Neuron operator-based chip thermal behavior prediction and self-thermal management method and system

Abstract

This invention relates to a method and system for predicting chip thermal behavior and implementing self-driven thermal management based on neural operators. The method includes the following steps: acquiring the input excitation function of the chip during operation; inputting the input excitation function as a continuous function into a neural operator model; predicting the temperature evolution function of the chip over the entire time domain using the neural operator model; generating at least one thermal management control strategy based on the predicted temperature evolution function; and self-adjusting the chip's operating state according to the thermal management control strategy to achieve predictive-driven closed-loop thermal management of chip thermal behavior. This not only overcomes the bottleneck of "high simulation accuracy but slow speed" and achieves real-time prediction, but also integrates physical laws and data-driven approaches to improve generalization ability and reliability. This solution supports forward-looking self-driven thermal management, elevating chip thermal management from "passive response" to a new stage of "active prevention."

Description

Technical Field

[0001] This invention relates to the field of chip thermal behavior prediction and self-driven thermal management technology, and more specifically, to a chip thermal behavior prediction and self-driven thermal management method and system based on neural operators. Background Technology

[0002] With the rapid development of new energy vehicles, 5G communications, and industrial power supplies, the demand for high power density and high efficiency in power electronic devices is becoming increasingly urgent. Gallium nitride (GaN), as a third-generation wide-bandgap semiconductor material, has become an ideal choice for realizing high-power-density chips due to its high electron mobility and high breakdown field strength. However, the high power density of GaN chips also brings serious thermal effect problems. 1. Local heat accumulation: When a chip operates under high voltage and high current, power loss (such as switching loss and conduction loss) will be converted into a large amount of heat, resulting in the formation of local high temperature areas (hot spots) inside the chip.

[0003] 2. Performance and Lifetime Decay: For every 10–15°C increase in temperature, the lifespan of GaN devices may be halved. At the same time, high temperatures can lead to a decrease in carrier mobility and threshold voltage drift, directly affecting the switching speed and efficiency of the chip.

[0004] Currently, chip thermal management mainly relies on the following two technical approaches, but both have significant shortcomings: (1) Offline simulation based on finite element analysis (FEA); Current technological status: The industry widely uses commercial simulation software such as Abaqus and COMSOL to model the thermal field of chips using the finite element method. This type of method, based on detailed physical parameters (such as material thermal conductivity, structural dimensions, and power consumption distribution), can achieve high accuracy.

[0005] limitation: High computational cost: FEA requires meshing the three-dimensional structure of the chip and solving complex partial differential equations (such as the heat conduction equation). A single simulation can take several hours or even days, which cannot meet the needs of real-time control.

[0006] High parameter dependence: The accuracy of simulation depends heavily on the accuracy of input parameters such as material properties and boundary conditions, while these parameters may be uncertain in actual working conditions (such as aging of packaging materials and changes in contact thermal resistance).

[0007] (2) Feedback control based on traditional PID; Current technology: PID controllers monitor chip temperature in real time and adjust the power of the cooling system (such as fans and liquid pumps) to maintain the temperature within the set range.

[0008] limitation: Hysteresis: Temperature sensors are usually located outside the chip, resulting in a delay in temperature feedback. In transient overload scenarios (such as rapid acceleration or sudden load increase), the PID controller may have difficulty responding in time, potentially causing the chip to overheat and be damaged.

[0009] Lack of foresight: PID control relies solely on the current temperature error and cannot predict future temperature change trends, making it difficult to achieve preventative thermal management.

[0010] To address the aforementioned issues, there is an urgent need for a new approach that can quickly and accurately predict chip thermal behavior and support proactive self-driven thermal management. Summary of the Invention

[0011] The technical problem to be solved by the present invention is to provide a chip thermal behavior prediction and self-driven thermal management method based on DeepONet, and to provide a chip thermal behavior prediction and self-driven thermal management system based on DeepONet, in view of the above-mentioned defects of the prior art.

[0012] The technical solution adopted by this invention to solve its technical problem is: A method for predicting chip thermal behavior and self-driven thermal management based on neural operators is constructed, comprising the following steps: Obtain the input excitation function of the chip during operation, wherein the input excitation function includes at least a current function, a voltage function, or a power function that varies with time; The input activation function is input as a continuous function into the neural operator model, which is used to learn the continuous operator mapping relationship between the input function space and the output function space, rather than the discrete point value mapping. The neural operator model is used to predict the temperature evolution function of the chip over the entire time domain, thereby obtaining the continuous time distribution results of the chip's thermal behavior. Based on the predicted temperature evolution function, at least one thermal management control strategy is generated, the control strategy including a dynamic adjustment strategy for chip operating parameters; According to the thermal management control strategy, the chip's operating state is automatically adjusted to achieve predictive closed-loop thermal management of the chip's thermal behavior.

[0013] The method described in this invention, wherein: the neural operator model adopts a neural operator structure based on the DeepONet architecture, including: A branch network is used to perform function-level encoding of the input activation function; The backbone network is used to encode continuous time domain data. By performing inner product or combination operations between the branch network output and the backbone network output, an operator mapping from the input excitation function to the temperature evolution function is constructed.

[0014] The method described in this invention, wherein: during training or inference, the neural operator model introduces physical constraints on heat conduction, so that the predicted temperature evolution function satisfies or approximately satisfies the chip heat conduction control equation, thereby improving the physical consistency and generalization ability of the model prediction results.

[0015] The method described in this invention, wherein: the neural operator model is further used to learn the operator mapping relationship under electro-thermal or electro-thermal-mechanical multi-physics coupling conditions, so as to realize the prediction of the thermal behavior of the chip under multi-physics coupling conditions.

[0016] The method described in this invention, wherein the temperature evolution function is a temperature function continuously defined over the entire time domain, rather than a prediction result based on discrete temperature points with a fixed time step.

[0017] The method described in this invention, wherein the generation process of the thermal management control strategy includes automatically generating a multi-level thermal management control strategy based on the predicted temperature evolution function's changing trend, peak characteristics, slope information, or duration information, rather than passively responding solely based on a temperature threshold.

[0018] The method of the present invention, wherein the thermal management control strategy includes at least one of the following: Dynamic adjustment of chip operating current or power; Adjusting the duty cycle or load distribution; Control of the timing of thermal management intervention (early or late); Coordinated scheduling of cooling resources.

[0019] The method described in this invention, wherein the self-driven adjustment process constitutes a prediction-decision-execution closed-loop control flow, transforming the chip thermal management process from passive response control to prediction-driven adaptive control.

[0020] A chip thermal behavior prediction and self-driven thermal management system based on neural operators, comprising: The input function acquisition module is used to acquire the input activation functions during chip operation. The neural operator prediction module is used to predict the chip temperature evolution function based on continuous operator mapping. The control strategy generation module is used to generate thermal management control strategies based on the prediction results; The execution control module is used to self-regulate the chip's operating state according to the control strategy.

[0021] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is used to implement the above-described method.

[0022] The beneficial effects of this invention are as follows: By applying the method of this application, it is possible not only to break through the bottleneck of "high simulation accuracy but slow speed" and achieve real-time prediction, but also to integrate physical laws and data-driven approaches to improve generalization ability and reliability. This solution supports forward-looking self-driven thermal management, which upgrades the thermal management of chips from "passive response" to a new stage of "active prevention". Moreover, the solution reduces the dependence on high-cost experimental data, shortens the model development cycle, and is compatible with multi-physics coupling characteristics to capture complex thermal effects. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort: Figure 1 This is a flowchart of a preferred embodiment of the chip thermal behavior prediction and self-driven thermal management method based on DeepONet of the present invention. Figure 2 This is a block diagram illustrating the principle of a chip thermal behavior prediction and self-driven thermal management system based on DeepONet, according to a preferred embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, a clear and complete description will be provided below in conjunction with the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0025] A preferred embodiment of the present invention is a chip thermal behavior prediction and self-driven thermal management method based on DeepONet, such as... Figure 1 As shown, the steps include: S01: Obtain the input excitation function of the chip during operation. The input excitation function includes at least a current function, voltage function, or power function that varies with time. Specifically, the following can be adopted: A thermo-electric coupling simulation model of the chip's operating environment was built using simulation software. A thermo-electric coupling simulation model of the chip's working environment is constructed. The model includes the chip, thermal interface material (TIM), substrate, and heat dissipation structure, and boundary conditions and material property parameters are set. By importing real operating current data, a simulation dataset of chip thermal behavior is obtained using simulation software; the simulation model imports operating current data as excitation and uses thermoelectric coupling function to simulate the thermal behavior of the chip during operation, obtaining a simulation dataset containing the current input sequence and the chip junction temperature change curve over time. S02: Input the activation function as a continuous function into the neural operator model. The neural operator model is used to learn the continuous operator mapping relationship between the input function space and the output function space, rather than the discrete point value mapping. The neural operator model adopts a neural operator structure based on the DeepONet architecture, including: Branch networks are used for function-level encoding of the input activation function; The backbone network is used to encode continuous time domain data. By performing inner product or combination operations between the branch network output and the backbone network output, an operator mapping from the input excitation function to the temperature evolution function is constructed.

[0026] During training or inference, the neural operator model introduces physical constraints on heat conduction, so that the predicted temperature evolution function satisfies or approximately satisfies the chip heat conduction control equation, thereby improving the physical consistency and generalization ability of the model prediction results.

[0027] The neural operator model is further used to learn the operator mapping relationship under electro-thermal or electro-thermal-mechanical multiphysics coupling conditions, so as to realize the prediction of the thermal behavior of the chip under multiphysics coupling conditions.

[0028] The temperature evolution function is a temperature function that is continuously defined over the entire time domain, rather than a prediction result based on discrete temperature points with a fixed time step.

[0029] Building the DeepONet deep learning model includes: Branch Net: Input: a discrete current sequence; Structure: MLP structure, width 128, depth 4; Output: A 128-dimensional feature vector b; Function: To extract features related to temperature prediction from the input current; Trunk Net: Input: scalar time t, where t is the time point at which the model needs to predict the temperature; Structure: MLP structure, width 128, depth 4; Output: A 128-dimensional feature vector τ; Function: To extract features from the time dimension and reflect the pattern of temperature change over time; Fusion layer: The predicted temperature is obtained by taking the dot product of the Branch Net's output b and the Trunk Net's output τ, and adding the bias term: ; The dot product operation enables the interaction of two feature vectors, allowing the model to flexibly predict temperature based on current input and time points.

[0030] Data preprocessing: Input normalization: Standardize the current u and temperature T: ; Where μ and σ are the mean and standard deviation, respectively; Unified Time Grid: Time is discretized into a linear grid of 0–120 seconds, with a total of 200 points. All input current sequences and temperature data are aligned to this time grid.

[0031] Loss function design: The loss function is determined by the data fitting loss. Physical consistency loss and regularization loss It consists of three parts, of which: Data fitting loss Mean squared error is used to measure the difference between the predicted temperature and the actual temperature label, ensuring that the model fits the observed data. Physical consistency loss To introduce residuals in the thermal equations, which are used to force the model output to conform to thermodynamic laws; the residuals are defined as: ; in: For heat capacity, For thermal resistance, For ambient temperature, This refers to chip power loss. residual It should approach 0, therefore ; Regularization loss L2 regularization penalizes excessively large model weights, prevents overfitting, and encourages a smooth output temperature field. The total loss function is: ; Where λ and α are hyperparameters used to balance the importance of the three losses.

[0032] Training parameters: Optimizer: Adam, Learning Rate ; Learning rate scheduling: Use ReduceLROnPlateau to automatically reduce the learning rate when the validation loss stops decreasing; Training epochs: 200–500 epochs, to ensure the model converges fully.

[0033] By changing the environment and heat dissipation conditions to generate diverse data, the DeepONet model can be further trained or fine-tuned. By changing the ambient temperature, replacing the TIM material or heat dissipation structure, and using liquid cooling or metal additive manufacturing of microchannel structures to change the cooling conditions, multivariate simulation data is generated, and the DeepONet model is further trained using Condition Net or fine-tuning strategies.

[0034] S03: By using a neural operator model, the temperature evolution function of the chip over the entire time domain is predicted, thereby obtaining the continuous-time distribution results of the chip's thermal behavior; S04: Based on the predicted temperature evolution function, generate at least one thermal management control strategy, including a dynamic adjustment strategy for chip operating parameters; The process of generating thermal management control strategies involves automatically generating multi-level thermal management control strategies based on the predicted temperature evolution function's changing trend, peak characteristics, slope information, or duration information, rather than passively responding solely based on temperature thresholds.

[0035] Thermal management control strategies include at least one of the following: Dynamic adjustment of chip operating current or power; Adjusting the duty cycle or load distribution; Control of the timing of thermal management intervention (early or late); Coordinated scheduling of cooling resources.

[0036] S05: Based on the thermal management control strategy, the chip's operating state is automatically adjusted to achieve predictive closed-loop thermal management of the chip's thermal behavior.

[0037] The self-driven adjustment process constitutes a closed-loop control flow of prediction-decision-execution, transforming the chip thermal management process from passive response control to predictive-driven adaptive control.

[0038] The method described in this application has at least the following advantages: 1. It can not only overcome the bottleneck of "high simulation accuracy but slow speed", but also achieve real-time prediction; Traditional FEA simulation: A single simulation takes several hours and cannot be used for real-time control.

[0039] DeepONet model: After training, forward prediction takes only milliseconds and can be embedded in chip controllers for real-time operation.

[0040] Beneficial effects: It realizes the transformation from "offline high-precision simulation to online rapid inference", making it possible to monitor chip temperature changes in real time and predict future thermal trends, providing a time window for proactive thermal management.

[0041] 2. Integrating physical laws with data-driven approaches enhances generalization ability and reliability; Physical consistency: The model forces the output to conform to the laws of thermodynamics through the residual loss of the thermal equation (Lphysics), avoiding the physically unreasonable results (such as sudden temperature changes) that may occur in a purely data-driven "black box" model.

[0042] Data augmentation: By changing the ambient temperature and heat dissipation conditions (such as fan speed and liquid cooling flow rate), multivariate data is generated, enabling the model to learn the thermal behavior under different operating conditions.

[0043] Beneficial effects: The model not only fits known data, but can also extrapolate to unseen working conditions (such as sudden overload or sudden changes in ambient temperature), adapting to the needs of actual complex working conditions.

[0044] 3. It supports forward-looking self-driven thermal management, which elevates the thermal management of chips from a "passive response" stage to a new stage of "active prevention"; Traditional PID control relies on current temperature feedback, which results in lag and makes it difficult to prevent transient overheating.

[0045] DeepONet prediction: It can predict temperature change trends in advance, from seconds to tens of minutes in advance.

[0046] Beneficial effects: Based on the prediction results, the controller can adjust the cooling strategy in advance (such as increasing fan power and limiting current peak) to prevent the chip from entering the dangerous temperature range, thus achieving "prevention before the event" and significantly improving chip life and reliability.

[0047] 4. The solution reduces reliance on high-cost experimental data and shortens the model development cycle; Traditional method: requires a large amount of measured temperature data (requires placing sensors inside the chip, which is costly and may interfere with chip operation).

[0048] This method generates a training set using simulation data, requiring only a small amount of real-world data for fine-tuning.

[0049] Beneficial effects: It significantly reduces data acquisition costs while avoiding incomplete data due to sensor deployment limitations, thus accelerating model development and iteration.

[0050] 5. It can be compatible with multi-physics coupling characteristics and capture complex thermal effects.

[0051] Thermoelectric coupling simulation: The training data comes from the thermoelectric coupling model and already includes the nonlinear relationship between current, power loss and temperature.

[0052] The DeepONet architecture uses Branch Net to process current sequences and Trunk Net to process time, making it naturally well-suited for learning the mapping from dynamic input (current) to dynamic output (temperature).

[0053] Beneficial effects: It can accurately capture key phenomena such as temperature spikes caused by transient current surges and the impact of changes in thermal capacity / thermal resistance on temperature curves, providing a foundation for precise thermal management.

[0054] The implementation example is explained below: Part 1: Simulation Environment Setup and Data Acquisition; Objective: To generate a high-fidelity, multi-physics coupled training dataset.

[0055] step: 1. Physical modeling: Use COMSOL or Abaqus to build a multi-layer structure model of the GaN chip, with the following parameters: GaN chip: 100–150 μm thickness, planar dimensions 3×3 mm or 4×6 mm, κ=130–180 W / m·K, ρ=6150 kg / m³, cp=490 J / kg·K.

[0056] Passivation layer: 2–5 μm thick.

[0057] Thermal interface material (TIM): thickness 50–200 μm, κ=3–7 W / m·K, contact thermal resistance Rc=1e-5–5e-4 m²·K / W.

[0058] DBC / Cu substrate: copper layer thickness 200–300 μm, ceramic layer (AlN / Al2O3) thickness 300–500 μm, κ=150–180 W / m·K.

[0059] PCB FR-4: thickness 1.0–1.5 mm, κ=0.3 W / m·K.

[0060] Heat dissipation base plate (aluminum): 2–5 mm thick.

[0061] Radiator fins: height 5–20 mm, material Al6061 (κ=170–210 W / m·K).

[0062] 2. Boundary condition settings: Ambient temperature: 25°C (standard) or 25–45°C (actual cabin temperature).

[0063] Convection heat transfer coefficient h: natural convection 5–10 W / m²·K, forced air cooling 20–80 W / m²·K, cold plate water cooling 500–5000 W / m²·K.

[0064] 3. Input electrical signal and data generation: Current waveform I(t): 0–4s linear rise to 30A → 4–60s constant current 30A → 60–65s drop to 10A → 65–120s constant current 10A.

[0065] Through thermoelectric coupling simulation, the power loss Ploss and temperature T are automatically exchanged to generate sample data: Input: Discrete current sequence ud=[I(t1),...,I(tn)] (n≈100–300 points).

[0066] Output: Corresponding temperature curves Tsim(t1),...,Tsim(tm) (m≈100–300 points).

[0067] Part 2: DeepONet Network Construction and Training; Objective: To build a physics-guided DeepONet model that learns the dynamic mapping between current and temperature.

[0068] step: 1. Network structure: Branch Net: Input discrete current sequence, MLP (width 128, depth 4), output 128-dimensional feature vector b.

[0069] Trunk Net: Input scalar time t, MLP (width 128, depth 4), output 128-dimensional feature vector τ.

[0070] Fusion layer: Tchip(t) = b·τ + bias.

[0071] 2. Data preprocessing: Current and temperature normalization: û=(u-μ) / σ.

[0072] Uniform time grid: 200 points are linearly divided from 0 to 120 seconds.

[0073] 3. Loss function: Total loss L = L_MSE + λL_physics + αL_reg.

[0074] L_physics: Thermal equation residual R = Cth·dT / dt + (T-Tamb) / Rth - Ploss (forced physical consistency).

[0075] L_reg: L2 regularization (prevents overfitting and encourages smooth output).

[0076] 4. Training parameters: Optimizer: Adam, initial learning rate lr=1e-3, with ReduceLROnPlateau scheduling.

[0077] Training epochs: 200–500 epochs.

[0078] Three: Multi-dimensional data expansion and fine-tuning; Objective: To improve the model's generalization ability under different operating conditions (environment, materials, heat dissipation structure).

[0079] step: 1. Generation of multivariate data: Import real-world temperature change data.

[0080] Replace the TIM material and heat sink structure.

[0081] Introducing metal additive manufacturing of microchannel heat sinks: Process: Laser powder bed melting (L-PBF), material is 15–45 μm OFHC copper powder, layer thickness is 25–35 μm, annealing temperature is 350–450°C.

[0082] Structure: AI-generated integrated microchannels that precisely match the shape of the GaN chip.

[0083] 2. Fast Fine-Tuning Strategy (Condition Net): Freeze the first few layers of Branch Net and Trunk Net, and only fine-tune the conditional network (input ambient temperature, material parameters, etc.) and fusion layer.

[0084] Convergence can be achieved quickly using a small amount of new operating data (such as 10 validation curves).

[0085] 4. Self-driven thermal management applications; Objective: To achieve proactive thermal control based on prediction results and avoid chip overheating.

[0086] step: 1. Model Deployment: Embed the trained DeepONet model into the chip control system.

[0087] 2. Real-time prediction: Collect real-time current waveforms and ambient temperature.

[0088] Input the model to predict the temperature curve Tpred(t) for a future time period (e.g., 0–5s).

[0089] 3. Self-driven control: If max(Tpred) > T_safe (safe temperature threshold), the controller automatically reduces the drive current (e.g., from 30A to the safe level).

[0090] Achieve closed-loop control of "prediction-decision-execution" to avoid overheating damage caused by lag.

[0091] Compared to existing technologies: Breakthrough in real-time performance: Inference time is approximately 15ms, which is about 1000 times faster than traditional FEA, meeting the requirements of high-frequency real-time control.

[0092] High precision guided by physics: Through residual loss of thermal equation, R²>0.98 is achieved in small samples, and the generalization error is reduced by 30%.

[0093] Rapid adaptation to multiple operating conditions: The Condition Net mechanism supports fine-tuning with few samples and can adapt to new heat dissipation structures (such as 3D printed microchannels) without retraining.

[0094] Active thermal management: Predicts temperature trends in advance, smoothly adjusts current, avoids passive power outages, and extends chip lifespan.

[0095] Accelerate the iteration of heat dissipation design: Use it as a surrogate model to replace physical experiments and quickly verify the performance of novel heat dissipation structures (such as biomimetic microchannels).

[0096] A chip thermal behavior prediction and self-driven thermal management system based on neural operators, such as Figure 2 As shown, it includes: The input function acquisition module 100 is used to acquire the input activation functions during chip operation; The neural operator prediction module 101 is used to predict the chip temperature evolution function based on continuous operator mapping. The control strategy generation module 102 is used to generate a thermal management control strategy based on the prediction results. The execution control module 103 is used to self-regulate the chip's operating state according to the control strategy.

[0097] By applying the methods of this application, we can not only break through the bottleneck of "high simulation accuracy but slow speed" and achieve real-time prediction, but also integrate physical laws and data-driven approaches to improve generalization ability and reliability. This solution supports forward-looking self-driven thermal management, which upgrades the thermal management of chips from "passive response" to a new stage of "active prevention". Moreover, the solution reduces the dependence on high-cost experimental data, shortens the model development cycle, and is compatible with multi-physics coupling characteristics to capture complex thermal effects.

[0098] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is used to implement the above-described method.

[0099] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for predicting chip thermal behavior and self-driven thermal management based on neural operators, characterized in that, Includes the following steps: Obtain the input excitation function of the chip during operation, wherein the input excitation function includes at least a current function, a voltage function, or a power function that varies with time; The input activation function is input as a continuous function into the neural operator model, which is used to learn the continuous operator mapping relationship between the input function space and the output function space, rather than the discrete point value mapping. The neural operator model is used to predict the temperature evolution function of the chip over the entire time domain, thereby obtaining the continuous time distribution results of the chip's thermal behavior. Based on the predicted temperature evolution function, at least one thermal management control strategy is generated, the control strategy including a dynamic adjustment strategy for chip operating parameters; According to the thermal management control strategy, the chip's operating state is automatically adjusted to achieve predictive closed-loop thermal management of the chip's thermal behavior.

2. The method according to claim 1, characterized in that: The neural operator model adopts a neural operator structure based on the DeepONet architecture, including: A branch network is used to perform function-level encoding of the input activation function; The backbone network is used to encode continuous time domain data. By performing inner product or combination operations between the branch network output and the backbone network output, an operator mapping from the input excitation function to the temperature evolution function is constructed.

3. The method according to claim 1 or 2, characterized in that: During training or inference, the neural operator model introduces physical constraints on heat conduction, so that the predicted temperature evolution function satisfies or approximately satisfies the chip heat conduction control equation, thereby improving the physical consistency and generalization ability of the model prediction results.

4. The method according to claim 3, characterized in that: The neural operator model is further used to learn the operator mapping relationship under electro-thermal or electro-thermal-mechanical multi-physics coupling conditions, so as to realize the prediction of the thermal behavior of the chip under multi-physics coupling conditions.

5. The method according to claim 1, characterized in that: The temperature evolution function is a temperature function continuously defined over the entire time domain, rather than a prediction result based on discrete temperature points with a fixed time step.

6. The method according to claim 1, characterized in that: The process of generating the thermal management control strategy includes automatically generating a multi-level thermal management control strategy based on the predicted temperature evolution function's changing trend, peak characteristics, slope information, or duration information, rather than passively responding based solely on temperature thresholds.

7. The method according to claim 6, characterized in that: The thermal management control strategy includes at least one of the following: Dynamic adjustment of chip operating current or power; Adjusting the duty cycle or load distribution; Control of the timing of thermal management intervention (early or late); Coordinated scheduling of cooling resources.

8. The method according to claim 1, characterized in that: The self-driven adjustment process constitutes a closed-loop control flow of prediction-decision-execution, transforming the chip thermal management process from passive response control to prediction-driven adaptive control.

9. A chip thermal behavior prediction and self-driven thermal management system based on neural operators, characterized in that, include: The input function acquisition module is used to acquire the input activation functions during chip operation. The neural operator prediction module is used to predict the chip temperature evolution function based on continuous operator mapping. The control strategy generation module is used to generate thermal management control strategies based on the prediction results; The execution control module is used to self-regulate the chip's operating state according to the control strategy.

10. A computer-readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the method of any one of claims 1 to 8.

Images