AI active flow control single-phase immersed heat dissipation system and data center

Abstract

The invention provides an AI active flow control single-phase immersed heat dissipation system and a data center. The system comprises a micro thermoelectric effect sensor array, a distributed acoustic transducer array, a micro electro mechanical system pump valve network and a system controller, wherein the micro thermoelectric effect sensor array and the distributed acoustic transducer array are arranged in a cabinet, and the micro electro mechanical system pump valve network is integrated in a flow channel. The controller cooperatively controls the acoustic transducer array to emit coded sound waves and receive echoes, and reconstructs a three-dimensional flow velocity field in the cabinet based on the computational tomography principle; meanwhile, reading of the thermoelectric effect sensor array is obtained and processed into a three-dimensional heat flow vector field; and finally, generating a control instruction based on the two three-dimensional fields, and independently driving each pump valve in the pump valve network of the micro-electro-mechanical system to perform targeted heat dissipation. The core problems of information asymmetry and control granularity mismatch in an existing immersed heat dissipation technology are fundamentally solved, and conversion from macroscopic passive heat dissipation to microcosmic active flow control is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of heat dissipation technology of high-performance computing equipment, and in particular to an AI active flow control single-phase immersion heat dissipation system and a data center. BACKGROUND

[0002] With the rapid development of artificial intelligence, high-performance computing and other technologies, single-chip power consumption has reached the level of kilowatts, and traditional air cooling heat dissipation methods have been unable to meet the heat dissipation needs. Single-phase full immersion liquid cooling technology has become the mainstream solution due to its high-efficiency heat dissipation capability. The principle is to immerse the heat-generating electronic components in non-conductive cooling liquid, and to remove heat through liquid circulation. However, the existing technology has the following defects: information asymmetry, macro control system cannot real-time perceive the micro heat field distribution, fluid dynamics and local hot spot conditions inside the cabinet; control granularity mismatch, the existing technology relies on a single total circulating pump for flow regulation, which cannot achieve precise and differentiated control of thousands of heat sources, high-heat areas are not sufficiently cooled, and low-heat areas waste energy; and there is a problem of power consumption and energy efficiency deterioration.

[0003] Therefore, there is an urgent need for a heat dissipation system that can realize micro perception, precise control, intelligent decision-making and safety guarantee integration to solve the above problems. SUMMARY

[0004] The primary purpose of the present application is to provide an AI active flow control single-phase immersion heat dissipation system to solve the problems of "information asymmetry" and "control granularity mismatch" in the prior art.

[0005] Another purpose of the present application is to provide a data center that, by deploying the above heat dissipation system, realizes global collaborative optimization and ultimate energy efficiency of computing and thermal resources.

[0006] In a first aspect, the present application provides an AI active flow control single-phase immersion heat dissipation system, comprising: a micro-thermoelectric effect sensor array arranged in a cabinet; a distributed acoustic transducer array arranged in the cabinet; a micro-electromechanical system pump valve network integrated in the flow channel beside the heat-generating element; a system controller connected to the micro-thermoelectric effect sensor array, the acoustic transducer array and the micro-electromechanical system pump valve network, respectively; wherein the system controller is configured to perform the following steps: reconstructing a three-dimensional flow velocity field inside the cabinet based on the echo signals received from the acoustic transducer array; reconstructing a three-dimensional heat flow vector field inside the cabinet based on the readings obtained from the micro-thermoelectric effect sensor array; generating control instructions based on the three-dimensional flow velocity field and the three-dimensional heat flow vector field; driving each pump valve in the micro-electro-mechanical system pump valve network independently according to the control instructions to perform targeted heat dissipation on the heat generating element.

[0007] In some embodiments, the step of reconstructing the three-dimensional flow velocity field in the cabinet includes: injecting a thermal pulse into the cooling liquid through an intelligent skin arranged on the inner wall of the cabinet; tracking the movement trajectory of the thermal pulse through the acoustic transducer array; calculating the flow velocity, diffusion coefficient, and turbulence intensity of the cooling liquid in a specific region based on the tracking results.

[0008] In some embodiments, the micro-electro-mechanical system pump valve network includes a plurality of piezoelectric film pumps and electrostatic micro valves; The step of driving each pump valve in the micro-electro-mechanical system pump valve network independently according to the control instructions includes generating focused jets or controllable vortices inside the cabinet by cooperatively controlling the pump valves in different regions of the micro-electro-mechanical system pump valve network.

[0009] In some embodiments, the system further includes a local microcontroller; the local microcontroller is arranged integrally with at least one pump valve in the micro-electro-mechanical system pump valve network and directly communicates with the adjacent micro-thermoelectric effect sensor array; The local microcontroller is configured to directly drive the at least one pump valve to perform an impulsive cooling program when a temperature jump signal is received from the adjacent micro-thermoelectric effect sensor array.

[0010] In some embodiments, a digital twin model runs in the system controller; the system controller is further configured to perform the following steps: continuously compare the sensor data acquired from the micro-thermoelectric effect sensor array and the acoustic transducer array with the predicted values of the digital twin model; based on the comparison results, calibrate the viscosity parameter of the cooling liquid in the digital twin model online through a Kalman filter.

[0011] In some embodiments, the system further includes a safety filtering module connected between the instruction output end of the system controller and the micro-electro-mechanical system pump valve network; the system controller is further configured to, after generating the control instructions, send the control instructions to the safety filtering module before driving the pump valve network according to the control instructions; the safety filtering module is configured to perform real-time verification and correction on the control instructions based on the control barrier function theory to output safety-verified instructions, ensuring that the system state is within the dynamic safety domain.

[0012] In some embodiments, the system further comprises a security authentication module configured to perform the following steps: injecting a challenge signal into the coolant through a micro-pump-valve randomly selected from the MEMS pump-valve network; capturing corresponding physical responses through the micro-thermoelectric effect sensor array and / or the acoustic transducer array; matching the captured physical responses against expected responses; and upon successful matching, allowing the security-verified instructions outputted by the security filtering module to be executed.

[0013] In some embodiments, the system controller is further configured to perform the following steps: compressing system operation data into a low-dimensional state embedding vector through a deep auto-encoder neural network; calculating geometric distance between the real-time state vector and a pre-stored healthy state manifold; and based on the geometric distance, quantifying the degree of system abnormality and issuing a fault warning.

[0014] In some embodiments, the system controller is further configured to perform the following steps: calculating and outputting a heat dissipation capability profile of the cabinet based on real-time operation data acquired from the micro-thermoelectric effect sensor array and the acoustic transducer array, the heat dissipation capability profile including instantaneous heat dissipation peak, heat dissipation margin, and heat dissipation efficiency; sending the heat dissipation capability profile to a data center level scheduling platform.

[0015] In a second aspect, the present application provides a data center comprising a plurality of computer cabinets, characterized in that at least one of the computer cabinets is equipped with the AI active flow control single-phase immersion cooling system of any one of the first aspect, and the data center further comprises a scheduling platform configured to perform joint task scheduling in three dimensions of computing resources, network resources, and heat dissipation resources based on heat dissipation capability profiles received from each cabinet.

[0016] The beneficial effects of the present application are: different from the prior art, the AI active flow control single-phase immersion cooling system provided by the present application comprises a micro-thermoelectric effect sensor array and a distributed acoustic transducer array arranged in a cabinet, a micro-electromechanical system pump valve network integrated in a flow channel, and a system controller. The controller controls the acoustic transducer array to emit coded sound waves and receive echoes, reconstructs a three-dimensional flow field in the cabinet based on the computed tomography principle; at the same time, the readings of the thermoelectric effect sensor array are obtained and processed into a three-dimensional heat flow vector field; finally, based on the two three-dimensional fields, control instructions are generated to independently drive each pump valve in the micro-electromechanical system pump valve network for targeted heat dissipation. The present application fundamentally solves the core problem of information asymmetry and control granularity mismatch in the existing immersion cooling technology, and realizes the transformation from macroscopic passive cooling to microscopic active flow control. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor. Among them: Figure 1 is a structural schematic diagram of an embodiment of the AI active flow control single-phase immersion cooling system provided by the present application; Figure 2 is a flow chart of an embodiment of the method for executing targeted heat dissipation control by the system controller provided by the present application; Figure 3 is a working flowchart of an embodiment of the security authentication module provided by the present application. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the drawings of the embodiments of the present application. It can be understood that the specific embodiments described here are only used to explain the present application, not to limit the present application. In addition, it should be noted that only parts related to the present application are shown in the drawings, not all structures. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0019] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0020] With the rapid development of technologies such as artificial intelligence and high-performance computing, the power consumption of single chips has approached the kilowatt level. Traditional air-cooling methods can no longer meet the heat dissipation requirements. Single-phase full-immersion liquid cooling technology has become the mainstream solution due to its high-efficiency heat dissipation capabilities. Its principle is to completely immerse the heat-generating electronic components in a non-conductive coolant, and remove heat through liquid circulation. However, existing technologies have the following drawbacks: information asymmetry, as the macro-control system cannot perceive the microscopic thermal field distribution, fluid dynamics, and local hot spots inside the cabinet in real time; control granularity mismatch, as existing technologies rely on a single main circulation pump for flow regulation, which cannot achieve precise and differentiated control of thousands of heat points, resulting in insufficient cooling of high-heat areas and wasted energy in low-heat areas; and issues with power consumption and energy efficiency degradation.

[0021] Therefore, the present invention aims to solve the aforementioned core technical challenges, namely: to construct an AI-driven active flow control single-phase immersive heat dissipation system capable of integrating microscopic sensing, precise control, intelligent decision-making, and safety assurance. See the following embodiments for details.

[0022] Firstly, this application provides an AI-driven active flow control single-phase immersion cooling system. In some embodiments, see Figure 1 , Figure 1 This is a schematic diagram of an embodiment of the AI ​​active flow control single-phase immersion cooling system provided by the present invention. The system includes: a coolant cabinet 101; a miniature thermoelectric effect sensor array 102 arranged within the cabinet; an acoustic transducer array 103 distributed within the cabinet; a microelectromechanical system (MEMS) pump-valve network 104 integrated in the bypass channel of the heating element 106; and a system controller 105, which is signal-connected to the miniature thermoelectric effect sensor array 102, the acoustic transducer array 103, and the MEMS pump-valve network 104.

[0023] Specifically: Rack 101: As the physical container of the system, it contains non-conductive coolant to completely submerge the heat-generating components 106 such as the server motherboard, CPU, and GPU.

[0024] Miniature thermoelectric effect sensor array 102: This array is arranged within the cabinet 101, integrated in high density on the surface of the package of the heat generating element 106 (only a part is schematically shown in the figure). The implementation basis of this array is the Seebeck effect, i.e. the direct measurement of the vector direction and intensity of the heat flow, instead of a single temperature scalar value, by sensing the thermoelectric potential gradient generated due to the micro-zone temperature difference.

[0025] Acoustic transducer array 103: This array is distributedly arranged within the cabinet 101 (only a part is schematically shown in the figure). The functional basis of this array is the physical propagation characteristics of acoustic waves, which is composed of a plurality of miniature acoustic transducers, capable of emitting a designed acoustic signal and receiving the echo formed by the reflection and scattering of the acoustic signal between the internal structures of the cabinet.

[0026] Micro-electro-mechanical system pump-valve network 104: This network is integrated in the flow channel beside the heat generating element 106 (only a part is schematically shown in the figure). The physical basis of this network is micro-electro-mechanical system technology, which contains a large number of miniaturized piezoelectric film pumps and electrostatic micro valves, each of which can be independently controlled.

[0027] System controller 105: The system controller 105 is signal connected with the miniature thermoelectric effect sensor array 102, the acoustic transducer array 103 and the micro-electro-mechanical system pump-valve network 104 respectively. The configuration basis of this controller is the control method defined by the following steps, used to coordinate the work of the whole system.

[0028] Referring to Figure 2 , Figure 2 is an embodiment of the method for the system controller provided by the present application to execute targeted heat dissipation control. The system controller is configured to execute the following steps: Step S1: Based on the echo signal received from the acoustic transducer array, reconstruct the three-dimensional flow velocity field in the cabinet; Step S2: Based on the readings obtained from the miniature thermoelectric effect sensor array, reconstruct the three-dimensional heat flow vector field in the cabinet; Step S3: Based on the three-dimensional flow velocity field and the three-dimensional heat flow vector field, generate control instructions; Step S4: According to the control instructions, independently drive each pump valve in the micro-electro-mechanical system pump-valve network to perform targeted heat dissipation on the heat generating element.

[0029] In this embodiment, first, in step S1, the system controller 105 drives the acoustic transducer array 103 to work, reconstructing the three-dimensional flow velocity field inside the cabinet based on the echo signals it receives. This step utilizes the principles of computed tomography and Doppler shift analysis, aiming to clearly "see" the flow pattern of the coolant in three-dimensional space, including its speed, direction, and possible existence of vortex or dead zone. Next, in step S2, the system controller 105 acquires the readings of the micro-thermoelectric effect sensor array 102 in parallel, and reconstructs the three-dimensional heat flow vector field inside the cabinet accordingly. This step solves the inverse problem of the heat conduction equation, inverting the heat source distribution inside the chip from the surface heat flow measurement, thus accurately "locating" the places where heat is generated and gathered.

[0030] After obtaining the "flow field" and "heat field" these two key physical fields, step S3 follows. The system controller 105 fuses the above-mentioned three-dimensional flow velocity field and three-dimensional heat flow vector field, forming a comprehensive system state panoramic view. Based on this fused panoramic view, the system generates a set of high-dimensional control instructions through intelligent algorithms, which precisely define the actions each micro-actuator needs to perform. Finally, in step S4, the control instructions are issued to the micro-electromechanical system pump-valve network 104. The system converts the instructions into physical actions by independently driving each pump-valve in the network. For example, by cooperatively controlling the pump-valve array in a specific area, a focused jet is generated to impact cooling at high-temperature hotspots, or a controllable vortex is created to enhance heat exchange efficiency, thus achieving precise and dynamic targeted heat dissipation for the heat-generating components.

[0031] In some embodiments, the step of reconstructing the three-dimensional flow velocity field inside the cabinet includes: Injecting a thermal pulse into the coolant through the intelligent skin set on the inner wall of the cabinet; Tracking the movement trajectory of the thermal pulse through the acoustic transducer array; Calculating the flow velocity, diffusion coefficient, and turbulence intensity of the coolant in a specific area based on the tracking results.

[0032] In this embodiment, first, a thermal pulse is injected into the cooling liquid by the intelligent skin provided on the inner wall of the cabinet. The intelligent skin is composed of micro piezoelectric and thermoelectric composite materials, and has both sensing and excitation functions. The system controller 105 selects a specific skin unit to act as a micro heat source, injecting a very low energy and very short duration thermal pulse into the cooling liquid. This thermal pulse will locally and instantaneously heat a small volume of cooling liquid, forming a "tracer cluster" with temperature heterogeneity. The advantage of this method is that the excitation signal (thermal pulse) is controllable and known, laying the foundation for subsequent high signal-to-noise ratio detection. Subsequently, the acoustic transducer array 103 is used to track the movement trajectory of the thermal pulse. At the same time of injecting the thermal pulse, the system controller 105 drives the aforementioned acoustic transducer array 103 into high-frequency scanning mode. Since the heated "tracer cluster" has a slight difference in sound speed and density from the surrounding cold cooling liquid, it will form a highlighted "signal point" in acoustic imaging that can be identified and moved over time. The acoustic transducer array 103 continuously emits coded acoustic waves and receives echoes to lock and track the spatial position changes of this "signal point" in multiple consecutive acoustic images with millisecond-level time resolution, thereby recording its complete movement trajectory. Finally, based on the tracking results, the flow velocity, diffusion coefficient and turbulence intensity of the cooling liquid in a specific region are calculated. That is, the system controller processes and analyzes the recorded movement trajectory data, including flow velocity calculation and diffusion coefficient and turbulence intensity calculation. The flow velocity calculation: according to the displacement vector of the "tracer cluster" in unit time, the instantaneous flow velocity vector of the cooling liquid in the region can be directly and unambiguously calculated. Diffusion coefficient and turbulence intensity calculation: by analyzing the morphological evolution and diffusion rate of the "tracer cluster" during movement, the turbulence intensity and thermal diffusion coefficient of the fluid in the region can be further calculated. For example, the more irregular the trajectory and the faster the diffusion, the higher the turbulence intensity.

[0033] Therefore, this embodiment realizes direct and high-precision measurement of local flow field parameters through the coherent technical chain of "active injection of characteristic markers (thermal pulse) → acoustic precise tracking → trajectory data analysis". This method effectively overcomes the interference of complex background flow field noise on passive acoustic detection, and provides key and reliable local data calibration and supplement for the "three-dimensional flow velocity field reconstruction" in step S1, thereby jointly constructing a more accurate global flow field image.

[0034] In some embodiments, the micro-electro-mechanical system pump valve network includes a plurality of piezoelectric film pumps and electrostatic micro valves; The step of independently driving each pump valve of the micro-electro-mechanical system pump valve network according to the control instruction includes: By cooperatively controlling the pump valves in different regions of the micro-electro-mechanical system pump valve network, focused jets or controllable vortices are generated inside the cabinet.

[0035] In this embodiment, the microelectromechanical system (MEMS) pump-valve network includes multiple piezoelectric diaphragm pumps and electrostatic microvalves. These micro-actuators, manufactured based on MEMS technology, are the physical basis for achieving precise fluid control. Their micrometer-scale size and millisecond-level response speed make it possible to decompose coolant flow into independently controllable microflow streams. The step of independently driving each pump and valve in the MEMS pump-valve network according to the control command embodies the advanced function of active flow field shaping. By coordinating the control of pumps and valves in different areas of the MEMS pump-valve network, focused jets or controllable vortices are generated within the cabinet.

[0036] in: How to generate a focused jet: When the system controller identifies a high-intensity heat source in a heat-generating component (such as a GPU core) based on the fused physical field information, it generates a corresponding coordinated control command. This command synchronously drives multiple piezoelectric thin-film pumps directly above the heat source to operate in high-power, in-phase pulse mode. The multiple microfluidic streams generated by these pumps converge and superimpose in space, ultimately forming a high-speed, concentrated focused jet that vertically impacts the chip surface. This jet can effectively penetrate the thermal boundary layer of the chip surface, achieving extremely high instantaneous heat transfer efficiency, representing a "precision strike" solution for localized hotspots.

[0037] Methods for generating controllable vortices: For scenarios requiring enhanced mixing or clearing of stagnant areas, the system controller employs an alternative driving strategy. It will differ... A controlled asymmetric flow drives a group of pumps and valves adjacent to each other in a specific area, for example, making one pump on one side active while the valve on the other side is partially closed. This controlled asymmetric flow introduces shear forces into the fluid, thereby actively generating a stable, rotating, controllable vortex. The center of rotation and intensity of this vortex can be precisely controlled, and its powerful entrainment and mixing effects can efficiently remove heat trapped in the corners of complex structures and disrupt the original stagnant boundary layer, achieving uniform heat dissipation.

[0038] In some embodiments, the system further includes a local microcontroller; The local microcontroller is integrated with at least one pump valve in the microelectromechanical system pump valve network and communicates directly with the adjacent micro thermoelectric effect sensor array. The local microcontroller is configured to directly drive the at least one pump valve to perform an impact cooling procedure when it receives a signal of a sudden temperature rise from the adjacent micro thermoelectric effect sensor array.

[0039] In this embodiment, to cope with the rapid temperature rise caused by the sudden surge in the instantaneous power of the computing chip, the system introduces a local microcontroller, which constitutes a super-fast response, locally autonomous reflex arc. The local microcontroller is integrated with at least one pump valve in the micro-electromechanical system pump valve network 104 and directly communicates with the adjacent micro-thermoelectric effect sensor array 102. The local microcontroller is a dedicated, computationally resource-reduced control chip that is physically integrated (e.g., co-packaged on the same small circuit board) with one or a small cluster of micro-electromechanical system pump valves, forming a distributed intelligent execution node. In terms of signal connection, it bypasses the system controller 105 and establishes a communication link directly with the micro-thermoelectric effect sensor array that is physically adjacent to it through a dedicated high-speed data channel. This layout ensures the shortest path from the perception signal to the execution instruction. The local microcontroller is configured to directly drive the at least one pump valve to execute the impulsive cooling program when it receives a temperature surge signal from the adjacent micro-thermoelectric effect sensor array. The local microcontroller has pre-stored simple judgment logic and execution program, the core of which is to continuously monitor the readings from the directly connected sensors. When the micro-thermoelectric effect sensor array detects that the temperature of the heat-generating element (such as a CPU core) adjacent to it rises sharply in a very short time (e.g., hundreds of microseconds) and exceeds the pre-set safety threshold, it will immediately send a temperature surge signal to the local microcontroller. Upon receiving this signal, the local microcontroller does not go through the upper-level system controller for complex global decision-making, but immediately triggers a pre-set, minimalist impulsive cooling program. The content of this program is to directly drive the micro-electromechanical system pump valve integrated with it to run at maximum or near-maximum power at a moment, and to direct a large amount of cooling liquid in the form of impulsive jet flow directly to the temperature surge point.

[0040] In some embodiments, a digital twin model is running in the system controller 105; the system controller is further configured to perform the following steps: Continuously compare the sensor data obtained from the micro-thermoelectric effect sensor array and the acoustic transducer array with the predicted values of the digital twin model; Based on the comparison results, calibrate the cooling liquid viscosity parameter in the digital twin model online through a Kalman filter.

[0041] In this embodiment, to ensure that the digital twin model in the system controller can accurately simulate the physical reality for a long time, The system introduces an online self-calibration mechanism based on real-time data. This mechanism ensures the continuous synchronization between the digital world and the physical world, and is the cornerstone of the system's accurate prediction and long-term reliable operation. The system controller runs a digital twin model inside; the digital twin model is a high-fidelity physical model driven by the Navier-Stokes equation and the heat conduction equation. During system operation, the system controller processes two-way data in parallel: one is the three-dimensional heat flow vector field and three-dimensional flow field data obtained in real time from the micro-thermoelectric effect sensor array and the acoustic transducer array; the other is the predicted value of the corresponding physical field calculated by the digital twin model under the same boundary conditions and current control instructions. The system continuously aligns and compares the two-way data in space and time, and this comparison process directly generates a bias data set representing the difference between the model prediction and the real measurement. Based on the bias, the cooling liquid viscosity parameter in the digital twin model is calibrated online through a Kalman filter. The system inputs this bias into a Kalman filter. As an optimal estimation algorithm, the filter treats the digital twin model itself as a dynamic system. The viscosity of the cooling liquid is defined as the key state variable to be estimated by the filter. Based on the input bias, the filter continuously and recursively updates the optimal estimate of the viscosity value and feeds back the updated parameter to the digital twin model in real time, replacing its original viscosity parameter.

[0042] In some embodiments, the system further comprises a safety filtering module connected between the instruction output of the system controller and the micro-electro-mechanical system pump-valve network; The system controller is further configured to, after generating the control instructions, send the control instructions to the safety filtering module before driving the pump-valve network according to the control instructions; The safety filtering module is configured to perform real-time verification and correction on the control instructions based on control barrier function theory to output safety-verified instructions, ensuring that the system state is within the dynamic safety domain.

[0043] In this embodiment, to ensure the absolute safety of the AI-driven single-phase immersion cooling system under any working condition and prevent aggressive strategies output by the intelligent algorithm from causing system overheating or instability, the system introduces a safety filtering module with the highest priority in the control loop. It is connected between the instruction output of the system controller and the micro-electro-mechanical system pump-valve network; The safety filter module is a standalone hardware logic unit or a protected software process, which is physically or logically concatenated in the command downlink channel of the system controller. All commands issued by the system controller, intended to drive the micro-electro-mechanical system pump-valve network, must first flow through this module, constituting an unskippable safety checkpoint.

[0044] Based on this, the system controller is further configured to, after generating the control instruction, send the control instruction to the safety filter module before driving the pump-valve network according to the control instruction. This step defines the timing of safety intervention. It explicitly defines the complete control chain as: the system controller first generates the theoretically optimal control instruction based on its intelligent algorithm (such as reinforcement learning, model predictive control), but before this instruction is actually executed, the system will forcibly send a copy of it to the safety filter module for final verification. This ensures that safety review occurs at the key node after decision-making and before execution.

[0045] Among them, the safety filter module is configured to perform real-time verification and correction on the control instruction based on the control barrier function theory to output the safety-verified instruction, ensuring that the system state is within the dynamic safety domain. Specifically, the safety filter module internally defines a dynamic safety domain for the system based on the mathematical tool of control barrier function. The safety domain is composed of all key safety constraints (such as the maximum junction temperature of each chip, the maximum pressure of the flow channel), and its boundary is not a fixed value, but a conservative state set dynamically calculated by combining real-time uncertainty information (such as temperature measurement error range) from the multi-modal sensing system. Finally, for each control instruction received, the module will use the mathematical conditions of the control barrier function to perform forward-looking verification. It will quickly predict whether all possible states of the system in the near future will always stay within the dynamic safety domain if the current system state executes this instruction. If the verification is safe: the instruction is released as is, becoming a safety-verified instruction.

[0046] If there is a risk: the module will not simply reject the instruction, but will immediately solve a quadratic programming problem to find a safe alternative instruction that is closest to the original instruction. This process is called minimal intervention correction. Finally, only this safe alternative instruction will be output as a safety-verified instruction to the micro-electro-mechanical system pump-valve network for execution.

[0047] In some embodiments, the system further includes a safety authentication module, see Figure 3 , Figure 3 is an embodiment of the safety authentication module provided by the application. The safety authentication module is configured to perform the following steps: Step S10: Injecting a challenge signal into the coolant through a micro-pump-valve randomly selected from the MEMS pump-valve network; Step S20: Capturing corresponding physical responses through the micro-thermoelectric effect sensor array and / or the acoustic transducer array; Step S30: Generating corresponding expected responses by simulating the propagation and evolution of the challenge signal in the coolant; Step S40: Matching the captured physical responses against the expected responses; Step S50: Allowing the execution of the security-checked instructions outputted by the security filter module upon successful matching.

[0048] In this embodiment, to cope with the potential risk of cyber-attack, the system ensures the integrity and authenticity of the control instructions by Following the security filtering module, a physical-based intrinsic security authentication module is introduced. This module utilizes the complex physical dynamics of the heat dissipation system itself as authentication credentials, realizing a coherent, five-step "physical identity" verification process. First, in step S10, a challenge signal is injected into the coolant through a micro-pump valve randomly selected from the microelectromechanical system's pump-valve network. This step is the starting point of authentication, aiming to proactively create an unpredictable physical interaction. The security authentication module randomly selects a micro-actuator as a physical excitation source, injecting a specific, weak pressure wave or thermal pulse into the coolant as a challenge, thereby "questioning" the physical system. Immediately afterwards, the corresponding physical response is captured by the micro-thermoelectric effect sensor array and / or the acoustic transducer array. After the challenge signal is injected, the system immediately activates its "senses" to "listen" to the physical world's "answer." The challenge signal evolves in the complex fluid and structural environment, forming a unique physical response field. At this point, the micro thermoelectric effect sensor array and the acoustic transducer array are synchronously activated to capture the physical response modulated by the system's own physical characteristics and convert it into a processable signal. While acquiring the real physical response, the system simulates the propagation and evolution of the challenge signal in the coolant to generate a corresponding expected response. To verify the authenticity of the "answer," the system performs parallel computation in the digital domain. This step, based on known system physical parameters and a built-in propagation model, simulates the theoretical expected response that the same challenge signal should produce under ideal conditions. This is equivalent to preparing a "standard answer" internally. Subsequently, the captured physical response is matched against the expected response. This is the core decision-making step for authentication. The system matches the captured "real answer" against the "standard answer." By calculating the similarity in features, it determines whether the physical system currently interacting with is a real, tamper-free, and legitimate system. Finally, after a successful match, the security-verified command output by the security filtering module is allowed to execute. This step is the final exit point of authentication and determines the fate of the command. Only after the aforementioned comparison is successful, proving physical identity, will the security authentication module release a permission signal, allowing the security-verified instruction, which has already been prepared by the security filtering module and is awaiting execution, to be sent to the microelectromechanical system pump-valve network for execution. Otherwise, the instruction will be intercepted, thus forming a complete security closed loop.

[0049] In some embodiments, the system controller is further configured to perform the following steps: The system's operational data is compressed into low-dimensional state embedding vectors using a deep autoencoder neural network. calculating a geometric distance between the real-time state vector and a pre-stored health state manifold; and based on the geometric distance, quantifying the abnormality degree of the system and triggering an early warning.

[0050] In this embodiment, to achieve the predictive health management of the system, the system controller is endowed with the ability of self-awareness and state assessment. With unsupervised learning techniques, the high-dimensional, noisy system operation data are transformed into concise and informative health indicators. The system controller is further configured to perform the following consecutive steps: first, compress the system operation data into a low-dimensional state embedding vector by a deep autoencoder neural network. This step is the basis of state awareness. The system controller continuously acquires massive, high-dimensional real-time operation data from the micro thermoelectric effect sensor array (providing heat flow vector data) and the acoustic transducer array (providing flow velocity field and structural information). These data are input into a pre-off-line trained deep autoencoder neural network. The encoder part of the network learns and extracts the most essential features in the data through multiple layers of nonlinear transformation, compressing the high-dimensional data into a low-dimensional state embedding vector with significantly reduced dimension but highly concentrated information. This vector can be understood as the "digital fingerprint" or "state identity card" of the system at the current time. Then, the geometric distance between the real-time state vector and a pre-stored health state manifold is calculated. In the offline training phase of the system, by injecting a large amount of historical data during normal operation into the autoencoder, the low-dimensional state embedding vectors corresponding to all health states will naturally cluster in the embedding space, forming a smooth, continuous low-dimensional manifold, i.e. the health state manifold, which defines the "map" of all normal operating conditions of the system. In online operation, the system projects the real-time state embedding vector obtained in step one onto this health state manifold and calculates the geometric distance (such as Euclidean distance or geodesic distance) from the nearest point on the manifold. This distance value intuitively represents the degree of deviation of the current state from the normal range. Finally, based on the geometric distance, the quantification of the abnormality degree of the system and the early warning of failure are realized. The geometric distance itself is a quantitative and robust system health index. When the distance is close to zero, it indicates that the system is running in a normal state. When the distance continues to increase or exceeds a threshold value, the accurate quantification of the abnormality degree of the system is realized. By continuously monitoring the trend of this distance, the system can trigger an early warning when it exceeds the warning threshold, realizing the early warning of failure. More importantly, by analyzing the deviation trajectory of the state vector on the manifold, the system can preliminarily judge the evolution direction and potential root cause of the abnormality (such as pump body performance degradation or local blockage of flow passage), providing deep insights for maintenance decisions.

[0051] In some embodiments, the system controller is further configured to perform the following steps: based on real-time operation data acquired from the micro-thermoelectric effect sensor array and the acoustic transducer array, calculate and output a heat dissipation capability image of the cabinet, the heat dissipation capability image including an instantaneous heat dissipation peak value, a heat dissipation margin, and a heat dissipation efficiency; send the heat dissipation capability image to a data center level scheduling platform.

[0052] In this embodiment, in order to realize global energy efficiency optimization at the data center level, the system controller is endowed with the ability of self-evaluation and resource reporting. This is achieved by converting the capability of the physical heat dissipation system into standardized digital assets that can be understood by the upper management system, and the core is to generate and report a heat dissipation capability image. First, based on real-time operation data acquired from the micro-thermoelectric effect sensor array and the acoustic transducer array, calculate and output a heat dissipation capability image of the cabinet. This step is the basis of capability quantification. The system controller continuously fuses real-time operation data such as three-dimensional heat flow vector field from the micro-thermoelectric effect sensor array and three-dimensional flow field from the acoustic transducer array. Based on these high-fidelity physical field information, through the built-in evaluation algorithm, dynamically calculate multiple key indicators representing the current heat dissipation capability of the cabinet, which together constitute the heat dissipation capability image. The image specifically includes: instantaneous heat dissipation peak value: based on the current flow field state and the physical properties of the cooling liquid, calculate the maximum heat load (unit: kW) that the cabinet can handle instantaneously under the premise of ensuring that all chips do not overheat. Heat dissipation margin: by comparing the above instantaneous heat dissipation peak value with the actual heat load of the cabinet, calculate the additional heat dissipation capability currently available, usually expressed in percentage or power value. Heat dissipation efficiency: by analyzing the total power consumption of the pump valve network and the actual heat carried away by the system, calculate the current energy efficiency coefficient or performance coefficient, which directly reflects the "cost performance" of the heat dissipation behavior. Subsequently, the heat dissipation capability image is sent to a data center level scheduling platform. After calculating the real-time heat dissipation capability image, the system controller sends this image to a global data center level scheduling platform through the data center internal network. This step establishes a key data bridge between the physical heat dissipation system and the computing task scheduling system. The image as a standardized digital asset makes the heat dissipation capability as traditional resources such as CPU computing power, memory capacity, and network bandwidth, become a key dimension that can be perceived, measured, and scheduled by the scheduling platform for global resource optimization.

[0053] In a second aspect, the present application provides a data center comprising a plurality of computer cabinets, at least one of the computer cabinets being equipped with the AI active flow control single-phase immersion cooling system according to any one of the embodiments of the first aspect, and the data center further comprising a scheduling platform configured to perform joint task scheduling in three dimensions of computing resources, network resources and cooling resources based on the received heat dissipation capability profiles of the computer cabinets. It can be imagined that the data center has all the beneficial effects of the AI active flow control single-phase immersion cooling system of the first aspect, which will not be repeated here.

[0054] The above description is only an embodiment of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation using the content of the specification and drawings, or direct or indirect application in other related technical fields, is also included in the patent protection scope of the present application.

Claims

1. An AI-driven active flow control single-phase immersion cooling system, comprising a cabinet containing coolant, characterized in that, Also includes: A miniature thermoelectric effect sensor array arranged inside a cabinet; A distributed acoustic transducer array deployed within a cabinet; Microelectromechanical system pump-valve network integrated in the bypass channel of the heating element; The system controller is connected to the micro thermoelectric effect sensor array, the acoustic transducer array, and the microelectromechanical system pump-valve network signal, respectively. The system controller is configured to perform the following steps: Based on the echo signal received from the acoustic transducer array, the three-dimensional flow field within the cabinet is reconstructed; Based on the readings obtained from the micro thermoelectric effect sensor array, the three-dimensional heat flow vector field within the cabinet is reconstructed; Based on the three-dimensional velocity field and the three-dimensional heat flux vector field, control commands are generated; According to the control command, each pump and valve in the microelectromechanical system pump-valve network is independently driven to control the heat-generating element. Targeted heat dissipation is applied to the components.

2. The AI ​​active flow control single-phase immersion cooling system according to claim 1, characterized in that, The step of reconstructing the three-dimensional flow field within the cabinet includes: Heat pulses are injected into the coolant through a smart skin installed on the inner wall of the cabinet; The acoustic transducer array is used to track the trajectory of the thermal pulse. The flow rate, diffusion coefficient, and turbulence intensity of the coolant in a specific region are calculated based on the tracking results.

3. The AI ​​active flow control single-phase immersion cooling system according to claim 1 or 2, characterized in that, The microelectromechanical system pump-valve network includes multiple piezoelectric diaphragm pumps and electrostatic microvalves; The step of independently driving each pump valve of the microelectromechanical system pump-valve network according to the control command includes: By coordinating the control of pumps and valves in different areas of the microelectromechanical system's pump-valve network, a focused jet is generated inside the cabinet. Or a controllable vortex.

4. The AI ​​active flow control single-phase immersion cooling system according to claim 1, characterized in that, It also includes local microcontrollers; The local microcontroller is integrated with at least one pump valve in the microelectromechanical system pump valve network and communicates directly with the adjacent micro thermoelectric effect sensor array. The local microcontroller is configured to directly drive the at least one pump valve to perform an impact cooling procedure when it receives a signal of a sudden temperature rise from the adjacent micro thermoelectric effect sensor array.

5. The AI ​​active flow control single-phase immersion cooling system according to claim 1, characterized in that, The system controller runs a digital twin model; the system controller is further configured to perform the following steps: The sensor data obtained from the micro thermoelectric effect sensor array and the acoustic transducer array are continuously compared with the predicted values ​​of the digital twin model; Based on the comparison results, the coolant viscosity parameters in the digital twin model are calibrated online using a Kalman filter.

6. The AI ​​active flow control single-phase immersion cooling system according to claim 1, characterized in that, It also includes a safety filter module, which is connected between the command output terminal of the system controller and the pump-valve network of the microelectromechanical system; The system controller is further configured to send the control command to the safety filter module after the control command is generated and before the pump-valve network is driven according to the control command; The security filtering module is configured to perform real-time verification and correction of the control commands based on the control barrier function theory, so as to output the security-verified commands and ensure that the system state is within the dynamic security domain.

7. The AI ​​active flow control single-phase immersion cooling system according to claim 6, characterized in that, Also includes security The authentication module, the security authentication module, is configured to perform the following steps: A challenge signal is injected into the coolant by randomly selecting a micropump valve from the microelectromechanical system's pump-valve network; The corresponding physical response is captured by the micro thermoelectric effect sensor array and / or the acoustic transducer array; The expected response is generated by simulating the propagation and evolution of the challenge signal in the coolant. The captured physical response is matched against the expected response; and upon successful matching, the security-verified instructions output by the security filtering module are allowed to be executed.

8. The AI ​​active flow control single-phase immersion cooling system according to claim 1, characterized in that, The system controller is further configured to perform the following steps: The system's operational data is compressed into low-dimensional state embedding vectors using a deep autoencoder neural network. Calculate the geometric distance between the real-time state vector and the pre-stored healthy state manifold; and based on the geometric distance, quantify the degree of system anomaly and provide early warning of faults.

9. The AI-driven active flow-controlled single-phase immersion cooling system according to claim 1, characterized in that, The system controller is further configured to perform the following steps: Based on real-time operating data obtained from the micro thermoelectric effect sensor array and the acoustic transducer array, the heat dissipation capacity profile of the cabinet is calculated and output, which includes instantaneous heat dissipation peak, heat dissipation margin and heat dissipation efficiency. The heat dissipation capability profile is sent to a data center-level scheduling platform.

10. A data center comprising multiple computer cabinets, characterized in that, At least one of the computer racks is equipped with an AI active flow control single-phase immersion cooling system as described in any one of claims 1-9, and the data center further includes a scheduling platform configured to perform joint task scheduling in three dimensions: computing resources, network resources, and cooling resources, based on the cooling capacity profiles received from each rack.

Images