Liquid cooling electronic equipment cavity flow field prediction method based on quantum physical neural network

Abstract

The invention discloses a liquid cooling electronic equipment cavity flow field prediction method based on a quantum physics neural network, and relates to the technical field of electronic equipment thermal management and computational fluid mechanics numerical prediction. Comprising the following steps: constructing a physical model of flow and heat transfer in the liquid cooling cavity; collecting or generating multi-source supervision data, and performing normalization processing; constructing a hybrid physical neural network structure with a quantum feature head; designing a total loss function containing multiple physical constraints and local weighted terms; carrying out iterative training by adopting a late time and hot plume region oversampling strategy; and utilizing the trained quantum physical neural network to quickly predict the flow field and the temperature field in the cavity under different working conditions. By constructing a hybrid physical neural network model and integrating multi-source supervision, partial differential equation residual error, boundary condition constraint and gradient sensing items of local thermal plume and thermocline in a loss function, high-precision prediction of a three-dimensional flow field and a temperature field in the liquid cooling cavity is realized.

Description

Technical Field

[0001] This invention relates to the fields of thermal management and computational fluid dynamics numerical prediction technology for electronic devices, and in particular to a method for predicting the flow field of a liquid-cooled electronic device cavity based on a quantum physics neural network, which is used to efficiently predict the velocity field, pressure field and temperature field distribution inside the liquid-cooled cavity. Background Technology

[0002] With the rapid development of high-power-density electronic devices and data centers, liquid-cooled heat dissipation chambers are widely used for thermal management of cabinets, power modules, and power electronic devices. Liquid-cooled chambers typically contain strong buoyancy, thermal plumes, localized high-temperature zones, and complex three-dimensional flow structures. The spatiotemporal distribution of their internal flow and temperature fields directly determines the junction temperature level and reliability of the devices.

[0003] Traditional numerical simulation methods are typically based on the incompressible Navier-Stokes equations and energy equations, solved using finite volume, finite element, or finite difference methods. While these methods offer high accuracy, their computational cost is extremely high in three-dimensional unsteady, multi-condition, and long-term time-series scenarios, making it difficult to support online optimization control and rapid design iteration.

[0004] In recent years, data-driven deep learning methods, such as convolutional neural networks, recurrent neural networks, and Transformers, have been used for flow and temperature field prediction. These methods fit the mapping between input conditions and output fields through supervised learning, but they lack constraints on the fundamental physical equations and are prone to problems such as unphysical oscillations, mass non-conservation, or energy non-conservation under extrapolated conditions, resulting in limited generalization ability.

[0005] Physical Information Neural Networks (PINNs) achieve end-to-end fitting of continuous space-time fields by incorporating the residuals of partial differential equations, boundary conditions, and initial conditions into the loss function. While existing PINN methods have been widely applied in steady-state or simple transient problems, they often encounter difficulties such as training challenges and large local errors in strongly buoyant turbulent plumes, thin thermoclines, and regions with severe local gradients.

[0006] Furthermore, the development of quantum machine learning has made it possible to encode and transform high-dimensional features using quantum circuits. Quantum feature heads can generate complex nonlinear features with a finite parameter scale by parameterizing the measurement expectation of the output state of the quantum circuit. However, among existing publicly available technologies, there are few schemes that deeply couple quantum features with physically constrained neural networks and perform loss design and sampling enhancement for local thermal plumes and thermoclines.

[0007] In summary, existing technologies for predicting the flow and temperature fields inside liquid-cooled electronic device cavities mainly suffer from the following problems: CFD simulation alone is computationally expensive and cannot meet the need for rapid prediction under multiple operating conditions; purely data-driven networks lack physical constraints, resulting in insufficient generalization ability and interpretability; existing PINNs exhibit significant local errors when handling strong buoyancy plumes and thin thermoclines, making it difficult to perform precise predictions for critical regions; and the integration of quantum neural networks and physically constrained neural networks is limited, lacking a systematic solution for liquid-cooled cavities.

[0008] Therefore, it is necessary to propose a liquid-cooled cavity flow field prediction method that combines quantum feature encoding, physical equation constraints, multi-source supervised data, and adaptive loss design and sampling strategies for thermal plumes and thermoclines. This method can improve prediction accuracy and generalization ability while ensuring physical interpretability and reducing dependence on large-scale labeled data. Summary of the Invention

[0009] The purpose of this invention is to provide a rapid prediction method for the three-dimensional flow field and temperature field of liquid-cooled electronic device cavities, addressing the aforementioned problems. This method uses incompressible flow and convection-diffusion heat transfer as physical consistency constraints, unifies a small amount of CFD field data, experimental observations, and operating parameters into supervisory information, and introduces a quantum feature encoding module to enhance the characterization capability of strongly nonlinear flow structures. This enables continuous mapping and prediction of velocity, pressure, and temperature fields under multiple operating conditions and long time scales.

[0010] The technical solution of the present invention is as follows: A method for predicting the flow field in the cavity of liquid-cooled electronic devices based on quantum physics neural networks includes the following steps: Establish governing equations and engineering boundary conditions, including Boussinesq buoyancy terms, for the liquid-cooled cavity; Acquire multi-source supervised samples with spatial-temporal coordinates and corresponding field quantities, and normalize the input and output; Construct a hybrid physical neural network of classical feature extraction, quantum feature injection, and classical regression decoding to achieve the transformation from normalized coordinates to ( , , , , Joint forecasts; A total loss function is constructed that includes supervised consistency, equation residuals, boundary constraints, incompressible constraints, interface enhancement, and microcalibration regularization. A sample generation procedure is used to intensively sample the late time period and key areas of the plume. Through iterative training, the network achieves consistency between supervised data and physical constraints. After training, the velocity, pressure and temperature field distribution inside the cavity are directly output for new operating conditions, which can be used for thermal design evaluation and operation optimization.

[0011] The above method provides a complete technical framework from physical modeling, data preparation, network construction, constraint training to final prediction. This method deeply integrates the advantages of quantum computing with physical information neural networks, systematically solving the problems of high computational cost of traditional CFD, poor physical consistency of pure data-driven methods, and insufficient accuracy of existing PINN in complex heat flow scenarios. It achieves efficient, high-precision, and physically interpretable end-to-end prediction of the flow field of liquid-cooled cavities.

[0012] Furthermore, the hybrid neural network uses a four-dimensional normalized space-time coordinate system. The input is jointly mapped to the output via a multilayer perceptron and a quantum feature head. The construction of a hybrid physical neural network with quantum feature heads specifically includes: Linear normalization is performed on the input and output: , , in, , The input numbers are respectively the first and second. The mean and standard deviation of the dimensional components. , To output the first The mean and standard deviation of the dimensional components.

[0013] Normalized coordinates First, the feature vector is obtained by encoding with a multilayer perceptron: ; from Extracting several components from the middle constitutes the quantum input. After linear preprocessing, the result is mapped to the qubit rotation angle: ; Quantum feature heads use a combination of... One quantum bit, Quantum circuitry for layer-parameterized rotation gates and entanglement gates; for each qubit Apply rotation: and in each layer Apply parameterization gates: Then apply the IsingXX entanglement gate: It acts on adjacent bit pairs; Forward By measuring the expectation of the Pauli-Z operator with each qubit, we obtain the quantum characteristics: ; By concatenating classical and quantum features along the channel dimension, a fused feature is obtained: ; The fused features are then mapped to a normalized output vector via a subsequent multilayer perceptron: , in, This is the weight matrix for linear jump connections.

[0014] By constructing quantum circuits containing parameterized rotation gates and entanglement gates, and generating features by measuring the expected values ​​of specific qubits, nonlinear quantum encoding of high-dimensional spatiotemporal input data was achieved. This structure significantly enhances the model's ability to represent and learn complex, nonlinear flow patterns with a relatively small number of classical parameters, thereby improving the overall expressive efficiency and prediction accuracy of the model.

[0015] Furthermore, the total loss function is a weighted superposition of the supervision loss, partial differential equation residual loss, boundary loss, incompressible constraints, interface enhancement term, and ambient temperature and heat flux micro-calibration regularization term: , in, , , , , , , , , and , All are non-negative weights; The interface enhancement terms are constructed based on the vertical temperature gradient. For supervised sample points, calculate: ; Constructing smooth weights: ; in, For the Sigmoid function, For gradient threshold, For smooth width; The interface gradient-aware loss is defined as: ; By dynamically allocating loss weights based on the vertical temperature gradient, the model can automatically focus on key physical regions with drastic gradient changes, such as thermoclines and the edges of thermal plumes, during training. This effectively guides the model optimization direction, significantly reduces local prediction errors in these hard-to-fit regions, and improves the accuracy of the overall temperature field, especially the prediction of thermal stratification structures.

[0016] The plume centerline temperature loss term included in the total loss function is calculated as follows: In the region near the centerline of the heat source and with a vertical height exceeding a certain threshold (satisfying...) Within this range, the average absolute value of the error between the predicted and actual temperature values ​​is calculated: , in, This is the sample set of the plume centerline region.

[0017] The method described above employs a regionalized loss function specifically designed for the core rising region of the hot plume, a critical channel directly impacting heat dissipation efficiency. This ensures the model accurately captures the temperature distribution and evolution along the plume centerline, providing direct engineering guidance for assessing hotspot locations and optimizing device layout, and enhancing the method's predictive reliability along the core heat dissipation path.

[0018] Furthermore, the supervision loss term adopts multi-channel weighted Huber loss, which is calculated as follows: For a sample set with supervised truth values The predicted value is denoted as Define the difference between each channel: ; The Huber function is defined as follows: ; For five output channels Set weights separately The multi-channel monitoring loss is obtained: , in, This represents the set of samples with supervised labels for that channel; The total loss function, which includes temperature physics errors and temperature uniformity constraints, is calculated as follows: Temperature physical error is expressed as mean absolute error: , Spatial average of predicted and actual temperature fields calculated at the same time section , Construct a uniform temperature constraint: .

[0019] Using the above method, Huber loss is employed as the basis for supervision, combined with a multi-channel independent weighting design, balancing training stability and accuracy control. Huber loss is less sensitive to outliers than mean squared error, improving training robustness; while setting independent weights for different physical quantities such as velocity, pressure, and temperature allows for fine-tuning based on the importance and magnitude of each field, optimizing overall training efficiency and multi-objective balance.

[0020] Furthermore, the strategy of oversampling the late time region and the thermal plume spatial region specifically includes: In the process of generating supervised samples and PDE samples, a "late-stage focusing" and "feather region local magnification" mechanism is introduced: Let the total time range be... Late time threshold Then at the sampling time Segmented pooled sampling is used:

[0021] For spatial coordinates, in a certain proportion of the sample, Limited to the plume area:

[0022] This increases the sampling density of the late thermal plume and the upper thermocline region, enhancing the model's ability to fit key regions.

[0023] By employing the above method and a piecewise non-uniform time sampling strategy, the training sample density is increased when the thermal flow reaches the quasi-steady state or fully developed stage. This helps the model better learn and predict the stable thermal stratification structure and flow pattern in the later stages of the system, avoiding the problem of overfitting the transient initial process and ignoring long-term steady-state characteristics, thus improving the physical rationality and engineering applicability of the prediction results. In the spatial sampling stage, sampling points are selectively densified above the heat source and in key areas where the plume develops. This strategy directly directs more computational resources and learning attention to local areas that are crucial to the overall heat dissipation performance, ensuring the model's ability to capture complex physical phenomena (such as buoyancy-driven and vortex structures) within these limited spaces. This significantly improves the detail resolution of predictions for key areas without significantly increasing the overall training cost.

[0024] Furthermore, to suppress non-physical oscillations at the initial temperature moment, the hard initial condition smoothing process performed on the temperature channel output by the hybrid physical neural network is as follows: set up This is the original temperature prediction value after network inverse normalization. Define dimensionless time for ambient temperature: And construct smooth weights: The corrected physical temperature is: ; Will The temperature output channel is renormalized and replaced to automatically meet the requirements as t→0. ≈ .

[0025] Furthermore, the energy equation PDE residuals included in the total loss function are calculated using the following method: At unsupervised physical sampling points At this point, the residuals of the energy equation are constructed by automatically differentiating and calculating the derivatives of temperature with respect to space and time: , The PDE loss is defined in Huber form: ; The Boussinesq vertical momentum residuals included in the total loss function are calculated as follows: For the vertical velocity component Construct the Boussinesq equation residuals: ; Corresponding loss: .

[0026] Furthermore, the incompressible constraints included in the total loss function are calculated as follows: On supervised samples, the divergence is calculated using automatic differentiation: , The corresponding loss is: .

[0027] Furthermore, the physical model for constructing the flow and heat transfer within the liquid-cooled cavity specifically includes: Treating the coolant within the cavity of liquid-cooled electronic equipment as an incompressible Newtonian fluid, the Boussinesq approximation is used to handle the buoyancy term, and its governing equations include the velocity field. Pressure field and temperature field : mass conservation equation: ; The momentum equation in the vertical direction In the direction of gravity: , in, For coolant density, For kinematic viscosity, It is the acceleration due to gravity. The coefficient of volume expansion, For reference temperature, It is a vertical unit vector; Energy equation: , in, For thermal diffusivity, Thermal conductivity, Specific heat capacity at constant pressure; The outer wall of the cavity adopts convective heat transfer boundary conditions: ; in, This represents the derivative along the outward normal. The equivalent heat transfer coefficient, Ambient temperature; Several equivalent heat source surfaces are set inside the cavity, and constant heat flux boundary conditions are adopted: , in, The nominal heat flux density per unit area, This is the heat flux ratio micro-calibration factor; The velocity of the inner wall of the liquid-cooled cavity satisfies the no-slip condition: .

[0028] Furthermore, the boundary condition loss included in the total loss function is calculated using the following method: For sampling points at the boundary between the outer wall and the heat source, construct residuals based on the boundary type: No slippage condition on the outer wall: ; External wall convection heat transfer conditions: ; Constant heat flux boundary conditions for heat source: ; Boundary loss is defined as: ; The environmental and thermal flux microcalibration regularization included in the total loss function is calculated as follows: Set a learnable ambient temperature bias and heat flux ratio factor Add L2 regularization terms respectively: and through and Weighting is applied.

[0029] By employing the methods described above, the no-slip condition, convective heat transfer condition, and constant heat flow condition are explicitly incorporated into the loss function in mathematical form as strong physical constraints. This ensures that the flow and temperature fields predicted by the neural network strictly satisfy basic physical conservation laws and realistic engineering boundary conditions, fundamentally guaranteeing the physical reliability of the prediction results and forming the basis for the model's good extrapolation and generalization capabilities.

[0030] Compared with existing technologies, the advantages of this invention are: 1. By explicitly introducing the Navier–Stokes equation, energy equation, Boussinesq buoyancy and boundary conditions into the loss function, the quantum physics neural network of the present invention has strong physical interpretability and can still maintain physical consistency under extrapolation conditions. 2. Introducing interface loss based on temperature gradient sensing and the loss of the plume centerline Furthermore, by oversampling the late-stage and plume regions during the sampling phase, this invention can significantly reduce local prediction errors in thermal plumes and near the thermocline. 3. By using quantum feature heads to perform nonlinear quantum encoding on the space-time input, this invention significantly improves the network's expressive power under complex flow patterns and improves the accuracy of flow field prediction under the premise of controllable parameter scale; 4. By adjusting the ambient temperature and heat flux ratio factor With its learnable micro-calibration, this invention can automatically absorb measurement and modeling errors, reducing manual parameter tuning workload and improving engineering applicability. 5. In the application of liquid-cooled electronic equipment cavities, this invention can quickly provide three-dimensional flow field and temperature field under different operating conditions based on existing limited CFD or experimental data, providing support for electronic equipment layout optimization, thermal safety assessment and control strategy design. Attached Figure Description

[0031] Figure 1 This is a flowchart of a method for predicting the flow field in the cavity of a liquid-cooled electronic device based on a quantum physics neural network.

[0032] Figure 2 This is a schematic diagram of the quantum physics neural network structure in this application.

[0033] Figure 3 This is a schematic diagram of the loss function composition of this application. Detailed Implementation

[0034] It should be noted that relational terms such as "first" and "second" are used merely 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 a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0035] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0036] Please see Figure 1-3 A method for predicting the flow field in the cavity of liquid-cooled electronic devices based on quantum physics neural networks, such as Figure 1 As shown, it includes the following steps: Physical model and boundary condition settings: First, based on the actual application scenario of the liquid cooling chamber, the density of the working fluid is selected. kinematic viscosity Thermal diffusivity Coefficient of volume expansion Thermal dilution and ambient temperature The incompressible Boussinesq model is used to describe the flow and heat transfer processes. The three-dimensional velocity components are denoted as follows: , , The pressure is The temperature is .

[0037] The conservation of mass is represented by the velocity divergence being zero: .

[0038] In the vertical direction, the momentum balance takes the form that includes a buoyancy term: , in, It is the acceleration due to gravity. This is a reference temperature.

[0039] The temperature field satisfies the convection-diffusion equation: .

[0040] Apply convective heat transfer boundary conditions to the outer wall of the cavity, using Describes the heat transfer between the wall's normal heat conduction and the environment, where The temperature gradient is along the outward normal direction of the wall. It is the equivalent heat transfer coefficient obtained by converting the heat conduction of the insulation layer and external convection.

[0041] Apply constant heat flux boundary conditions to the surface of the equivalent heat source, using Describes the normal heat flux density, where The nominal heat flux density per unit area, A learnable heat flux scaling factor for fine-tuning the actual heat input. No-slip conditions are applied to all solid walls. , , Initially, the fluid can be considered to be stationary and the temperature distribution to be approximately uniform.

[0042] Multi-source sample construction and unified normalization: Given a physical model and boundary conditions, velocity, pressure, and temperature data at several spatial-temporal points within the cavity were obtained through CFD numerical simulation and experimental measurements, forming a supervised sample set. .

[0043] Simultaneously, a batch of sample points were collected on the boundary to record the spatial-temporal coordinates. The boundary type (outer wall or heat source surface) and optional temperature or heat flow information form a boundary sample set.

[0044] To improve network training stability, the input coordinates and output physical quantities are linearly normalized. This is done using spatial coordinates. For example, the normalization method is: ,in and These are the mean and standard deviation of the sample, respectively. Similarly, for... Calculate the mean and standard deviation separately, and construct the normalized vector. .

[0045] Hybrid physical neural network construction: according to Figure 2 The structure shown constructs a hybrid neural network. The input layer receives normalized coordinates. First, intermediate feature vectors are obtained through a multilayer perceptron.

[0046] from Alternatively, several components can be selected from the intermediate features, and the qubit rotation angle can be obtained through linear transformation, then input into the parameterized quantum circuit. Each qubit in the quantum circuit is first subjected to a rotation angle proportional to the input. The axis is rotated, and then learnable rotation gates and entanglement gates are applied in a multi-layer structure. Finally, the Pauli-Z expectation value is measured on several qubits to obtain the quantum eigenvector.

[0047] Classical and quantum features are concatenated along the channel dimension, and a normalized output is obtained through a subsequent fully connected layer. Furthermore, a linear jump connection from input to output is superimposed, making the overall mapping both nonlinearly expressive and easy to train and converge.

[0048] A hard initial condition smoothing function is introduced into the temperature channel: the normalized temperature output by the network is restored to the physical temperature. At the current time With maximum prediction duration The ratio of the two values ​​is used as a dimensionless time. ,structure And set the correction temperature to Then, normalize the data again using the mean and standard deviation and write it back into the temperature channel to ensure... When the temperature approaches zero, it naturally becomes closer to the ambient or initial temperature.

[0049] Total loss function design: like Figure 3 As shown, the multi-channel supervision loss is calculated on the supervised samples. The difference between the predicted value and the true value for any channel is calculated. Using Huber loss with threshold The supervised loss is obtained by averaging across all samples. Meanwhile, the mean absolute error of temperature is calculated on samples with temperature labels, and the square of the difference between the predicted temperature field and the spatial mean of the reference temperature field at each time section is used to constrain the overall energy level.

[0050] At the physical sampling points, the first and second derivatives of temperature with respect to time and spatial coordinates are calculated automatically using differentiation. Substituting into the Huber function and averaging, we obtain the residual terms of the energy equation. Similarly, ... Substitute the values ​​into the Huber function and calculate the average to constrain the vertical momentum equation.

[0051] On the supervised samples, calculate the velocity divergence based on the incompressibility condition. The average value of Huber is used as an incompressible constraint term.

[0052] On the boundary samples, the wall velocity component itself is treated as a residual, and the outer wall temperature boundary is treated as follows: As residuals, the boundary losses are obtained by averaging them using the Huber function, which are used to ensure no-slip, heat transfer, and constant heat flow conditions.

[0053] On temperature-monitored samples, based on the absolute value of the vertical temperature gradient... Interface weights are constructed, and regions with large gradients are assigned higher weights using a smoothing function. The average of the squared temperature error is then used to form the interface gradient perception loss. Sample points are selected in the rising region near the centerline of the heat source, and the average of the absolute temperature error is used to form the plume centerline loss.

[0054] By treating the ambient temperature bias and heat flux scaling factor as learnable parameters, and weighting the squares of their deviations from the reference values, a micro-calibration regularization term is formed.

[0055] Sampling strategy and network training: In each training iteration, a batch of samples is randomly drawn from the supervised sample set and the boundary sample set, and a batch of PDE sampling points is randomly generated within the physical domain region. Temporal sampling employs an "early-late" hybrid strategy: a portion of the time is sampled uniformly across the entire interval, while the other portion is concentrated in the late interval near the final moment, to enhance the model's learning of the stable phase of the thermal plume and the temperature stratification structure. In spatial sampling, samples near the centerline of the equivalent heat source and the upper region are appropriately densified to improve the prediction accuracy near the main plume channel and the thermocline.

[0056] An optimization algorithm based on adaptive estimation (such as AdamW) is used to update the network parameters and micro-calibration parameters. Mixed-precision training and gradient clipping are combined to improve numerical stability. In each iteration, the loss of each item is calculated, and the total loss is obtained by summing them according to the weights. This total loss is then backpropagated and the parameters are updated until the validation set error converges or the preset number of training rounds is reached.

[0057] Flow field prediction and application: After training, freeze the network parameters. For new operating conditions, only the boundary conditions of the liquid-cooled cavity, the heat source power, and the target prediction time need to be given, and the spatial grid coordinates at the corresponding time points need to be provided. Normalization By inputting the trained quantum physics neural network, a normalized output can be obtained. After inverse normalization and temperature smoothing correction, the three-dimensional velocity field, pressure field, and temperature field distribution under this working condition can be obtained.

[0058] By comparing with a small number of CFD simulations or experimental measurement results, the prediction accuracy of this application in terms of overall flow structure, thermal plume morphology and temperature stratification location can be verified. This method can be used in application scenarios such as structural design optimization, operation strategy formulation and thermal safety assessment of liquid-cooled electronic device cavities.

[0059] The embodiments described above merely illustrate specific implementation methods of this application, and while the descriptions are detailed and specific, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the technical solution of this application, and these modifications and improvements all fall within the scope of protection of this application.

Claims

1. A method for predicting the flow field in the cavity of liquid-cooled electronic devices based on quantum physics neural networks, characterized in that, Includes the following steps: A physical model of flow and heat transfer within the liquid-cooled cavity is constructed. The governing equations of the physical model include the mass conservation equation, the momentum equation containing the Boussinesq buoyancy term, and the energy equation. Acquire multi-source monitoring data containing spatial-temporal coordinates and corresponding flow field physical quantities, and normalize the input coordinates and output physical quantities; A hybrid physical neural network is constructed, comprising: a classical network for extracting input features, a quantum feature module for injecting quantum parameterization circuits into the features, and a regression decoding module for outputting velocity field, pressure field, and temperature field. The output of the quantum feature module is concatenated or fused with the output of the classical network and then input to the regression decoding module. Design a total loss function, which includes at least a supervision loss term based on supervised data, a partial differential equation residual loss term based on the physical model control equation, and a boundary condition loss term. A training sample set is generated by employing a strategy that includes oversampling of the late time region and the thermal plume spatial region. The hybrid physical neural network is then iteratively trained using the sample set to minimize the total loss function. The space-time coordinates under the target working condition are input into the trained hybrid physical neural network, which outputs the velocity field, pressure field and temperature field inside the cavity.

2. The method for predicting the cavity flow field of liquid-cooled electronic devices based on quantum physics neural networks according to claim 1, characterized in that, The hybrid neural network uses a four-dimensional normalized space-time coordinate system. The input is jointly mapped to the output via a multilayer perceptron and a quantum feature head. The construction of a hybrid physical neural network with quantum feature heads specifically includes: Linear normalization is performed on the input and output: ， ， in, , The input numbers are respectively the first and second. The mean and standard deviation of the dimensional components. , To output the first The mean and standard deviation of the dimensional components; Normalized coordinates First, the feature vector is obtained by encoding with a multilayer perceptron: ； from Extracting several components from the middle constitutes the quantum input. After linear preprocessing, the result is mapped to the qubit rotation angle: ; Quantum feature heads use a combination of... One quantum bit, Quantum circuitry for layer-parameterized rotation gates and entanglement gates; for each qubit Apply rotation: and in each layer Apply parameterization gates: Then apply the IsingXX entanglement gate: It acts on adjacent bit pairs; Forward By measuring the expectation of the Pauli-Z operator with each qubit, we obtain the quantum characteristics: ； By concatenating classical and quantum features along the channel dimension, a fused feature is obtained: ； The fused features are then mapped to a normalized output vector via a subsequent multilayer perceptron: ， in, This is the weight matrix for linear jump connections.

3. The method for predicting the cavity flow field of liquid-cooled electronic devices based on quantum physics neural networks according to claim 1, characterized in that, The total loss function is a weighted superposition of the supervision loss, partial differential equation residual loss, boundary loss, incompressible constraints, interface enhancement term, and ambient temperature and heat flux micro-calibration regularization term. ， in, , , , , , , , , and , All are non-negative weights; The interface enhancement terms are constructed based on the vertical temperature gradient. For supervised sample points, calculate: ; Constructing smooth weights: ; in, For the Sigmoid function, For gradient threshold, For smooth width; The interface gradient-aware loss is defined as: ; The plume centerline temperature loss term included in the total loss function is calculated as follows: In the region near the centerline of the heat source and with a vertical height exceeding a certain threshold (satisfying...) Within this range, the average absolute value of the error between the predicted and actual temperature values ​​is calculated: ， in, This is the sample set of the plume centerline region.

4. The method for predicting the cavity flow field of liquid-cooled electronic devices based on quantum physics neural networks according to claim 3, characterized in that, The monitoring loss term adopts multi-channel weighted Huber loss, and its calculation method is as follows: For a sample set with supervised truth values The predicted value is denoted as Define the difference between each channel: ； The Huber function is defined as follows: ; For five output channels Set weights separately The multi-channel monitoring loss is obtained: ， in, This represents the set of samples with supervised labels for that channel; The total loss function, which includes temperature physics errors and temperature uniformity constraints, is calculated as follows: Temperature physical error is expressed as mean absolute error: ， Spatial average of predicted and actual temperature fields calculated at the same time section , Construct a uniform temperature constraint: 。 5. The method for predicting the cavity flow field of liquid-cooled electronic devices based on quantum physics neural networks according to claim 1, characterized in that, The strategy of oversampling the late time region and the thermal plume spatial region specifically includes: In the process of generating supervised samples and PDE samples, a "late-stage focusing" and "feather region local magnification" mechanism is introduced: Let the total time range be... Late time threshold Then at the sampling time Segmented pooled sampling is used: For spatial coordinates, in a certain proportion of the sample, Limited to the plume area: This increases the sampling density of the late thermal plume and the upper thermocline region, enhancing the model's ability to fit key regions.

6. The method for predicting the cavity flow field of liquid-cooled electronic devices based on quantum physics neural networks according to claim 1, characterized in that, The hard initial condition smoothing process performed on the temperature channel output by the hybrid physical neural network is as follows: set up This is the original temperature prediction value after network inverse normalization. Define dimensionless time for ambient temperature: And construct smooth weights: The corrected physical temperature is: ; Will The temperature output channel is renormalized and replaced to automatically meet the requirements as t→0. ≈ .

7. The method for predicting the cavity flow field of liquid-cooled electronic devices based on quantum physics neural networks according to claim 1, characterized in that, The energy equation PDE residuals included in the total loss function are calculated using the following method: At unsupervised physical sampling points At this point, the residuals of the energy equation are constructed by automatically differentiating and calculating the derivatives of temperature with respect to space and time: ， The PDE loss is defined in Huber form: ; The Boussinesq vertical momentum residuals included in the total loss function are calculated as follows: For the vertical velocity component Construct the Boussinesq equation residuals: ； Corresponding loss: .

8. The method for predicting the cavity flow field of liquid-cooled electronic devices based on quantum physics neural networks according to claim 3 or 4, characterized in that, The Boussinesq vertical momentum residual included in the total loss function is calculated as follows: For the vertical velocity component Construct the Boussinesq equation residuals: ； Corresponding loss: ; The incompressible constraints included in the total loss function are calculated as follows: On supervised samples, the divergence is calculated using automatic differentiation: , The corresponding loss is: .

9. The method for predicting the cavity flow field of liquid-cooled electronic devices based on quantum physics neural networks according to claim 1, characterized in that, The physical model for the flow and heat transfer within the liquid-cooled cavity specifically includes: Treating the coolant within the cavity of liquid-cooled electronic equipment as an incompressible Newtonian fluid, the Boussinesq approximation is used to handle the buoyancy term, and its governing equations include the velocity field. Pressure field and temperature field : mass conservation equation: ; The momentum equation in the vertical direction In the direction of gravity: ， in, For coolant density, For kinematic viscosity, It is the acceleration due to gravity. The coefficient of volume expansion. For reference temperature, It is a vertical unit vector; Energy equation: , in, For thermal diffusivity, Thermal conductivity, Specific heat capacity at constant pressure; The outer wall of the cavity adopts convective heat transfer boundary conditions: ; in, This represents the derivative along the outward normal. The equivalent heat transfer coefficient, Ambient temperature; Several equivalent heat source surfaces are set inside the cavity, and constant heat flux boundary conditions are adopted: , in, The nominal heat flux density per unit area, This is the heat flux ratio micro-calibration factor; The velocity of the inner wall of the liquid-cooled cavity satisfies the no-slip condition: .

10. The method for predicting the cavity flow field of liquid-cooled electronic devices based on quantum physics neural networks according to claim 3, characterized in that, The boundary condition loss included in the total loss function is calculated using the following method: For sampling points at the boundary between the outer wall and the heat source, construct residuals based on the boundary type: No slippage condition on the outer wall: ; External wall convection heat transfer conditions: ; Constant heat flux boundary conditions for heat source: ; Boundary loss is defined as: ； The environmental and thermal flux microcalibration regularization included in the total loss function is calculated as follows: Set a learnable ambient temperature bias and heat flux ratio factor Add L2 regularization terms respectively: and through and Weighting is applied.

Images