A method and system for dynamic control of supercritical CO2 two-phase cooling

Abstract

This invention discloses a dynamic control method and system for supercritical CO2 two-phase cooling, belonging to the field of high-efficiency heat exchange technology. The invention collects and standardizes cooling system operating data, constructs a reinforcement learning input space, and outputs piezoelectric valve control signals via network computation to adjust the medium flow rate of the supercritical CO2 two-phase circulation in real time. Based on the adjusted operating data, a reward value is calculated and network parameters are updated. Combined with digital twin technology, virtual mapping of the cooling system and fault scenario simulation are achieved, realizing closed-loop intelligent control of the system's operating state. This invention can stabilize the chip cooling process, improve heat exchange efficiency, enhance system fault response capabilities, adapt to the heat dissipation requirements of high-density chips, and features fast control response and strong operational stability. It effectively solves the problems of slow response and weak fault adaptability in traditional cooling methods, optimizes the overall operating effect of the cooling system, and meets the dual requirements of efficient chip heat dissipation and continuous stable operation.

Description

Technical Field

[0001] The present invention relates to the technical field of efficient heat transfer, and particularly to a dynamic regulation method and system for supercritical CO2 two-phase cooling. Background Art

[0002] As an important technical direction in the field of efficient heat transfer, the supercritical CO2 two-phase cooling technology has been gradually popularized and applied in multiple industrial scenarios such as high-density electronic chip cooling, data center cooling systems, and new energy vehicle electronic control component cooling. In the industry, a large amount of research and practical work has been carried out in the directions of phase change heat transfer characteristics of supercritical CO2, construction of circulation pipeline paths, optimization of heat exchange component structures, and integration of cooling systems. Technically, a complete technical system covering links such as supercritical fluid heat transfer, circulation flow regulation, processing of heat exchange components, acquisition of system operation state data, and basic regulation execution has been formed. The application of various temperature, flow, and pressure sensing components has been gradually popularized, and the adaptability and universality of core execution components such as flow regulation valves and heat exchange modules have been gradually improved. Intelligent regulation technologies such as digital twin and deep reinforcement learning have also begun to be gradually integrated into the regulation process of cooling systems. Relevant technologies have completed preliminary application verification in laboratory environments and some industrial application scenarios, and the structure construction and basic regulation process of cooling systems have gradually become standardized. The advantages of supercritical CO2 two-phase cooling, such as efficient heat transfer and environmental protection without loss, have also been widely recognized in the industry.

[0003] At present, in the actual operation regulation process of relevant cooling systems, there is a lack of a standardized processing process in the acquisition and integration link of operation state data. It is difficult to ensure the timestamp matching and format unity of various sensing data, and it is impossible to provide complete and accurate basic data support for subsequent regulation links; there is a lack of standardized filtering and normalization processing processes for the operation state data of cooling systems, and the degree of data standardization is insufficient, making it difficult to construct an input state space suitable for intelligent regulation networks, affecting the accuracy of subsequent regulation decisions; the flow regulation link of cooling systems cannot achieve dynamic intelligent regulation in combination with the real-time operation state, and the adaptability of regulation actions to the actual operation state of the system is poor. The real-time performance and accuracy of flow regulation are difficult to meet the requirements of efficient heat transfer; there is a lack of a dynamic update mechanism for combining regulation effects in the parameter update link of the intelligent regulation network. The network parameters cannot be continuously optimized according to the actual regulation effects of the system, and the iterative upgrade ability of regulation strategies is insufficient, making it difficult to ensure the coherence and stability of the overall regulation process. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a dynamic regulation method and system for supercritical CO2 two-phase cooling.

[0005] The purpose of the present invention is achieved through the following technical solutions: A method for dynamic control of supercritical CO2 two-phase cooling is provided, which includes the following steps: S1. Acquire cooling system operating status data, including chip temperature data, supercritical CO2 flow rate data, pipeline pressure data, and ambient temperature and humidity data; S2. Filter and normalize the cooling system operating status data to construct the input state space of the deep reinforcement learning network; S3. Input the input state space into a deep reinforcement learning network, and output a valve opening control signal through the decision of the deep reinforcement learning network. Adjust the opening of the piezoelectric valve according to the valve opening control signal, thereby adjusting the flow rate of the supercritical CO2 main circulation. S4. Obtain the adjusted cooling system operating status data, calculate the reward value of the deep reinforcement learning network based on the adjusted cooling system operating status data, and update the parameters of the deep reinforcement learning network based on the reward value through the experience replay mechanism.

[0006] Furthermore, step S1 includes the following sub-steps: S1.1. Temperature data of the chip surface is collected by temperature sensors arranged on the chip surface. The temperature sensors correspond to the area of ​​the chip, and the collected temperature data includes the location identifier of the corresponding area and the collection time information. S1.2. Flow and pressure data of the supercritical CO2 main circulation pipeline are collected by flow sensors and pressure sensors installed at the location of the supercritical CO2 main circulation pipeline. The flow sensors collect the mass flow rate data of the supercritical CO2, and the pressure sensors collect the pressure data of the first set pipeline section and the second set pipeline section respectively. S1.3. Temperature and humidity data of the environment in which the cooling system is located are collected by temperature and humidity sensors installed in the external environment of the cooling system, and the collection frequency is consistent with the collection frequency of chip temperature data; S1.4. Align all collected data according to a unified timestamp, remove data with mismatched timestamps, convert the aligned data into a unified data format, and generate cooling system operating status data containing all collected parameters.

[0007] Furthermore, step S2 includes the following sub-steps: S2.1. Perform sliding window mean filtering on each type of data in the cooling system operating status data. The size of the sliding window is set according to the data acquisition frequency. The time series characteristics of the data are preserved during the filtering process. S2.2. Normalize each type of cooling system operating status data after filtering. Use linear normalization to map each type of data to a set value range. During the normalization process, retain the relative size relationship between the data. S2.3. Combine all the normalized data in a preset order to generate an input state space in the form of a one-dimensional vector. The dimension of the input state space is consistent with the number of parameters of the cooling system operating state data.

[0008] Furthermore, step S3 includes the following sub-steps: S3.1. Input the input state space into the online network of the deep reinforcement learning network. The online network of the deep reinforcement learning network extracts features from the input state space through multiple fully connected layers and calculates the action value corresponding to each optional action. The fully connected layers use a linear activation function for output. S3.2. A greedy strategy is adopted to select the action with the highest action value from all available actions. The action space is the opening range of the piezoelectric valve. The opening range is discretized into gears, and each gear corresponds to an independent action. The valve opening control signal corresponding to the gear is generated. S3.3. Convert the valve opening control signal into a piezoelectric ceramic actuator driving voltage signal, send the driving voltage signal to the piezoelectric ceramic actuator of the piezoelectric valve, drive the piezoelectric ceramic actuator to generate displacement, drive the valve core to move and adjust to the corresponding opening degree, and regulate the flow rate of the supercritical CO2 main circulation.

[0009] Furthermore, step S4 includes the following sub-steps: S4.1. After the piezoelectric valve opening is adjusted and a set time interval has elapsed, re-collect the temperature data on the chip surface and the pressure data of the supercritical CO2 main circulation pipeline; S4.2. Calculate the temperature reward based on the difference between the adjusted chip temperature data and the set temperature threshold, calculate the pressure penalty based on the difference between the adjusted pipeline pressure data and the set pressure threshold, and sum the temperature reward and pressure penalty together to calculate the reward value of the deep reinforcement learning network. S4.3. Combine the input state space, valve opening control signal, reward value and new state space into an experience sample, store the experience sample in the experience playback pool, and prioritize the experience sample according to the size of the reward value during the storage process; S4.4. Extract a set number of experience samples from the experience replay pool according to priority, and update the online network parameters of the deep reinforcement learning network using gradient descent. S4.5. According to the set training step size, copy the online network parameters of the deep reinforcement learning network to the target network of the deep reinforcement learning network to complete the synchronous update of the target network parameters of the deep reinforcement learning network.

[0010] Furthermore, in step S1, a digital twin platform is pre-constructed. The digital twin platform is a virtual mapping carrier built based on a physical information neural network. The physical information neural network includes an input layer, a hidden layer, and an output layer. The input layer receives the structural and operational parameters of the physical cooling system. The hidden layer contains multiple fully connected layers, which are linearly connected. The output layer outputs the predicted operating state of the physical cooling system. The physical information neural network embeds fluid dynamics control equations and thermodynamic control equations. The structural and operational parameters of the physical cooling system are input into the physical information neural network for training. The training steps include inputting a training dataset, calculating a loss function, updating the parameters of the physical information neural network using gradient descent, and generating a virtual mapping model of the physical cooling system. The operating state data of the physical cooling system is synchronized to the digital twin platform in real time according to a set sampling rate. During the synchronization process, the data is verified and abnormal data is removed.

[0011] Furthermore, in step S3, fault scenarios are injected into the virtual mapping model of the physical cooling system through the fault injection module of the digital twin platform. The fault scenarios include pump failure, pipeline leakage and voltage surge. Different fault severity parameters are set for the fault scenarios. The operating state of the virtual mapping model of the physical cooling system after the fault is injected is deduced based on the physical information neural network, and the execution content corresponding to the fault scenario is generated. The execution content includes the valve opening adjustment method at different times. The generated execution content is fused with the valve opening control signal output by the deep reinforcement learning network to adjust the final output valve opening control signal.

[0012] Furthermore, in step S3, the operating path of the supercritical CO2 main cycle is as follows: supercritical CO2 enters the microcavity evaporation module, and through the microcavity array of the microcavity evaporation module, it directly contacts the chip to be cooled, absorbs the heat generated by the chip, and undergoes a phase change, changing from liquid to gas. The gaseous supercritical CO2 after absorbing heat enters the compressor for pressurization, and the pressurized supercritical CO2 enters the condenser tower, where it exchanges heat with the outside air through the fins of the condenser tower, releasing heat and condensing. The supercritical CO2 after releasing heat and condensation enters the expander for depressurization, and the depressurized supercritical CO2 returns to the liquid state and returns to the microcavity evaporation module to complete the closed-loop operation.

[0013] Furthermore, in step S3, based on the temperature distribution data of different areas of the chip, the flow rate of the supercritical CO2 main circulation is allocated to the corresponding area of ​​the microcavity evaporation module. When the area temperature reaches the set temperature value, the corresponding flow rate is allocated. At the same time, the ventilation volume of the condenser is adjusted according to the ambient temperature data. When the ambient temperature reaches the set temperature value, the condenser is adjusted to the set ventilation volume value to control the condensation temperature of the supercritical CO2.

[0014] A supercritical CO2 two-phase cooling dynamic control system is provided. The system includes a microcavity evaporation module, a biomimetic condenser tower, a supercritical CO2 main circulation pipeline, a piezoelectric valve assembly, a deep reinforcement learning controller, and a digital twin platform. The surface of the microcavity evaporation module is provided with a microcavity array, and the microcavity evaporation module is in direct contact with the chip to be cooled. The surface of the biomimetic condenser tower is provided with needle-shaped fins, and the biomimetic condenser tower is connected to a first predetermined pipeline section of the supercritical CO2 main circulation pipeline. The supercritical CO2 main circulation pipeline is sequentially connected to a compressor, a gas cooler, an expander, and a liquid storage tank. The piezoelectric valve assembly is located in the first predetermined pipeline section of the supercritical CO2 main circulation pipeline. The deep reinforcement learning controller is electrically connected to the piezoelectric valve assembly, and the digital twin platform is communicatively connected to the deep reinforcement learning controller.

[0015] The beneficial effects of this invention are: (1) By collecting the operating status data of the cooling system and performing filtering and normalization processing, a deep reinforcement learning network input space is constructed and flow regulation and network parameter updates are completed to realize real-time dynamic adjustment of supercritical CO2 cooling flow and ensure the continuous stability of chip cooling process; (2) The structure of two-phase cooling and direct contact heat exchange is adopted to optimize the heat transfer and air condensation process, reduce heat transfer loss during the cooling process, reduce the dependence of the cooling system on water resources, and improve the environmental adaptability of the cooling system. (3) By using digital twin virtual mapping and fault scenario simulation, combined with rapid response control components and convenient disassembly and maintenance structure, the fault prediction and handling capabilities of the cooling system are improved, the operation and maintenance process is simplified, and the overall operation and maintenance cost of the system is reduced. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the steps of a dynamic control method for supercritical CO2 two-phase cooling; Figure 2 The following is a flowchart illustrating the specific steps of a dynamic control method for supercritical CO2 two-phase cooling, provided as an example. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1 See Figure 1 This embodiment provides a method for dynamic control of supercritical CO2 two-phase cooling, which includes the following steps: S1. Acquire cooling system operating status data, including chip temperature data, supercritical CO2 flow rate data, pipeline pressure data, and ambient temperature and humidity data; S2. Filter and normalize the cooling system operating status data to construct the input state space of the deep reinforcement learning network; S3. Input the input state space into a deep reinforcement learning network, and output a valve opening control signal through the decision of the deep reinforcement learning network. Adjust the opening of the piezoelectric valve according to the valve opening control signal, thereby adjusting the flow rate of the supercritical CO2 main circulation. S4. Obtain the adjusted cooling system operating status data, calculate the reward value of the deep reinforcement learning network based on the adjusted cooling system operating status data, and update the parameters of the deep reinforcement learning network based on the reward value through the experience replay mechanism.

[0019] In some embodiments, step S1 includes the following sub-steps: S1.1. Temperature data of the chip surface is collected by temperature sensors arranged on the chip surface. The temperature sensors correspond to the area of ​​the chip, and the collected temperature data includes the location identifier of the corresponding area and the collection time information. S1.2. Flow and pressure data of the supercritical CO2 main circulation pipeline are collected by flow sensors and pressure sensors installed at the location of the supercritical CO2 main circulation pipeline. The flow sensors collect the mass flow rate data of the supercritical CO2, and the pressure sensors collect the pressure data of the first set pipeline section and the second set pipeline section respectively. S1.3. Temperature and humidity data of the environment in which the cooling system is located are collected by temperature and humidity sensors installed in the external environment of the cooling system, and the collection frequency is consistent with the collection frequency of chip temperature data; S1.4. Align all collected data according to a unified timestamp, remove data with mismatched timestamps, convert the aligned data into a unified data format, and generate cooling system operating status data containing all collected parameters.

[0020] In some embodiments, step S2 includes the following sub-steps: S2.1. Perform sliding window mean filtering on each type of data in the cooling system operating status data. The size of the sliding window is set according to the data acquisition frequency. The time series characteristics of the data are preserved during the filtering process. S2.2. Normalize each type of cooling system operating status data after filtering. Use linear normalization to map each type of data to a set value range. During the normalization process, retain the relative size relationship between the data. S2.3. Combine all the normalized data in a preset order to generate an input state space in the form of a one-dimensional vector. The dimension of the input state space is consistent with the number of parameters of the cooling system operating state data.

[0021] In some embodiments, step S3 includes the following sub-steps: S3.1. Input the input state space into the online network of the deep reinforcement learning network. The online network of the deep reinforcement learning network extracts features from the input state space through multiple fully connected layers and calculates the action value corresponding to each optional action. The fully connected layers use a linear activation function for output. S3.2. A greedy strategy is adopted to select the action with the highest action value from all available actions. The action space is the opening range of the piezoelectric valve. The opening range is discretized into gears, and each gear corresponds to an independent action. The valve opening control signal corresponding to the gear is generated. S3.3. Convert the valve opening control signal into a piezoelectric ceramic actuator driving voltage signal, send the driving voltage signal to the piezoelectric ceramic actuator of the piezoelectric valve, drive the piezoelectric ceramic actuator to generate displacement, drive the valve core to move and adjust to the corresponding opening degree, and regulate the flow rate of the supercritical CO2 main circulation.

[0022] In some embodiments, step S4 includes the following sub-steps: S4.1. After the piezoelectric valve opening is adjusted and a set time interval has elapsed, re-collect the temperature data on the chip surface and the pressure data of the supercritical CO2 main circulation pipeline; S4.2. Calculate the temperature reward based on the difference between the adjusted chip temperature data and the set temperature threshold, calculate the pressure penalty based on the difference between the adjusted pipeline pressure data and the set pressure threshold, and sum the temperature reward and pressure penalty together to calculate the reward value of the deep reinforcement learning network. S4.3. Combine the input state space, valve opening control signal, reward value and new state space into an experience sample, store the experience sample in the experience playback pool, and prioritize the experience sample according to the size of the reward value during the storage process; S4.4. Extract a set number of experience samples from the experience replay pool according to priority, and update the online network parameters of the deep reinforcement learning network using gradient descent. S4.5. According to the set training step size, copy the online network parameters of the deep reinforcement learning network to the target network of the deep reinforcement learning network to complete the synchronous update of the target network parameters of the deep reinforcement learning network.

[0023] In some embodiments, in step S1, a digital twin platform is pre-constructed. The digital twin platform is a virtual mapping carrier based on a physical information neural network. The physical information neural network includes an input layer, a hidden layer, and an output layer. The input layer receives the structural parameters and operating parameters of the physical cooling system. The hidden layer includes multiple fully connected layers with linear connections between them. The output layer outputs the predicted operating state values ​​of the physical cooling system. The physical information neural network embeds fluid dynamics control equations and thermodynamic control equations. The structural parameters and operating parameters of the physical cooling system are input into the physical information neural network for training. The training steps include inputting a training dataset, calculating a loss function, updating the parameters of the physical information neural network using gradient descent, and generating a virtual mapping model of the physical cooling system. The operating state data of the physical cooling system is synchronized to the digital twin platform in real time according to a set sampling rate. During the synchronization process, the data is verified and abnormal data is removed.

[0024] In some embodiments, in step S3, a fault scenario is injected into the virtual mapping model of the physical cooling system through the fault injection module of the digital twin platform. The fault scenario includes pump failure, pipeline leakage and voltage surge. Different fault severity parameters are set for the fault scenario. The operating state of the virtual mapping model of the physical cooling system after the fault is injected is deduced based on the physical information neural network, and the execution content corresponding to the fault scenario is generated. The execution content includes the valve opening adjustment method at different times. The generated execution content is fused with the valve opening control signal output by the deep reinforcement learning network to adjust the final output valve opening control signal.

[0025] In some embodiments, in step S3, the operating path of the supercritical CO2 main cycle is as follows: supercritical CO2 enters the microcavity evaporation module, and through the microcavity array of the microcavity evaporation module, it directly contacts the chip to be cooled, absorbs the heat generated by the chip, and undergoes a phase change, changing from liquid to gas. The gaseous supercritical CO2 after absorbing heat enters the compressor for pressurization, and the pressurized supercritical CO2 enters the condenser tower, where it exchanges heat with the outside air through the fins of the condenser tower, releasing heat and condensing. The supercritical CO2 after releasing heat and condensation enters the expander for depressurization, and the depressurized supercritical CO2 returns to the liquid state and returns to the microcavity evaporation module to complete the closed operation.

[0026] In some embodiments, in step S3, the flow rate of the supercritical CO2 main circulation is allocated to the corresponding area of ​​the microcavity evaporation module according to the temperature distribution data of different areas of the chip. When the temperature of the area reaches the set temperature value, the corresponding flow rate is allocated. At the same time, the ventilation volume of the condenser is adjusted according to the ambient temperature data. When the ambient temperature reaches the set temperature value, the condenser is adjusted to the set ventilation volume value to control the condensation temperature of the supercritical CO2.

[0027] A supercritical CO2 two-phase cooling dynamic control system is provided. The system includes a microcavity evaporation module, a biomimetic condenser tower, a supercritical CO2 main circulation pipeline, a piezoelectric valve assembly, a deep reinforcement learning controller, and a digital twin platform. The surface of the microcavity evaporation module is provided with a microcavity array, and the microcavity evaporation module is in direct contact with the chip to be cooled. The surface of the biomimetic condenser tower is provided with needle-shaped fins, and the biomimetic condenser tower is connected to a first predetermined pipeline section of the supercritical CO2 main circulation pipeline. The supercritical CO2 main circulation pipeline is sequentially connected to a compressor, a gas cooler, an expander, and a liquid storage tank. The piezoelectric valve assembly is located in the first predetermined pipeline section of the supercritical CO2 main circulation pipeline. The deep reinforcement learning controller is electrically connected to the piezoelectric valve assembly, and the digital twin platform is communicatively connected to the deep reinforcement learning controller.

[0028] Example 2 This embodiment provides a specific implementation process for a dynamic control method of supercritical CO2 two-phase cooling. This process is carried out systematically in four stages: acquisition and integration of cooling system operating status, processing of operating data and construction of network input space, intelligent control of cooling medium flow rate, and calculation of reward value and updating of network parameters. Simultaneously, it is accompanied by supporting operations such as digital twin virtual mapping construction, fault scenario simulation, cooling circulation path control, and allocation of flow and heat exchange parameters, thus fully realizing the dynamic control of the entire process of the supercritical CO2 two-phase cooling system. Figure 2 As shown, the specific implementation process is as follows: Step 1. Collect and integrate cooling system operating status data: Step 1.1. Collect temperature data of the chip surface using temperature sensors positioned on the chip surface. The temperature sensor corresponds to a region of the chip, and the collected temperature data includes the location identifier of the corresponding region and the collection time information. The temperature sensor is an existing sensing component used to detect the surface temperature of an object. In this embodiment, the component is attached to the chip surface and continuously collects temperature information of the chip surface. The location identifier is used to distinguish the temperature data of different regions of the chip, and the collection time information is used to mark the time when the temperature data was acquired, providing a basis for subsequent data integration. The collected temperature data can truly reflect the heat distribution state of the chip, providing basic data support for subsequent regulation.

[0029] Step 1.2. Flow and pressure data of the supercritical CO2 main circulation pipeline are collected by flow sensors and pressure sensors installed at the locations of the supercritical CO2 main circulation pipeline. The flow sensors collect the mass flow rate data of the supercritical CO2, and the pressure sensors collect the pressure data of the first and second designated pipeline sections respectively. The flow sensors are existing sensing components used to detect the flow rate of fluid transmission. In this embodiment, the component collects the mass flow rate data of the supercritical CO2, which reflects the transmission rate of the cooling medium. The pressure sensors are existing sensing components used to detect the internal pressure of the pipeline. In this embodiment, the component collects the pressure data of the two designated pipeline sections in the supercritical CO2 main circulation pipeline respectively. The pipeline pressure data reflects the operating status of the cooling cycle. The data collected by the two types of sensing components can completely reflect the transmission status of the cooling medium.

[0030] Step 1.3. Collect temperature and humidity data of the environment in which the cooling system is located by temperature and humidity sensors installed in the external environment of the cooling system. The collection frequency is consistent with the collection frequency of chip temperature data: The ambient temperature sensor is used to detect the temperature information of the environment around the cooling system, and the ambient humidity sensor is used to detect the humidity information of the environment around the cooling system. Ambient temperature and humidity data will affect the heat exchange effect of the cooling system. In this embodiment, the collection frequency of the two types of data is unified, which can ensure the time matching of environmental data with chip temperature data and pipeline medium data, avoid data misalignment caused by different collection frequencies, and improve the continuity of subsequent data processing.

[0031] Step 1.4. Align all collected data according to a unified timestamp, remove data with mismatched timestamps, and convert the aligned data into a unified data format to generate cooling system operating status data containing all collected parameters: The timestamp is an existing identifier used to mark the time of data collection. In this embodiment, all collected data are aligned based on a unified timestamp, invalid data with mismatched time dimensions are removed, and different types of sensor data are converted into data content of the same format and integrated to form complete cooling system operating status data. This data includes four types of parameters: chip temperature, supercritical CO2 flow rate, pipeline pressure, and ambient temperature and humidity, which can comprehensively reflect the real-time operating status of the cooling system.

[0032] In some embodiments, when collecting cooling system operating status data, a synchronous acquisition module can be used to uniformly trigger the acquisition actions of various sensors. By using synchronous triggering, the acquisition time deviation between different sensors can be reduced, further improving the efficiency of data alignment processing. The synchronous acquisition module is an existing control component used to uniformly trigger sensor acquisition actions and is adapted to the acquisition triggering logic of various sensor components.

[0033] Step 2. Process runtime data and construct the network input space: Step 2.1. Perform sliding window mean filtering on each type of data in the cooling system operating status data. The size of the sliding window is set according to the data acquisition frequency. The time series characteristics of the data are retained during the filtering process. Sliding window mean filtering is an existing data filtering method. This method calculates the mean of the data by setting a window range to remove random noise in the data. In this embodiment, the size of the sliding window is adjusted according to the data acquisition frequency to ensure that the filtering process matches the data acquisition rhythm. The time series characteristics of the data are retained after filtering. The time series characteristics can reflect the continuous change trend of the cooling system operating status and retain effective information for subsequent data processing.

[0034] Step 2.2. Normalize each type of cooling system operating status data after filtering. Linear normalization is used to map each type of data to a set numerical range. The relative size relationship between data is preserved during the normalization process. Linear normalization is an existing data standardization processing method. This method can convert data with different dimensions and different numerical ranges into values ​​within a unified range. In this embodiment, this method is used to eliminate the numerical differences between different types of data. The relative size relationship between data is preserved during the processing to ensure that the differences in operating status reflected by the data are not changed, thus providing a standardized data foundation for subsequent network input.

[0035] Step 2.3. Combine all the normalized data in a preset order to generate a one-dimensional vector input state space. The dimension of the input state space is consistent with the number of parameters in the cooling system operating state data. The input state space is the input data carrier of the deep reinforcement learning network. In this embodiment, the normalized data are integrated in a preset order to form a one-dimensional vector input state space. The spatial dimension corresponds to the number of parameters in the cooling system operating state data, which can ensure that the deep reinforcement learning network can completely receive all operating state parameters and provide complete input information for network decision-making.

[0036] In some embodiments, when filtering the running status data, median filtering can be used instead. Median filtering is an existing data filtering method that can adapt to scenarios where there is sudden noise in the data. While preserving the time series characteristics of the data, it can effectively remove noise data of sudden types and improve the adaptability of data processing.

[0037] Step 3. Regulate the flow rate of the cooling medium based on a deep reinforcement learning network: Step 3.1. Input the input state space into the online network of the deep reinforcement learning network. The online network extracts features from the input state space through multiple fully connected layers and calculates the action value corresponding to each optional action. The fully connected layers use linear activation functions for output. The deep reinforcement learning network is an existing network structure that combines deep feature extraction and reinforcement learning decision-making. This network includes an online network and a target network. The online network is the core decision-making component of the deep reinforcement learning network. The fully connected layers are the basic hierarchical structure of existing neural networks. Multiple fully connected layers are linearly connected, and the linear activation functions are existing neural network activation functions used to output the hierarchical calculation results. In this embodiment, the online network extracts feature information from the input state space through multiple fully connected layers and calculates the action value corresponding to the piezoelectric valve opening adjustment action based on the feature information. The action value is used to measure the adaptability of different opening adjustment methods.

[0038] Step 3.2. A greedy strategy is used to select the action with the highest action value from all available actions. The action space is the opening range of the piezoelectric valve. The opening range is discretized into levels, and each level corresponds to an independent action. A valve opening control signal for the corresponding level is generated. The greedy strategy is an existing professional decision-making strategy in the field of reinforcement learning. This strategy is used to select the action with the highest adaptability in the current state. The action space is the set of actions that can be executed by the deep reinforcement learning network. In this embodiment, the action space is the opening range of the piezoelectric valve. The continuous opening range is discretized into multiple independent levels, and each level corresponds to an opening adjustment action. The level action with the highest action value is selected according to the greedy strategy, and the corresponding valve opening control signal is generated. The control signal is an electrical signal carrier used to drive the piezoelectric valve action.

[0039] Step 3.3. Convert the valve opening control signal into a driving voltage signal for the piezoelectric ceramic actuator, and send the driving voltage signal to the piezoelectric ceramic actuator of the piezoelectric valve. The piezoelectric ceramic actuator is driven to generate displacement, which moves the valve core to adjust to the corresponding opening, thereby regulating the flow rate of the supercritical CO2 main circulation. The piezoelectric ceramic actuator is an existing drive component that achieves displacement drive based on the piezoelectric effect. This component can convert voltage signals into mechanical displacement. The piezoelectric valve is an existing valve component that adjusts the opening through piezoelectric drive. The valve core is the component inside the piezoelectric valve that regulates the flow rate. In this embodiment, the valve opening control signal is converted into a driving voltage signal adapted to the piezoelectric ceramic actuator. The driving voltage signal drives the piezoelectric ceramic actuator to generate displacement. The displacement moves the valve core to the corresponding opening. By adjusting the opening, the flow cross section of the supercritical CO2 main circulation is changed, thereby realizing the regulation of the cooling medium flow rate.

[0040] Before implementing step 3, a digital twin platform is pre-constructed. The digital twin platform is a virtual mapping carrier built based on a physical information neural network. The physical information neural network is an existing neural network structure that embeds physical control equations. This neural network includes an input layer, a hidden layer, and an output layer. The input layer receives the structural and operational parameters of the physical cooling system. The hidden layer contains multiple fully connected layers, which are linearly connected. The output layer outputs the predicted operating state values ​​of the physical cooling system. The physical information neural network embeds fluid dynamics control equations and thermodynamic control equations. The fluid dynamics control equations and thermodynamic control equations are existing physical equations that reflect the laws of fluid transport and heat exchange. The structural and operational parameters of the physical cooling system are input into the physical information neural network for training. The training steps include inputting a training dataset, calculating a loss function, updating the parameters of the physical information neural network using gradient descent, and generating a virtual mapping model of the physical cooling system. The operating state data of the physical cooling system is synchronized to the digital twin platform in real time according to a set sampling rate. During the synchronization process, the data is verified and abnormal data is removed.

[0041] In step 3, fault scenarios are injected into the virtual mapping model of the physical cooling system through the fault injection module of the digital twin platform. The fault scenarios include pump failure, pipeline leakage, and voltage surge. Different fault severity parameters are set for the fault scenarios. Based on the physical information neural network, the operating state of the virtual mapping model of the physical cooling system after the fault is injected is deduced, and corresponding execution content is generated. The execution content includes the valve opening adjustment method at different times. The generated execution content is fused with the valve opening control signal output by the deep reinforcement learning network to adjust the final output valve opening control signal. The fault injection module of the digital twin platform is an existing functional module in the digital twin platform used to simulate fault scenarios. It can input fault parameters into the virtual mapping model to restore the system operating state under real faults.

[0042] In step 3, the supercritical CO2 main circulation path is executed. Supercritical CO2 enters the microcavity evaporation module, and through the microcavity array of the microcavity evaporation module, it directly contacts the chip to be cooled, absorbing the heat generated by the chip and undergoing a phase change from liquid to gas. The gaseous supercritical CO2 after absorbing heat enters the compressor for pressurization. The pressurized supercritical CO2 enters the condenser, where it exchanges heat with the outside air through the fins of the condenser, releasing heat and condensing. The supercritical CO2 after releasing heat and condensation enters the expander for depressurization. The depressurized supercritical CO2 returns to the liquid state and returns to the microcavity evaporation module to complete the closed-loop operation. The microcavity evaporation module is an existing heat exchange component with a microcavity array, which is used to increase the contact area between the chip and the cooling medium. The compressor is an existing power component used to increase the fluid pressure. The condenser is an existing heat exchange device that achieves fluid condensation through air heat exchange. The expander is an existing pressure-reducing component used to reduce the fluid pressure. The closed-loop operation can realize the recycling of supercritical CO2 and continuously complete the transfer and dissipation of heat from the chip.

[0043] During step 3, based on the temperature distribution data of different areas of the chip, the flow rate of the supercritical CO2 main circulation is allocated to the corresponding areas of the microcavity evaporation module. When the temperature of a region reaches a set temperature value, the corresponding flow rate is allocated. At the same time, the ventilation volume of the condenser tower is adjusted according to the ambient temperature data. When the ambient temperature reaches a set temperature value, the ventilation volume of the condenser tower is adjusted to the set ventilation volume value to control the condensation temperature of the supercritical CO2. The flow rate allocation ensures that the heat of different areas of the chip is evenly dissipated, and the ventilation volume adjustment can adapt to changes in ambient temperature, stabilize the condensation effect of supercritical CO2, and maintain the stable operation of the cooling cycle.

[0044] In some embodiments, the online network of a deep reinforcement learning network can be augmented with a batch normalization layer. The batch normalization layer is an existing neural network hierarchical structure that can standardize the output data of the fully connected layer, improve the stability of network feature extraction, reduce the impact of data fluctuations on action value calculation, and make network decisions more consistent with the actual operating state of the cooling system.

[0045] In some specific implementations, addressing the specific problems of insufficient heat flux, excessive interfacial thermal resistance, and severe water resource dependence in arid regions, a structured parameter setup and operation of a microcavity evaporation module and a biomimetic condensation tower were implemented. A 355nm ultraviolet laser was used for microcavity fabrication, with the laser power set to 20W and the scanning speed set to 5mm / s. A honeycomb microcavity array with a diameter between 80μm and 150μm and a depth of 50μm was etched onto the chip surface. Then, a nanoporous titanium film was prepared using a high-vacuum magnetron sputtering device, controlling the base vacuum at 5×10⁻⁻⁻⁻⁶. 5At a pressure of 0.8 Pa and an argon gas pressure of 200 W, a deposition time of 60 min was used to form a nanoporous titanium film with a thickness of 10 μm and a pore size of less than 1 μm. This structure can eliminate the interfacial thermal resistance between the chip and the cooling medium, improve heat transfer efficiency, and increase the system heat flux to 1200 W / cm², solving the problem of insufficient heat dissipation for high-density chips. Simultaneously, a biomimetic sea buckthorn needle cooling tower was fabricated. Needle-shaped fins were formed by stamping with a hydraulic press at a pressure of 200 T, controlling the fin height to 15 mm and the base thickness to 0.5 mm. The fin spacing gradually increased from 0.8 mm at the inlet end to the outlet end. A 1.2mm thick alumina layer, formed through anodizing, increases the heat exchange surface area of ​​the condensing tower by 300% and the air heat exchange efficiency by 40%. This optimizes the critical temperature of supercritical CO2 to 35℃ and reduces the condensing pressure from 10MPa to 7.5MPa. It replaces traditional water-cooled towers to achieve zero-water-consumption operation, solving the problems of high water consumption and high operation and maintenance costs in data centers in arid regions. The implementation of the above structural parameters is synchronized with the cooling cycle process. The microcavity array directly contacts the chip to complete heat absorption, and the biomimetic condensing tower completes air heat exchange and condensation, forming a two-phase cooling closed loop with no water resource consumption.

[0046] In some specific implementations, to address the issues of high dynamic response delay and temperature control overshoot in traditional cooling systems, parameterized control operations using piezoelectric valve assemblies and deep reinforcement learning networks are implemented. Piezoelectric valves are assembled using piezoelectric ceramic actuators with a response time of less than 1ms, controlling the flow regulation accuracy to ±0.1%. A deep reinforcement learning network with a dual-DQN structure is constructed, setting the parameter synchronization step size between the online network and the target network. The state space is defined with four parameters: chip temperature, supercritical CO2 flow rate, pipeline pressure drop, and ambient temperature and humidity. The action space discretizes the piezoelectric valve opening from 0 to 100% into 51 levels, employing a greedy strategy for action selection. A reward function incorporating temperature and pressure differences is constructed, and network parameters are updated through an experience replay mechanism, controlling the response time of the entire control process to be less than 200ms. This technology addresses the temperature control lag problem caused by the response delay of traditional mechanical valves exceeding 30 seconds. Simultaneously, it embeds a physical information neural network into a digital twin platform, fully inputting the Navier-Stokes equations into the network. This ensures that the state prediction error of the virtual mapping model of the physical cooling system is less than ±0.3℃. Through the fault injection module of the digital twin platform, it simulates over 50 fault scenarios, including pump failure, pipeline leakage, and voltage surges. After fault deduction in the virtual mapping model, the control parameters are synchronized to the physical cooling system, increasing the system's fault prediction rate to 92%. This solves the problems of incomplete fault prediction in traditional cooling systems and the ease with which sudden failures can cause equipment downtime. The rapid response of the piezoelectric valve and the real-time decision-making of the deep reinforcement learning network work together to allow the cooling system to adapt to dynamic changes in chip load and maintain a stable chip temperature.

[0047] Step 4. Calculate the reward value and update the parameters of the deep reinforcement learning network: Step 4.1. After the piezoelectric valve opening is adjusted and a set time interval has elapsed, re-collect the temperature data of the chip surface and the pressure data of the supercritical CO2 main circulation pipeline: The set time interval is the time period used to wait for the system operating state to stabilize. After the piezoelectric valve opening is adjusted, the operating state of the cooling system needs a certain amount of time to reach stability. In this embodiment, waiting for the set time interval before re-collecting the chip temperature data and pipeline pressure data can ensure that the collected data can truly reflect the system operating state after the opening adjustment, providing accurate data basis for the calculation of reward value.

[0048] Step 4.2. Calculate the temperature reward based on the difference between the adjusted chip temperature data and the set temperature threshold, and calculate the pressure penalty based on the difference between the adjusted pipeline pressure data and the set pressure threshold. Then, perform a weighted summation of the temperature reward and pressure penalty to calculate the reward value of the deep reinforcement learning network. The temperature threshold is set as the reference temperature value for chip operation, and the pressure threshold is set as the reference pressure value for cooling pipeline operation. The temperature reward is used to measure the adaptability of chip temperature control, and the pressure penalty is used to measure the compliance of pipeline pressure control. The weighted summation is the existing numerical calculation method. By setting weights to integrate the temperature reward and pressure penalty, the reward value of the deep reinforcement learning network is obtained. The reward value is used to evaluate the rationality of the current opening adjustment action.

[0049] Step 4.3. Combine the input state space, valve opening control signal, reward value, and new state space into an experience sample. Store the experience sample in the experience replay pool. During storage, prioritize the experience samples according to the reward value: An experience sample is a complete data unit containing network input, executed action, reward result, and new state. The experience replay pool is an existing data storage medium for storing experience samples. Priority sorting is an existing method of dividing sample storage levels according to the reward value. In this embodiment, sorting the experience samples according to the reward value ensures that high-value samples are called first, improving the effectiveness of network parameter updates.

[0050] Step 4.4. Extract a set number of experience samples from the experience replay pool according to priority, and update the online network parameters of the deep reinforcement learning network using gradient descent: Gradient descent is a calculation method for updating existing neural network parameters. This method can reduce the prediction error of the network through numerical calculation. In this embodiment, high-priority experience samples are extracted from the experience replay pool, and the online network parameters of the deep reinforcement learning network are adjusted by gradient descent to optimize the network's feature extraction and action decision-making capabilities, so that the control signal output by the network is more adapted to the regulation requirements of the cooling system.

[0051] Step 4.5. Copy the online network parameters of the deep reinforcement learning network to the target network of the deep reinforcement learning network according to the set training step size, and complete the synchronous update of the target network parameters of the deep reinforcement learning network: The target network is the auxiliary network used for stable training in the deep reinforcement learning network, and the training step size is the set period for network parameter synchronization. In this embodiment, the updated parameters of the online network are copied to the target network according to the set training step size to complete the synchronous update of the target network parameters, avoid parameter fluctuations during network training, and ensure the stability of the deep reinforcement learning network training process.

[0052] In some embodiments, experience samples can be stored in a fixed-capacity experience replay pool. When the storage capacity of the experience replay pool reaches the set capacity, the experience samples with lower priority in the pool are replaced to ensure that the experience replay pool always stores high-value experience samples, thereby improving the efficiency and effectiveness of network parameter updates.

[0053] In some embodiments, the training of the physical information neural network can adopt an incremental training method, continuously receiving new operating data of the physical cooling system, and iteratively updating the virtual mapping model of the physical cooling system, so that the virtual mapping model of the physical cooling system always fits the actual operating state of the physical cooling system, thereby improving the fit between fault inference and state prediction.

[0054] In some specific implementations, addressing the issues of low maintenance efficiency, poor sealing performance, and high operation and maintenance costs of traditional cooling systems, a magnetic locking assembly and nano-sealing structure are used for assembly and maintenance. Neodymium iron boron (NdFeB) magnetic locking rings are made, with an attraction force exceeding 800N. These rings, combined with 304 stainless steel clips, form a rotary locking structure. Locking and unlocking operations can be completed with a 90° rotation angle, reducing module maintenance time from 30 minutes to less than 5 minutes. This solves the problems of cumbersome disassembly and assembly of traditional threaded fasteners and the need for specialized tools. Simultaneously, a nanoporous titanium membrane is used as the sealing component, with a pressure resistance of 12MPa, and the system leakage rate is controlled at 0.1%. The following features enhance sealing performance and reduce the risk of cooling medium leakage by replacing traditional sealing rings. During system operation and maintenance, the magnetic locking assembly enables quick assembly and disassembly of the microcavity evaporation module, condenser tower, and piping components. The nanoporous titanium membrane does not require frequent replacement, reducing operation and maintenance steps and lowering system operation and maintenance costs by 40%. The parameterized structural design and operation and maintenance are adapted to the normal operation requirements of supercritical CO2 two-phase cooling systems, solving the specific problems of complex maintenance, seal failure, and high operation and maintenance costs of traditional cooling systems. The magnetic quick-release structure and the nano-sealing structure work together to ensure system sealing performance while improving the convenience of component assembly and disassembly and maintenance, making system operation and maintenance more simplified.

[0055] The supercritical CO2 two-phase cooling dynamic control system used in this embodiment supports the implementation of the above-mentioned control method. This system includes a microcavity evaporation module, a biomimetic condenser tower, a supercritical CO2 main circulation pipeline, a piezoelectric valve assembly, a deep reinforcement learning controller, and a digital twin platform. The microcavity evaporation module has a microcavity array on its surface and is in direct contact with the chip to be cooled. The biomimetic condenser tower has needle-like fins on its surface and is connected to a first predetermined section of the supercritical CO2 main circulation pipeline. The supercritical CO2 main circulation pipeline is sequentially connected to a compressor and a gas cooler. The system includes an expander, a storage tank, and a piezoelectric valve assembly located in the first designated section of the supercritical CO2 main circulation pipeline. The deep reinforcement learning controller is electrically connected to the piezoelectric valve assembly, and the digital twin platform is communicatively connected to the deep reinforcement learning controller. The deep reinforcement learning controller is an existing control component used to run a deep reinforcement learning network. It can output control signals to drive the piezoelectric valve assembly. The communication connection is an existing data transmission connection method, which can realize data interaction between the digital twin platform and the deep reinforcement learning controller. The various components of the system cooperate with each other to provide hardware and carrier support for the implementation of the control method.

[0056] This embodiment achieves complete dynamic control of supercritical CO2 two-phase cooling by collecting and integrating cooling system operating status data, processing the data to construct a network input space, regulating the cooling medium flow rate based on a deep reinforcement learning network, and calculating reward values ​​to update network parameters. Combined with digital twin virtual mapping, fault simulation, cyclic path control, and system hardware support, this embodiment fully realizes dynamic control of supercritical CO2 two-phase cooling. Through the collection and processing of operating status data, this embodiment provides accurate basic information for flow regulation. Real-time regulation is then achieved through a deep reinforcement learning network, keeping the cooling system's operating status within a stable range and reducing fluctuations in chip temperature and pipeline pressure. The virtual mapping model constructed through a physical information neural network allows for the simulation of system operating status and pre-adaptation to control operations under fault scenarios. The closed-loop circulation and flow distribution of supercritical CO2 improves the uniformity of heat transfer, and combined with the adjustment of the condenser tower's ventilation volume, stabilizes the operating efficiency of the cooling cycle. Parameter updates through the deep reinforcement learning network continuously optimize the control strategy, allowing flow regulation actions to gradually adapt to the operating patterns of the cooling system. The control process in this embodiment requires no manual intervention. All steps are implemented based on existing technology and specialized components, which can adapt to the normal dynamic control requirements of supercritical CO2 two-phase cooling systems and improve the stability and continuity of the cooling system operation.

[0057] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for dynamic control of supercritical CO2 two-phase cooling, characterized in that, Includes the following steps: S1. Acquire cooling system operating status data, including chip temperature data, supercritical CO2 flow rate data, pipeline pressure data, and ambient temperature and humidity data; S2. Filter and normalize the cooling system operating status data to construct the input state space of the deep reinforcement learning network; S3. Input the input state space into a deep reinforcement learning network, and output a valve opening control signal through the decision of the deep reinforcement learning network. Adjust the opening of the piezoelectric valve according to the valve opening control signal, thereby adjusting the flow rate of the supercritical CO2 main circulation. S4. Obtain the adjusted cooling system operating status data, calculate the reward value of the deep reinforcement learning network based on the adjusted cooling system operating status data, and update the parameters of the deep reinforcement learning network based on the reward value through the experience replay mechanism.

2. The method according to claim 1, characterized in that, Step S1 includes the following sub-steps: S1.

1. Temperature data of the chip surface is collected by temperature sensors arranged on the chip surface. The temperature sensors correspond to the area of ​​the chip, and the collected temperature data includes the location identifier of the corresponding area and the collection time information. S1.

2. Flow and pressure data of the supercritical CO2 main circulation pipeline are collected by flow sensors and pressure sensors installed at the location of the supercritical CO2 main circulation pipeline. The flow sensors collect the mass flow rate data of the supercritical CO2, and the pressure sensors collect the pressure data of the first set pipeline section and the second set pipeline section respectively. S1.

3. Temperature and humidity data of the environment in which the cooling system is located are collected by temperature and humidity sensors installed in the external environment of the cooling system, and the collection frequency is consistent with the collection frequency of chip temperature data; S1.

4. Align all collected data according to a unified timestamp, remove data with mismatched timestamps, convert the aligned data into a unified data format, and generate cooling system operating status data containing all collected parameters.

3. The method according to claim 1, characterized in that, Step S2 includes the following sub-steps: S2.

1. Perform sliding window mean filtering on each type of data in the cooling system operating status data. The size of the sliding window is set according to the data acquisition frequency. The time series characteristics of the data are preserved during the filtering process. S2.

2. Normalize each type of cooling system operating status data after filtering. Use linear normalization to map each type of data to a set value range. During the normalization process, retain the relative size relationship between the data. S2.

3. Combine all the normalized data in a preset order to generate an input state space in the form of a one-dimensional vector. The dimension of the input state space is consistent with the number of parameters of the cooling system operating state data.

4. The method according to claim 1, characterized in that, Step S3 includes the following sub-steps: S3.

1. Input the input state space into the online network of the deep reinforcement learning network. The online network of the deep reinforcement learning network extracts features from the input state space through multiple fully connected layers and calculates the action value corresponding to each optional action. The fully connected layers use a linear activation function for output. S3.

2. A greedy strategy is adopted to select the action with the highest action value from all available actions. The action space is the opening range of the piezoelectric valve. The opening range is discretized into gears, and each gear corresponds to an independent action. The valve opening control signal corresponding to the gear is generated. S3.

3. Convert the valve opening control signal into a piezoelectric ceramic actuator driving voltage signal, send the driving voltage signal to the piezoelectric ceramic actuator of the piezoelectric valve, drive the piezoelectric ceramic actuator to generate displacement, drive the valve core to move and adjust to the corresponding opening degree, and regulate the flow rate of the supercritical CO2 main circulation.

5. The method according to claim 1, characterized in that, Step S4 includes the following sub-steps: S4.

1. After the piezoelectric valve opening is adjusted and a set time interval has elapsed, re-collect the temperature data on the chip surface and the pressure data of the supercritical CO2 main circulation pipeline; S4.

2. Calculate the temperature reward based on the difference between the adjusted chip temperature data and the set temperature threshold, calculate the pressure penalty based on the difference between the adjusted pipeline pressure data and the set pressure threshold, and sum the temperature reward and pressure penalty together to calculate the reward value of the deep reinforcement learning network. S4.

3. Combine the input state space, valve opening control signal, reward value and new state space into an experience sample, store the experience sample in the experience playback pool, and prioritize the experience sample according to the size of the reward value during the storage process; S4.

4. Extract a set number of experience samples from the experience replay pool according to priority, and update the online network parameters of the deep reinforcement learning network using gradient descent. S4.

5. According to the set training step size, copy the online network parameters of the deep reinforcement learning network to the target network of the deep reinforcement learning network to complete the synchronous update of the target network parameters of the deep reinforcement learning network.

6. The method according to claim 1, characterized in that, In step S1, a digital twin platform is pre-constructed. The digital twin platform is a virtual mapping carrier based on a physical information neural network. The physical information neural network includes an input layer, a hidden layer, and an output layer. The input layer receives the structural and operational parameters of the physical cooling system. The hidden layer contains multiple fully connected layers, which are linearly connected. The output layer outputs the predicted operating state of the physical cooling system. The physical information neural network embeds fluid dynamics control equations and thermodynamic control equations. The structural and operational parameters of the physical cooling system are input into the physical information neural network for training. The training steps include inputting a training dataset, calculating a loss function, updating the parameters of the physical information neural network using gradient descent, and generating a virtual mapping model of the physical cooling system. The operating state data of the physical cooling system is synchronized to the digital twin platform in real time according to a set sampling rate. During the synchronization process, the data is verified and abnormal data is removed.

7. The method according to claim 6, characterized in that, In step S3, fault scenarios are injected into the virtual mapping model of the physical cooling system through the fault injection module of the digital twin platform. The fault scenarios include pump failure, pipeline leakage and voltage surge. Different fault severity parameters are set for the fault scenarios. The operating state of the virtual mapping model of the physical cooling system after the fault is injected is deduced based on the physical information neural network, and the execution content corresponding to the fault scenario is generated. The execution content includes the valve opening adjustment method at different times. The generated execution content is fused with the valve opening control signal output by the deep reinforcement learning network to adjust the final output valve opening control signal.

8. The method according to claim 1, characterized in that, In step S3, the operating path of the supercritical CO2 main cycle is as follows: supercritical CO2 enters the microcavity evaporation module, and through the microcavity array of the microcavity evaporation module, it directly contacts the chip to be cooled, absorbs the heat generated by the chip, and undergoes a phase change, changing from liquid to gas. The gaseous supercritical CO2 after absorbing heat enters the compressor for pressurization. The pressurized supercritical CO2 enters the condenser, and through the fins of the condenser, it exchanges heat with the outside air, releases heat, and condenses. The supercritical CO2 after releasing heat and condensation enters the expander for depressurization. The depressurized supercritical CO2 returns to the liquid state and returns to the microcavity evaporation module to complete the closed operation.

9. The method according to claim 8, characterized in that, In step S3, based on the temperature distribution data of different areas of the chip, the flow rate of the supercritical CO2 main circulation is allocated to the corresponding area of ​​the microcavity evaporation module. When the temperature of the area reaches the set temperature value, the corresponding flow rate is allocated. At the same time, the ventilation volume of the condenser is adjusted according to the ambient temperature data. When the ambient temperature reaches the set temperature value, the condenser is adjusted to the set ventilation volume value to control the condensation temperature of the supercritical CO2.

10. A supercritical CO2 two-phase cooling dynamic control system, used to execute the method as described in any one of claims 1-9, characterized in that, The system includes a microcavity evaporation module, a biomimetic condenser tower, a supercritical CO2 main circulation pipeline, a piezoelectric valve assembly, a deep reinforcement learning controller, and a digital twin platform. The microcavity evaporation module has a microcavity array on its surface and is in direct contact with the chip to be cooled. The biomimetic condenser tower has needle-like fins on its surface and is connected to the first designated section of the supercritical CO2 main circulation pipeline. The supercritical CO2 main circulation pipeline is sequentially connected to a compressor, a gas cooler, an expander, and a liquid storage tank. The piezoelectric valve assembly is located in the first designated section of the supercritical CO2 main circulation pipeline. The deep reinforcement learning controller is electrically connected to the piezoelectric valve assembly, and the digital twin platform is communicatively connected to the deep reinforcement learning controller.

Images