Rack cold plate magnetorheological flow channel edge ai cooperative heat dissipation control method and device

Abstract

The application discloses a rack cold plate magnetorheological flow channel edge AI cooperative heat dissipation control method and device, and relates to the technical field of cold plate heat dissipation control. The rack cold plate magnetorheological flow channel edge AI cooperative heat dissipation control method and device comprise the following steps: S1, collecting edge thermal control state data, and performing pretreatment on the edge thermal control state data; S2, identifying a thermal variation unit and constructing a thermal load cluster, evaluating the cooling regulation priority of the thermal variation unit, and generating a thermal load space partition; S3, constructing a high-thermal-response unit set, evaluating the magnetorheological regulation response strength of the high-thermal-response unit, and hierarchically generating a regional regulation strategy; S4, driving microstructure deformation, electromagnetic excitation and flow regulation; and evaluating the regulation response hysteresis after execution, and correcting the control instruction and re-driving execution if the hysteresis is out of standard. The application solves the problem that, under the condition of severe fluctuation of chip thermal load, the traditional structure cannot dynamically regulate flow and thermal resistance, resulting in uneven heat dissipation and waste of cooling resources.

Description

Technical Field

[0001] This invention relates to the field of cold plate heat dissipation control technology, specifically to a method and device for AI-coordinated heat dissipation control at the edge of the magnetorheological flow channel of a rack cold plate. Background Technology

[0002] In order to cope with the ever-increasing demand for thermal power consumption in high-performance computing devices, data centers and AI chip platforms, the industry has proposed a variety of innovative heat dissipation solutions. Some of these solutions combine new mechanisms such as magnetic cooling, thermoelectric conversion and structural optimization to improve system heat dissipation efficiency and responsiveness.

[0003] For example, the invention with publication number CN117979630A relates to the field of heat dissipation technology, aiming to solve the problems of high energy consumption, high noise, and short lifespan of traditional heat dissipation methods. It provides a magnetic cooling heat dissipation cover plate, including a cover plate body on which a magnetic cooling material is integrated. A thermoelectric power generation device is disposed outside the magnetic cooling material. The thermoelectric power generation device is used to convert the heat energy released by the magnetic cooling material into electrical energy. A control system is disposed outside the thermoelectric power generation device to adjust the magnetic field strength and direction. This invention's magnetic cooling heat dissipation cover plate uses magnetic cooling technology to absorb and convert the heat generated by electronic devices into electrical energy, thereby achieving a highly efficient heat dissipation effect. By converting the heat generated by electronic devices into electrical energy for storage or power supply, dependence on external power sources can be reduced, resulting in energy saving and environmental protection. It also requires no large amount of energy to operate, resulting in low energy consumption.

[0004] For example, invention CN112739188B discloses a high-heat-dissipation electromagnetic shielding device and heat dissipation method, including a fixed base plate, an electromagnetic shielding chamber, a negative pressure chamber, a square air outlet pipe, a cooling plate, an air supply pipe, and a water collection box. The fixed base plate is installed on the inner bottom wall of the electromagnetic shielding chamber, the negative pressure chamber is installed on the back of the electromagnetic shielding chamber, the cooling plate is installed inside the negative pressure chamber, a square air outlet pipe is connected to one side surface of the electromagnetic shielding chamber, an air supply pipe is connected to the other side surface of the electromagnetic shielding chamber, an air compressor is installed on the inner bottom wall of the negative pressure chamber, a water collection box is installed at the bottom of the negative pressure chamber, and a storage cabinet is installed inside the electromagnetic shielding chamber. This invention, by setting a fixed base plate, allows the airflow inside the electromagnetic shielding chamber to compress through the extension and retraction of the electric telescopic rods on both sides, thereby agitating the hot air in the internal corners, facilitating omnidirectional extraction by the negative pressure fan, and achieving efficient and comprehensive heat dissipation and cooling.

[0005] However, despite the progress made by the aforementioned technologies in improving heat dissipation efficiency and optimizing structural design, they are still limited to overall heat dissipation strategies and lack intelligent identification mechanisms for the spatial heterogeneity of chip-level thermal loads, making it difficult to achieve dynamic zoning and control of local high-heat areas. Furthermore, the lack of microscale magnetorheological control mechanisms and feedback closed-loop path adjustment capabilities makes it difficult to meet the real-time, fine-grained control requirements under complex thermal dynamic scenarios.

[0006] Therefore, in order to address the above problems, there is an urgent need for a method and device for AI-assisted heat dissipation control at the edge of the magnetorheological flow channel of the rack cold plate. Summary of the Invention

[0007] Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides a method and device for AI-coordinated heat dissipation control at the edge of the magnetorheological flow channel of a rack cold plate. This solves the problem that traditional flow channel structures cannot dynamically adjust flow rate and thermal resistance under drastic fluctuations in chip thermal load, resulting in uneven heat dissipation and wasted cooling resources.

[0009] Technical solution

[0010] To achieve the above objectives, the present invention provides the following technical solution: a method and apparatus for AI-coordinated heat dissipation control of the edge of a rack cold plate magnetorheological channel, comprising: S1, collecting edge thermal control state data, performing structural integrity verification, disturbance smoothing, feature alignment, and scale normalization on the edge thermal control state data to obtain preprocessed edge thermal control state data; S2, based on the preprocessed edge thermal control state data, identifying thermal variation units in the chip region and constructing thermal load clusters, evaluating the cooling control priority of thermal variation units, and generating thermal load spatial partitions; S3, extracting thermal variation units in high-heat areas to construct a set of high-thermal-response units, evaluating the magnetorheological control response intensity of high-thermal-response units, generating regional control strategies in a hierarchical manner, and constructing a control strategy dataset; S4, parsing the control strategy dataset, generating a path-level control instruction set, driving microstructure deformation, electromagnetic excitation, and flow regulation; evaluating the control response hysteresis after execution, and correcting the control instructions and re-driving execution if the hysteresis exceeds the standard, to achieve closed-loop control.

[0011] Further, edge thermal control status data is collected, and structural integrity verification, perturbation smoothing, feature alignment, and scale normalization are performed on the edge thermal control status data. The specific steps to obtain the preprocessed edge thermal control status data are as follows: Edge thermal control status data is collected, including chip temperature, coolant inlet temperature, coolant outlet temperature, chip thermal power density, coolant flow rate, coolant density, coolant specific heat capacity, electromagnetic coil current, magnetic field strength, chip power consumption, thermal interface contact pressure, and microchannel pressure. The structural integrity of the edge thermal control status data is verified using a dynamic boundary consistency detection algorithm to remove abnormal data points caused by abnormal acquisition. Local perturbation reduction is performed on the edge thermal control status data using a multi-scale trend filtering algorithm to smooth high-frequency perturbation waveforms caused by load mutations and control drives. Feature alignment is performed on the edge thermal control status data using a distributed normalization renormalization algorithm to unify the numerical distribution. Scale normalization is performed on the edge thermal control status data using an interval linear mapping function.

[0012] Further, based on the preprocessed edge thermal control state data, the specific steps for identifying thermal variation units in the chip region and constructing thermal load clusters are as follows: Based on the preprocessed edge thermal control state data, a coolant heat exchange power algorithm based on energy conservation is called, with coolant flow rate, coolant inlet temperature, and outlet temperature as inputs. Coolant density and specific heat capacity are introduced as constant factors, and heat exchange power calculation is performed to obtain the cooling environment reference power; the difference between the chip temperature and the coolant inlet temperature is calculated and divided by the chip thermal power density to obtain the thermal interface thermal resistance; a thermal load expression vector containing the cooling environment reference power, thermal interface thermal resistance, and edge thermal control state data is constructed; the spatial index of the corresponding sensor in the edge thermal control state data is extracted, and a mapping relationship between the thermal load expression vector and the cold plate position is established; a sliding window is set, the local gradient of the chip temperature in each window is calculated, and data points whose local gradient exceeds the thermal gradient identification threshold are selected as thermal variation units; thermal variation units in adjacent steps are connected to form thermal load clusters.

[0013] Further, the specific steps for evaluating the cooling control priority of thermal variation units and generating thermal load spatial partitions are as follows: Multiply the chip thermal power density by the difference between the chip temperature and the coolant inlet temperature to obtain the thermal load temperature rise; square the difference between the coolant outlet temperature and the inlet temperature and multiply it by the sum of the thermal interface thermal resistance and the minimum term to obtain the comprehensive heat dissipation difficulty factor; divide the thermal load temperature rise by the comprehensive heat dissipation difficulty factor to obtain the chip thermal pressure intensity; divide the chip power consumption by the sum of the cooling environment reference power and the minimum term, add one to the result and take the natural logarithm to obtain the heat source load adjustment factor; multiply the chip thermal pressure intensity by the heat source load adjustment factor to obtain the cooling control priority evaluation value; summarize the cooling control priority evaluation values ​​corresponding to each thermal variation unit in each thermal load cluster, calculate the mean as the cluster representative value, and compare the cluster representative value with the thermal load partition threshold in real time: when the cluster representative value is greater than the thermal load partition threshold, the thermal load cluster is marked as a high-heat area, and the remaining thermal load clusters are marked as low-heat areas.

[0014] Furthermore, the specific steps for extracting thermal variation units from high-heat regions to construct a high-thermal-response unit set and evaluating the magnetorheological control response intensity of the high-thermal-response units are as follows: Extract all thermal variation units from the thermal load cluster labeled as high-heat regions to construct a high-thermal-response unit set; calculate the difference in microchannel inlet and outlet pressure data based on the corresponding positions of the high-thermal-response units to obtain the microchannel pressure difference; extract the contact pressure data at the start and end positions of the thermal variation units in the thermal interface contact pressure array and calculate the difference to obtain the thermal interface contact pressure difference; divide the cooling control priority evaluation value by the cluster representative value to obtain the relative intensity of cooling control. The following steps are performed: 1) Multiply the chip power consumption by the difference between the chip temperature and the coolant inlet temperature to construct a power consumption-temperature rise coupling factor; 2) Multiply the difference between the coolant outlet temperature and the inlet temperature by the sum of the thermal interface thermal resistance and a minimum term to obtain a comprehensive heat dissipation difficulty factor; 3) Divide the power consumption-temperature rise coupling factor by the comprehensive heat dissipation difficulty factor to obtain a thermal load gradient adjustment factor; 4) Add the microchannel pressure difference and the thermal interface contact pressure difference, divide by the product of the electromagnetic coil magnetic field strength and current plus a minimum term to obtain a magnetic control adjustment driving force factor; 5) Multiply the relative intensity of cooling regulation, the thermal load gradient adjustment factor, and the magnetic control adjustment driving force factor to obtain a magnetorheological response evaluation value.

[0015] Further, the specific steps for generating hierarchical regional control strategies and constructing a control strategy dataset are as follows: Compare the magnetorheological response evaluation value with the control grading threshold to generate hierarchical regional control strategies: When the magnetorheological response evaluation value is less than or equal to the first-level control grading threshold, the high thermal response unit is divided into a stable maintenance region to maintain the current original microchannel structure and electromagnetic excitation; when the magnetorheological response evaluation value is greater than the first-level control grading threshold but less than the second-level control grading threshold, it is divided into a collaborative compensation region to increase the magnetic particle aggregation density and shrink the flow channel cross-sectional area; when the magnetorheological response evaluation value is greater than or equal to the second-level control grading threshold, it is divided into an active intervention region unit to execute maximum power excitation of the electromagnetic coil, construct a high-resistance chain microchannel structure, and increase the coolant inlet flow rate; extract the magnetorheological response evaluation value, control strategy, and spatial index of each high thermal response unit to construct a control strategy dataset.

[0016] Furthermore, the specific steps for parsing the control strategy dataset and generating a path-level control instruction set to drive microstructure deformation, electromagnetic excitation, and flow regulation are as follows: Receive the control strategy dataset, call the hardware control rule set, and map the control strategy and spatial index into a path-level control instruction set, including micro-area electromagnetic excitation control instructions, microstructure deformation driving instructions, and inlet flow scheduling instructions; send the control instructions to the corresponding micropipe nodes to drive the execution of local electromagnetic excitation, electromagnetic coil current modulation, microstructure deformation, and inlet flow regulation operations.

[0017] Further, the specific steps for evaluating the hysteresis of the control response after execution are as follows: Based on the edge thermal control state data before and after the control execution, extract the trigger timestamp when the chip temperature first exceeds the temperature threshold, and the corresponding response timestamp when the chip temperature first shows a stable downward trend, and calculate the difference to obtain the response time delay; multiply the difference between the chip temperature before and after control by the sum of the chip thermal power density and the minimum term, and then multiply by the response time delay to obtain the temperature rise hysteresis load response work; add the microchannel pressure difference and the thermal interface contact pressure difference after control to obtain the composite pressure difference; divide the chip power consumption after control by the sum of the cooling environment reference power and the minimum term, add one to the ratio and then take the natural logarithm to obtain the control benefit function; divide the temperature rise hysteresis load response work by the product of the composite pressure difference and the control benefit function to obtain the thermal control response hysteresis evaluation value.

[0018] Furthermore, if the hysteresis exceeds the limit, the control command is corrected and re-driven to achieve closed-loop control. The specific steps are as follows: Real-time comparison of the thermal control response hysteresis evaluation value and the hysteresis tolerance threshold. When the thermal control response hysteresis evaluation value is less than or equal to the hysteresis tolerance threshold, the current path-level control strategy is maintained. When the thermal control response hysteresis evaluation value is greater than the hysteresis tolerance threshold, a correction operation based on the artificial intelligence control strategy is executed: increasing the magnetic particle aggregation density of the target flow channel section; calculating the ratio of the composite pressure difference to the pressure difference threshold. If the ratio is less than a constant, the electromagnetic coil current is reduced and the branch pump speed is lowered; otherwise, the electromagnetic coil current is increased and the branch pump speed is raised; based on the control changes of magnetic particle aggregation density, electromagnetic coil current, and branch pump speed, a correction control command set is generated, replacing the original command and issued to the corresponding micropipe node, and the control operation is re-executed to achieve closed-loop control response.

[0019] The second aspect of this invention provides an AI-coordinated heat dissipation control device for the edge of a rack-mounted cold plate magnetorheological channel, comprising: a data acquisition and preprocessing module, a thermal load identification and partitioning module, a hierarchical control strategy generation module, and a control command execution feedback module, wherein: the data acquisition and preprocessing module is used to acquire edge thermal control status data, and perform structural integrity verification, disturbance smoothing, feature alignment, and scale normalization processing on the edge thermal control status data to obtain preprocessed edge thermal control status data; the thermal load identification and partitioning module is used to identify thermal variation units in the chip region and construct based on the preprocessed edge thermal control status data. The system employs several mechanisms: thermal load clustering to evaluate the cooling control priority of thermally variable units and generate thermal load spatial partitions; a hierarchical control strategy generation module to extract thermally variable units in high-heat regions to construct a set of high-thermal-response units, evaluate the magnetorheological control response intensity of high-thermal-response units, generate hierarchical regional control strategies, and construct a control strategy dataset; and a control command execution feedback module to parse the control strategy dataset, generate a path-level control command set, drive microstructure deformation, electromagnetic excitation, and flow regulation; and an evaluation of the control response hysteresis after execution. If the hysteresis exceeds the limit, the control command is corrected and re-driven to achieve closed-loop control.

[0020] Beneficial effects

[0021] The present invention has the following beneficial effects:

[0022] (1) The AI ​​collaborative heat dissipation control method and device for the edge of the cold plate magnetorheological channel of the rack significantly improves the stability and availability of the original thermal control data by performing preprocessing on the edge thermal control status data. On this basis, the thermal variation unit in the chip area is dynamically identified by combining artificial intelligence algorithm, and the thermal load space partition is automatically constructed based on the clustering model. The area of ​​drastic thermal load fluctuation in the chip surface or local structure is accurately captured, and its partitioning relationship is updated in real time, providing basic support for the intelligent formulation of regional directional cooling strategy, thereby significantly improving the response accuracy and environmental adaptability of thermal regulation.

[0023] (2) The AI-coordinated heat dissipation control method and device for the edge of the magnetorheological flow channel of the rack cold plate extracts typical thermal variation units marked as high-heat zones based on the spatial partitioning results of the heat load, and constructs a set of high-heat response units. The magnetorheological regulation response intensity of each response unit is further evaluated, and the corresponding cooling strategy priority is set according to different heat load levels to generate a graded regional regulation scheme. This strategy can not only effectively improve the priority cooling efficiency of high heat density units, but also suppress redundant flow of resources in low-heat zones, realizing the intelligent transformation of heat dissipation regulation from average distribution to heat perception-driven.

[0024] (3) The AI-coordinated heat dissipation control method and device for the edge of the magnetorheological flow channel of the cold plate of the rack introduces an artificial intelligence model to participate in strategy execution and correction judgment. By calculating the thermal control response hysteresis evaluation value and comparing it with the hysteresis tolerance threshold in real time, it intelligently determines whether to execute strategy correction. When excessive hysteresis is detected, the control variables such as magnetic particle aggregation density, electromagnetic coil current and micro-pump speed are dynamically corrected based on AI decision logic, and the correction control command is regenerated to drive the closed-loop execution of the control action, effectively improving the system's response sensitivity and control robustness.

[0025] (4) The AI-coordinated heat dissipation control method and device for the edge of the magnetorheological flow channel of the rack cold plate automatically generates multi-dimensional control commands, including active deformation of microstructures, excitation of electromagnetic coils and flow regulation, by constructing a path-level control command set and combining it with a regional regulation strategy dataset, thereby achieving precise regulation of the cooling path. By making full use of the tunable rheological properties of magnetorheological materials and combining them with the deformable channel structure of microstructures, the flow resistance of coolant and the pressure distribution of heat exchange interface can be flexibly adjusted at local locations. Compared with traditional rigid fixed structures, it has higher regulation accuracy, faster dynamic response and lower energy consumption. Attached Figure Description

[0026] Figure 1 Flowchart of AI-coordinated heat dissipation control method for the edge of magnetorheological channel of rack cold plate;

[0027] Figure 2 Structural diagram of the AI-coordinated heat dissipation control device at the edge of the magnetorheological flow channel of the rack cold plate;

[0028] Figure 3 The graph shows the thermal control response hysteresis evaluation of the high thermal response unit.

[0029] Figure 4 This is a schematic diagram of the overall structure of the magnetorheological flow channel on the cold plate of the frame;

[0030] Figure 5 This is a schematic diagram of the magnetorheological control structure inside the cold plate.

[0031] In the diagram, 1 is the two-phase cold plate; 2 is the chip; 3 is the coolant inlet pipe; 4 is the electromagnet; and 5 is the microchannel wall. Detailed Implementation

[0032] 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.

[0033] Please see Figures 1-5 This invention provides a technical solution: a method and apparatus for AI-coordinated heat dissipation control of the edge of a rack cold plate magnetorheological channel, comprising: S1, collecting edge thermal control state data, performing structural integrity verification, disturbance smoothing, feature alignment and scale normalization processing on the edge thermal control state data to obtain preprocessed edge thermal control state data; S2, based on the preprocessed edge thermal control state data, identifying thermal variation units in the chip 2 region and constructing thermal load clusters, evaluating the cooling regulation priority of thermal variation units, and generating thermal load spatial partitions; S3, extracting thermal variation units in high-heat areas to construct a high-thermal-response unit set, evaluating the magnetorheological regulation response intensity of high-thermal-response units, generating hierarchical regional regulation strategies, and constructing a regulation strategy dataset; S4, parsing the regulation strategy dataset, generating a path-level control instruction set, driving microstructure deformation, electromagnetic excitation and flow regulation; evaluating the regulation response hysteresis after execution, and correcting the control instructions and re-driving execution if the hysteresis exceeds the standard to achieve closed-loop regulation.

[0034] Specifically, the steps for acquiring edge thermal control status data and performing structural integrity verification, perturbation smoothing, feature alignment, and scale normalization on the edge thermal control status data are as follows: The edge thermal control status data includes twelve categories of thermal control parameters: chip temperature, coolant inlet temperature, coolant outlet temperature, chip thermal power density, coolant flow rate, coolant density, coolant specific heat capacity, electromagnetic coil current, magnetic field strength, chip power consumption, thermal interface contact pressure, and microchannel pressure. All parameters are periodically and in real-time acquired through edge distributed sensing nodes. Specifically, the chip temperature is acquired using a miniature thermoelectric sensor deployed on the surface of chip 2; and the chip temperature is acquired using digital thermal sensors deployed at the coolant inlet pipe 3 and the coolant outlet pipe. The system acquires the inlet and outlet temperatures of the coolant using sensors; obtains the chip's thermal power density using an integrated heat flux density meter located beneath the chip 2's packaging layer; acquires coolant flow rate data using an electromagnetic flow meter installed in the liquid cooling pipeline; acquires coolant density using an online density meter located within the coolant flow channel; acquires the coolant specific heat capacity using an insertion-type heat capacity measurement probe; acquires the electromagnetic coil current using a current sensor connected to the magnetically controlled excitation circuit; acquires magnetic field strength data using a Hall element placed within the magnetic field's area of ​​action; acquires chip power consumption using a dual-channel voltage and current acquisition unit integrated into the chip 2's power input terminal; acquires the thermal interface contact pressure using a piezoelectric thin-film sensor located at the interface between chip 2 and the cold plate; and acquires micro-channel pressure using micro-differential pressure sensors located at both ends of the microchannel. Then, a dynamic boundary consistency detection algorithm is used to verify the structural integrity of the edge thermal control status data. Constraint consistency analysis is performed on individual and combined structural parameters of chip temperature, coolant inlet temperature, coolant outlet temperature, chip thermal power density, coolant flow rate, coolant density, coolant specific heat capacity, electromagnetic coil current, magnetic field strength, chip power consumption, thermal interface contact pressure, and microchannel pressure. Abnormal data points caused by edge sensing node failures, communication delays, or sudden disturbances are identified and eliminated. A multi-scale trend filtering algorithm is used to reduce local disturbances in the edge thermal control status data. Focusing on chip thermal power density, chip temperature, coolant flow rate, and magnetic field strength, multi-time-scale trend extraction and reconstruction are performed to smooth the disturbances in the edge thermal control status data caused by transient load fluctuations and rapid actuator responses. Finally, a distributional normalization renormalization algorithm is used to align the features of the edge thermal control status data, eliminating feature offsets caused by differences in physical dimensions between thermal control parameters and unifying the numerical distribution. The edge thermal control status data is scaled and normalized by using an interval linear mapping function. Based on the theoretical value range of each thermal control parameter and the rated operating range of the equipment, each edge thermal control status data is converted into a dimensionless value within a unified numerical range, forming preprocessed edge thermal control status data with a unified measurement standard.

[0035] In this implementation plan, by finely defining the edge thermal control status data acquisition items, the accurate perception and full coverage of each thermal control parameter in the physical space are ensured. At the same time, by matching them with specific sensor types and deployment locations, the acquisition accuracy, timeliness and consistency of edge thermal control status data are effectively improved, providing a high-quality data foundation for subsequent control strategy identification, response evaluation and control execution, and enhancing the feasibility and credibility of edge AI thermal control perception and decision-making.

[0036] Specifically, based on the preprocessed edge thermal control state data, the specific steps for identifying thermal variation units in the chip 2 region and constructing thermal load clusters are as follows: Based on the preprocessed edge thermal control state data, a coolant heat exchange power algorithm based on energy conservation is called, with coolant flow rate, coolant inlet temperature, and coolant outlet temperature as inputs, and coolant density and specific heat capacity as constant factors. Heat exchange power calculation is performed to obtain the cooling environment reference power; the difference between the chip temperature and the coolant inlet temperature is calculated and divided by the chip thermal power density to obtain the thermal interface resistance, which is used to characterize the heat conduction capability between the chip 2 surface and the coolant; a thermal load containing the cooling environment reference power, thermal interface resistance, and various measurement parameters from the edge thermal control state data is constructed. The thermal load expression vector is used to ensure that the vector remains synchronized in time and has spatial continuity in physical location. The spatial index of the corresponding sensor is extracted from the edge thermal control status data to establish a one-to-one mapping relationship between the thermal load expression vector and the physical installation area of ​​the cold plate, ensuring that the spatial information is not distorted. A sliding calculation window with a fixed window length and step size is set to calculate the local gradient change amplitude of the chip temperature in each window. When the change rate exceeds the thermal gradient identification threshold, it is marked as a thermal variation unit. The thermal variation units that appear consecutively in adjacent steps are connected to form a high thermal distribution region. Based on the similarity of various parameters in the thermal load expression vector in multi-dimensional space, thermal load clustering is constructed to realize the structured expression and dynamic positioning of the high thermal variation region in chip 2.

[0037] In this implementation scheme, a thermal load representation vector is constructed based on coolant flow rate, coolant inlet temperature, coolant outlet temperature, coolant density, coolant specific heat capacity, chip temperature, chip thermal power density, and spatial index. Combined with the cooling environment reference power and thermal interface thermal resistance, the thermal variation units in the chip 2 region are accurately identified. Furthermore, a sliding window is used to filter regions with significant changes in local chip temperature gradient, forming thermal load clusters. This significantly improves the perception resolution and representation completeness of abnormal thermal load regions in chip 2, and enhances the spatial orientation and dynamic tracking capability of the heat dissipation control object.

[0038] Specifically, the steps for evaluating the cooling control priority of the thermal variation unit and generating the thermal load spatial partition are as follows: Multiply the chip thermal power density by the difference between the chip temperature and the coolant inlet temperature, and combine this with the trend of heat flux per unit area to obtain the thermal load temperature rise, which reflects the instantaneous thermal intensity of a local area of ​​chip 2; Square the difference between the coolant outlet temperature and the inlet temperature, multiply this by the sum of the thermal interface thermal resistance and the minimum term, and further add the influence of the contact thermal resistance change caused by the unit pressure difference to obtain the comprehensive heat dissipation difficulty factor, which measures the comprehensive resistance characteristics of the liquid heat dissipation capacity in the cooling channel; where the minimum term is a very small but non-zero positive real number used to avoid numerical instability caused by division by zero during the calculation process, and its value range is... arrive Unless otherwise specified, all subsequent minima will use this definition and value range. The heat load temperature rise is divided by the comprehensive heat dissipation difficulty factor to obtain the chip thermal pressure intensity, reflecting the heat output intensity per unit of heat dissipation capacity. The chip power consumption is divided by the sum of the cooling environment reference power and the introduced minima, and the result is incremented by one and the natural logarithm is taken to suppress the amplification effect of outliers, resulting in the heat source load adjustment factor, reflecting the heat source power deviation trend per unit of cooling capacity. The chip thermal pressure intensity is multiplied by the heat source load adjustment factor to comprehensively reflect the urgency of cooling demand in local areas, obtaining the priority evaluation value for cooling regulation. The priority evaluation values ​​for cooling regulation corresponding to each thermal variation unit within each heat load cluster are summarized, and the mean is calculated based on the moving average model as the cluster representative value. The cluster representative value and the heat load partition threshold are compared in real time: when the cluster representative value is greater than the heat load partition threshold, the heat load cluster is marked as a high-heat area, indicating a priority regulation area; the remaining heat load clusters are marked as low-heat areas, indicating areas where regulation can be delayed, thus completing the intelligent delineation of heat load spatial partitions.

[0039] The specific formula for calculating the priority evaluation value of cooling regulation is as follows:

[0040] ;

[0041] In the formula, S represents the priority evaluation value for cooling regulation, and Q represents the chip's thermal power density. Indicates chip temperature. Indicates the coolant inlet temperature. This indicates the coolant outlet temperature, and R represents the thermal resistance of the heat exchanger. Indicates chip power consumption. Indicates the reference power for the cooling environment. Indicates a minus term.

[0042] In this implementation scheme, a calculation model is constructed based on chip thermal power density, chip temperature, coolant inlet temperature, coolant outlet temperature, coolant flow rate, thermal interface thermal resistance, chip power consumption, and cooling environment reference power. This model integrates thermal load temperature rise, comprehensive heat dissipation difficulty factor, chip thermal pressure intensity, and heat source load adjustment factor to form a priority evaluation value for cooling regulation. This value can accurately quantify the cooling urgency of each thermal variation unit under the current thermal load. By dynamically comparing the thermal load cluster mean of the priority evaluation value for cooling regulation in real time, and automatically calibrating the boundary between high-heat and low-heat areas using thermal load partitioning thresholds, the thermal load spatial partitioning results become more responsive and discriminative, thereby effectively improving the regional targeting of the regulation strategy and the allocation efficiency of heat dissipation control resources.

[0043] Specifically, the steps for extracting thermal variation units from high-heat regions to construct a high-thermal-response unit set and evaluating the magnetorheological regulation response intensity of the high-thermal-response units are as follows: Extract all thermal variation units from the thermal load cluster labeled as high-heat regions to construct a high-thermal-response unit set; calculate the difference between the inlet and outlet pressures of the microchannels corresponding to the high-thermal-response units to obtain the microchannel pressure difference, which characterizes the flow resistance change in the local cooling path; extract the thermal interface contact pressure data corresponding to the start and end positions of the thermal variation units in the thermal interface contact pressure array, calculate the difference between the two, and obtain the thermal interface contact pressure difference, which characterizes the stress response change trend of the thermal coupling interface under the influence of high thermal load; divide the cooling regulation priority evaluation value of each thermal variation unit by the cluster representative value to obtain the relative intensity of cooling regulation, which reflects the relative ranking of the cooling urgency of the thermal variation units within the region; and combine the chip power consumption with the chip temperature and coolant inlet temperature... The difference in temperature is multiplied to construct a power consumption temperature rise coupling factor, which is used to describe the transient heat accumulation level in the local area of ​​chip 2. The difference between the coolant outlet temperature and the inlet temperature is multiplied by the sum of the thermal resistance of the thermal interface and the minimum term to obtain a comprehensive heat dissipation difficulty factor, which serves as a composite resistance characterization of the chip's heat dissipation path. The power consumption temperature rise coupling factor is divided by the comprehensive heat dissipation difficulty factor to obtain a thermal load gradient adjustment factor, which is used to quantify the degree of thermal gradient mismatch between the heat source and the cooling channel. The pressure difference of the microchannel and the contact pressure difference of the thermal interface are added to obtain the structural flow thermal coupling pressure difference, which is then divided by the product of the electromagnetic coil magnetic field strength and current plus the minimum term to obtain a magnetic control adjustment driving force factor, which is used to evaluate the effective driving capability of magnetic field changes on the flow channel behavior under the current structural conditions. The relative intensity of cooling regulation, the thermal load gradient adjustment factor, and the magnetic control adjustment driving force factor are multiplied to obtain a magnetorheological response evaluation value, which serves as the basis for judging the regulation response intensity of the high thermal response unit.

[0044] The specific formula for calculating the magnetorheological response evaluation value is as follows:

[0045] ;

[0046] In the formula, G represents the magnetorheological response evaluation value, and S represents the cooling control priority evaluation value. This represents the cluster representative value. Indicates chip power consumption. Indicates chip temperature. Indicates the coolant inlet temperature. This indicates the coolant outlet temperature, and R represents the thermal resistance of the heat exchanger. Indicates the pressure difference in the microchannel. B represents the contact pressure difference at the thermal interface, B represents the magnetic field strength of the electromagnetic coil, and I represents the current in the electromagnetic coil. Indicates a minus term.

[0047] This implementation scheme systematically evaluates key parameters such as microchannel pressure difference, thermal interface contact pressure difference, chip power consumption, chip temperature, coolant inlet temperature, coolant outlet temperature, thermal interface thermal resistance, electromagnetic coil magnetic field strength, electromagnetic coil current, and cooling control priority evaluation value. It then constructs intermediate physical quantities such as the relative intensity of cooling control, power consumption-temperature rise coupling factor, thermal load gradient adjustment factor, and magnetic control driving force factor to quantitatively determine the magnetorheological control response capability of high-thermal-response units. The calculation of the magnetorheological response evaluation value provides a precise basis for the subsequent graded execution and dynamic adjustment of the magnetorheological control strategy, effectively improving the targeting and adjustment efficiency of cooling control in high-heat areas, thereby enhancing the dynamic matching capability and local response sensitivity of the system's heat dissipation performance.

[0048] Specifically, the steps for generating hierarchical regional control strategies and constructing a control strategy dataset are as follows: Comparing the magnetorheological response evaluation value with the control level threshold, and based on the thermal control scheduling requirements and cooling response capabilities of the high thermal response unit, hierarchical regional control strategies are generated: When the magnetorheological response evaluation value is less than or equal to the first-level control level threshold, the high thermal response unit is divided into a stable maintenance zone, where the region is in a cooling equilibrium and stable stage, maintaining the original structure state and electromagnetic excitation parameters of the current microchannel, ensuring that the local thermal flow field is not disturbed; when the magnetorheological response evaluation value is greater than the first-level control level threshold but less than the second-level control level threshold, it is divided into a collaborative compensation zone, where the regional heat load is in a subcritical growth state, and the local magnetorheological material is enhanced by increasing the magnetic particle aggregation density. The structural viscosity of the material induces the shrinkage of the cross-sectional area of ​​the microchannels, regulating the local flow velocity distribution of the coolant and achieving enhanced heat flow transport. When the magnetorheological response evaluation value is greater than or equal to the secondary control grading threshold, it is classified as an active intervention zone unit. The region is in a high-thermal-intensity disturbance state, and the maximum power excitation of the electromagnetic coil is executed to instantaneously enhance the local magnetic field strength, guiding the magnetic particles to form a high-resistance chain-type microchannel structure, rapidly increasing the shear resistance of the microstructure wall, and simultaneously increasing the coolant inlet flow rate, driving enhanced cooling response. The magnetorheological response evaluation value, regional control strategy, microchannel spatial index, and geographical location information of each high-thermal-response unit are extracted to construct a control strategy dataset, which serves as the input basis for the subsequent path-level control instruction set generation, ensuring the continuity of control logic and the accuracy of execution actions.

[0049] In this implementation scheme, a hierarchical mechanism for regional control strategies based on magnetorheological response evaluation values ​​is constructed, significantly improving the cooling accuracy and response efficiency of high thermal response units. By setting primary and secondary control thresholds, and rationally dividing the stable maintenance zone, collaborative compensation zone, and active intervention zone, hierarchical control of the original structural state of the microchannel, electromagnetic excitation parameters, magnetic particle aggregation density, microchannel cross-sectional area, and coolant inlet flow rate is achieved, effectively enhancing the adaptability and responsiveness of the control strategy. The construction of the control strategy dataset further realizes the dynamic matching between the magnetorheological response evaluation values ​​of each high thermal response unit, regional control strategies, and spatial indices, providing highly reliable data support for the generation of path-level control instruction sets.

[0050] Specifically, the steps for parsing the control strategy dataset and generating a path-level control instruction set to drive microstructure deformation, electromagnetic excitation, and flow regulation are as follows: The control strategy dataset is received, and a hardware control rule set is invoked. This hardware control rule set defines the operational boundaries and control accuracy requirements for micro-pipe nodes to perform micro-region electromagnetic excitation, microstructure deformation, and flow regulation. The control strategy and spatial index are mapped according to the control rule set to generate a path-level control instruction set. This set includes micro-region electromagnetic excitation control instructions, microstructure deformation driving instructions, and inlet flow scheduling instructions. These three types of instructions are used to adjust the magnetic particle aggregation density in the target area, the magnitude and direction of the electromagnetic coil current, the geometry of the microstructure flow channel, and the amplitude of the coolant inlet flow scheduling, respectively. The path-level control instruction set is then sent to the corresponding micro-pipe node addresses to drive the execution of heat dissipation control operations, including local electromagnetic excitation, electromagnetic coil current modulation, reversible microstructure deformation, and quantitative adjustment of the coolant inlet flow, ensuring the accuracy of the spatial mapping of the control strategy and the real-time performance of the response.

[0051] In this implementation scheme, by parsing the control strategy dataset and combining it with the spatial index and control rule set, a path-level control command set is generated, which enables precise mapping of control strategies to physical execution actions. Through the joint issuance of micro-area electromagnetic excitation control commands, microstructure deformation driving commands, and inlet flow scheduling commands, each micro-pipe node is driven to accurately respond to magnetic particle aggregation density adjustment, electromagnetic coil current modulation, microstructure geometric deformation, and coolant inlet flow scheduling. This effectively improves the spatial accuracy and real-time performance of control response during thermal control regulation, thereby enhancing dynamic adaptability and collaborative execution efficiency under high heat load distribution conditions.

[0052] Specifically, the steps for evaluating the hysteresis of the control response after execution are as follows: Based on the edge thermal control state data before and after the control execution, extract the trigger timestamp when the chip temperature first exceeds the temperature threshold, and monitor the response timestamp when the chip temperature change sequence first shows a stable downward trend. Calculate the difference between the trigger timestamp and the response timestamp to obtain the response time delay, which characterizes the time interval from the execution of the thermal control strategy to the thermal response of chip 2. Use the difference between the chip temperature before and after control as a measure of temperature rise mitigation, multiply it by the sum of the chip heat power density and the minimum term, and then multiply it by the response time delay to obtain the temperature rise hysteresis load response work, which is used to reflect the coordination between temperature rise improvement and response speed. The system employs a combination of control and regulation capabilities. It extracts the inlet and outlet pressure data of the microchannels after regulation and calculates the difference, adding it to the contact pressure difference at the thermal interface after regulation to construct a composite pressure difference index, serving as a measure of fluid mechanical load during the regulation response process. The system divides the chip power consumption after regulation by the sum of the cooling environment reference power and the minimum term to form the load ratio of the heat source to the heat dissipation capacity. Adding one to the result and taking the natural logarithm yields the regulation benefit function, used to characterize the degree to which the regulation method improves the efficiency of heat load absorption. Finally, the system divides the work done by the temperature rise hysteresis load response by the product of the composite pressure difference and the regulation benefit function to obtain the thermal control response hysteresis evaluation value, used to comprehensively evaluate the overall performance of the regulation measures in terms of timeliness, energy efficiency, and cooling load adaptability.

[0053] The specific formula for calculating the thermal control response hysteresis evaluation value is as follows:

[0054] ;

[0055] In the formula, E represents the thermal control response hysteresis evaluation value. This indicates the chip temperature before adjustment. This indicates the chip temperature after adjustment, and Q represents the chip's thermal power density. This indicates a delay in response time. This indicates the pressure difference in the microchannel after regulation. This indicates the pressure difference at the heat interface after adjustment. This indicates the chip power consumption after adjustment, where H represents the reference power of the cooling environment. Indicates a minus term.

[0056] In this embodiment, Table 1 is a data table of thermal control response hysteresis evaluation values, listing the key thermal control parameter values ​​for five high thermal response units during the microchannel magnetorheological cooling control process. These parameters include chip temperature before control, chip temperature after control, chip heat power density, response time delay, microchannel pressure difference, thermal interface contact pressure difference, chip power consumption after control, cooling environment reference power, and thermal control response hysteresis evaluation values. Minor items are uniformly set to... In unit U1, the chip temperature before regulation is 95°C, and after regulation, the chip temperature is 85°C. The chip heat power density is 0.85, the response time delay is 0.45, the microchannel pressure difference is 12, the thermal interface contact pressure difference is 8.0, the chip power consumption after regulation is 85, the cooling environment reference power is 100, and the calculated thermal control response hysteresis evaluation value is 0.31. In unit U2, the chip temperature before regulation is 98°C, and after regulation, the chip temperature is 90°C. The chip heat power density is 0.92, the response time delay is 0.50, the microchannel pressure difference is 15, the thermal interface contact pressure difference is 9.0, the chip power consumption after regulation is 95, the cooling environment reference power is 110, and the thermal control response hysteresis evaluation value is 0.25. In unit U3, the chip temperature before regulation is 96°C, and after regulation, the chip temperature is 88°C. The chip heat power density is 0.88, the response time delay is 0.42, the microchannel... In unit U4, the chip temperature before adjustment is 100℃, and after adjustment it is 92℃. The chip heat power density is 0.95, the response time delay is 0.60, the microchannel pressure difference is 18℃, the thermal interface contact pressure difference is 10.0℃, the chip power consumption after adjustment is 100℃, the cooling environment reference power is 120℃, and the thermal control response hysteresis evaluation value is 0.22. In unit U5, the chip temperature before adjustment is 94℃, and after adjustment it is 87℃. The chip heat power density is 0.90, the response time delay is 0.48, the microchannel pressure difference is 14℃, the thermal interface contact pressure difference is 8.2℃, the chip power consumption after adjustment is 92℃, the cooling environment reference power is 108℃, and the thermal control response hysteresis evaluation value is 0.22.

[0057] Table 1. Evaluation Values ​​of Thermal Control Response Hysteresis

[0058]

[0059] like Figure 3 The figure shows the thermal control response hysteresis evaluation values ​​of five high thermal response units during the thermal control response process, and determines whether to implement control strategy correction. The figure uses a bar chart to indicate the thermal control response hysteresis evaluation values ​​of each high thermal response unit, with a blue dashed line indicating the hysteresis tolerance threshold. The bar color distinguishes the correction status: red indicates correction is required, and green indicates no correction is required. Specifically, the hysteresis evaluation value of U1 exceeds the hysteresis tolerance threshold, requiring correction using an artificial intelligence control strategy; the evaluation values ​​of U2, U3, U4, and U5 are all below the hysteresis tolerance threshold, and are determined to require no correction. Figure 3 This intuitively demonstrates the intelligent decision-making capability of implementing differentiated control strategies based on the thermal control response hysteresis assessment mechanism.

[0060] This implementation scheme constructs a method for evaluating the hysteresis of control response based on edge thermal control state data. It comprehensively quantifies the dynamic feedback process of thermal control after the execution of the control strategy by combining core data such as chip temperature trigger timestamp, response timestamp, chip temperature, chip thermal power density, response time delay, microchannel pressure difference, thermal interface contact pressure difference, chip power consumption, and cooling environment reference power. By calculating the load response work under temperature rise hysteresis, the composite pressure difference, and the control benefit function, the scheme effectively evaluates the timeliness, energy efficiency, and thermal load adaptability of the control actions. This significantly improves the accuracy of judging the control effect and the closed-loop regulation capability under dynamic thermal load scenarios, enhancing the accuracy and adaptability of the control strategy to actual thermal state changes.

[0061] Specifically, if the hysteresis exceeds the limit, the control command is corrected and re-driven to achieve closed-loop control. The specific steps are as follows: Real-time comparison of the thermal control response hysteresis evaluation value and the hysteresis tolerance threshold. When the thermal control response hysteresis evaluation value is less than or equal to the hysteresis tolerance threshold, the current path-level control strategy is maintained. The path-level control strategy includes a joint instruction set of microstructure deformation, electromagnetic excitation control, and inlet flow scheduling. When the thermal control response hysteresis evaluation value is greater than the hysteresis tolerance threshold, a correction operation based on an artificial intelligence control strategy is executed. The artificial intelligence control strategy learns response efficiency and optimizes control variables based on historical control effect samples and current edge thermal control state data: First, the magnetic particle aggregation density of the target flow channel section is increased to enhance the deformability of the local flow channel magnetic response. Then, the ratio of the current composite pressure difference to the pressure difference threshold is calculated. If the ratio is less than a constant, the current flow resistance control is deemed too strong, requiring a reduction in the electromagnetic coil current to weaken the magnetic field excitation intensity and a decrease in the branch pump speed to reduce the coolant flow rate. Otherwise, the current cooling flow capacity is deemed insufficient, requiring an increase in the electromagnetic coil current to enhance the magnetic field response capability and an increase in the branch pump speed to increase the liquid supply rate. Based on the real-time adjustment and changes in magnetic particle aggregation density, electromagnetic coil current, and branch pump speed, a correction control instruction set is generated. The correction control instruction set is constructed at the path level, covering all control commands corresponding to the micropipe node positions, replacing the original path-level control commands, and sending them to the corresponding micropipe nodes to re-drive the execution of electromagnetic excitation modulation, flow channel structure deformation, and inlet flow adjustment operations, thereby achieving a closed-loop control response and enhancing the dynamic adaptability of the control strategy under scenarios of severe thermal load fluctuations.

[0062] In this implementation scheme, by comparing the thermal control response hysteresis evaluation value with the hysteresis tolerance threshold in real time, the magnetic particle aggregation density, electromagnetic coil current, and branch pump speed are dynamically corrected based on an artificial intelligence control strategy, and a corrected control instruction set is generated to re-drive the execution path-level control operation. This method effectively improves the coordination between microchannel pressure difference, thermal interface contact pressure difference, electromagnetic excitation control, and liquid flow regulation, enhances the adaptive capability of the thermal control response, and ensures that key variables such as chip temperature, chip power consumption, coolant temperature, and thermal load gradient can still be accurately and dynamically controlled under conditions of severe thermal load fluctuations. This significantly improves the real-time performance, robustness, and energy efficiency matching of the control response.

[0063] like Figure 2 As shown, the second aspect of the present invention provides an AI-coordinated heat dissipation control device for the edge of the magnetorheological channel of a rack cold plate, comprising: a data acquisition and preprocessing module, a thermal load identification and partitioning module, a hierarchical control strategy generation module, and a control command execution feedback module, wherein: the data acquisition and preprocessing module is used to acquire edge thermal control status data, and perform structural integrity verification, disturbance smoothing, feature alignment, and scale normalization processing on the edge thermal control status data to obtain preprocessed edge thermal control status data; the thermal load identification and partitioning module is used to identify thermal variation units in the chip 2 region based on the preprocessed edge thermal control status data and A thermal load clustering module is constructed to evaluate the cooling regulation priority of thermal variation units and generate thermal load spatial partitions. A hierarchical regulation strategy generation module is used to extract thermal variation units in high-heat regions to construct a set of high thermal response units, evaluate the magnetorheological regulation response intensity of high thermal response units, generate regional regulation strategies hierarchically, and construct a regulation strategy dataset. A control command execution feedback module is used to parse the regulation strategy dataset, generate a path-level control command set, and drive microstructure deformation, electromagnetic excitation, and flow regulation. After execution, the regulation response hysteresis is evaluated. If the hysteresis exceeds the limit, the control command is corrected and re-driven to achieve closed-loop regulation.

[0064] like Figure 4 The diagram illustrates the overall layout of a rack-mounted heat dissipation structure based on a two-phase cold plate configuration. The two-phase cold plate 1 is the main component for heat load regulation, achieving adaptive adjustment of regional heat flux density through an internal magnetorheological fluid flow channel structure. The chip 2 is the primary heat source for edge thermal control status data, and its real-time operating status is crucial for driving the regulation strategy. The coolant inlet pipe 3 is the starting point of the external coolant delivery path, connecting the coolant circulation loop and handling inlet flow scheduling. This structural diagram clearly shows the arrangement of multiple cold plates stacked vertically within the rack structure, with the coolant inlet pipe 3 guiding the coolant into each cold plate layer, achieving layered coverage of the chip 2's heat dissipation requirements. The structural design supports the integrated deployment of the entire process, including thermal variation unit identification, cooling regulation strategy generation, and path-level command issuance.

[0065] like Figure 5 As shown, the controllable configuration layout of the magnetorheological flow channel inside the rack cold plate is illustrated, wherein:

[0066] Two-phase cold plate 1, as a key component integrating heat load bearing and cooling channel, has liquid-gas two-phase cooling capability; electromagnet 4 is used to generate a magnetic field in a designated area to implement aggregation control of magnetic particles and adjust the local microstructure state; micro-channel wall 5 constitutes the basic boundary of the coolant flow channel, and can undergo micro-scale deformation under the action of magnetic field, which is used to dynamically control the flow channel opening and closing state and flow resistance distribution. Figure 5 The array arrangement of electromagnets 4 around the microchannel wall 5 is clearly demonstrated. By precisely driving the microstructure to respond to changes in the magnetic field, flow behavior control at the cooling path level is achieved. Based on the thermophysical properties of the two-phase cold plate 1, this structure makes the magnetorheological control mechanism more real-time, responsive, and spatially flexible, providing an efficient execution basis for edge AI collaborative heat dissipation control strategies.

[0067] In this implementation scheme, an AI-coordinated heat dissipation control device for the edge of the rack cold plate magnetorheological channel is constructed, including a data acquisition and preprocessing module, a heat load identification and partitioning module, a hierarchical control strategy generation module, and a control command execution feedback module. This comprehensively achieves high-precision acquisition and processing of edge thermal control status data. Through structural integrity verification, multi-scale disturbance smoothing, feature alignment, and scale normalization, the stability and usability of the original data are enhanced. Based on this, according to the coolant heat exchange power, thermal interface thermal resistance, and spatial temperature gradient identification, thermal variation units in the chip 2 region are accurately identified, and heat load clustering is constructed to complete the thermal load spatial partitioning. Furthermore, thermal variation units in the high-heat zone are extracted based on magnetorheological... The response evaluation value, combined with the cooling control priority evaluation value, thermal load gradient adjustment factor, and magnetic control driving force factor, accurately generates regional control strategies and constructs a control strategy dataset. Subsequently, the control strategy dataset is parsed to generate a path-level control command set covering micro-region electromagnetic excitation, microstructure deformation, and inlet flow scheduling, which is then issued to micro-pipeline nodes for execution. After the control is completed, based on the comparison between the thermal control response hysteresis evaluation value and the hysteresis tolerance threshold, it is determined whether there is a response hysteresis. If the hysteresis exceeds the standard, the magnetic particle aggregation density, electromagnetic coil current, and branch pump speed are dynamically corrected, and the corrected control commands are regenerated and the control operation closed-loop execution is driven, effectively enhancing the real-time adaptability and target accuracy of the control behavior.

[0068] 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.

[0069] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for AI-coordinated heat dissipation control at the edge of a magnetorheological flow channel on a rack cold plate, characterized in that, Includes the following steps: S1, collect edge thermal control status data, perform structural integrity verification, disturbance smoothing, feature alignment and scale normalization on the edge thermal control status data, and obtain preprocessed edge thermal control status data; S2, based on the preprocessed edge thermal control state data, identifies thermal variation units in the chip region and constructs thermal load clusters, evaluates the cooling control priority of thermal variation units, and generates thermal load spatial partitions. The specific steps for evaluating the cooling control priority of the thermal variation unit and generating the thermal load space partition are as follows: Multiply the chip's thermal power density by the difference between the chip temperature and the coolant inlet temperature to obtain the thermal load temperature rise; square the difference between the coolant outlet temperature and the inlet temperature and multiply it by the sum of the thermal interface thermal resistance and the minimum term to obtain the comprehensive heat dissipation difficulty factor; divide the thermal load temperature rise by the comprehensive heat dissipation difficulty factor to obtain the chip's thermal stress intensity; divide the chip's power consumption by the sum of the cooling environment reference power and the minimum term, add one to the result and take the natural logarithm to obtain the heat source load adjustment factor; multiply the chip's thermal stress intensity by the heat source load adjustment factor to obtain the cooling control priority evaluation value. The cooling control priority evaluation values ​​corresponding to each thermal variation unit within each heat load cluster are summarized, and the mean value is calculated as the cluster representative value. The cluster representative value and the heat load partition threshold are compared in real time: when the cluster representative value is greater than the heat load partition threshold, the heat load cluster is marked as a high heat zone, and the other heat load clusters are marked as low heat zones. S3. Extract thermal variation units from high-thermal-region areas to construct a set of high-thermal-response units, evaluate the magnetorheological modulation response intensity of high-thermal-response units, generate regional modulation strategies in a hierarchical manner, and construct a modulation strategy dataset. S4 analyzes the control strategy dataset, generates a path-level control instruction set, and drives microstructure deformation, electromagnetic excitation, and flow regulation. After execution, the control response lag is evaluated. If the lag exceeds the limit, the control command is corrected and the drive is restarted to achieve closed-loop control.

2. The rack cold plate magnetorheological channel edge AI collaborative heat dissipation control method according to claim 1, characterized in that: The specific steps for collecting edge thermal control status data and performing structural integrity verification, perturbation smoothing, feature alignment, and scale normalization on the edge thermal control status data to obtain preprocessed edge thermal control status data are as follows: Collect edge thermal control status data, including chip temperature, coolant inlet temperature, coolant outlet temperature, chip thermal power density, coolant flow rate, coolant density, coolant specific heat capacity, electromagnetic coil current, magnetic field strength, chip power consumption, thermal interface contact pressure, and microchannel pressure. The edge thermal control status data is structurally verified using a dynamic boundary consistency detection algorithm to remove abnormal data points caused by acquisition anomalies. A multi-scale trend filtering algorithm is used to reduce local disturbances in the edge thermal control status data, smoothing high-frequency disturbance waveforms caused by load mutations and control drives. A distributional normalization renormalization algorithm is used to align the features of the edge thermal control status data, unifying the numerical distribution. Finally, an interval linear mapping function is used to normalize the scale of the edge thermal control status data.

3. The rack cold plate magnetorheological channel edge AI collaborative heat dissipation control method according to claim 1, characterized in that: The specific steps for identifying thermal variation units in the chip region and constructing thermal load clusters based on the preprocessed edge thermal control state data are as follows: Based on the preprocessed edge thermal control status data, a coolant heat exchange power algorithm based on energy conservation is invoked. The coolant flow rate, coolant inlet temperature, and outlet temperature are input, and the coolant density and specific heat capacity are introduced as constant factors to perform heat exchange power calculation and obtain the cooling environment reference power. The difference between the chip temperature and the coolant inlet temperature is calculated and divided by the chip thermal power density to obtain the thermal interface thermal resistance. A thermal load expression vector containing the cooling environment reference power, thermal interface thermal resistance, and edge thermal control status data is constructed. Extract the spatial index of the corresponding sensor from the edge thermal control status data, and establish the mapping relationship between the thermal load expression vector and the cold plate position; set a sliding window, calculate the local gradient of chip temperature in each window, and select data points whose local gradient exceeds the thermal gradient identification threshold as thermal variation units; connect the thermal variation units in adjacent steps to form thermal load clusters.

4. The rack cold plate magnetorheological channel edge AI collaborative heat dissipation control method according to claim 1, characterized in that: The specific steps for constructing a set of high thermal response units by extracting thermal variation units from high-thermal regions and evaluating the magnetorheological modulation response intensity of the high thermal response units are as follows: Extract all thermal variation units from the thermal load cluster labeled as high-heat zone to construct a set of high thermal response units; calculate the difference based on the inlet and outlet pressure data of the micro-channel at the corresponding location of the high thermal response unit to obtain the micro-channel pressure difference; Extract the contact pressure data at the start and end positions of the thermal variation units in the thermal interface contact pressure array, and calculate the difference to obtain the thermal interface contact pressure difference; The relative intensity of cooling regulation is obtained by dividing the priority evaluation value of cooling regulation by the cluster representative value; the power consumption and the difference between chip temperature and coolant inlet temperature are multiplied to construct the power consumption temperature rise coupling factor; the difference between coolant outlet temperature and inlet temperature is multiplied by the sum of thermal interface thermal resistance and minimum term to obtain the comprehensive heat dissipation difficulty factor; the power consumption temperature rise coupling factor is divided by the comprehensive heat dissipation difficulty factor to obtain the heat load gradient adjustment factor; the pressure difference of microchannel and the contact pressure difference of thermal interface are added together and divided by the product of electromagnetic coil magnetic field strength and current plus minimum term to obtain the magnetic control adjustment driving force factor. The magnetorheological response evaluation value is obtained by multiplying the relative intensity of cooling regulation, the thermal load gradient adjustment factor, and the magnetic control driving force factor.

5. The rack cold plate magnetorheological channel edge AI collaborative heat dissipation control method according to claim 1, characterized in that: The specific steps for generating hierarchical regional regulation strategies and constructing the regulation strategy dataset are as follows: By comparing the magnetorheological response evaluation value with the control grading threshold, a graded control strategy is generated for the region: when the magnetorheological response evaluation value is less than or equal to the first-level control grading threshold, the high thermal response unit is divided into a stable maintenance region to maintain the original structure and electromagnetic excitation of the current microchannel; when the magnetorheological response evaluation value is greater than the first-level control grading threshold but less than the second-level control grading threshold, it is divided into a collaborative compensation region to increase the magnetic particle aggregation density and shrink the cross-sectional area of ​​the flow channel. When the magnetorheological response evaluation value is greater than or equal to the secondary control grading threshold, it is divided into an active intervention zone unit, which executes the maximum power excitation of the electromagnetic coil, constructs a high resistance chain micro-pipe structure, and increases the coolant inlet flow rate. Extract the magnetorheological response evaluation value, control strategy and spatial index of each high thermal response unit to construct a control strategy dataset.

6. The rack cold plate magnetorheological channel edge AI collaborative heat dissipation control method according to claim 1, characterized in that: The specific steps for generating a path-level control instruction set from the analytical control strategy dataset to drive microstructure deformation, electromagnetic excitation, and flow regulation are as follows: The system receives the control strategy dataset, calls the hardware control rule set, and maps the control strategy and spatial index to a set of path-level control instructions, including micro-area electromagnetic excitation control instructions, microstructure deformation driving instructions, and inlet flow scheduling instructions. The control instructions are then sent to the corresponding micro-pipe nodes to drive the execution of local electromagnetic excitation, electromagnetic coil current modulation, microstructure deformation, and inlet flow regulation operations.

7. The rack cold plate magnetorheological channel edge AI collaborative heat dissipation control method according to claim 1, characterized in that: The specific steps for evaluating the lag in the control response after implementation are as follows: Based on the edge thermal control state data before and after the regulation is executed, the trigger timestamp when the chip temperature first exceeds the temperature threshold and the response timestamp when the chip temperature first shows a stable downward trend are extracted. The difference is calculated to obtain the response time delay. The difference between the chip temperature before and after regulation is multiplied by the sum of the chip thermal power density and the minimum term, and then multiplied by the response time delay to obtain the temperature rise hysteresis load response power. The differential pressure of the microchannel after regulation is added to the differential pressure of the thermal interface to obtain the composite differential pressure. The power consumption of the chip after regulation is divided by the sum of the reference power of the cooling environment and the minimum term. The ratio is increased by one and then the natural logarithm is taken to obtain the regulation benefit function. The power consumption of the load response with temperature rise hysteresis is divided by the product of the composite differential pressure and the regulation benefit function to obtain the thermal control response hysteresis evaluation value.

8. The rack cold plate magnetorheological channel edge AI collaborative heat dissipation control method according to claim 1, characterized in that: The specific steps for correcting the control command and re-driving execution if the lag exceeds the limit to achieve closed-loop control are as follows: The thermal control response hysteresis assessment value and the hysteresis tolerance threshold are compared in real time. When the thermal control response hysteresis assessment value is less than or equal to the hysteresis tolerance threshold, the current path-level control strategy is maintained. When the thermal control response hysteresis assessment value is greater than the hysteresis tolerance threshold, a correction operation based on artificial intelligence control strategy is executed: increase the magnetic particle aggregation density of the target flow channel section; Calculate the ratio of the composite pressure difference to the pressure difference threshold. If the ratio is less than a constant, reduce the electromagnetic coil current and decrease the branch pump speed; otherwise, increase the electromagnetic coil current and increase the branch pump speed. Based on the changes in magnetic particle aggregation density, electromagnetic coil current, and branch pump speed, a modified control instruction set is generated, which replaces the original instructions and is sent to the corresponding micro-pipeline nodes to re-execute the control operation, thereby achieving a closed-loop control response.

9. A rack-mounted cold plate magnetorheological channel edge AI-coordinated heat dissipation control device, used to implement the rack-mounted cold plate magnetorheological channel edge AI-coordinated heat dissipation control method as described in any one of claims 1-8, characterized in that: include: The module comprises a data acquisition and preprocessing module, a heat load identification and zoning module, a hierarchical control strategy generation module, and a control command execution feedback module, among which: The data acquisition and preprocessing module is used to acquire edge thermal control status data, and perform structural integrity verification, disturbance smoothing, feature alignment and scale normalization on the edge thermal control status data to obtain preprocessed edge thermal control status data. The thermal load identification and partitioning module is used to identify thermal variation units in the chip region and construct thermal load clusters based on preprocessed edge thermal control status data, evaluate the cooling control priority of thermal variation units, and generate thermal load spatial partitions. The hierarchical control strategy generation module is used to extract thermal variation units in high-heat regions to construct a set of high-thermal-response units, evaluate the magnetorheological control response intensity of high-thermal-response units, generate regional control strategies hierarchically, and construct a control strategy dataset. The control command execution feedback module is used to parse the control strategy dataset, generate a path-level control command set, drive microstructure deformation, electromagnetic excitation and flow regulation; after execution, it evaluates the control response hysteresis, and if the hysteresis exceeds the standard, it corrects the control command and re-drives the execution to achieve closed-loop control.

Images