Single-phase and two-phase immersion liquid cooling method and system based on AI intelligent decision-making

Abstract

The application relates to the field of data center heat dissipation technology, and particularly provides a single-phase and two-phase immersion liquid cooling method and system based on AI intelligent decision-making, which comprises a data injection dynamic feature extraction engine after coupling, outputs a thermodynamic state evolution tensor containing features such as temperature change rate and load-heat flux density coupling coefficient, inputs the thermodynamic state evolution tensor into a deep neural network model, calculates the temperature and pressure adapted to the current thermodynamic state evolution tensor, generates a closed-loop control instruction set executable by equipment, and injects the closed-loop control instruction set into an actuator group; the actuator executes power reconstruction and flow channel switching according to the instruction; gaseous fluorinated liquid in a two-phase mode is liquefied backflow through an efficient condenser, and heat dissipation mode adaptive switching and heat cycle reconstruction are realized. The system comprises a server, an AI algorithm controller, a cooling liquid storage device, a condenser, a circulating pump, an electric valve, a pressure relief valve and a temperature sensor. The application significantly improves heat dissipation efficiency and system reliability.

Description

Technical Field

[0001] This invention relates to the field of data center heat dissipation technology, and in particular to a single-phase and two-phase immersion liquid cooling method and system based on AI intelligent decision-making. Background Technology

[0002] Traditional air cooling technology, limited by air's specific heat capacity and heat dissipation efficiency, struggles to meet ever-increasing heat dissipation demands, leading to the emergence and rapid development of liquid cooling technology. With the rapid advancements in artificial intelligence, big data, and other technologies, data centers are continuously expanding in scale, and server power density is surging, presenting unprecedented challenges to heat dissipation. Traditional air cooling technology is no longer sufficient for high heat flux density scenarios, making immersion liquid cooling technology a focal point in the industry. Existing single-phase / two-phase immersion liquid cooling systems largely rely on simple threshold control, determining mode switching solely based on the coolant temperature, which cannot accurately handle complex operating conditions. For example, during sudden peak computing power in data centers (such as the instantaneous startup of AI model training), server load fluctuates dramatically in a short period, resulting in uneven and rapidly changing heat flux distribution. Traditional control modes are prone to switching lag or over-switching, leading to localized overheating that affects equipment lifespan or increasing system energy consumption due to frequent switching. Furthermore, traditional systems lack comprehensive consideration of the coupled effects of multiple factors such as coolant characteristics, real-time server load, and ambient temperature and humidity, making it difficult to achieve an optimal balance between heat dissipation efficiency and energy consumption, thus limiting the efficient and stable operation of data centers.

[0003] Prior art 1, Chinese Patent Application No. 202211150784.7, discloses an immersion liquid cooling method, an immersion liquid cooling structure, and a liquid cooling system. The immersion liquid cooling method includes: immersing the heat exchange fluid of the liquid cooling system into the heat dissipation components of the liquid cooling system; and using a fluid-driven component to move the heat exchange fluid upstream of the heat dissipation components, increasing the flow velocity of the heat exchange fluid through the heat dissipation components, wherein the fluid-driven component is positioned horizontally opposite to the heat dissipation components. Although immersing the heat exchange fluid into the heat dissipation components and using a fluid-driven component positioned horizontally opposite to the heat dissipation components to move the heat exchange fluid upstream of the heat dissipation components, thereby increasing the flow velocity of the heat exchange fluid through the heat dissipation components, effectively improves heat exchange efficiency and increases the amount of heat exchanged; however, relying solely on mechanical fluid drive to increase the flow velocity lacks the ability to respond to dynamic changes in heat flux density, cannot adaptively adjust the cooling mode (single-phase / two-phase) according to load fluctuations, and does not establish a closed-loop correlation between thermodynamic state and execution control.

[0004] Prior art two, Chinese patent application number 202410606085.1, discloses a temperature control system and method for an immersion liquid cooling device. The temperature control system includes an immersion liquid cooling device and a temperature control device. The immersion liquid cooling device is filled with immersion liquid, which flows to the temperature control device through an inlet pipe and flows back to the immersion liquid cooling device through a return pipe. The temperature control device includes: a temperature detection unit disposed on the inlet and return pipes of the immersion liquid to detect the inlet and return temperatures of the immersion liquid; a variable frequency liquid pump connected to the inlet or return pipe of the immersion liquid to drive the flow of the immersion liquid; a coolant circuit assembly including a heat exchanger for heat exchange between the immersion liquid and the coolant and a coolant pipeline for supplying coolant to the heat exchanger, with a throttle valve connected to the coolant pipeline; and a controller unit electrically connected to the temperature detection unit, the variable frequency liquid pump, and the throttle valve to control the flow of the immersion liquid based on the data measured by the temperature detection unit. Although the operating frequency of the variable frequency liquid pump and the opening of the throttle valve are controlled to regulate the temperature of the immersion liquid, the reliance on traditional PID temperature control logic cannot handle nonlinear thermodynamic processes. It only adjusts the flow rate and valve opening, and does not solve the problems of phase change working fluid recovery and thermal cycle reconstruction. The temperature detection unit has a simple layout and is difficult to reflect local hot spots on the server.

[0005] Prior art three, Chinese patent application number 202410642116.9, discloses a data center immersion liquid cooling system and control method. The cooling system includes dual cold sources, a mixed cooling capacity distribution module, an air-cooled module, and an immersion liquid cooling module. The dual cold source module is located outside the data center server room and has three modes: completely natural cooling, mechanical refrigeration, and dual cold source cooling. The mixed cooling capacity distribution module, air-cooled module, and immersion liquid cooling module inside the server room constitute a mixed air-liquid cooling system. During operation, the secondary-side immersion coolant circuit and air-cooled circuit remove heat from the equipment and exchange heat with the primary-side chilled water circuit. The heat is then transferred to the outdoor environment through the mixed cooling capacity distribution module. Although the energy consumption of the cold source is reduced by making full use of the natural cold source, and the cooling efficiency of the immersion liquid cooling system is improved by coupling active and passive heat transfer enhancement technologies, and then the efficient and low-carbon operation of the data center dual-cold source air-liquid mixed cooling system is achieved by using intelligent operation control strategies, the mixed cooling system has a complex structure, the mode switching is lagging, the utilization of the natural cold source and the immersion liquid cooling lack deep collaborative control, and the optimization of condensation reflux under two-phase flow conditions is not addressed.

[0006] Current technologies 1, 2, and 3 suffer from problems such as rigid cooling modes, sluggish dynamic response, two-phase flow control instability, and poor adaptability to abnormal operating conditions. Therefore, this invention provides a single-phase and two-phase immersion liquid cooling method and system based on AI intelligent decision-making; based on AI intelligent decision-making, a multi-dimensional data fusion mode switching and heat dissipation optimization system is constructed to accurately adapt to complex operating conditions and break through the bottlenecks of traditional control. Summary of the Invention

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] One aspect of the present invention provides a single-phase and two-phase immersion liquid cooling method based on AI intelligent decision-making, comprising the following steps:

[0009] The thermodynamic state evolution tensor, which includes the characteristics of temperature change rate and load-heat flux density coupling coefficient, is input into a deep neural network model to calculate the temperature and pressure that the current thermodynamic state evolution tensor is adapted to. The output of the deep neural network model is dynamically corrected by a reinforcement learning compensator based on the deviation between the current actual pressure and temperature and the target value. The corrected instructions are then converted into a set of closed-loop control instructions that can be executed by the equipment through a physical signal converter.

[0010] The closed-loop control instruction set is injected into the actuator group. The actuator executes the power reconfiguration and flow channel switching according to the instructions. The single-phase and two-phase flow channel opening and closing combination is triggered by the heat flux density gradient threshold. When the flow channel switching is completed, the condensation-reflux closed loop is activated simultaneously. In the two-phase mode, the gaseous fluorinated liquid is liquefied and refluxed through the high-efficiency condenser to realize the adaptive switching of heat dissipation mode and thermal cycle reconstruction.

[0011] In one alternative implementation, the process of generating a set of closed-loop control instructions executable by the device includes the following steps:

[0012] Based on the output thermodynamic state evolution tensor, multimodal feature decoupling is performed; the temperature change rate feature contained in the thermodynamic state evolution tensor and the load-heat flux density coupling coefficient are separated into independent control channels;

[0013] The decoupled feature input dual-channel deep neural network model performs joint inference. The temperature channel network uses historical pressure fluctuation patterns as constraints to output the phase change critical point prediction value; the load channel network combines the filtered real-time load data to generate the optimal heat exchange efficiency parameters.

[0014] The joint inference results are fed into reinforcement learning compensation, which uses the deviation between the temperature and pressure data measured by the current sensors and the output values ​​of the deep neural network model to construct a dynamic compensation matrix; the compensated control parameters form a preliminary command vector.

[0015] The physical signal conversion stage maps the initial command vector to the device control topology. Based on the spatial parameters provided by the server heat flux density distribution model, the abstract control quantities in the initial command vector are converted into the physical dimensions of the specific execution unit.

[0016] In one alternative implementation, the process of constructing the dynamic compensation matrix includes the following steps:

[0017] An initial deviation field is established based on the predicted critical point of phase transition in the temperature channel and the heat exchange efficiency parameter of the load channel output by the dual-channel neural network. The measured temperature and pressure data from the sensor and the predicted values ​​from the deep neural network are then subjected to a three-dimensional difference calculation to form the original deviation vector field with physical dimensions.

[0018] The original deviation vector field is input into the trend backtracking filter for processing. The generated thermodynamic state evolution tensor is used as the historical benchmark, and the time-series patterns of temperature change rate and load coupling coefficient are extracted as filter weights. The filtered deviation field becomes the steady-state deviation characterization.

[0019] Steady-state deviation characterization is incorporated into spatial mapping. Based on the spatial topological relationship provided by the heat flux density distribution model, the two-dimensional deviation data is reprojected onto the actual physical coordinates of the thermally sensitive area of ​​the server chip, forming a deviation distribution matrix with spatial resolution.

[0020] The compensation matrix is ​​generated by dynamic weight fusion. Based on the deviation distribution matrix after spatial mapping, the heat flux density gradient information shared by the intermediate layer of the dual-channel neural network is superimposed as the adjustment coefficient. Each element value of the fused matrix represents the compensation intensity required at a specific spatial location, thus completing the transformation from the original deviation to the compensation parameter.

[0021] In one optional implementation, the process of superimposing the shared heat flux density gradient information of the intermediate layers of the dual-channel neural network as an adjustment coefficient includes the following steps:

[0022] The deviation distribution matrix is ​​decomposed into its eigendomain, and the obtained spatial projection results are separated into temperature channel deviation components and load channel deviation components to form two orthogonal temperature channel submatrices and load channel submatrices.

[0023] For the temperature channel submatrix, extract the temperature sensitivity coefficient from the heat flux density gradient information shared by the intermediate layer of the dual-channel neural network; nonlinearly combine the temperature sensitivity coefficient with the confidence index of the phase transition critical point prediction value to generate a temperature compensation weight field.

[0024] For the load channel submatrix, the spatial correlation strength in the heat flux density gradient information is used as the modulation factor, carrying the load dynamic characteristics in the thermodynamic state evolution tensor. Combined with the historical fluctuation data of the heat exchange efficiency parameter output by the load channel network, a load compensation weight field is generated through sliding window variance analysis.

[0025] The compensation matrix is ​​constructed through tensor product operation of dual-channel weighted fields. The operation results are superimposed using spatial topological rules of physical signal conversion to form a complete dynamic compensation matrix.

[0026] In one optional implementation, the process of generating the temperature compensation weight field and the load compensation weight field includes the following steps:

[0027] Temperature sensitivity coefficients are extracted from the heat flux density gradient information shared by the intermediate layers of a dual-channel neural network. These temperature sensitivity coefficients contain the instantaneous temperature change rate characteristics and historical phase transition behavior patterns in the thermodynamic state evolution tensor.

[0028] The temperature sensitivity coefficient is dynamically coupled with the confidence index of the output phase transition critical point prediction value. The confidence index of the phase transition critical point prediction value reflects the confidence of the neural network in the current phase transition boundary.

[0029] The output is a temperature sensitivity-confidence fused field;

[0030] The trend backtracking filter, established by the temperature sensitivity-confidence fusion field input, uses the time series pattern of temperature change rate in the historical thermodynamic state evolution tensor to calculate the time decay coefficient of the compensation weight.

[0031] The time decay coefficient is nonlinearly superimposed with the fusion field to generate a temperature-compensated weight field. Its spatial distribution is aligned with the deviation distribution region of the mapping, forming a high-weight gradient band in the phase transition critical region.

[0032] In one optional implementation, the process of nonlinearly superimposing the time decay coefficient with the fusion field includes the following steps:

[0033] The time decay coefficient output from the trend backtracking filter is calculated based on the time-series pattern of the rate of temperature change in the historical thermodynamic state evolution tensor. The time decay coefficient is dynamically adjusted as the rate of temperature change changes.

[0034] The temperature sensitivity-confidence fusion field already contains the weight distribution information of the active phase transition region. For each weight value in the temperature sensitivity-confidence fusion field, a nonlinear mapping is used to adjust it so that the weights exhibit a smooth transition under the influence of the time decay coefficient, rather than a nonlinear abrupt change.

[0035] The time decay coefficient acts on the weight distribution of the fusion field, making the weight adjustment process conform to the thermal inertia law; the resulting dynamic compensation weight field forms a high-weight gradient band in the phase transition critical region.

[0036] In one optional implementation, the process of making the weight adjustment process conform to the thermal inertia law of the system includes the following steps:

[0037] The temporal fluctuation characteristics of the rate of temperature change are extracted from the historical thermodynamic evolution tensor to generate a time decay coefficient with spatiotemporal continuity.

[0038] The preset weight distribution and time decay coefficient are deeply coupled in the temperature sensitivity-confidence fusion field;

[0039] The adjusted weight field ultimately forms a dynamic gradient structure with thermodynamic adaptability: in the active phase transition region, the weight difference between adjacent grid points is controlled within a dynamic threshold determined by the decay coefficient, and the dynamic threshold changes negatively with the thermal inertia intensity of the system; while in the region where thermal disturbances occur frequently, the weight change rate is constrained within the product of the time decay coefficient and the local temperature sensitivity.

[0040] In one optional implementation, the process of setting the heat flux density gradient threshold includes the following steps:

[0041] The load-heat flux density coupling coefficient is separated from the thermodynamic state evolution tensor output by the dynamic feature extraction engine and fused with the temperature change rate predicted by the deep neural network model to generate a regional heat flux dynamic index, which comprehensively reflects the dynamic evolution trend of heat load per unit area under the current operating conditions.

[0042] The pressure-temperature deviation correction output from the reinforcement learning compensator and the regional heat flow dynamic index are input together into the nonlinear mapping module to generate the initial gradient threshold benchmark.

[0043] The initial gradient threshold is adaptively calibrated using historical data of the actuator's flow channel switching. When the continuous opening time of the single-phase flow channel exceeds the critical point of condenser efficiency decay, the initial gradient threshold is reduced proportionally according to the real-time efficiency coefficient of the condensation-reflux closed loop. If the heat flux density oscillation amplitude is detected to exceed the stability margin during the operation of the two-phase flow channel, the initial gradient threshold is dynamically increased according to the reciprocal relationship of the oscillation frequency.

[0044] In one optional implementation, the real-time data streams of temperature, pressure, and load are filtered out for outliers using the 3σ principle. The filtered data streams are then spatiotemporally aligned and coupled with the server's heat flux density distribution model. The coupled data is then injected into a dynamic feature extraction engine, which outputs a thermodynamic state evolution tensor.

[0045] Another aspect of the present invention provides a single-phase and two-phase immersion liquid cooling system based on AI intelligent decision-making, for implementing the aforementioned single-phase and two-phase immersion liquid cooling method based on AI intelligent decision-making, comprising: a server, an AI algorithm controller, a coolant reservoir, a condenser, a circulating pump, an electric valve, a pressure relief valve, and a temperature sensor.

[0046] The top of the server is connected to the inlet of the coolant reservoir via a coolant pipe. The outlet of the coolant reservoir is connected to the inlet of the condenser via a coolant pipe. The outlet of the condenser is connected to the inlet of the circulation pump via a coolant pipe. The outlet of the circulation pump is connected to one end of an electric valve, and the other end of the electric valve is connected to the bottom of the server. A pressure relief valve is installed on the right side of the top of the server. The pressure relief valve is connected to the AI ​​algorithm controller via a data cable. The AI ​​algorithm controller is connected to a temperature sensor via a data cable. The temperature sensor is embedded in the server.

[0047] This invention utilizes AI-driven temperature-pressure coordinated control to fully leverage the characteristics of this boiling point fluorinated liquid in both single-phase (stable at ambient pressure) and two-phase (efficient phase change under reduced pressure) modes. This results in improved heat dissipation efficiency compared to traditional single-mode operation, and faster response times to effectively handle scenarios such as sudden changes in server load and environmental fluctuations. The model's self-learning mechanism enables the system to continuously optimize over time, in response to changes in environment and business needs, eliminating the need for frequent manual adjustments. Significant energy consumption optimization is achieved through dynamic pressure adjustment and precise mode matching, reducing energy consumption during off-peak hours and lowering overall lifecycle operating costs. Attached Figure Description

[0048] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0049] Figure 1 This is a flowchart of the single-phase and two-phase immersion liquid cooling method based on AI intelligent decision-making provided in Embodiment 1 of the present invention;

[0050] Figure 2 This is a schematic diagram of the single-phase and two-phase immersion liquid cooling method based on AI intelligent decision-making provided in Embodiment 1 of the present invention;

[0051] Figure 3 This is a process diagram of the output thermodynamic state evolution tensor provided in Embodiment 2 of the present invention;

[0052] Figure 4 This is a process diagram of generating a set of closed-loop control instructions executable by the device, as provided in Embodiment 4 of the present invention;

[0053] Figure 5 This is a diagram illustrating the process of setting the heat flux density gradient threshold provided in Embodiment 10 of the present invention;

[0054] Figure 6 This is a schematic diagram of the single-phase and two-phase immersion liquid cooling system based on AI intelligent decision-making provided in Embodiment 11 of the present invention;

[0055] Figure 7 This is a schematic diagram of the AI ​​algorithm controller process provided in Embodiment 11 of the present invention;

[0056] Figure 8 This is a schematic diagram of the AI ​​algorithm controller instruction conversion provided in Embodiment 11 of the present invention;

[0057] Figure 9 A block diagram of the electronic device provided by the present invention;

[0058] Figure 10 A block diagram of a computer-readable storage medium provided for this invention. Detailed Implementation

[0059] The technical solutions of the present invention will now be described with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0060] Hereinafter, the terms "first," "second," etc., are used for descriptive convenience only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0061] In this invention, unless otherwise explicitly specified and limited, the term "connection" should be interpreted broadly. For example, "connection" can be a fixed mechanical connection, a detachable mechanical connection, or an integral part; or, "connection" can be a direct connection or an indirect connection through an intermediate medium. Furthermore, unless otherwise explicitly specified and limited, the term "coupling" should be interpreted broadly. For example, "coupling" can be a direct electrical connection, such as physical contact and electrical conduction between two components; it can also be understood as an electrical connection between different components in a circuit structure through physical lines capable of transmitting electrical signals, such as copper foil or wires on a printed circuit board (PCB), to transmit electrical signals; or, "coupling" can be an indirect electrical connection between two components through an intermediate medium; or, "coupling" can be an electrical connection between two components in a non-contact manner, such as an electrical connection between two components using capacitive coupling to transmit electrical signals.

[0062] In this embodiment of the invention, directional terms such as "up," "down," "left," and "right" may be defined relative to the orientation of the components shown in the accompanying drawings. It should be understood that these directional terms can be relative concepts, used for relative description and clarification, and can change accordingly depending on the orientation of the components in the accompanying drawings.

[0063] Example 1: As Figure 1 As shown, this embodiment of the invention provides a single-phase and two-phase immersion liquid cooling method based on AI intelligent decision-making, comprising the following steps:

[0064] Step S100: The real-time data streams of temperature, pressure and load are filtered out for outliers using the 3σ principle. The filtered data streams are spatiotemporally aligned and coupled with the server's heat flux density distribution model. The coupled data is injected into the dynamic feature extraction engine, which outputs a thermodynamic state evolution tensor containing features such as the rate of temperature change and the load-heat flux density coupling coefficient.

[0065] Step S200: The thermodynamic state evolution tensor is input into the deep neural network model to calculate the temperature and pressure that the current thermodynamic state evolution tensor is adapted to. The output of the deep neural network model is dynamically corrected by a reinforcement learning compensator based on the deviation between the current actual pressure and temperature and the target value. The corrected instructions are then converted into a set of closed-loop control instructions that the equipment can execute, including pump power setpoints, valve opening instructions, etc., by a physical signal converter.

[0066] Step S300: The closed-loop control instruction set is injected into the actuator group. The actuator executes the power reconstruction and flow channel switching according to the instructions. The single-phase and two-phase flow channel opening and closing combination is triggered by the heat flux density gradient threshold. When the flow channel switching is completed, the condensation-reflux closed loop is activated simultaneously. In the two-phase mode, the gaseous fluorinated liquid is liquefied and refluxed through the high-efficiency condenser to realize the adaptive switching of heat dissipation mode and thermal cycle reconstruction.

[0067] In the above embodiments, the appendix Figure 2This is a schematic diagram of single-phase and two-phase immersion liquid cooling methods based on AI intelligent decision-making. This embodiment achieves intelligent dynamic control of the immersion liquid cooling system through the systematic coupling of three key steps. The combined effect of its technical features is as follows: 3σ outlier screening ensures data reliability; spatiotemporal alignment coupling eliminates the temporal / spatial deviation between sensor data and the server heat flux density distribution model (the server heat flux density distribution model is constructed by fusing computational fluid dynamics simulation data and historical infrared thermal imaging data, using a three-dimensional convolutional neural network to spatially encode the internal chip layout of the server, establishing a nonlinear mapping relationship between heat source location and heat dissipation path, and introducing a temporal convolutional layer to capture transient heat load change characteristics, ultimately outputting a heat flux density gradient field with spatiotemporal continuity. During the training phase, the model enhances data generalization ability through adversarial generative networks, ensuring high-precision modeling of the thermal distribution characteristics of different server architectures). The dynamic feature extraction engine compresses multidimensional data into a thermodynamic state tensor containing key parameters such as temperature change rate and load-heat flux coupling coefficient, providing high-fidelity input for subsequent decision-making. A deep neural network model (emphasizing a dual-channel heterogeneous architecture: the temperature prediction channel uses a temporal convolutional network to extract the temperature change rate features from the thermodynamic state evolution tensor; the load prediction channel uses a graph neural network to model the spatial correlation of the heat flux density coupling coefficient; the two channels share heat flux density gradient information through a cross-modal attention mechanism; and historical pressure fluctuation mode constraints and real-time load data features are fused at the output layer. Finally, the physical properties of the original tensor are preserved through residual connections, forming an intelligent decision-making model with both temporal dynamics and spatial correlation) predicts optimal temperature and pressure parameters based on the thermodynamic tensor. A reinforcement learning compensator dynamically corrects the model output error through real-time deviation feedback, and a physical signal converter converts the correction instructions into an executable control instruction set (pump power, valve opening, etc.), forming a control closed loop with online adaptability. The actuator dynamically adjusts the flow channel topology according to the instructions: single-phase / two-phase flow channel switching is triggered by the heat flux density gradient threshold to achieve adaptive selection of the cooling mode. In two-phase mode, the condensation-reflux closed loop is activated synchronously, and the working fluid balance is maintained through the liquefaction and recovery of gaseous fluorinated liquid. The coordinated control of dynamic reconfiguration and flow channel switching achieves rapid reconstruction of the thermal cycle system.

[0068] In summary, this embodiment improves the system's response accuracy to transient load changes through a three-level architecture of abnormal data processing, dynamic modeling, and reinforcement learning compensation; achieves dynamic optimal matching of heat dissipation capacity and energy consumption based on a dual-mode switching mechanism using heat flux density gradient threshold; ensures the stability of the working fluid circulation under two-phase conditions through coordinated control of the condensation reflux system and the flow channel topology; and forms a complete intelligent control closed loop from data perception to execution feedback, significantly improving heat dissipation efficiency and system reliability.

[0069] Example 2: Figure 3As shown, based on Example 1, the process of outputting the thermodynamic state evolution tensor in step S100 of this embodiment of the invention includes the following steps:

[0070] Step S101: Input the real-time data streams of temperature, pressure and load into the dynamically filtered boundary field generated based on the historical operating condition database, and perform thermodynamic outlier annihilation;

[0071] Step S102: The annihilated data stream and the thermal conduction phase field of the server chip material are compensated for thermal hysteresis. The compensation amount = heat flux density × material relaxation time, thus completing the spatiotemporal energy alignment.

[0072] Step S103: Inject energy-aligned data into the intrinsic mode decomposition cavity and output the thermodynamic state evolution tensor.

[0073] In the above embodiments, this embodiment identifies and eliminates abnormal data points by dynamically filtering boundary fields, ensuring that the input data conforms to the normal operating conditions of the system and providing a high-quality data foundation for subsequent processing. The thermal hysteresis compensation module corrects the timing deviation between the measured data and the actual thermal state caused by the thermal conductivity characteristics of the chip material, achieving strict spatiotemporal synchronization between sensor data and physical processes. The intrinsic mode decomposition cavity decomposes the synchronized multidimensional data into a state tensor containing the essential thermodynamic characteristics of the system. This tensor fully characterizes core features such as dynamic temperature changes and the coupling relationship between load and heat dissipation. The three-stage processing forms a serial data processing chain: anomaly filtering, timing correction, and feature extraction. The final output thermodynamic state evolution tensor has the following characteristics: it eliminates the interference of sensor noise and material thermal inertia, establishes a direct correlation between load changes and thermal flow response, and provides an adapted data structure for standardized neural network input.

[0074] In summary, this embodiment achieves a reliable conversion from raw sensor data to decisionable thermodynamic characteristics, providing accurate state input for subsequent intelligent control.

[0075] Example 3: Based on Example 2, the process of injecting energy-aligned data into the intrinsic mode decomposition cavity in step S103 of this embodiment of the invention includes the following steps:

[0076] Step S1031: Based on the output spatiotemporal energy alignment data, perform multidimensional thermodynamic field reconstruction; map the compensated temperature gradient distribution, pressure fluctuation characteristics and load dynamic changes in a three-dimensional field to form a spatiotemporally continuous heat flux density-material response joint distribution field.

[0077] Step S1032: The heat flux density-material response joint distribution field enters feature decoupling. Based on the topology of the heat conduction path inside the server chip, the field data is decomposed into axial conduction components and radial diffusion components. The axial component reflects the heat accumulation characteristics of the chip in the vertical direction, and the radial component characterizes the planar heat dissipation capability. The weights are automatically adjusted by the proportion of historical operating conditions recorded in the dynamically filtered boundary field.

[0078] Step S1033: The axial and radial components are nonlinearly superimposed in the boundary transition region. During the superposition process, the calculated heat flux density compensation is compared in real time to ensure feature fusion under the energy conservation constraint; a transition state thermodynamic feature matrix is ​​generated, the dimension of which corresponds to the spatial resolution of the chip's thermally sensitive area.

[0079] Step S1034: Generate a thermodynamic state evolution tensor through variable-scale feature distillation. Based on the transition state matrix, slide and extract feature windows in the time dimension. Perform the following operations in each window: extract the pressure fluctuation pattern corresponding to the extreme point of temperature change, associate it with the load change rate of the current window, combine it with the data distribution characteristics retained after outlier annihilation, and output a tensor slice with spatiotemporal correlation. Stack the continuous slices according to the thermal relaxation period to form a complete thermodynamic state evolution tensor.

[0080] In the above embodiments, the thermodynamic state evolution tensor is essentially a computable state characterization carrier formed by multi-level field reconstruction and feature distillation, based on the thermal response hysteresis characteristics of the chip material, the energy rebalancing process under dynamic load, and the pure data features filtered for abnormal operating conditions. Using the energy distribution corrected for material relaxation time as the input benchmark, feature decoupling and recombination are achieved based on the chip's thermal conduction topology, and a sliding window distillation mechanism is employed to maintain temporal continuity. The entire process forms an end-to-end transformation chain from energy-aligned data to the state tensor.

[0081] Example 4: Figure 4 As shown, based on Embodiment 1, the process of generating a device-executable closed-loop control instruction set in step S200 of this embodiment of the invention includes the following steps:

[0082] Step S201: Based on the output thermodynamic state evolution tensor, perform multimodal feature decoupling; separate the temperature change rate feature and load-heat flux density coupling coefficient contained in the thermodynamic state evolution tensor into independent control channels, wherein the temperature change rate channel is associated with the stability of the phase change process, and the coupling coefficient channel corresponds to the heat dissipation efficiency optimization requirements.

[0083] Step S202: Decouple the feature input dual-channel deep neural network model for joint inference. The temperature channel network uses historical pressure fluctuation patterns as constraints to output the predicted value of the phase transition critical point; the load channel network combines the filtered real-time load data to generate the optimal heat exchange efficiency parameters; the intermediate layer of the two networks shares the heat flux density gradient information through a feature cross attention mechanism.

[0084] Step S203: The joint inference results are fed into reinforcement learning compensation. The deviation between the temperature and pressure data measured by the current sensors and the output values ​​of the deep neural network model is used to construct a dynamic compensation matrix. The update strategy of the dynamic compensation matrix is ​​based on the historical trend component in the extracted thermodynamic state evolution tensor. The compensation amount is adaptively adjusted by time backtracking comparison. The compensated control parameters form a preliminary command vector.

[0085] Step S204: In the physical signal conversion stage, the initial command vector is mapped to the device control topology. Based on the spatial parameters provided by the server heat flux density distribution model, the abstract control quantities in the initial command vector are converted into the physical dimensions of the specific execution unit. Among them, the pump power setpoint is generated by the load channel output parameters after being corrected by the compensation matrix, and the valve opening command is determined by combining the temperature channel prediction value and the real-time phase change monitoring data. The conversion process retains the reliable data boundary formed after filtering out outliers using the 3σ principle.

[0086] In the above embodiments, the closed-loop control instruction set generated in this embodiment essentially transforms the system dynamic characteristics represented by the thermodynamic state tensor into a spatiotemporally coordinated sequence of execution commands through four stages of continuous processing: feature decoupling, dual-channel inference, deviation compensation, and physical dimension mapping. Based on the dual-channel decoupling control strategy using the heat flux density distribution characteristics, it utilizes a strengthened compensation mechanism based on the historical trend of the state evolution tensor, as well as an instruction conversion method that maintains physical constraints. Each processing stage strictly relies on the data characteristics and control parameters output from the preceding steps, forming a complete closed loop from state representation to execution instructions.

[0087] Example 5: Based on Example 4, the process of constructing the dynamic compensation matrix in step S203 of this embodiment of the invention includes the following steps:

[0088] Step S2031: Based on the predicted value of the phase transition critical point of the temperature channel and the heat exchange efficiency parameter of the load channel output by the dual-channel neural network, establish an initial deviation field; perform three-dimensional difference calculation between the sensor measured temperature and pressure data and the predicted value of the deep neural network, wherein the temperature dimension difference is associated with the decoupled temperature change rate feature channel, and the pressure dimension difference is bound to the load-heat flux density coupling coefficient channel to form an original deviation vector field with physical dimensions.

[0089] Step S2032: The original deviation vector field is input into the trend backtracking filter for processing. The generated thermodynamic state evolution tensor is used as the historical benchmark, and the time-series patterns of temperature change rate and load coupling coefficient are extracted as filtering weights. During the filtering process, the current deviation vector field is matched with the historical trend data through a sliding window correlation matching. Deviation components that conform to the system inertial characteristics are retained, and sudden interference signals are filtered out. The filtered deviation field becomes the steady-state deviation characterization.

[0090] Step S2033: Steady-state deviation characterization enters spatial mapping. Based on the spatial topological relationship provided by the heat flux density distribution model, the two-dimensional deviation data is reprojected onto the actual physical coordinates of the heat-sensitive area of ​​the server chip. During the projection process, the load channel deviation is preferentially mapped to the high heat flux density area, while the temperature channel deviation is mainly distributed in the active phase transition area, forming a deviation distribution matrix with spatial resolution.

[0091] Step S2034: A compensation matrix is ​​generated by dynamic weight fusion. Based on the deviation distribution matrix after spatial mapping, the heat flux density gradient information shared by the intermediate layer of the dual-channel neural network is superimposed as the adjustment coefficient. The compensation weight of the temperature channel is inversely proportional to the reliability of the phase transition critical point prediction value, and the compensation weight of the load channel is adaptively adjusted according to the historical fluctuation range of the heat exchange efficiency parameter. Each element value of the fused matrix represents the compensation intensity required for a specific spatial location, completing the conversion from the original deviation to the compensation parameter.

[0092] In the above embodiments, this embodiment uses a bias filtering method based on the historical trend of thermodynamic state tensor, combined with the projection mapping rule of heat flux density spatial characteristics, and a dynamic weight allocation strategy involving features of the intermediate layer of the neural network. The entire process inherits the feature decoupling results, joint inference output, and spatiotemporal alignment characteristics, forming a closed-loop compensation system with physical interpretability. The compensation matrix, as a key transformation layer connecting intelligent decision-making and execution control, retains the predictive advantages of the neural network while ensuring the physical rationality of control commands through multi-dimensional bias processing.

[0093] Example 6: Based on Example 5, the process of superimposing the shared heat flux density gradient information of the intermediate layer of the dual-channel neural network as an adjustment coefficient in step S2034 of this embodiment includes the following steps:

[0094] Step S20341: Decompose the deviation distribution matrix into its characteristic domain and separate the obtained spatial projection results according to the temperature channel deviation component and the load channel deviation component to form two orthogonal temperature channel sub-matrices and load channel sub-matrices.

[0095] Step S20342: For the temperature channel submatrix, extract the temperature sensitivity coefficient from the heat flux density gradient information shared by the intermediate layers of the dual-channel neural network; perform a nonlinear combination of the temperature sensitivity coefficient and the confidence index of the phase transition critical point prediction value to generate a temperature compensation weight field.

[0096] For the load channel submatrix, the spatial correlation strength in the heat flux density gradient information is used as the modulation factor, carrying the load dynamic characteristics in the thermodynamic state evolution tensor. Combined with the historical fluctuation data of the heat exchange efficiency parameter output by the load channel network, a load compensation weight field is generated through sliding window variance analysis.

[0097] Step S20343: The compensation matrix is ​​constructed by tensor product operation of dual-channel weight fields. The temperature compensation weight field and the temperature channel submatrix are multiplied by Hadamard, and the product result reflects the stability requirements of the phase transition process. The load compensation weight field and the load channel submatrix are multiplied by Kronecker, and the operation result represents the heat dissipation efficiency optimization constraint. The two operation results are superimposed by the spatial topology rules of physical signal conversion to form a complete dynamic compensation matrix.

[0098] In the above embodiments, this embodiment is based on the weight generation mechanism of the intermediate layer features of the neural network and the behavior of the historical system, the compensation field distribution strategy that follows the thermodynamic evolution law, and the matrix fusion method that maintains the uniformity of physical dimensions.

[0099] Example 7: Based on Example 6, the process of generating the temperature compensation weight field and the load compensation weight field in step S20342 of this embodiment of the invention includes the following steps:

[0100] Step S203421: Extract the temperature sensitivity coefficient from the heat flux density gradient information shared by the intermediate layers of the dual-channel neural network. The temperature sensitivity coefficient includes the instantaneous temperature change rate characteristics and historical phase transition behavior patterns in the thermodynamic state evolution tensor.

[0101] The temperature sensitivity coefficient is dynamically coupled with the confidence index of the output phase transition critical point prediction value. The confidence index of the phase transition critical point prediction value reflects the confidence of the neural network in the current phase transition boundary. The coupling process adopts an adaptive weighting strategy, which enhances the compensation weight in the low confidence region and converges the compensation weight in the high confidence region.

[0102] The output is a temperature sensitivity-confidence fusion field, which shows a smooth transition in the active phase transition region and maintains low gain characteristics in the stable region.

[0103] Step S203422: The trend backtracking filter established by the temperature sensitivity-confidence fusion field input is used to calculate the time decay coefficient of the compensation weight by utilizing the time series pattern of temperature change rate in the historical thermodynamic state evolution tensor.

[0104] The time decay coefficient is nonlinearly superimposed with the fusion field to ensure that the dynamic adjustment of the weight field conforms to the thermal inertia law of the system and avoids compensation oscillations caused by high-frequency disturbances;

[0105] Finally, a temperature-compensated weight field is generated, whose spatial distribution is aligned with the mapping deviation distribution region, forming a high-weight gradient band in the phase transition critical region.

[0106] In the above embodiments, this embodiment constructs a dynamic adaptive compensation mechanism for precisely adjusting the temperature and pressure control of the immersion liquid cooling system. Temperature sensitivity coefficient extraction and dynamic coupling of confidence: Integrating real-time temperature change characteristics of the thermodynamic state evolution tensor with the prediction confidence of the neural network for the phase transition critical point, a physically constrained compensation weight distribution is formed. This ensures that the compensation strategy strengthens regulation in the phase transition critical region and reduces intervention in the stable region, avoiding overcompensation. Trend backtracking filtering and time decay coefficient superposition: Historical temperature change rate patterns are used to inertially constrain the compensation weights, making the dynamic adjustment conform to the thermodynamic transient characteristics of the system, suppressing control oscillations caused by high-frequency disturbances, and improving closed-loop stability. Spatial alignment and gradient band distribution optimization: The final temperature compensation weight field is strictly matched with the deviation distribution region, forming a high-weight gradient band in the active phase transition region, ensuring that the compensation effect is accurately applied to key heat dissipation areas, improving the system's thermal management efficiency under transient load changes.

[0107] In summary, this embodiment achieves an adaptive compensation strategy by integrating real-time sensor data, neural network prediction confidence, and historical thermodynamic evolution patterns. This strategy can optimize the heat dissipation efficiency in the phase transition critical region while maintaining system stability.

[0108] Example 8: Based on Example 7, the process of nonlinearly superimposing the time decay coefficient and the fusion field in step S203422 of this embodiment of the invention includes the following steps:

[0109] Step S2034221: The time decay coefficient output from the trend backtracking filter is calculated based on the time-series pattern of the rate of temperature change in the historical thermodynamic state evolution tensor, reflecting the thermal inertia characteristics of the system at different operating stages; the time decay coefficient is dynamically adjusted with the trend of the rate of temperature change. If historical data shows that the temperature fluctuates drastically, the decay coefficient will increase accordingly; if the temperature change is stable, the decay coefficient will decrease accordingly.

[0110] Step S2034222: The temperature sensitivity-confidence fusion field already contains the weight distribution information of the active phase transition region, where the weights of the high confidence region have converged and the weights of the low confidence region have been enhanced; for each weight value in the temperature sensitivity-confidence fusion field, a nonlinear mapping is used to adjust it so that the weights exhibit a smooth transition under the influence of the time decay coefficient, rather than a nonlinear abrupt change;

[0111] Step S2034223: The time decay coefficient acts on the weight distribution of the fusion field, so that the weight adjustment process conforms to the thermal inertia law of the system. If the decay coefficient is large (the system has strong thermal inertia), the weight adjustment amplitude is reduced to avoid overcompensation; if the decay coefficient is small (the system has weak thermal inertia), the weight adjustment amplitude is appropriately increased to ensure rapid response. The finally generated dynamic compensation weight field forms a high weight gradient band in the phase transition critical region, while maintaining stability in the high-frequency disturbance region, ensuring that the thermal management process of the system is both sensitive and reliable.

[0112] In the above embodiments, the time decay coefficient is calculated from the historical temperature change trend, reflecting the thermal inertia characteristics of the system; the fusion field provides the initial weight distribution, and combined with the dynamic adjustment of the time decay coefficient, the adaptability of the compensation strategy is optimized; the final weight field not only meets the mapping requirements of the deviation distribution area, but also avoids compensation oscillations, improving the stability and control accuracy of the system.

[0113] Example 9: Based on Example 8, the process in step S2034223 of this embodiment of the invention that makes the weight adjustment process conform to the thermal inertia law of the system includes the following steps:

[0114] Step S20342231: Extract the temporal fluctuation characteristics of the temperature change rate from the historical thermodynamic evolution tensor, generate a time decay coefficient with spatiotemporal continuity, reflecting the cumulative effect of the system's thermal inertia - when the standard deviation of the temperature change rate exceeds the critical threshold, the coefficient value increases exponentially with the cube root of the fluctuation amplitude.

[0115] Step S20342232: Deeply couple the preset weight distribution and time decay coefficient in the temperature sensitivity-confidence fusion field; the coupling process is achieved through the nonlinear transformation of the hyperbolic tangent function: for each grid point in the fusion field, obtain the relative deviation value between its original weight and the mean confidence value of its neighborhood, and substitute the relative deviation value into the adaptive adjustment function containing the time decay coefficient - when the decay coefficient is large, the function output value is compressed to the square root range of the original value; when the coefficient is small, the output value is amplified to 1.5 times the reference value in an inverse relationship;

[0116] Step S20342233: The adjusted weight field eventually forms a dynamic gradient structure with thermodynamic adaptability: In the active phase transition region, the weight difference between adjacent grid points is controlled within the dynamic threshold determined by the decay coefficient, and the dynamic threshold changes negatively with the thermal inertia intensity of the system; while in the region where thermal disturbances occur frequently, the weight change rate is constrained within the product of the time decay coefficient and the local temperature sensitivity.

[0117] In the above embodiments, this embodiment ensures that the system can both track rapid changes in heat flux and suppress high-frequency noise interference. This cascaded parameter transfer mechanism makes each processing step strictly dependent on the characteristic quantities output by the preceding step, forming a closed-loop optimized thermodynamic response system.

[0118] Example 10: As Figure 5 As shown, based on Example 1, the process of setting the heat flux density gradient threshold in step S300 of this embodiment of the invention includes the following steps:

[0119] Step S301: Separate the load-heat flux density coupling coefficient from the thermodynamic state evolution tensor output by the dynamic feature extraction engine, fuse it with the temperature change rate predicted by the deep neural network model, and generate a regional heat flux dynamic index to comprehensively reflect the dynamic evolution trend of the heat load per unit area under the current operating conditions.

[0120] Step S302: The pressure-temperature deviation correction output by the reinforcement learning compensator and the regional heat flow dynamic index are input together into the nonlinear mapping module to generate an initial gradient threshold benchmark. The initial gradient threshold benchmark increases logarithmically with the increase of the pressure-temperature deviation correction and is simultaneously adjusted by the quadratic term of the heat flow dynamic index.

[0121] Step S303: The initial gradient threshold is adaptively calibrated using historical data of the flow channel switching of the actuator; when the continuous opening time of the single-phase flow channel exceeds the critical point of condenser efficiency decay, the initial gradient threshold is reduced proportionally according to the real-time efficiency coefficient of the condensation-reflux closed loop; if the heat flux density oscillation amplitude is detected to exceed the stability margin during the operation of the two-phase flow channel, the initial gradient threshold is dynamically increased according to the reciprocal relationship of the oscillation frequency.

[0122] In the above embodiments, the setting process of this embodiment ensures that the threshold is always within the balance range between the thermodynamic state evolution tensor prediction range and the actual response capability of the actuator, forming a coherent decision chain from feature extraction to control closed loop.

[0123] Example 11: As Figure 6As shown, based on Examples 1-10, the single-phase and two-phase immersion liquid cooling system based on AI intelligent decision-making provided in this embodiment of the invention includes: server 1, AI algorithm controller 2, coolant storage tank 3, condenser 4, circulation pump 5, electric valve 6, pressure relief valve 7, and temperature sensor 8.

[0124] The top of server 1 is connected to the inlet of coolant reservoir 3 via a coolant pipe. The outlet of coolant reservoir 3 is connected to the inlet of condenser 4 via a coolant pipe. The outlet of condenser 4 is connected to the inlet of circulating pump 5 via a coolant pipe. The outlet of circulating pump 5 is connected to one end of electric valve 6. The other end of electric valve 6 is connected to the bottom of server 1. A pressure relief valve 7 is installed on the right side of the top of server 1. The pressure relief valve 7 is connected to AI algorithm controller 2 via a data cable. AI algorithm controller 2 is connected to temperature sensor 8 via a data cable. Temperature sensor 8 is embedded in server 1.

[0125] In the above embodiments, the formation of the immersion liquid cooling command is a process in which the AI ​​algorithm controller 2 calculates the optimal temperature-pressure coordination strategy based on real-time operating data and a pre-trained model. The core is the progressive logic of "multi-source data fusion - feature mapping - multi-objective optimization - decision output". The system switches operating modes through the AI ​​algorithm controller 2. It maintains a single-phase immersion liquid cooling mode at low temperatures or low heat flux densities, and switches to a two-phase mode when the temperature set value is exceeded. The coolant used in the system is a mixture, both of which are coexisting fluorinated liquids. A suitable boiling point is selected (e.g., 70°C). When the boiling point is exceeded, the two phases of coolant undergo a phase change, carrying away more heat, which is then returned to the liquid cooling box through circulation cooling. The AI ​​algorithm controller 2 collects multi-source data such as temperature, heat flux density, server load, environmental parameters, and system pressure in real time. After data cleaning and feature extraction, the data is input into the pre-trained deep neural network model. The model outputs the optimal operating mode (single-phase / two-phase) and equipment control parameters (pump power, valve opening, pressure value), driving the circulating pump, intelligent valve group, and pressure regulating device to work together to achieve precise heat dissipation.

[0126] In specific applications (see attached principle reference) Figure 7 and attached Figure 8System initialization: Technicians inject fluorinated liquid into the coolant reservoir 3 to ensure that server 1 is completely submerged. A deep neural network model trained on over 100,000 complex operating data points (including scenarios with different seasons, workloads, and heat flux density distributions) is imported into the AI ​​algorithm controller 2. The model has self-learning capabilities, fine-tuning weights daily based on newly collected data from the previous day. The intelligent valve group (electric valve 6, pressure relief valve 7) defaults to single-phase flow channel opening, maintaining atmospheric pressure initially. Single-phase liquid cooling mode operation: The system defaults to single-phase mode upon startup. Circulating pump 5 drives the fluorinated liquid circulation; the AI ​​algorithm controller 2 continuously receives data from temperature sensor 8 (fluorinated liquid absorbs sensible heat in liquid form and is then cooled by the cooling tower), which flows back to the water tank containing server 1 to complete the circulation.

[0127] The AI ​​algorithm controller 2 first receives multi-dimensional real-time data from the system and processes it accordingly, providing "high-quality raw materials" for the deep neural network model's decision-making. Specifically, this includes: core monitoring data: temperature data: coolant inlet and outlet temperature difference (ΔT), server 1 chip surface temperature, water tank internal average temperature, and ambient temperature; pressure data: current system absolute pressure, circulating pump 5 outlet pressure, and two-phase flow channel pressure difference; related parameters: server 1 CPU / GPU load rate (0-100%), heat flux density, circulating pump 5 current power, and valve real-time opening.

[0128] Data preprocessing: Outlier removal: Using the 3σ principle or the isolated forest algorithm, outlier values ​​caused by temperature sensor 8 failures are removed (such as anomalies where the temperature suddenly spikes to 100°C but the load is 0); Feature engineering: Dynamic features such as temperature change rate and load fluctuation rate are extracted, and coupled features such as load rate-heat flux density ratio and temperature difference-pressure are extracted. These features can more accurately reflect the operating condition trend. A sudden increase in load leads to a rapid increase in heat flux density, and pressure adjustment needs need to be predicted in advance.

[0129] The commands are not statically output, but are adjusted in real time through closed-loop feedback: the deep neural network model receives feedback data (such as actual pressure and actual temperature) from temperature sensor 8 every 10ms, calculates the deviation from the target value, and dynamically corrects the next round of commands through reinforcement learning. If the actual pressure is too high, the opening command of the pressure reducing device is increased. The command conversion is the process of converting the "abstract control parameters" (such as "target pressure 0.08MPa") output by the AI ​​model into physical signals (such as electrical signals and mechanical signals) that the equipment can recognize. The converted signals act on the equipment through drive circuits (such as relays and power amplifiers), triggering specific actions, and real-time feedback ensures accurate command execution.

[0130] AI-driven switch to two-phase mode: When the load on server 1 suddenly increases, the AI ​​model makes a decision: After feature extraction (such as increased heat flux density gradient, load change rate exceeding the threshold), the data is input into the model to determine that the current working condition is a complex "high heat flux density-high load-high temperature environment", and outputs a switching command; the actuator responds: the AI ​​algorithm controller 2 increases the power of the circulation pump, and at the same time sends a signal to the intelligent valve group to close the single-phase main channel, open the two-phase dedicated flow channel, and start the pressure reducing device; the gaseous fluorinated liquid is liquefied by the built-in high-efficiency condenser 4 of the cooling tower and flows back to the bottom of the tank to complete the two-phase circulation.

[0131] AI-driven switchback to single-phase mode: When the load on server 1 decreases: The AI ​​model decides: it determines the operating condition as "low heat flux density - low load - normal temperature environment" and outputs a switchback command; the actuator responds: the power of circulation pump 5 decreases, the intelligent valve group closes the two-phase flow channel and opens the single-phase main channel, the pressure regulating device raises the system pressure back to normal pressure, the fluorinated liquid stops phase change at this time, and the system returns to the single-phase sensible heat dissipation cycle. Model self-learning and optimization: The system records "operating condition - mode - energy consumption - heat dissipation effect" data every 15 minutes, and uses new data to fine-tune the AI ​​model every day from 2-4 am (using an online learning algorithm). For example, during the high-temperature period in summer, the model automatically increases the weight of ambient temperature on the pressure regulation parameter; for specific business periods (such as peak financial trading periods), it optimizes the correlation characteristics between load changes and heat flux density to ensure the accuracy and efficiency of the system's long-term operation.

[0132] Figure 9 A block diagram of an exemplary electronic device suitable for implementing embodiments of the present invention is shown.

[0133] The electronic device may include a central processing unit / microprocessor / main control chip, etc. 9; and a storage medium 10, coupled to the central processing unit / microprocessor / main control chip, etc. 9, and storing computer-executable instructions therein for performing the steps of various methods of embodiments of the present invention when executed by the processor.

[0134] The central processing unit / microprocessor / main control chip, etc., can include, but are not limited to, one or more processors or microprocessors.

[0135] Storage medium 10 may include, but is not limited to, random access memory (RAM), read-only memory (ROM), flash memory, EPROM memory, EEPROM memory, registers, computer storage media (e.g., hard disk, floppy disk, solid-state drive, removable disk, CD-ROM, DVD-ROM, Blu-ray disc, etc.).

[0136] In addition, the electronic device may also include (but is not limited to) a data bus 11, an input / output bus / external bus / device bus 12, a display 13, and input / output devices 14 (e.g., keyboard, mouse, speaker, etc.).

[0137] The central processing unit / microprocessor / main control chip, etc., 9 can communicate with external devices (13, 14, etc.) via I / O bus 12 through wired or wireless network (not shown).

[0138] Storage medium 10 may also store at least one computer-executable instruction for performing the steps of various functions and / or methods in the embodiments described herein when the central processing unit / microprocessor / main control chip, etc., 9 is running.

[0139] In one embodiment, the at least one computer-executable instruction may also be compiled into or comprise a software product, wherein one or more computer-executable instructions are executed by a processor to perform the steps of the various functions and / or methods in the embodiments described herein.

[0140] Figure 10 A schematic diagram of a computer-readable storage medium according to an embodiment of the present invention is shown.

[0141] like Figure 10 As shown, instructions, such as computer-readable instructions 15, are stored on the non-transitory computer-readable storage medium 16. When the computer-readable instructions 15 are executed by a processor, the various methods described above can be performed. The non-transitory computer-readable storage medium includes, but is not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory. Non-transitory non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. For example, the non-transitory computer-readable storage medium 16 can be connected to a computing device such as a computer, and then, when the computing device executes the computer-readable instructions 15 stored on the computer-readable storage medium 16, the various methods described above can be performed.

[0142] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0143] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0144] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0145] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for executing all or part of the steps of the methods of the various embodiments of this invention through a computer device (which may be a personal computer, server, or network device, etc.). The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0146] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A single-phase and two-phase immersion liquid cooling method based on AI intelligent decision-making, characterized in that, Includes the following steps: The thermodynamic state evolution tensor, which includes the characteristics of temperature change rate and load-heat flux density coupling coefficient, is input into a deep neural network model to calculate the temperature and pressure that the current thermodynamic state evolution tensor is adapted to. The output of the deep neural network model is dynamically corrected by a reinforcement learning compensator based on the deviation between the current actual pressure and temperature and the target value. The revised instructions are converted into a set of closed-loop control instructions that the device can execute, through a physical signal converter. The closed-loop control instruction set is injected into the actuator group. The actuator executes the power reconfiguration and flow channel switching according to the instructions. The single-phase and two-phase flow channel opening and closing combination is triggered by the heat flux density gradient threshold. When switching from a single-phase flow channel to a two-phase flow channel, the condensation-reflux closed loop is activated simultaneously when the flow channel switching is completed. In the two-phase mode, the gaseous fluorinated liquid is liquefied and refluxed through the high-efficiency condenser to realize the adaptive switching of heat dissipation mode and thermal cycle reconstruction.

2. The single phase and two phase immersion liquid cooling method based on AI intelligent decision as claimed in claim 1, wherein, The process of generating a set of closed-loop control instructions executable by the device includes the following steps: Based on the output thermodynamic state evolution tensor, multimodal feature decoupling is performed; the temperature change rate feature contained in the thermodynamic state evolution tensor and the load-heat flux density coupling coefficient are separated into independent control channels; The decoupled feature input dual-channel deep neural network model performs joint inference, and the temperature channel network is constrained by historical pressure fluctuation patterns, outputting the predicted value of the phase transition critical point. The load channel network combines filtered real-time load data to generate optimal heat exchange efficiency parameters. The joint inference results are fed into reinforcement learning compensation, which uses the deviation between the temperature and pressure data measured by the current sensors and the output values ​​of the deep neural network model to construct a dynamic compensation matrix; the compensated control parameters form a preliminary command vector. The physical signal conversion stage maps the initial command vector to the device control topology. Based on the spatial parameters provided by the server heat flux density distribution model, the abstract control quantities in the initial command vector are converted into the physical dimensions of the specific execution unit.

3. The single-phase and two-phase immersion liquid cooling method based on AI intelligent decision-making as described in claim 2, characterized in that, The process of constructing a dynamic compensation matrix includes the following steps: An initial deviation field is established based on the predicted critical point of phase transition in the temperature channel and the heat exchange efficiency parameter of the load channel output by the dual-channel neural network. The measured temperature and pressure data from the sensor and the predicted values ​​from the deep neural network are then subjected to a three-dimensional difference calculation to form the original deviation vector field with physical dimensions. The original deviation vector field is input into the trend backtracking filter for processing. The generated thermodynamic state evolution tensor is used as the historical benchmark, and the time-series patterns of temperature change rate and load coupling coefficient are extracted as filter weights. The filtered deviation field becomes a representation of the steady-state deviation; Steady-state deviation characterization is incorporated into spatial mapping. Based on the spatial topological relationship provided by the heat flux density distribution model, the two-dimensional deviation data is reprojected onto the actual physical coordinates of the thermally sensitive area of ​​the server chip, forming a deviation distribution matrix with spatial resolution. The compensation matrix is ​​generated by dynamic weight fusion. Based on the deviation distribution matrix after spatial mapping, the heat flux density gradient information shared by the intermediate layer of the dual-channel neural network is superimposed as the adjustment coefficient. Each element value of the fused matrix represents the compensation intensity required at a specific spatial location, thus completing the transformation from the original deviation to the compensation parameter.

4. The single-phase and two-phase immersion liquid cooling method based on AI intelligent decision-making as described in claim 3, characterized in that, The process of superimposing the shared heat flux density gradient information of the intermediate layers of a dual-channel neural network as an adjustment coefficient includes the following steps: The deviation distribution matrix is ​​decomposed into its eigendomain, and the obtained spatial projection results are separated into temperature channel deviation components and load channel deviation components to form two orthogonal temperature channel submatrices and load channel submatrices. For the temperature channel submatrix, extract the temperature sensitivity coefficient from the heat flux density gradient information shared by the intermediate layer of the dual-channel neural network; nonlinearly combine the temperature sensitivity coefficient with the confidence index of the phase transition critical point prediction value to generate a temperature compensation weight field. For the load channel submatrix, the spatial correlation strength in the heat flux density gradient information is used as the modulation factor, carrying the load dynamic characteristics in the thermodynamic state evolution tensor. Combined with the historical fluctuation data of the heat exchange efficiency parameter output by the load channel network, a load compensation weight field is generated through sliding window variance analysis. The compensation matrix is ​​constructed through tensor product operation of dual-channel weighted fields. The operation results are superimposed using spatial topological rules of physical signal conversion to form a complete dynamic compensation matrix.

5. The single-phase and two-phase immersion liquid cooling method based on AI intelligent decision-making as described in claim 4, characterized in that, The process of generating the temperature compensation weight field and the load compensation weight field includes the following steps: Temperature sensitivity coefficients are extracted from the heat flux density gradient information shared by the intermediate layers of a dual-channel neural network. These temperature sensitivity coefficients contain the instantaneous temperature change rate characteristics and historical phase transition behavior patterns in the thermodynamic state evolution tensor. The temperature sensitivity coefficient is dynamically coupled with the confidence index of the output phase transition critical point prediction value. The confidence index of the phase transition critical point prediction value reflects the confidence of the neural network in the current phase transition boundary. The output is a temperature sensitivity-confidence fused field; The trend backtracking filter, established by the temperature sensitivity-confidence fusion field input, uses the time series pattern of temperature change rate in the historical thermodynamic state evolution tensor to calculate the time decay coefficient of the compensation weight. The time decay coefficient is nonlinearly superimposed with the fusion field to generate a temperature-compensated weight field. Its spatial distribution is aligned with the deviation distribution region of the mapping, forming a high-weight gradient band in the phase transition critical region.

6. The single-phase and two-phase immersion liquid cooling method based on AI intelligent decision-making as described in claim 5, characterized in that, The process of nonlinearly superimposing the time decay coefficient and the fusion field includes the following steps: The time decay coefficient output from the trend backtracking filter is calculated based on the time-series pattern of the rate of temperature change in the historical thermodynamic state evolution tensor. The time decay coefficient is dynamically adjusted as the rate of temperature change changes. The temperature sensitivity-confidence fusion field already contains the weight distribution information of the active phase transition region. For each weight value in the temperature sensitivity-confidence fusion field, a nonlinear mapping is used to adjust it so that the weights exhibit a smooth transition under the influence of the time decay coefficient, rather than a nonlinear abrupt change. The time decay coefficient acts on the weight distribution of the fusion field, making the weight adjustment process conform to the thermal inertia law; the resulting dynamic compensation weight field forms a high-weight gradient band in the phase transition critical region.

7. The single-phase and two-phase immersion liquid cooling method based on AI intelligent decision-making as described in claim 6, characterized in that, The process of ensuring that the weight adjustment conforms to the thermal inertia of the system includes the following steps: The temporal fluctuation characteristics of the rate of temperature change are extracted from the historical thermodynamic evolution tensor to generate a time decay coefficient with spatiotemporal continuity. The preset weight distribution and time decay coefficient are deeply coupled in the temperature sensitivity-confidence fusion field; The adjusted weight field ultimately forms a dynamic gradient structure with thermodynamic adaptability: in the active phase transition region, the weight difference between adjacent grid points is controlled within a dynamic threshold determined by the decay coefficient, and the dynamic threshold changes negatively with the thermal inertia intensity of the system; while in the region where thermal disturbances occur frequently, the weight change rate is constrained within the product of the time decay coefficient and the local temperature sensitivity.

8. The single-phase and two-phase immersion liquid cooling method based on AI intelligent decision-making as described in claim 1, characterized in that, The process of setting the heat flux density gradient threshold includes the following steps: The load-heat flux density coupling coefficient is separated from the thermodynamic state evolution tensor output by the dynamic feature extraction engine and fused with the temperature change rate predicted by the deep neural network model to generate a regional heat flux dynamic index, which comprehensively reflects the dynamic evolution trend of heat load per unit area under the current operating conditions. The pressure-temperature deviation correction output from the reinforcement learning compensator and the regional heat flow dynamic index are input together into the nonlinear mapping module to generate the initial gradient threshold benchmark. The initial gradient threshold is adaptively calibrated using historical data of the actuator's flow channel switching. When the continuous opening time of the single-phase flow channel exceeds the critical point of condenser efficiency decay, the initial gradient threshold is reduced proportionally according to the real-time efficiency coefficient of the condensation-reflux closed loop. If the heat flux density oscillation amplitude is detected to exceed the stability margin during the operation of the two-phase flow channel, the initial gradient threshold is dynamically increased according to the reciprocal relationship of the oscillation frequency.

9. The single-phase and two-phase immersion liquid cooling method based on AI intelligent decision-making as described in claim 1, characterized in that, The real-time data streams of temperature, pressure, and load are filtered out for outliers using the 3σ principle. The filtered data streams are then spatiotemporally aligned and coupled with the server's heat flux density distribution model. The coupled data is injected into the dynamic feature extraction engine, which outputs a thermodynamic state evolution tensor.

10. A single-phase and two-phase immersion liquid cooling system based on AI intelligent decision-making, used to implement the single-phase and two-phase immersion liquid cooling method based on AI intelligent decision-making as described in any one of claims 1 to 9, characterized in that, Includes: server, AI algorithm controller, coolant reservoir, condenser, circulation pump, electric valve, pressure relief valve, temperature sensor; The top of the server is connected to the inlet of the coolant reservoir via a coolant pipe. The outlet of the coolant reservoir is connected to the inlet of the condenser via a coolant pipe. The outlet of the condenser is connected to the inlet of the circulation pump via a coolant pipe. The outlet of the circulation pump is connected to one end of an electric valve, and the other end of the electric valve is connected to the bottom of the server. A pressure relief valve is installed on the right side of the top of the server. The pressure relief valve is connected to the AI ​​algorithm controller via a data cable. The AI ​​algorithm controller is connected to a temperature sensor via a data cable. The temperature sensor is embedded in the server.

Images