comprise: The model analysis module is used for analyzing the mobile phone three-dimensional model which can be subjected to finite element analysis by the preprocessor module of the CAE software to obtain a basic simulation environment, and the basic simulation environment comprises a structured data set of the following information, specifically comprising: complete finite element grid data, material attribute pointers corresponding to each unit, boundary conditions defined on the surface, and a thermal contact model defined on the component contact surface; The model script loading module is used for loading the trained machine learning agent model and the throttling logic script, and defining and inputting an initial power consumption profile into the basic simulation environment, wherein the throttling logic script defines a mobile phone temperature control strategy; The temperature prediction module is used for inputting a power consumption vector at the current simulation time into the trained machine learning agent model to obtain a predicted temperature vector; The new target power generation module is used for inputting the predicted temperature vector into the throttling logic script to obtain a new target power consumption vector; The simulation temperature field generation module is used for inputting the new target power consumption vector as a heat source and the temperature field at the last simulation time as an initial condition into the CAE solver to obtain the temperature field at the current simulation time; The simulation time length judgment module is used for judging whether the current simulation time is less than the total simulation time, if the current simulation time is less than the total simulation time, returning to the temperature prediction module to the simulation temperature field generation module; if the current simulation time is greater than or equal to the total simulation time, ending the transient simulation.

Images