Micro-channel liquid cooling method for heat dissipation of integrated circuit chip

Abstract

The invention relates to the technical field of integrated circuit heat dissipation, and discloses a micro-channel liquid cooling method for heat dissipation of an integrated circuit chip. According to the method, a chip working time sequence temperature signal and a power consumption fluctuation signal under a corresponding timestamp are captured in real time, and a chip operation heat source profile is formed after alignment and integration. The method comprises the following steps: performing space grid decomposition on a heat source profile, generating a temperature field grid and a power consumption mapping grid covering a three-dimensional structure of a chip, and identifying thermal coupling characteristics such as an instantaneous heat flow direction, a heat density concentration ratio and a heat propagation trajectory through cross correlation analysis; and inputting the characteristics into a dynamic response model trained based on historical heat dissipation data, and calculating to obtain a heat management control instruction containing the expected flow rate of the cooling liquid, the pressure distribution in the flow channel and the liquid flow switching opportunity. According to the instruction, cooling liquid circulation in a micro-channel structure integrated on the surface of the chip is regulated and controlled in real time. According to the invention, accurate sensing and self-adaptive heat dissipation control of the dynamic thermal behavior of the chip are realized.

Description

Technical Field

[0001] This invention relates to the field of integrated circuit heat dissipation technology, specifically to a microchannel liquid cooling method for heat dissipation of integrated circuit chips. Background Technology

[0002] Traditional microchannel liquid cooling technology for integrated circuits relies heavily on the chip's steady-state temperature monitoring points and preset constant or typical power consumption models. This approach struggles to accurately reflect the dynamic characteristics of the chip's power consumption and temperature, which change rapidly and drastically over time during actual operation. When the cooling system makes decisions based on such static or lagging information, a mismatch arises between its response and the chip's actual transient thermal load, limiting its cooling efficiency.

[0003] Existing liquid cooling control strategies typically set the coolant flow rate based on measured temperatures at a limited number of points on the chip surface or the overall average temperature. This method essentially treats the chip as one or a few homogeneous thermal zones, failing to address the non-uniform distribution of heat sources and dynamic thermal coupling relationships within the chip's three-dimensional space. Because it cannot identify instantaneous hotspot locations, heat flow paths, and heat density concentration in real time, the allocation of cooling resources is not spatially precise and lacks temporal foresight, making it difficult to achieve optimal coordination with the chip's actual and dynamic thermal state.

[0004] The core problem this invention aims to solve is how to achieve real-time and accurate mapping of dynamic heat sources in a chip, and based on this, perform fine-grained heat dissipation control in three-dimensional space. This requires the technical solution to simultaneously capture and integrate time-series temperature and power consumption fluctuation signals to construct a heat source profile that truly reflects the chip's operating state; and on this basis, to perform three-dimensional mesh deconstruction and thermo-mechanical coupling feature identification on the chip, ultimately driving the coolant flow within the microchannels to achieve adaptive and fine-grained dynamic control. Summary of the Invention

[0005] The purpose of this invention is to provide a microchannel liquid cooling method for heat dissipation of integrated circuit chips, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a microchannel liquid cooling method for heat dissipation of integrated circuit chips, the method comprising: Real-time capture of the timing temperature signal and the power consumption fluctuation signal at the corresponding timestamp generated during chip operation; alignment and integration of the timing temperature signal and the power consumption fluctuation signal to form a chip operation heat source profile. The spatial mesh is decomposed on the heat source profile of the chip to generate a temperature field mesh and a power consumption mapping mesh covering the three-dimensional geometry of the chip. By performing cross-correlation analysis between the temperature field grid and the power consumption mapping grid, the thermal coupling characteristics inside the chip are identified. The thermal coupling characteristics include instantaneous heat flow direction, heat density concentration, and heat propagation trajectory. The thermal coupling characteristics are input into a dynamic response model trained based on historical heat dissipation data to calculate thermal management control instructions for the current state of the chip. The thermal management control instructions include at least the expected flow rate of the coolant, the pressure distribution in the flow channel, and the timing of the flow switching. Based on the thermal management control instructions, real-time regulation of the coolant circulation within the microchannel structure integrated on the chip surface is performed.

[0007] Preferably, spatial mesh decomposition is performed on the heat source profile of the chip to generate a temperature field mesh and a power consumption mapping mesh covering the three-dimensional geometry of the chip, specifically including: Based on the three-dimensional structural data of the chip package, a reference computing grid is constructed within the physical boundary of the chip; temperature signal segments corresponding to each grid cell are extracted from the heat source profile of the chip operation. For each extracted temperature signal segment, time-domain features are extracted, and the thermal inertia index and temperature oscillation frequency of each grid cell are calculated. The thermal inertia index represents the lag of temperature change in the grid cell, and the temperature oscillation frequency represents the number of temperature fluctuations per unit time. The power consumption fluctuation signal component corresponding to the timestamp of each grid cell is synchronously extracted from the heat source profile of the chip operation. Based on the power consumption fluctuation signal components and the material properties of the grid cells, the instantaneous thermal power contribution value of each grid cell is calculated. The instantaneous thermal power contribution value reflects the contribution intensity of the grid cell to the overall heat generation at a specific moment. The thermal inertia index, temperature oscillation frequency, and instantaneous thermal power contribution value of each grid cell are respectively assigned to the grid cell to form a temperature attribute grid and a power consumption attribute grid. The temperature attribute grid and the power consumption attribute grid together constitute the temperature field grid and the power consumption mapping grid.

[0008] Preferably, the temperature field grid and the power consumption mapping grid are cross-correlation analyzed to identify the thermal coupling characteristics inside the chip, including: Traverse all grid cells and compare the temperature and power consumption attributes of each grid cell for synchronization. Identify grid cells where the temperature rise and power consumption increase are synchronized in time and mark them as active heat source cells. For the active heat source cells, analyze the temperature attribute differences between them and their adjacent grid cells and calculate the heat flow diffusion vector. The heat flow diffusion vector includes directional information from the high temperature cell to the low temperature cell and heat conduction intensity information. The heat flow diffusion vector under continuous time series is tracked and analyzed to identify the heat flow path that exists continuously in multiple time intervals, and the heat flow path is defined as the heat propagation trajectory; the number of active heat source units passed through each heat propagation trajectory is counted, and the heat density concentration of the heat propagation trajectory is calculated by combining the instantaneous heat power contribution value of the active heat source unit. The set of all heat flux diffusion vectors at each time point is extracted, and the peak value of the vector density in a specific direction is determined by statistical analysis. The direction and intensity corresponding to the peak value of the vector density are defined as the instantaneous heat flux direction.

[0009] Preferably, the thermal coupling characteristics are input into a dynamic response model trained based on historical heat dissipation data to calculate thermal management control instructions for the current state of the chip, including: Separate the geometric coordinate data of the heat propagation trajectory and the numerical sequence of heat density concentration from the aforementioned thermal coupling characteristics; The geometric coordinate data is overlaid and mapped with the topology diagram of the microchannel network to find the microchannel segment closest to the heat propagation trajectory in space, which is marked as the critical cooling channel. The temporal variation data of the instantaneous heat flow direction is extracted from the thermo-coupling characteristics and input into the velocity prediction module of the dynamic response model. The velocity prediction module outputs the expected velocity curve of the coolant in the critical cooling channel that meets the heat dissipation requirements of the instantaneous heat flow direction. The distribution curve of heat density concentration on the corresponding path of the critical cooling channel is extracted from the thermo-coupling characteristics and input into the pressure control module of the dynamic response model. The pressure control module outputs the pressure distribution function along the critical cooling channel required to match the distribution curve. Combining the temporal variation data of the instantaneous heat flow direction and the stability judgment result of the heat propagation trajectory, the decision module of the dynamic response model calculates the liquid flow switching timing of the coolant switching between different microchannel branches.

[0010] Preferably, according to the thermal management control command, real-time regulation of coolant circulation within the microchannel structure integrated on the chip surface is performed, including: The expected flow rate curve in the thermal management control command is obtained and converted into a pulse width modulation signal for the micro-circulation pump. The output flow rate of the micro-circulation pump is controlled by adjusting the duty cycle of the pulse width modulation signal so that the flow rate of the coolant flowing through the key cooling channel matches the expected flow rate curve. The system receives the pressure distribution function from the thermal management control command and establishes a pressure gradient field that matches the pressure distribution function in the flow channel by adjusting the opening of multiple micro pressure regulating valves set at the inlet and branch nodes of the microchannel network. According to the liquid flow switching timing in the thermal management control command, an on / off control signal is sent to the micro solenoid valve group located at the bifurcation point of the microchannel network. The on / off control signal controls the opening and closing state of the micro solenoid valve group, thereby guiding the main flow of coolant to the branch network containing the key cooling channel at a specified time.

[0011] Preferably, the process of adjusting the opening of multiple micro pressure regulating valves located at the inlet and branch nodes of the microchannel network further includes: Real-time monitoring of readings from multiple pressure sensors along key cooling channels to obtain the actual pressure distribution within the flow channels; The actual pressure distribution is compared with the pressure distribution function point by point to calculate the pressure deviation value of each monitoring point; Based on the pressure deviation value, an incremental proportional control algorithm is used to generate an incremental adjustment command for the opening of the micro pressure regulating valve at the corresponding position. The incremental adjustment command is used to reduce the pressure deviation value, so that the actual pressure distribution converges to the target pressure distribution function.

[0012] Preferably, before inputting the thermal coupling features into the dynamic response model trained based on historical heat dissipation data, the method further includes a real-time adaptation and update process for the dynamic response model, specifically: During the liquid cooling control process, the actual temperature change curves of multiple monitoring points on the chip surface are continuously collected; Extract the temperature decay rate and final steady-state temperature after the implementation of the control measures from the actual temperature change curve; The collected temperature decay rate and final steady-state temperature are compared with the expected temperature decay rate and expected steady-state temperature predicted by the dynamic response model under the same input features to generate a model prediction error set. Based on the model prediction error set, the weight parameters inside the dynamic response model are adjusted using the backpropagation algorithm, so that the model's response output to subsequent thermal coupling features is closer to the measured heat dissipation effect, thus completing the online adaptation and update of the model.

[0013] Preferably, the specific steps for adjusting the internal weight parameters of the dynamic response model using the backpropagation algorithm based on the model prediction error set include: From the set of model prediction errors, the prediction errors concerning the expected flow rate, the prediction errors concerning the pressure distribution, and the prediction errors concerning the effect of fluid switching timing are separated. For the prediction error regarding the expected flow velocity, calculate its gradient with respect to the weight parameters of the flow velocity prediction module in the dynamic response model, and fine-tune the weight parameters in the opposite direction of the gradient. To address the prediction error regarding pressure distribution, the error is correlated with the activation values ​​of the output layer nodes of the pressure control module. The chain rule is then used to calculate the impact of the error on the weights of each layer of the pressure control module, and the weights are corrected accordingly. To address the prediction error regarding the effect of fluid switching timing, the correlation between the error and the timing logic judgment parameters in the decision module is analyzed, and the threshold of the timing logic judgment parameters is adjusted to optimize the judgment of switching timing.

[0014] Preferably, during the real-time regulation of coolant circulation, adaptive adjustments to the coolant's own state are also performed, including: Conductivity and viscosity sensors are installed in the coolant circulation loop to monitor the online conductivity and real-time viscosity of the coolant. The monitored conductivity value is compared with a preset pure coolant conductivity benchmark value. If it exceeds a preset threshold, it is determined that the coolant ion concentration has increased, and a coolant replacement prompt command is generated. The monitored real-time viscosity value is compared with the standard viscosity value at the current coolant temperature. If it deviates beyond the allowable range, it is determined that the coolant physical properties have deteriorated, and a coolant performance alarm command is generated. Combining the coolant replacement prompt command and the coolant performance alarm command, the opening strategy of the coolant replenishment valve is automatically adjusted or a maintenance warning is triggered.

[0015] Preferably, the step of calculating the instantaneous thermal power contribution value of each grid cell based on the power consumption fluctuation signal components and the material properties of the grid cells specifically includes: Extract the current intensity sequence and voltage fluctuation sequence within a specified time window from the power consumption fluctuation signal component; multiply the current intensity sequence and voltage fluctuation sequence point by point to obtain the instantaneous power value at each sampling moment; Obtain the thermal resistivity and specific heat capacity parameters of the chip material corresponding to each grid cell; Multiplying the instantaneous electrical power value by the thermal resistance coefficient yields the theoretical temperature rise contribution of the grid cell at the heating level. The theoretical temperature rise contribution is weighted and fused with the specific heat capacity parameter to calculate the heat value emitted by the grid cell to the surrounding environment per unit time. The heat value is normalized to a preset baseline power consumption level to obtain a dimensionless instantaneous thermal power contribution value, which is then assigned to the grid cell.

[0016] Compared with the prior art, the beneficial effects of the present invention are: By capturing and precisely aligning time-series temperature signals with power consumption fluctuation signals at corresponding timestamps in real time, a dynamic profile of the chip's operating heat source is formed. This technology directly constructs a heat source model based on the chip's actual operating timing data, overcoming the limitations of traditional methods that rely on preset or steady-state models. The thermal control system can thus acquire thermal load information that is strictly synchronized with the chip's instantaneous operating state, shifting its decision-making basis from approximate estimation to real feedback, thereby improving the timeliness and fidelity of the system's perception of changes in thermal state.

[0017] Spatial mesh decomposition is performed on the integrated heat source profile to generate a temperature field and power consumption mapping mesh covering the chip's three-dimensional geometry. Specific instantaneous thermo-mechanical coupling characteristics are identified through cross-correlation analysis. A dynamic response model trained based on historical heat dissipation data receives these characteristic parameters and directly calculates control commands including expected flow rate, channel pressure distribution, and fluid flow switching timing. This achieves a closed loop from three-dimensional transient thermal feature identification to control strategy generation. The coolant flow distribution and flow direction switching can adaptively adjust according to the finely and dynamically changing thermal field inside the chip, suppressing local overheating while optimizing the overall system flow resistance and pumping energy consumption. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the working principle of the microchannel liquid cooling method for heat dissipation of integrated circuit chips according to the present invention. Figure 2 A flowchart for generating the temperature field mesh and power consumption mapping mesh; Figure 3 A flowchart for identifying the thermal coupling characteristics inside a chip; Figure 4 The thermal distribution of the temperature decay rate error at the monitoring point; Figure 5 Heatmap of coolant condition monitoring trigger count. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0020] Please see Figure 1This invention provides a microchannel liquid cooling method for heat dissipation of integrated circuit chips. The method includes: real-time capture of time-series temperature signals and power consumption fluctuation signals generated during chip operation; alignment and integration of the time-series temperature signals and power consumption fluctuation signals to form a chip operating heat source profile; spatial mesh decomposition of the chip operating heat source profile to generate a temperature field mesh and a power consumption mapping mesh covering the three-dimensional geometry of the chip; cross-correlation analysis of the temperature field mesh and the power consumption mapping mesh to identify the internal thermal coupling characteristics of the chip, including instantaneous heat flow direction, heat density concentration, and heat propagation trajectory; inputting the thermal coupling characteristics into a dynamic response model trained based on historical heat dissipation data to obtain a thermal management control command for the current state of the chip, which includes at least the expected flow rate of the coolant, the pressure distribution in the flow channel, and the timing of liquid flow switching; and real-time regulation of the coolant circulation within the microchannel structure integrated on the chip surface according to the thermal management control command.

[0021] In one embodiment of the present invention, see [reference] Figure 2 Based on the three-dimensional structural data of the chip package, a benchmark computational grid is constructed within the physical boundary of the chip. Temperature signal segments corresponding to each grid cell are extracted from the chip's operating heat source profile. Time-domain feature extraction is performed on each extracted temperature signal segment to calculate the thermal inertia index and temperature oscillation frequency of each grid cell. The thermal inertia index represents the lag of temperature change in the grid cell, and the temperature oscillation frequency represents the number of temperature fluctuations per unit time. The power consumption fluctuation signal component corresponding to the timestamp of each grid cell is synchronously extracted from the chip's operating heat source profile. Based on the power consumption fluctuation signal component and the material properties of the grid cell, the instantaneous thermal power contribution value of each grid cell is calculated. The instantaneous thermal power contribution value reflects the contribution intensity of the grid cell to the overall heat generation at a specific moment. The thermal inertia index, temperature oscillation frequency, and instantaneous thermal power contribution value of each grid cell are assigned to the grid cell to form a temperature attribute grid and a power consumption attribute grid. The temperature attribute grid and the power consumption attribute grid together constitute the temperature field grid and the power consumption mapping grid. When calculating the instantaneous thermal power contribution, the current intensity sequence and voltage fluctuation sequence within a specified time window are extracted from the power consumption fluctuation signal component. The current intensity sequence and voltage fluctuation sequence are multiplied point by point to obtain the instantaneous power value at each sampling moment. The thermal resistance coefficient and specific heat capacity parameter of the chip material corresponding to each grid cell are obtained. The instantaneous power value is multiplied by the thermal resistance coefficient to obtain the theoretical temperature rise contribution of the grid cell at the heating level. The theoretical temperature rise contribution and the specific heat capacity parameter are weighted and fused to calculate the heat value emitted by the grid cell to the surrounding environment per unit time. The heat value is normalized to a preset benchmark power consumption level to obtain a dimensionless instantaneous thermal power contribution value and assigned to the grid cell.

[0022] In its implementation, the benchmark computing grid is constructed within the physical boundary of the chip based on the three-dimensional structural data of the chip package. The cell division of the benchmark computing grid is adaptively encrypted according to the physical layout and thermal coupling strength of the internal functional modules of the chip. Temperature signal segments corresponding to each grid cell are extracted from the chip's operating heat source profile. The temperature signal segments are continuous time-series temperature data associated with the spatial coordinates of the center point of the grid cell. The thermal inertia index and temperature oscillation frequency of each grid cell are calculated by performing time-domain feature extraction on each extracted temperature signal segment. The thermal inertia index is quantified by fitting the time constant of the exponential decay process of the temperature signal segment after the power consumption pulse ends. The temperature oscillation frequency is determined by identifying the fundamental frequency component of the temperature signal segment on the spectrum.

[0023] In some embodiments, power consumption fluctuation signal components corresponding to the timestamps of each grid cell are synchronously extracted from the chip's operating heat source profile. The timestamps of the power consumption fluctuation signal components and the timestamps of the temperature signal segments are strictly aligned using the system's global clock. The instantaneous thermal power contribution value of each grid cell is calculated based on the power consumption fluctuation signal components and the material properties of the grid cells. The calculation process first extracts the current intensity sequence and voltage fluctuation sequence within a specified time window from the power consumption fluctuation signal components. The current intensity sequence and voltage fluctuation sequence are multiplied point by point to obtain the instantaneous power value at each sampling moment. The thermal resistance coefficient and specific heat capacity parameters of the chip material corresponding to each grid cell are obtained. The thermal resistance coefficient and specific heat capacity parameters are derived from the chip material property parameter database.

[0024] Optionally, the instantaneous electrical power value is multiplied by the thermal resistivity to obtain the theoretical temperature rise contribution of the grid cell at the heating level. The theoretical temperature rise contribution characterizes the temperature rise of the grid cell due to electrical power consumption under adiabatic conditions. The theoretical temperature rise contribution is then weighted and fused with the specific heat capacity parameter to calculate the heat dissipated by the grid cell to the surrounding environment per unit time. The weighted fusion process is achieved through a dimensionally harmonious linear formula, which is in the form of: in: It is the calculated heat dissipation value of the grid cell per unit time. and It is a weighting coefficient reflecting the proportion of contributions from the Joule heating effect and the heat capacity effect. It is the instantaneous power value. It is a parameter of specific heat capacity. It is the density of the material block represented by the grid cell. It is the volume of the mesh cell. It is the theoretical contribution to temperature rise. It is a characteristic time constant used to convert temperature rise into a rate of change characteristic.

[0025] It can be understood that the thermal inertia index, temperature oscillation frequency, and instantaneous thermal power contribution value of each grid cell are respectively assigned to the grid cell to form a temperature attribute grid and a power consumption attribute grid. Each data point in the temperature attribute grid contains a pair of values, namely the thermal inertia index and the temperature oscillation frequency. Each data point in the power consumption attribute grid stores a scalar, namely the instantaneous thermal power contribution value. The temperature attribute grid and the power consumption attribute grid together constitute the temperature field grid and the power consumption mapping grid. The temperature field grid and the power consumption mapping grid serve as a spatial discretization model that characterizes the transient thermal behavior and power consumption distribution of the chip.

[0026] In one embodiment of the present invention, see [reference] Figure 3 The process iterates through all grid cells, comparing the temperature and power consumption attributes of each cell for synchronization. Grid cells whose temperature rise and power consumption increase are synchronized in time are identified and marked as active heat source cells. For active heat source cells, the temperature attribute differences between them and their adjacent grid cells are analyzed to calculate the heat flow diffusion vector. The heat flow diffusion vector contains directional information from high-temperature cells to low-temperature cells and information on heat conduction intensity. The heat flow diffusion vectors in continuous time series are tracked and analyzed to identify heat flow paths that exist continuously over multiple time intervals and are defined as heat propagation trajectories. The number of active heat source cells traversed on each heat propagation trajectory is counted, and the heat density concentration of the heat propagation trajectory is calculated by combining the instantaneous heat power contribution value of the active heat source cells. The set of all heat flow diffusion vectors at each time point is extracted, and the peak vector density in a specific direction is determined through statistical analysis. The direction and intensity corresponding to the peak vector density are defined as the instantaneous heat flow direction. Geometric coordinate data of the heat propagation trajectory and numerical sequence of heat density concentration are separated from the thermo-coupling characteristics. The geometric coordinate data is overlaid and mapped with the topology diagram of the microchannel network to find the microchannel segment that is spatially closest to the heat propagation trajectory and marked as the critical cooling channel. The temporal variation data of instantaneous heat flow direction is extracted from the thermo-coupling characteristics and input into the velocity prediction module of the dynamic response model. The velocity prediction module outputs the expected velocity curve of the coolant in the critical cooling channel that meets the heat dissipation requirements of the instantaneous heat flow direction. The distribution curve of heat density concentration on the corresponding path of the critical cooling channel is extracted from the thermo-coupling characteristics and input into the pressure control module of the dynamic response model. The pressure control module outputs the pressure distribution function along the critical cooling channel required to match the distribution curve. Combining the temporal variation data of instantaneous heat flow direction and the stability judgment result of the heat propagation trajectory, the decision module of the dynamic response model calculates the liquid flow switching timing of the coolant switching between different microchannel branches.

[0027] In its implementation, the process iterates through all grid cells, comparing the temperature and power consumption attributes of each cell for synchronization. This synchronization is achieved by calculating the cross-correlation coefficient between the temperature and power consumption time series of each grid cell. Grid cells whose temperature rise and power consumption increase are synchronized in time are identified as active heat source cells. Grid cells with a cross-correlation coefficient exceeding a preset synchronization threshold are also designated as active heat source cells. For active heat source cells, the process analyzes the temperature attribute differences between them and their adjacent grid cells to calculate the heat flux diffusion vector. The calculation is based on the temperature gradient and spatial distance between each active heat source cell and its directly adjacent grid cells.

[0028] In some embodiments, the heat flux diffusion vector includes directional information from the high-temperature cell to the low-temperature cell and thermal conductivity information, where the thermal conductivity information is determined by the product of the temperature difference and the thermal conductivity of the material between the mesh cells. Tracking and analyzing the heat flux diffusion vector over a continuous time series identifies heat flux paths that persist across multiple time intervals. The criterion for persistence is that the direction of the heat flux diffusion vector remains consistent for more than a preset proportion within consecutive time frames. The heat flux path is defined as a heat propagation trajectory, and each heat propagation trajectory is represented by a series of mesh cell coordinate sequences arranged in chronological order.

[0029] Optionally, the heat density concentration of the heat propagation trajectory can be calculated by counting the number of active heat source units along each heat propagation trajectory and combining this with the instantaneous heat power contribution value of the active heat source units. in: Indicates the concentration of heat density. Indicates the number of active heat source units on the trajectory. This represents the instantaneous thermal power contribution value of the u-th active heat source unit. This represents the total number of grid cells traversed by the trajectory. This is the trajectory length coefficient used for normalization. The set of all heat flux diffusion vectors at each time point is extracted, and the peak vector density in a specific direction is determined through statistical analysis. The statistical analysis uses a direction histogram method to statistically analyze the direction angles of all heat flux diffusion vectors across the entire chip. The direction and intensity corresponding to the peak vector density are defined as the instantaneous heat flux direction, which is a two-dimensional data point containing angle and amplitude values.

[0030] In practical implementation, the geometric coordinate data of the heat propagation trajectory and the numerical sequence of heat density concentration are separated from the thermal coupling characteristics. The geometric coordinate data is an ordered set of coordinate points in three-dimensional space. The geometric coordinate data is overlaid and mapped with the topology diagram of the microchannel network to find the microchannel segment that is spatially closest to the heat propagation trajectory. Proximity is evaluated by calculating the average of the shortest distances from each point on the trajectory to the centerlines of all microchannel segments.

[0031] It can be understood that the temporal variation data of instantaneous heat flow direction extracted from the thermo-coupling characteristics is input into the velocity prediction module of the dynamic response model. The temporal variation data is a triplet sequence containing timestamps, orientation angles, and intensity values. The velocity prediction module outputs the expected velocity curve of the coolant in the critical cooling channel that meets the heat dissipation requirements of the instantaneous heat flow direction. The expected velocity curve is a function of the velocity value changing over time. The distribution curve of heat density concentration along the corresponding path of the critical cooling channel extracted from the thermo-coupling characteristics is input into the pressure control module of the dynamic response model. The distribution curve describes the heat density concentration value changing along the axial position of the critical cooling channel. The pressure control module outputs the pressure distribution function along the critical cooling channel required to match the distribution curve. The pressure distribution function defines the relationship of pressure value change along the channel length. Combining the temporal variation data of instantaneous heat flow direction with the stability judgment result of the heat propagation trajectory, the decision module of the dynamic response model calculates the liquid flow switching timing of the coolant switching between different microchannel branches. The stability judgment is completed by analyzing the rate of change of the geometric shape of the heat propagation trajectory within the most recent time window. The decision module of the dynamic response model calculates the timing for coolant flow switching between different microchannel branches by receiving temporal changes in instantaneous heat flow direction and stability assessment results of the heat propagation trajectory. The stability assessment process is based on the rate of change of the geometric shape of the heat propagation trajectory within the most recent time window. It evaluates the persistence and predictability of the heat propagation path by analyzing the spatiotemporal evolution characteristics of the trajectory point coordinate sequence, such as the rate of change of trajectory length, curvature, or orientation angle. The decision module integrates these input features and utilizes pre-trained parameter weights to output a switching time point dynamically matched to the heat source, ensuring that coolant flow direction adjustments are synchronized with changes in chip thermal load.

[0032] In one embodiment of the present invention, the expected flow rate curve in the thermal management control command is obtained and converted into a pulse width modulation signal for the micro-circulation pump. The output flow rate of the micro-circulation pump is controlled by adjusting the duty cycle of the pulse width modulation signal so that the flow rate of the coolant flowing through the key cooling channel matches the expected flow rate curve. The pressure distribution function in the thermal management control command is received, and the opening degree of multiple micro pressure regulating valves set at the inlet and branch nodes of the microchannel network is adjusted to establish a pressure gradient field in the flow channel that matches the pressure distribution function. According to the liquid flow switching timing in the thermal management control command, an on / off control signal is sent to the micro solenoid valve group set at the bifurcation point of the microchannel network. The opening and closing state of the micro solenoid valve group is controlled by the control signal to guide the main flow of coolant to the branch network containing the key cooling channel at a specified time point. The actual pressure distribution within the flow channel is obtained by real-time monitoring of the readings of multiple pressure sensors along the critical cooling channel during the process of adjusting the opening of multiple micro pressure regulating valves set at the inlet and branch nodes of the microchannel network. The actual pressure distribution is compared with the pressure distribution function point by point to calculate the pressure deviation value of each monitoring point. Based on the pressure deviation value, an incremental proportional control algorithm is used to generate incremental adjustment commands for the opening of the micro pressure regulating valves at the corresponding positions. The incremental adjustment commands are used to reduce the pressure deviation value so that the actual pressure distribution converges to the target pressure distribution function.

[0033] In practice, the system executes the expected flow rate curve from the thermal management control command and converts it into a pulse width modulation (PWM) signal for the micro-circulation pump. This conversion is based on a micro-circulation pump flow-duty cycle calibration lookup table stored in the controller. By adjusting the duty cycle of the PWM signal, the output flow rate of the micro-circulation pump is controlled to match the coolant flow rate through the critical cooling channel with the expected flow rate curve. An ultrasonic Doppler flow meter installed at the inlet of the critical cooling channel provides real-time flow rate feedback data.

[0034] In some embodiments, the pressure distribution function received in the thermal management control command establishes a pressure gradient field within the flow channel that conforms to the pressure distribution function by adjusting the opening of multiple micro pressure regulating valves located at the inlet and branch nodes of the microchannel network. The pressure distribution function defines the mapping relationship between the axial coordinates along the centerline of the critical cooling channel and the target static pressure value. Based on the fluid flow switching timing in the thermal management control command, on / off control signals are sent to a group of micro solenoid valves located at the bifurcation points of the microchannel network. The on / off control signals include the address code of the target solenoid valve, the action command, and the execution time window. The opening and closing states of the micro solenoid valve group are controlled by the control signals to guide the main coolant flow to the branch network containing the critical cooling channel at a specified time point. The action response time of the micro solenoid valve group is less than ten milliseconds.

[0035] In practice, the actual pressure distribution within the flow channel is obtained by real-time monitoring of multiple pressure sensors along the critical cooling channel during the adjustment of the opening of several miniature pressure regulating valves located at the inlet and branch nodes of the microchannel network. The pressure sensors operate at a sampling frequency of 100 Hz. The pressure deviation value at each monitoring point is calculated by comparing the actual pressure distribution with the pressure distribution function point by point. The pressure deviation value is the arithmetic difference between the target pressure value and the sensor-measured pressure value under the same axial coordinate.

[0036] Optionally, based on the pressure deviation value, an incremental proportional control algorithm is used to generate incremental adjustment commands for the opening of the micro pressure regulating valve at the corresponding position. The calculation of the incremental adjustment commands follows the formula: in: This represents the increment of the micro pressure regulating valve opening calculated in the nth control cycle. It is the differential gain coefficient in the control algorithm. This represents the pressure deviation value calculated in the nth control cycle. This represents the pressure deviation value during the (n-1)th control cycle. It is the proportional gain coefficient in the control algorithm. The incremental adjustment command is used to reduce the pressure deviation value so that the actual pressure distribution converges to the target pressure distribution function. In each control cycle, the output of the digital potentiometer is updated according to the calculated valve opening increment to drive the stepper motor of the miniature pressure regulating valve.

[0037] It can be understood that the entire real-time control process is executed cyclically with a fixed control cycle. The drive signal of the micro-circulation pump, the opening commands of each micro pressure regulating valve, and the on / off commands of the micro solenoid valve group are synchronously refreshed at the beginning of each control cycle according to the latest thermal management control commands. The states of the micro-circulation pump, micro pressure regulating valve, and micro solenoid valve group constitute a coordinated operation set of the actuators, which together realize the real-time control of the coolant circulation within the microchannel structure integrated on the chip surface.

[0038] In one embodiment of the present invention, during the liquid cooling control process, the actual temperature change curves of multiple monitoring points on the chip surface are continuously collected. The temperature decay rate and the final steady-state temperature after the control measures are implemented are extracted from the actual temperature change curves. The collected temperature decay rate and the final steady-state temperature are compared with the expected temperature decay rate and the expected steady-state temperature predicted by the dynamic response model under the same input characteristics to generate a model prediction error set. Based on the model prediction error set, the backpropagation algorithm is used to adjust the internal weight parameters of the dynamic response model so that the model's response output to subsequent thermal coupling characteristics is closer to the measured heat dissipation effect, thus completing the online adaptation and update of the model. When adjusting the model weight parameters, prediction errors related to expected flow velocity, pressure distribution, and the timing of fluid switching are separated from the model prediction error set. For the prediction error related to expected flow velocity, the gradient of its effect on the weight parameters of the flow velocity prediction module in the dynamic response model is calculated, and the weight parameters are fine-tuned according to the reverse direction of the gradient. For the prediction error related to pressure distribution, it is correlated with the activation values ​​of the output layer nodes of the pressure control module. The chain rule is used to calculate the impact of the error on the weights of each layer of the pressure control module, and the weights are corrected accordingly. For the prediction error related to the timing of fluid switching, its correlation with the timing logic judgment parameters in the decision module is analyzed, and the threshold of the timing logic judgment parameters is adjusted to optimize the judgment of the switching timing.

[0039] In practice, during the liquid cooling control process, the actual temperature change curves of multiple monitoring points on the chip surface are continuously collected. These monitoring points are evenly distributed across the chip surface and their positions are fixed. The temperature decay rate and the final steady-state temperature after the control measures are implemented are extracted from the actual temperature change curves. The temperature decay rate is obtained by linearly fitting the temperature data within a specific time window after the control takes effect, yielding the slope. The collected temperature decay rate and final steady-state temperature are compared with the expected temperature decay rate and expected steady-state temperature predicted by the dynamic response model under the same input characteristics to generate a model prediction error set. The model prediction error set records the difference between the predicted value and the measured value for each monitoring point in each control cycle.

[0040] In some embodiments, the structure of the model prediction error set is organized and managed in tabular form. Referring to Table 1, detailed data comparing each model inference with the actual heat dissipation effect are recorded.

[0041] Table 1: Record of Prediction Errors in Dynamic Response Model Record Identifier Timestamp Monitoring point ID Expected temperature decay rate (°C / s) Measured temperature decay rate (°C / s) Expected steady-state temperature (°C) Measured steady-state temperature (°C) Error type label ERR_001 T1 S01 -15.2 -14.1 68.5 69.8 Flow rate and steady state ERR_002 T1 S02 -12.7 -13.0 70.1 69.5 steady state ERR_003 T2 S01 -18.5 -16.0 65.0 67.2 Flow rate and steady state Based on the model prediction error set, the backpropagation algorithm is used to adjust the internal weight parameters of the dynamic response model, making the model's response output to subsequent thermal coupling characteristics closer to the measured heat dissipation effect, thus completing the online adaptation update of the model. The online adaptation update is triggered during system idle time or in a background thread to avoid interfering with real-time control tasks.

[0042] In practice, when adjusting the model weight parameters, prediction errors related to expected flow velocity, pressure distribution, and the timing of fluid flow switching are separated from the model prediction error set. This separation process is based on the error type label and the flow channel region where the monitoring point is located. For the prediction error related to expected flow velocity, the gradient of its effect on the weight parameters of the flow velocity prediction module in the dynamic response model is calculated, and the weight parameters are fine-tuned according to the reverse direction of the gradient. The gradient calculation is based on a chain rule expansion of the partial derivatives of the error with respect to the final output layer nodes of the flow velocity prediction module.

[0043] Optionally, the prediction error regarding pressure distribution is correlated with the activation values ​​of the output layer nodes of the pressure control module. The chain rule is then used to calculate the impact of the error on the weights of each layer of the pressure control module, and the weights are adjusted accordingly. in: This represents the correction amount of the weight matrix of the l-th layer of the pressure control module. It's the learning rate. It is a loss function relating to pressure distribution. This is the gradient of the loss function with respect to the weights of the l-th layer. To analyze the prediction error regarding the effect of fluid switching timing, its correlation with the timing logic judgment parameters in the decision module is investigated. The thresholds of the timing logic judgment parameters are adjusted to optimize the judgment of switching timing. The correlation analysis is performed by statistically analyzing the relationship between the temperature change trend before and after the switching action and the prediction error.

[0044] See Figure 4The thermal distribution of temperature decay rate error at monitoring points is presented in the online adaptation and update process of the dynamic response model. This heatmap visually shows the distribution of temperature decay rate prediction errors for each monitoring point (S01-S08) at different time stamps (T1-T5). Specifically, the value of each cell in the figure represents the difference (unit: ℃ / s) between the measured temperature decay rate and the expected temperature decay rate predicted by the model at the corresponding monitoring point at the corresponding time stamp. The color gradient (red→blue) corresponds to the change of error from positive (measured faster than predicted) to negative (measured slower than predicted). From the distribution characteristics, the error of S07 at time stamp T5 (-1.99℃ / s) is the largest negative value within the stage, reflecting that the model's prediction of the temperature decay rate of S07 at this time is significantly faster; while the error of S07 at time stamp T1 (1.95℃ / s) is the largest positive value, indicating that the prediction is significantly slower at this time. Furthermore, the errors for S08 at each time stamp are all positive (0.88~1.61℃ / s), indicating that the model predictions for the area corresponding to this monitoring point are generally conservative. The errors for S03 are consistently negative from time stamps T1 to T3, reflecting a systematic bias towards a faster decay rate prediction for this area during this period. These error distributions can provide a directional basis for adjusting model weights: for example, for the extreme errors in S07, the weight gradient correction of the velocity prediction module for its corresponding flow channel area needs to be strengthened; for the persistent positive errors in S08, the activation value association logic of the pressure control module needs to be optimized.

[0045] In one embodiment of the present invention, a conductivity sensor and a viscosity sensor are installed on the coolant circulation loop to monitor the conductivity and real-time viscosity of the coolant online. The monitored conductivity value is compared with a preset pure coolant conductivity benchmark value. If it exceeds a preset threshold, it is determined that the coolant ion concentration has increased, and a coolant replacement prompt command is generated. The monitored real-time viscosity value is compared with a standard viscosity value at the current coolant temperature. If it deviates from the allowable range, it is determined that the physical properties of the coolant have decreased, and a coolant performance alarm command is generated. The opening strategy of the coolant replenishment valve is automatically adjusted or a maintenance warning is triggered by combining the coolant replacement prompt command and the coolant performance alarm command.

[0046] In practice, conductivity and viscosity sensors are installed on the coolant circulation loop to monitor the coolant's conductivity and real-time viscosity online. Both sensors are installed in the bypass detection chamber of the main circulation pipe to obtain representative fluid samples. The monitored conductivity value is compared with a preset benchmark value for pure coolant conductivity. If it exceeds a preset threshold, an increase in coolant ion concentration is detected, generating a coolant replacement prompt. The preset threshold is set based on the coolant's chemical composition and the upper limit of permissible impurity ion concentration.

[0047] In some embodiments, the monitored real-time viscosity value is compared with the standard viscosity value at the current coolant temperature. If the deviation exceeds the allowable range, it is determined that the physical properties of the coolant have deteriorated, generating a coolant performance alarm command. The standard viscosity value at the current coolant temperature is obtained by querying a pre-stored coolant viscosity-temperature characteristic curve table. The deviation of the real-time viscosity value from the standard viscosity value is calculated. in: Indicates viscosity deviation. This represents the real-time viscosity value monitored by the viscosity sensor. This represents the standard viscosity value of the coolant at the current temperature T, obtained by looking up a table. If the deviation is greater than the preset allowable deviation coefficient, it is determined that the deviation exceeds the allowable range.

[0048] Optionally, the opening strategy of the coolant replenishment valve can be automatically adjusted or a maintenance warning can be triggered by combining the coolant replacement prompt command and the coolant performance alarm command. The adjustment of the coolant replenishment valve opening strategy includes increasing the replenishment frequency or the duration of a single replenishment. The automatic adjustment logic executes predefined operations based on the combined state of the coolant replacement prompt command and the coolant performance alarm command. The operation logic is implemented through a decision matrix.

[0049] It is understandable that the adaptive adjustment process of the coolant state is independent of but parallel to the main heat dissipation control cycle. Data from the conductivity and viscosity sensors are sampled at fixed intervals and synchronized with the system clock. The criteria for generating coolant displacement prompts and coolant performance alarms are based on continuous monitoring data rather than single sample values. A sliding window averaging method is typically used to process sensor readings to improve the robustness of the judgment.

[0050] See Figure 5 In the adaptive adjustment monitoring system for coolant condition, the heatmap uses conductivity (normal / exceeding standard) and viscosity (normal / exceeding standard) as two dimensions to quantify the historical trigger counts under different state combinations. Specifically, when both conductivity and viscosity are normal, the trigger count reaches 128 times, the highest among all combinations; when conductivity is normal but viscosity exceeds the standard, the trigger count is 27 times; when conductivity exceeds the standard but viscosity is normal, the trigger count is 35 times; and when both conductivity and viscosity exceed the standard, the trigger count is the lowest, at 8 times. The color gradient scale (20-120) on the right side of the graph corresponds to the numerical range of the trigger count; the darker the color, the higher the trigger frequency, intuitively reflecting that "normal conductivity + normal viscosity" is the main monitoring scenario for coolant condition, while the abnormal scenario of both indicators exceeding the standard has a low probability of occurrence.

[0051] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0052] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A microchannel liquid cooling method for heat dissipation of integrated circuit chips, characterized in that, The method includes: Real-time capture of the timing temperature signal and the power consumption fluctuation signal at the corresponding timestamp generated during chip operation; alignment and integration of the timing temperature signal and the power consumption fluctuation signal to form a chip operation heat source profile. The spatial mesh is decomposed on the heat source profile of the chip to generate a temperature field mesh and a power consumption mapping mesh covering the three-dimensional geometry of the chip. By performing cross-correlation analysis between the temperature field grid and the power consumption mapping grid, the thermal coupling characteristics inside the chip are identified. The thermal coupling characteristics include instantaneous heat flow direction, heat density concentration, and heat propagation trajectory. The thermal coupling characteristics are input into a dynamic response model trained based on historical heat dissipation data to calculate thermal management control instructions for the current state of the chip. The thermal management control instructions include at least the expected flow rate of the coolant, the pressure distribution in the flow channel, and the timing of the flow switching. Based on the thermal management control instructions, real-time regulation of the coolant circulation within the microchannel structure integrated on the chip surface is performed.

2. The microchannel liquid cooling method for heat dissipation of integrated circuit chips according to claim 1, characterized in that, The chip's operating heat source profile is subjected to spatial mesh decomposition to generate a temperature field mesh and a power consumption mapping mesh covering the chip's three-dimensional geometry, specifically including: Based on the three-dimensional structural data of the chip package, a reference computing grid is constructed within the physical boundary of the chip; temperature signal segments corresponding to each grid cell are extracted from the heat source profile of the chip operation. For each extracted temperature signal segment, time-domain features are extracted, and the thermal inertia index and temperature oscillation frequency of each grid cell are calculated. The thermal inertia index represents the lag of temperature change in the grid cell, and the temperature oscillation frequency represents the number of temperature fluctuations per unit time. The power consumption fluctuation signal component corresponding to the timestamp of each grid cell is synchronously extracted from the heat source profile of the chip operation. Based on the power consumption fluctuation signal components and the material properties of the grid cells, the instantaneous thermal power contribution value of each grid cell is calculated. The instantaneous thermal power contribution value reflects the contribution intensity of the grid cell to the overall heat generation at a specific moment. The thermal inertia index, temperature oscillation frequency, and instantaneous thermal power contribution value of each grid cell are respectively assigned to the grid cell to form a temperature attribute grid and a power consumption attribute grid. The temperature attribute grid and the power consumption attribute grid together constitute the temperature field grid and the power consumption mapping grid.

3. The microchannel liquid cooling method for heat dissipation of integrated circuit chips according to claim 2, characterized in that, By performing cross-correlation analysis between the temperature field grid and the power consumption mapping grid, the thermal coupling characteristics inside the chip are identified, including: Traverse all grid cells and compare the temperature and power consumption attributes of each grid cell for synchronization. Identify grid cells where the temperature rise and power consumption increase are synchronized in time and mark them as active heat source cells. For the active heat source cells, analyze the temperature attribute differences between them and their adjacent grid cells and calculate the heat flow diffusion vector. The heat flow diffusion vector includes directional information from the high temperature cell to the low temperature cell and heat conduction intensity information. The heat flow diffusion vector under continuous time series is tracked and analyzed to identify the heat flow path that exists continuously in multiple time intervals, and the heat flow path is defined as the heat propagation trajectory; the number of active heat source units passed through each heat propagation trajectory is counted, and the heat density concentration of the heat propagation trajectory is calculated by combining the instantaneous heat power contribution value of the active heat source unit. The set of all heat flux diffusion vectors at each time point is extracted, and the peak value of the vector density in a specific direction is determined by statistical analysis. The direction and intensity corresponding to the peak value of the vector density are defined as the instantaneous heat flux direction.

4. The microchannel liquid cooling method for heat dissipation of integrated circuit chips according to claim 1, characterized in that, The thermal coupling characteristics are input into a dynamic response model trained based on historical heat dissipation data to calculate thermal management control instructions for the current state of the chip, including: Separate the geometric coordinate data of the heat propagation trajectory and the numerical sequence of heat density concentration from the aforementioned thermal coupling characteristics; The geometric coordinate data is overlaid and mapped with the topology diagram of the microchannel network to find the microchannel segment closest to the heat propagation trajectory in space, which is marked as the critical cooling channel. The temporal variation data of the instantaneous heat flow direction is extracted from the thermo-coupling characteristics and input into the velocity prediction module of the dynamic response model. The velocity prediction module outputs the expected velocity curve of the coolant in the critical cooling channel that meets the heat dissipation requirements of the instantaneous heat flow direction. The distribution curve of heat density concentration on the corresponding path of the critical cooling channel is extracted from the thermo-coupling characteristics and input into the pressure control module of the dynamic response model. The pressure control module outputs the pressure distribution function along the critical cooling channel required to match the distribution curve. Combining the temporal variation data of the instantaneous heat flow direction and the stability judgment result of the heat propagation trajectory, the decision module of the dynamic response model calculates the liquid flow switching timing of the coolant switching between different microchannel branches.

5. The microchannel liquid cooling method for heat dissipation of integrated circuit chips according to claim 1, characterized in that, Based on the thermal management control instructions, real-time regulation of coolant circulation within the microchannel structure integrated on the chip surface is performed, including: The expected flow rate curve in the thermal management control command is obtained and converted into a pulse width modulation signal for the micro-circulation pump. The output flow rate of the micro-circulation pump is controlled by adjusting the duty cycle of the pulse width modulation signal so that the flow rate of the coolant flowing through the key cooling channel matches the expected flow rate curve. The system receives the pressure distribution function from the thermal management control command and establishes a pressure gradient field that matches the pressure distribution function in the flow channel by adjusting the opening of multiple micro pressure regulating valves set at the inlet and branch nodes of the microchannel network. According to the liquid flow switching timing in the thermal management control command, an on / off control signal is sent to the micro solenoid valve group located at the bifurcation point of the microchannel network. The on / off control signal controls the opening and closing state of the micro solenoid valve group, thereby guiding the main flow of coolant to the branch network containing the key cooling channel at a specified time.

6. The microchannel liquid cooling method for heat dissipation of integrated circuit chips according to claim 5, characterized in that, The process of adjusting the opening of multiple miniature pressure regulating valves located at the inlet and branch nodes of the microchannel network also includes: Real-time monitoring of readings from multiple pressure sensors along key cooling channels to obtain the actual pressure distribution within the flow channels; The actual pressure distribution is compared with the pressure distribution function point by point to calculate the pressure deviation value of each monitoring point; Based on the pressure deviation value, an incremental proportional control algorithm is used to generate an incremental adjustment command for the opening of the micro pressure regulating valve at the corresponding position. The incremental adjustment command is used to reduce the pressure deviation value, so that the actual pressure distribution converges to the target pressure distribution function.

7. The microchannel liquid cooling method for heat dissipation of integrated circuit chips according to claim 4, characterized in that, Before inputting the thermal coupling features into the dynamic response model trained based on historical heat dissipation data, the process also includes real-time adaptation and updating of the dynamic response model, specifically: During the liquid cooling control process, the actual temperature change curves of multiple monitoring points on the chip surface are continuously collected; Extract the temperature decay rate and final steady-state temperature after the implementation of the control measures from the actual temperature change curve; The collected temperature decay rate and final steady-state temperature are compared with the expected temperature decay rate and expected steady-state temperature predicted by the dynamic response model under the same input features to generate a model prediction error set. Based on the model prediction error set, the weight parameters inside the dynamic response model are adjusted using the backpropagation algorithm, so that the model's response output to subsequent thermal coupling features is closer to the measured heat dissipation effect, thus completing the online adaptation and update of the model.

8. The microchannel liquid cooling method for heat dissipation of integrated circuit chips according to claim 7, characterized in that, Based on the model prediction error set, the specific steps for adjusting the internal weight parameters of the dynamic response model using the backpropagation algorithm include: From the set of model prediction errors, the prediction errors concerning the expected flow rate, the prediction errors concerning the pressure distribution, and the prediction errors concerning the effect of fluid switching timing are separated. For the prediction error regarding the expected flow velocity, calculate its gradient with respect to the weight parameters of the flow velocity prediction module in the dynamic response model, and fine-tune the weight parameters in the opposite direction of the gradient. To address the prediction error regarding pressure distribution, the error is correlated with the activation values ​​of the output layer nodes of the pressure control module. The chain rule is then used to calculate the impact of the error on the weights of each layer of the pressure control module, and the weights are corrected accordingly. To address the prediction error regarding the effect of fluid switching timing, the correlation between the error and the timing logic judgment parameters in the decision module is analyzed, and the threshold of the timing logic judgment parameters is adjusted to optimize the judgment of switching timing.

9. The microchannel liquid cooling method for heat dissipation of integrated circuit chips according to claim 1, characterized in that, During the real-time regulation of coolant circulation, adaptive adjustments to the coolant's own state are also performed, including: Conductivity and viscosity sensors are installed in the coolant circulation loop to monitor the online conductivity and real-time viscosity of the coolant. The monitored conductivity value is compared with a preset pure coolant conductivity benchmark value. If it exceeds a preset threshold, it is determined that the coolant ion concentration has increased, and a coolant replacement prompt command is generated. The monitored real-time viscosity value is compared with the standard viscosity value at the current coolant temperature. If it deviates beyond the allowable range, it is determined that the coolant physical properties have deteriorated, and a coolant performance alarm command is generated. Combining the coolant replacement prompt command and the coolant performance alarm command, the opening strategy of the coolant replenishment valve is automatically adjusted or a maintenance warning is triggered.

10. The microchannel liquid cooling method for heat dissipation of integrated circuit chips according to claim 2, characterized in that, The instantaneous thermal power contribution value of each grid cell is calculated based on the power consumption fluctuation signal components and the material properties of the grid cells, specifically including: Extract the current intensity sequence and voltage fluctuation sequence within a specified time window from the power consumption fluctuation signal component; multiply the current intensity sequence and voltage fluctuation sequence point by point to obtain the instantaneous power value at each sampling moment; Obtain the thermal resistivity and specific heat capacity parameters of the chip material corresponding to each grid cell; Multiplying the instantaneous electrical power value by the thermal resistance coefficient yields the theoretical temperature rise contribution of the grid cell at the heating level. The theoretical temperature rise contribution is weighted and fused with the specific heat capacity parameter to calculate the heat value emitted by the grid cell to the surrounding environment per unit time. The heat value is normalized to a preset baseline power consumption level to obtain a dimensionless instantaneous thermal power contribution value, which is then assigned to the grid cell.

Images