Topological optimization improvement method of initial layout for extracting mainstream features under complex heat dissipation working condition, medium and program product

Abstract

The invention discloses a topological optimization improvement method of an initial layout for extracting mainstream characteristics under a complex heat dissipation working condition, a medium and a program product, and belongs to the technical field of intersection of thermal management and structural topological optimization, and the method is designed for a liquid cooling heat dissipation structure under the conditions of high Reynolds number turbulence and conjugate heat transfer. The method comprises the following steps: constructing a two-dimensional array type parameterized initial layout, generating high-fidelity sample data based on CFD simulation, training a neural network prediction model with an SE attention mechanism, carrying out initialization scheme screening in combination with a correction objective function under a flow field freezing hypothesis and an ETO optimization algorithm, and carrying out prediction on the flow field freezing hypothesis. And topological optimization evolution based on a variable density method is implemented on the basis of the optimization initial state. According to the method, the final effect of topological optimization can be improved while the calculation complexity is reduced, collaborative optimization of the heat dissipation efficiency and the flowing performance is achieved, and an effective means is provided for conjugate heat transfer design of the liquid cooling heat dissipation structure under the high Reynolds number turbulence complex working condition.

Description

Technical Field

[0001] The present invention belongs to the cross - technical field of thermal management and structural topology optimization, relates to the optimization technology of liquid - cooled conjugate heat transfer under high Reynolds number turbulence, and particularly relates to a topology optimization improvement method, medium and program product for an initial layout extracting mainstream features under complex heat dissipation conditions. Background Art

[0002] Currently, the heat dissipation strategies of electronic devices mainly include passive methods such as natural convection and phase - change cooling, and active methods such as forced air cooling and forced liquid cooling. Compared with other methods, liquid - cooled radiators using liquid media with high thermal conductivity have relatively high compactness, higher heat transfer rates, and lower costs, and have become one of the mainstream choices in high - heat - flux - density scenarios.

[0003] The structural parameters and geometric shapes of cooling channels in a liquid - cooled heat dissipation system are the main factors affecting the flow state, heat and mass transfer characteristics, and the distribution of related physical state variables such as temperature and flow velocity. Existing technologies use bionic or fractal methods to improve micro - channel structures. However, such methods rely on design experience and trial - and - error strategies and lack universality. In addition, although some cooling channels use optimization algorithms to improve performance and design efficiency, their essence is still to change the distribution of the physical field by optimizing the structure and find the optimum according to the change of performance parameters. This iterative calculation process still requires a large amount of time and resources.

[0004] In this context, the topology optimization (TO) method is considered an excellent choice for the target design of cooling channels due to its high design freedom, low experience dependence, and strong physical orientation. The construction and basic principle of the TO model are based on optimization algorithms. On the basis of given loads, governing equations, boundary constraints, and optimization variables, it seeks the optimal structural configuration of the material positions and density values of grid nodes within the design domain, so as to maximize or minimize the objective function value. Using the TO method to design cooling channels takes grid cells as the optimization object, has a very high design freedom, and allows simultaneous optimization of size, shape, and topology from a global perspective throughout the design domain.

[0005] However, there are still challenges when applying the TO method to fluid-solid conjugate heat transfer problems involving high Reynolds number turbulence. Most of the existing research on fluid topology optimization focuses on laminar flow at low Reynolds numbers. For high Reynolds number turbulence, due to the extreme complexity of its flow mechanism, direct topology optimization is very likely to cause strong numerical oscillations, convergence difficulties, and even physically distorted optimization results. To circumvent these problems, existing technologies usually adopt two simplified approaches: one is to use simplified models (such as the Darcy model) to characterize turbulence, but this will sacrifice the accuracy of the physical model and limit the performance upper limit of the final optimization result; the other is to stabilize the optimization process through complex penalty coefficient adjustment and interpolation method settings, but this requires a lot of debugging work, which is time-consuming and labor-intensive. In addition, some attempts to use preset structures (such as three-periodic minimal surfaces) to fill the intermediate density area have reintroduced reliance on human experience, and the performance adaptability under variable working conditions is questionable.

[0006] In the topology optimization process of conjugate heat transfer, in order to balance computing resources and results, some technical solutions try to use artificial intelligence methods, hoping to directly generate results under new models through a large amount of existing data, or to construct a reasonable initial layout by exploring various bionic methods, hoping to generate better results. However, the topology optimization of turbulent conjugate heat transfer under high Reynolds numbers is itself a difficult problem. There are very few available databases, and most of them are composed of low-fidelity optimization, which limits its optimization space. Traditional bionic methods generally have their limitations. The given initial layout has considerable human factors and requires multiple trial and error. Moreover, only when the bionic object and the research object are similar to each other to some extent can a relatively good effect be achieved, and there is no good universality.

[0007] In summary, under the conditions of high heat flux and high Reynolds number flow, how to effectively identify and guide the dominant structural characteristics of heat dissipation, improve the physical rationality and computational stability in the topology optimization process, and thus obtain a better heat transfer-flow synergistic structure within limited computing resources is a technical problem that needs to be solved urgently in the current complex heat dissipation topology optimization design. Summary of the invention

[0008] 1. Purpose of the invention Aiming at the above defects and deficiencies of the prior art, the present invention aims to provide a topology optimization improvement method, medium and program product for the initialization layout of extracting mainstream features under complex heat dissipation conditions. By combining the advantages of artificial intelligence methods and advanced optimization algorithms, first, a universal array layout framework is constructed to directly improve the initial flow field, and then the optimized initial layout is used to guide the topology direction of turbulent conjugate heat transfer under complex flow conditions. By integrating the Squeeze-and-Excitation Net (SE-Net) and the Exponential-Trigonometric Optimization (ETO) algorithm, a framework for multi-objective topology optimization of the initial layout of a turbulent thermal fluid system is established, which can effectively avoid a large number of trial-and-error costs and strong numerical problems during the optimization process. Under the condition of a relatively low number of grids and without a large number of continuous changes in topology parameters during the topology process, quite good results can be obtained. The present invention is of great significance for innovating the traditional optimization method of turbulent conjugate heat transfer, and can provide a feasible path and method reference for the selection of multi-objective topology optimization, the establishment of a benchmark model, and its application in conjugate heat transfer optimization.

[0009] (II)Technical Solution To achieve the object of the invention and solve its technical problems, the present invention adopts the following technical solutions: The first object of the present invention is to provide a topology optimization improvement method for the initialization layout of extracting mainstream features under complex heat dissipation conditions, which is used for conjugate heat transfer topology optimization design of a liquid cooling structure under high Reynolds number turbulent conditions, improving the heat dissipation efficiency and flow uniformity of the cooling channel structure, and realizing the collaborative optimization design of multi-objective heat-flow performance. The mainstream features refer to the structural paths or regions that mainly affect the heat-flow performance of the system in the regions with high heat flux density and pressure drop sensitivity under complex flow conditions. Different from the traditional bionic layout, the present invention only uses a small amount of initial solid domain to reshape the initial flow field to guide the optimization direction of the topology. The method used consists of a reduced-order surrogate model based on SE-Net, an ETO optimization algorithm, a turbulent conjugate heat transfer topology optimization algorithm, and an objective corrected by the relative standard deviation under the flow field freezing assumption. The method at least includes the following steps when implemented: S100. Construct a parametric structural array layout: In the liquid cooling cavity design domain of the liquid cooling plate radiator, construct a two-dimensional array-type initial layout formed by arranging multiple structural units along the mainstream direction and the lateral direction, set one or more of the radius, spacing, and arrangement mode of the structural units as design variables, and establish an adjustable parameter space; S200. Generate a high-fidelity training sample data set: Random sampling is carried out within the adjustable parameter space to generate an initial layout sample with a predetermined number and different combinations of geometric parameters; for each sample, computational fluid dynamics (CFD) numerical simulation is carried out under high Reynolds number conditions to solve the flow field and temperature field under the preset high Reynolds number turbulence and conjugate heat transfer boundary conditions, and macroscopic performance indicators are extracted, including at least the average or maximum temperature characterizing the heat dissipation performance and the inlet and outlet pressure drop characterizing the flow performance, to construct a high-fidelity training dataset containing the combination of initial layout geometric parameters and their corresponding macroscopic performance indicators; S300. Establish and train a mainstream feature prediction model: Based on the high-fidelity training dataset, a convolutional neural network (SE-Net) reduced-order surrogate model with SE attention mechanism is constructed and trained. Through feature squeezing, excitation and scaling operations, a mapping prediction relationship between the combination of initial layout geometric parameters and the output of macroscopic performance indicators under high Reynolds number conditions is established, enabling the model to quickly evaluate the macroscopic performance of the initial layout without explicitly analyzing the flow field; S400. Construct a modified objective function under the assumption of flow field freezing: Define a modified objective function under the assumption of flow field freezing for comprehensively evaluating the macroscopic performance of the initial layout to achieve the collaborative optimization of heat dissipation performance and pressure drop characteristics. The modified objective function is a multi-objective evaluation function, including at least the average or maximum temperature, the inlet and outlet pressure drop, and the structural unit area ratio index, and is constructed by combining an empirical correction factor and a normalized weight; S500. ETO global optimization to screen the optimal initial layout: Sampling is carried out within the adjustable parameter space to generate a large-scale potential initial layout sample. The ETO algorithm is used to iteratively optimize the potential initial layout sample space. When evaluating the quality of candidate solutions in each iteration, the trained SE-Net reduced-order surrogate model is called for rapid prediction. Through large-scale sample screening, the initial layout that meets the optimal solution of the objective function is selected as the initial state of topology optimization; S600. Perform topology optimization based on the mainstream-guided initial layout: Taking the optimal initial layout obtained by screening as the initial field of topology optimization, a topology optimization method based on density field interpolation is adopted, combined with the conjugate heat transfer control equation, the k-ε turbulence model at high Reynolds number, and the fluid momentum equation corrected by inverse magnetic permeability, and combined with the boundary conditions of the convection-conduction double domain of the structural field. Under the comprehensive performance objective function, the topology optimization iteration process is performed to optimize the density distribution function to achieve the optimal configuration of materials within the design domain, and to achieve the global optimal evolution of the topology structure under the thermal-fluid synergy objective.

[0010] The second object of the present invention is to provide a computer program product, including computer instructions for executing the above-mentioned method for improving the topological optimization of the initial layout of extracting the main features of the present invention under complex heat dissipation conditions.

[0011] The third object of the present invention is to provide a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it realizes the above-mentioned method for improving the topological optimization of the initial layout of extracting the main features of the present invention under complex heat dissipation conditions.

[0012] (III) Technical effects Compared with the prior art, the method, medium and program product for improving the topological optimization of the initial layout of extracting the main features of the present invention under complex heat dissipation conditions have the following beneficial and remarkable technical effects: (1) By constructing a convolutional neural network reduced-order surrogate model with SE attention mechanism and an ETO global optimization algorithm, the present invention establishes a framework for multi-objective topological optimization of the initial layout of the turbulent heat fluid system, which can effectively avoid a large number of trial-and-error costs and strong numerical problems during the optimization process, and obtain quite good results without a large number of continuous changes of topological parameters during the topological process. (2) Based on the mainstream-guided structure initialization, the present invention introduces multi-physics field coupling control equations, material interpolation models, Helmholtz filters and hyperbolic tangent projection equations during the topological optimization process to form a topological optimization framework with full-process coordination. While realizing the dual optimization of thermal performance and flow performance, this method effectively suppresses problems such as topological pathologies, numerical oscillations and flow separations, and significantly improves the thermal-fluid coordination performance, structural manufacturability and simulation calculation stability of the optimization results. Description of the drawings

[0013] Figure 1 The figure shows a flowchart of the method for improving the topological optimization of the initial layout of extracting the main features of the present invention under complex heat dissipation conditions; Figure 2 The figure shows a flowchart of the convolutional neural network algorithm in the present invention; Figure 3 The figure shows a flowchart of the ETO optimization algorithm in the present invention; Figure 4 The figure shows a flowchart of the topological optimization of the initial layout in the present invention; Figure 5 The figure shows a schematic diagram of the topological optimization area in Embodiment 2 of the present invention; Figure 6 The figure shows a flow field diagram at an inlet Reynolds number of 14000 when the radius of the cylinder is 0.005 meters; Figure 7The figure shows the schematic diagram of the training results of the SE-Net reduced-order surrogate model for the average temperature, where (a) is the comparison of the prediction results of the training set, and (b) is the comparison of the prediction results of the test set; Figure 8 The figure shows the schematic diagram of the training results of the SE-Net reduced-order surrogate model for the pressure drop at the inlet and outlet, where (a) is the comparison of the prediction results of the training set, and (b) is the comparison of the prediction results of the test set; Figure 9 The figure shows the schematic diagram of the ETO optimization results under the modified objective function, where (a) is the target array one, and (b) is the target array two; Figure 10 The figure shows the comparison diagram of the flow field distribution of different channels under the unified heat-flow boundary conditions. In the figure: (a) the bionic topology optimization channel; (b) the topology channel one guided by the mainstream characteristics; (c) the topology channel two guided by the mainstream characteristics; (d) the regular rectangular channel; (e) the cylindrical array channel; (f) the square column array channel; Figure 11 The figure shows the temperature field distribution diagrams of different channels. In the figure: (a) the topology channel one guided by the mainstream characteristics; (b) the topology channel two guided by the mainstream characteristics; (c) the square column array channel; (d) the cylindrical array channel; (e) the bionic topology optimization channel; (f) the regular rectangular channel. Detailed implementation manners

[0014] The present invention aims to provide a topology optimization improvement method, medium and program product for the initialization layout of extracting mainstream characteristics under complex heat dissipation conditions, which is used for conjugate heat transfer topology optimization design of liquid cooling structures under high Reynolds number turbulent conditions, improving the heat dissipation efficiency and flow uniformity of the cooling channel structure, and realizing the collaborative optimization design of multi-objective heat-flow performance. The mainstream characteristics refer to the structural paths or regions that mainly affect the heat-flow performance of the system in the regions with high heat flux density and pressure drop sensitivity under complex flow conditions. To make the purpose, technical solution and advantages of the implementation of the present invention clearer, the technical solutions in the embodiments of the present invention will be described in more detail below with reference to the accompanying drawings in the embodiments of the present invention.

[0015] Embodiment 1: Topology optimization improvement method As a specific example, as Figure 1 shown, the topology optimization improvement method for the initialization layout of extracting mainstream characteristics under complex heat dissipation conditions provided by the embodiment of the present invention mainly includes six key steps: parametric array construction, high-fidelity dataset generation, SE-Net reduced-order surrogate model training, modified objective function design, ETO global optimization, and final topology optimization when implemented. The cooperation of each step ensures obtaining a heat dissipation structure with excellent performance while ensuring the computational efficiency. Specifically: S100. Construct a parametric structure array layout: Within the liquid cooling cavity design domain of the liquid cooling plate radiator, a two-dimensional array-type initial layout formed by arranging multiple structural units in the mainstream direction and the lateral direction is constructed. One or more of the radius, spacing, and arrangement mode of the structural units are set as design variables to establish an adjustable parameter space.

[0016] Preferably, the design domain is a symmetric or asymmetric structure, the structural unit is a cylinder or an equivalent geometric body, and the two-dimensional array-type initial layout is an orthogonal or non-orthogonal array composed of multiple rows and columns of structural units. By adjusting the design variables of the structural units, the velocity distribution and temperature gradient of the initial flow field are affected to guide the gradient direction of the objective function in the early stage of topology optimization and provide guiding initial conditions for subsequent topology optimization.

[0017] S200. Generate a high-fidelity training sample dataset: Perform random sampling within the adjustable parameter space to generate a predetermined number of initial layout samples with different geometric parameter combinations. For each sample, conduct CFD numerical simulation under high Reynolds number conditions, solve the flow field and temperature field under the preset high Reynolds number turbulence and conjugate heat transfer boundary conditions, and extract its macroscopic performance indicators, including at least the average or maximum temperature characterizing the heat dissipation performance and the inlet and outlet pressure drops characterizing the flow performance, to construct a high-fidelity training dataset containing the geometric parameter combinations of the initial layout and their corresponding macroscopic performance indicators.

[0018] In the embodiments of the present invention, the generation of the high-fidelity training sample dataset at least includes the following sub-steps: S201. Perform Latin hypercube sampling: Use the Latin hypercube sampling method to perform random sampling within the adjustable parameter space to generate several groups of training samples and test samples, and each group of samples corresponds to a configuration of the geometric parameter combination of an initial layout; S202. Set the CFD numerical simulation boundary conditions: Construct a corresponding two-dimensional or three-dimensional computational domain model for each initial layout sample, and configure unified high Reynolds number turbulence and conjugate heat transfer boundary conditions, including at least the inlet high Reynolds number, inlet temperature, outlet pressure, heat source intensity, and characteristic length; S203. Perform multi-physics field coupling numerical solution: Under the preset boundary conditions, perform CFD numerical simulation at high Reynolds number for each initial layout sample, use the k-ε turbulence model to describe the turbulent flow characteristics at high Reynolds number, and at the same time consider the conjugate heat transfer coupling effect between the solid domain and the fluid domain, and calculate the velocity field, pressure field distribution in the fluid domain and the temperature field distribution of the solid domain and the fluid domain simultaneously; S204. Extract key macroscopic performance indicators: Extract macroscopic performance indicators from the numerical solution results, including at least the average or maximum temperature indicator characterizing the heat dissipation performance and the inlet and outlet pressure drop indicator characterizing the flow performance, to form a structured data record containing the one-to-one correspondence between the initial layout geometric parameter combination configuration and the macroscopic performance indicator output; S205. Construct a high-fidelity training dataset: Use the initial layout geometric parameter combination as the input feature vector and the corresponding macroscopic performance indicator as the output label vector to construct a high-fidelity training dataset containing the input-output mapping relationship, and divide part of the data into training data and the rest into test data.

[0019] S300. Establish and train a mainstream feature prediction model: Based on the high-fidelity training dataset, construct and train a reduced-order proxy model of a convolutional neural network (SE-Net) with SE attention mechanism. Through feature squeezing, excitation, and impulse calibration operations, establish the mapping prediction relationship between the initial layout geometric parameter combination and the macroscopic performance indicator output under high Reynolds number conditions, enabling the model to quickly evaluate the macroscopic performance of the initial layout without explicitly analyzing the flow field.

[0020] In the embodiments of the present invention, as Figure 2 shown, the establishment and training of the SE-Net reduced-order proxy model include: S301. Construct the SE-Net architecture: Construct a convolutional neural network structure with SE attention mechanism, including an input layer, a convolutional operation layer, a feature channel SE attention branch, a backbone activation branch, and an output layer. The input layer receives the feature input tensor encoded by the initial layout geometric parameters, extracts local spatial features through the first-layer convolutional operation, and sends them to the parallel SE attention path and the backbone feature path; S302. Construct the SE attention mechanism branch: In the SE attention path, perform squeezing and excitation operations in sequence. The squeezing operation squeezes the features of the initial layout geometric parameters through global average pooling, compresses the spatial dimension information of each feature channel into a scalar value, and generates a compact feature representation with a channel dimension descriptor; the excitation operation uses a gating mechanism composed of two fully connected layers to perform excitation operations on the squeezed feature descriptor. The first fully connected layer performs dimensionality reduction through the ReLU activation function, and the second fully connected layer generates a channel weight coefficient between 0 and 1 through the sigmoid activation function. This weight coefficient reflects the relative importance degree of each feature channel for the prediction target; S303. Extract and fuse the backbone features: In the main path, a convolutional operation is followed by a ReLU activation function, and this is repeated and stacked several times to extract deep features. Subsequently, an element-wise weighted fusion is performed with the output of the attention branch to achieve feature calibration based on importance weights, highlighting the feature channels that contribute highly to the prediction task and suppressing unimportant feature information. The fused feature map is fed into a multi-layer fully connected network; S304. Deep network training and optimization: The fused output passes through a multi-layer fully connected network in sequence. The ReLU activation function is used between layers, and a Dropout layer (such as discarding 5% of the neurons) is added in the middle to prevent overfitting. The final output layer is a continuous value output node for predicting the response of the macroscopic performance indicators of the initial layout. The SE-Net network is trained end-to-end using the backpropagation algorithm, with the optimization goal of minimizing the mean square error between the predicted value and the true value. The network parameters are continuously updated through the gradient descent method until convergence; S305. Model verification and generalization evaluation: Use an independent test dataset to verify the performance of the trained SE-Net reduced-order surrogate model, and statistically calculate the root mean square error RMSE and the coefficient of determination R 2 of the relevant macroscopic performance indicators, and evaluate the prediction stability of the model in each region of the sample space to ensure that it has good generalization ability.

[0021] S400. Construct a modified objective function under the assumption of frozen flow field: Define a modified objective function under the assumption of frozen flow field for comprehensively evaluating the macroscopic performance of the initial layout to achieve the coordinated optimization of heat dissipation performance and pressure drop characteristics. The modified objective function is a multi-objective evaluation function, which at least includes the average or maximum temperature, inlet and outlet pressure drop, and the proportion of the structural unit area index, and is constructed by combining an empirical correction factor and a normalized weight.

[0022] In the embodiment of the present invention, the modified objective function is the objective function corrected relative to the sample standard under the assumption of frozen flow field, and its mathematical expression is: Among them, Π is the comprehensive performance objective function, are respectively the average or maximum temperature and the inlet and outlet pressure drop under the current initial layout sample, are respectively the temperature and pressure drop reference values of the reference sample, w i and w j are respectively the weight coefficients of the temperature and pressure drop targets and w i + w j = 1, S i is thei The area of a structural unit, n is the number of structural units, S all is the maximum total area allowed for the structural units; b 1 is the temperature difference correction factor and , are the average temperatures of the solid region and the fluid region, respectively; M is the normalized standard deviation of the current sample structure parameters and , x i is the i th size parameter (such as radius) of the structural unit, is the average value of the sizes of all structural units, L max is the maximum allowable size of the structural unit.

[0023] S500. ETO global optimization to screen the optimal initial layout: Sampling is carried out in the parameter space to generate a large number of potential initial layout samples. The ETO algorithm is used to iteratively optimize the potential initial layout sample space. When evaluating the quality of candidate solutions in each iteration, the trained SE-Net reduced-order surrogate model is called for rapid prediction. Through large-scale sample screening, the initial layout that satisfies the optimal solution of the objective function is selected as the initial state of topology optimization.

[0024] Preferably, the ETO algorithm adopts a global search strategy based on the combination of exponential function and trigonometric function. By iteratively optimizing in the large-scale potential initial layout sample space, the dynamic adjustment of the search range is controlled by the exponential decay factor, and the comprehensive exploration of the solution space is realized by combining the periodic characteristics of the trigonometric function. The SE-Net reduced-order surrogate model is called for rapid performance evaluation during the iteration process to reduce the number of CFD simulation calls and achieve efficient iterative optimization.

[0025] In the embodiments of the present invention, as Figure 3 shown, when globally optimizing and screening the optimal initial layout based on the ETO algorithm, it at least includes the following sub-steps: S501. Initialize candidate solutions: A large number of potential initial layout samples are generated in the parameter space in a uniform or random manner as the initial candidate solution population, and each candidate solution corresponds to a set of initial layout geometric parameter combinations; at the same time, the key control parameters of the ETO algorithm are initialized, at least including the maximum number of iterations T max, the current iteration number t (initially set to 0), the exploration factor CM (a dynamic parameter that controls the balance between global exploration and local exploitation), the convergence control parameter CEi (a specific iteration threshold used to trigger the update of the search space), and the phase switching parameter Ti (the iteration boundary point that distinguishes the exploration phase from the exploitation phase); S502. Create initial parameters and calculate fitness and the best value: Call the SE-Net reduced-order surrogate model to perform fast performance prediction on each candidate solution, obtain the corresponding macroscopic performance indicators, calculate the fitness value of each candidate solution using the modified objective function, and identify the solution with the optimal fitness value in the current population as the global optimal solution F best , and initialize the position update control parameters α 1 (the first parameter in the exploration phase, which controls the large-scale search intensity), α 2 (the second parameter in the exploration phase, which adjusts the diversity of the search direction), α 3 (the parameter in the exploitation phase, which controls the local fine search range) for subsequent exponential trigonometric transformation operations; S503. Judge the convergence condition and perform CM update: Check whether the current iteration number t reaches the maximum iteration number T max limit. If t ≥ T max then jump to step S507 to return the best solution. If t < T max then dynamically update the exploration factor CM according to the current iteration progress, and achieve a smooth transition of the algorithm from global exploration to local exploitation through the adaptive adjustment of the CM value; S504. Execute the conditional judgment and parameter calculation branch: First, judge whether the current iteration number t is equal to the convergence control parameter CEi. If t = CEi, update the search space range by reconstruction or perturbation, and then continue with the subsequent judgment. Then, judge whether the exploration factor CM is greater than 1. If CM > 1, enter the exploration phase and calculate the position update parameters α 1 and α 2 for large-scale global search and exploration of the new solution space. If CM ≤ 1, enter the exploitation phase and calculate the update parameter α3 for fine-grained local search and solution improvement near the current optimal solution; S505. Implement the phased position update strategy: Perform the corresponding phase position update according to the judgment result of step S504: In the exploration phase, if t < Ti, use the parameters α 1 and α 2 combined with the current best solution F best to update the individual positions in the first and second exploration phases in sequence, and achieve large-scale solution space exploration through the combination of exponential and trigonometric functions; In the exploitation phase, if t < Ti, use the parameterα 3 Update the individual positions in the first and second development stages in sequence, and achieve the in-depth optimization of the solution through the local fine search around F best ; S506. Calculate the fitness and update the optimal solution: Re - call the SE - Net reduced - order surrogate model for performance prediction and fitness evaluation of all candidate solutions after position update, compare the fitness value of the new solution with the fitness of the current global optimal solution F best ; if a better solution is found, update F best to the new optimal solution, and record the corresponding geometric parameter combination configuration at the same time; S507. Execute the stop condition judgment and return the best solution: Repeat the iterative loop of steps S503 - S506 until the preset convergence criterion is met, including that the current iteration number t reaches the maximum iteration number T max , the fitness improvement amplitude of the global optimal solution F best is less than the preset threshold for several consecutive generations and other termination conditions, and finally return the geometric parameter combination configuration corresponding to F best with the optimal objective function value as the initial layout of the topology optimization.

[0026] S600. Execute the topology optimization based on the mainstream - guided initial layout: Use the obtained optimal initial layout as the initial field of the topology optimization, adopt the topology optimization method based on density field interpolation, combine the conjugate heat transfer control equation, the k - ε turbulence model at high Reynolds numbers and the fluid momentum equation modified by inverse magnetic permeability, and combine the boundary conditions of the convection - heat conduction dual domain of the structure field to execute the topology optimization iteration process, optimize the density distribution function to achieve the optimal configuration of materials in the design domain, and realize the global optimal evolution of the topology structure under the goal of heat - flow synergy.

[0027] In the embodiment of the present invention, as Figure 4 shown, the topology optimization of the initial layout includes the following sub - steps: S601. Construct the initial density field of the topology optimization: Take the optimal initial layout obtained by the ETO algorithm as the input of the initial density field distribution of the topology optimization, map the geometric parameter combination configuration of the structural unit to the initial material density value of the corresponding grid cell in the topology design domain, construct the continuous material density distribution function ρ(x), where ρ ∈ [0, 1], where ρ(x)=0 corresponds to the completely fluid region, ρ(x)=1 corresponds to the completely solid region, and the intermediate value represents the porous medium or the transition region, and smooth the density value near the cylinder boundary through the distance function; S602. Construct the multi - physical - field coupling analysis framework: Establish a multi - physical field analysis framework for high - Reynolds - number complex thermal - fluid coupling, including: the Navier - Stokes momentum conservation equation for solving the velocity distribution and pressure distribution of fluid in the design domain; the turbulence transport equation set based on the standard k - ε model, which is used to describe the energy change and dissipation process of turbulent perturbations at high Reynolds numbers; the convective - conductive coupling equation set based on energy conservation, which is used to describe the temperature field distribution in the solid domain and fluid domain and the heat exchange process between the two.

[0028] S603. Construct a thermal - fluid collaborative topology optimization model: Aiming to achieve the collaborative optimum of heat dissipation performance and flow performance, construct a thermal - fluid collaborative topology optimization mathematical model with the material density function ρ(x) as the design variable. The constraint conditions at least include the material volume fraction limit, the fluid domain connectivity constraint, and the fluid domain boundary integrity constraint; and based on the material density interpolation (Solid Isotropic Material with Penalization, SIMP, Rational Approximation of Material Properties, RAMP) model, establish the mapping relationship between physical properties (such as thermal conductivity, flow resistance) and design variables to ensure the differentiability of structure - fluid properties during the topology evolution process and their applicability to the gradient iteration algorithm. S604. Set boundary conditions and apply loads: Set unified conjugate boundary conditions, at least including the inlet velocity or pressure boundary, the outlet free pressure boundary, the adiabatic boundary or temperature boundary of the structure wall, and apply a uniform volume heat source term Q in the specified area inside the structure to simulate the heating element under actual working conditions. S605. Define the comprehensive performance objective function and perform sensitivity analysis and gradient calculation: First, define the comprehensive performance objective function, and its mathematical expression is , where F is the comprehensive performance objective function, are the average temperature and the inlet - outlet pressure drop of the current topology respectively, a , b are the corresponding weight coefficients and a + b = 1, are the average temperature and the inlet - outlet pressure drop at the first step when the initial density field is 0.5 respectively; considering the influence of the penalty factor in the SIMP and RAMP interpolation models on sensitivity calculation, use the adjoint method for topology sensitivity analysis, obtain the gradient information of topology optimization by solving the adjoint temperature field equation and the adjoint flow field equation, and calculate the gradient of the objective function F with respect to the material density ρ(x) using the chain - rule of differentiation. , which is used to guide the optimization direction of the material distribution in the iterative step, where T is the temperature and u is the velocity; S606. Implement density filtering and projection processing: Perform density filtering operation on the design variables to eliminate the checkerboard phenomenon and mesh dependence problems. Use the Helmholtz filter ( , where r is the filtering radius, γ is the material density design variable, is the filtered design variable, Ω is the design domain, is the boundary of the design domain, n is the boundary normal vector) to smooth the material density distribution. Subsequently, apply the hyperbolic tangent projection function ( , where is the density value after projection processing, is the filtered material density value, β and η are the projection parameter and threshold respectively) to project the filtered density value onto the 0 - 1 interval, and achieve a gradual transition from a fuzzy boundary to a clear boundary by adjusting the projection parameter and threshold; S607. Update the design variables and check the constraint conditions: Based on the sensitivity analysis results, use a gradient - based mathematical programming algorithm, such as the Method of Moving Asymptotes (MMA), to update the material density distribution design variables. The iterative formula is , where α is the step - size parameter, are the material densities at the iterative steps k and k + 1 respectively. In each iteration, check whether the constraint conditions are satisfied, and correct the design variables that violate the constraints. Use the penalty function method to handle inequality constraints, such as using the Lagrange multiplier method for correction. The algorithm formula is , where are the Lagrange multipliers at the iterative steps k and k + 1 respectively, μ is the penalty parameter, is the inequality constraint function; S608. Judge the convergence and output the optimization result: Set the convergence judgment criteria, including the threshold of the objective function change rate, the maximum number of iterations T, or the density gradient convergence criterion; if the convergence conditions are met, terminate the optimization process and output the final topology optimization result. If not converged, return to step S602 to continue iterative optimization, and finally obtain the topology structure of the material distribution that is co - optimally balanced between heat dissipation performance and flow performance under the given constraint conditions; S609. Topology structure extraction and post - processing: Perform threshold segmentation on the final density distribution result to form a clear topological boundary, and extract the manufacturable cooling channel geometry; perform CFD simulation based on the obtained structure to verify its temperature distribution, pressure drop characteristics, and mainstream uniformity, and evaluate its thermal-fluid coupling performance under complex conditions of high Reynolds number.

[0029] Preferably, in step S602, in the momentum conservation equation, an artificial porous medium Darcy-type penalty term related to the material density is introduced and corrected through an inverse magnetic permeability function to reflect the change in flow resistance under different material densities; in the convective-conductive coupling equations, the heat conduction and convective heat transfer mechanisms are defined in the solid domain and the fluid domain respectively, and the heat capacity, density, and thermal conductivity are interpolated through the material density function to achieve continuous regulation of the thermal physical properties of the structure.

[0030] Specifically, the fluid flow control equations include: Continuity equation (mass conservation): , where u is the fluid velocity vector.

[0031] Momentum conservation equation (introducing porous medium resistance term): Among them, ρ is the fluid density, I is the unit tensor, μ is the molecular viscosity, μ T is the turbulent viscosity coefficient, α ( γ ) is the Darcy damping coefficient associated with the material density design variable γ ; Turbulent kinetic energy k transport equation: Among them, k is the turbulent kinetic energy, σ k is the Prandtl number of the turbulent kinetic energy (standard value is 1.0), P k is the turbulent kinetic energy generation term, ε is the turbulent dissipation rate, α f ( γ ) is the Darcy-type artificial resistance coefficient for the turbulent kinetic energy equation; Turbulent dissipation rate ε Transport equation: In the formula, σ ε is the Prandtl number of the turbulent dissipation rate (standard value is 1.3), C 1ε 、 C2ε are model constants (standard values are 1.44 and 1.92 respectively), α ε ( γ ) is the artificial drag coefficient for the turbulent dissipation rate equation; Turbulent viscosity calculation formula: , C μ is a turbulent model constant (standard value is 0.09); Turbulent kinetic energy generation term: , where ":" represents the tensor double dot product operation.

[0032] In addition, to achieve continuous expression of structure-fluid properties, an inverse magnetic permeability correction form of the artificial porous medium is introduced into the above momentum and turbulence equations. The interpolation expression of the inverse magnetic permeability type parameters is as follows: In the formula, γ is the material density field, α (·) represents the inverse magnetic permeability of the artificial porous medium for different equation terms in the solid / fluid domain, α fmax , α kmax , α εmax are the maximum drag coefficients in the solid region of the momentum equation, turbulent kinetic energy equation, and turbulent dissipation rate equation respectively, α fmin , α kmin , α εmin are the minimum drag coefficients in the fluid region of the corresponding equations respectively. γ is the material density design variable (value range 0 - 1), p α is the penalty factor (usually takes the value 0.01).

[0033] In terms of heat conduction, the conjugate heat transfer model is used to describe the heat conduction and convective heat transfer mechanism between the solid and fluid regions. Its control equation is as follows: Energy conservation equation (conjugate heat transfer model): Among them, T is the temperature field, Q is the volume heat source term, ρ ( γ )、 c ( γ )、 k ( γ ) are the density, specific heat capacity, and thermal conductivity related to the density design variable γ respectively. Material property interpolation function: , ,

[0034] Wherein: ρ f is the fluid density, ρ s is the solid density, p 1 is the penalty factor for density interpolation; k f is the fluid thermal conductivity, k s is the solid thermal conductivity, p 2 is the penalty factor for thermal conductivity interpolation; c f is the fluid specific heat capacity, c s is the solid specific heat capacity, p 3 is the penalty factor for specific heat capacity interpolation ( p 1, p 2, p 3 usually takes values from 1 to 5); The thermal boundary conditions are set as follows: Among them, the two expressions respectively represent the inlet temperature boundary and the adiabatic boundaries of the outlet and the wall. Γ in , Γ out , Γ w respectively represent the inlet, outlet and wall boundaries, T0 is the inlet temperature, and n is the boundary normal vector.

[0035] Further preferably, in step S600, the initial layout obtained by optimization is used to guide the direction of topology optimization. By jointly optimizing the mainstream velocity field distribution and the temperature field distribution, the target value is continuously decreased in the conventional weighted evaluation function, so as to continuously update the topology until it is stable.

[0036] In summary, Example 1 realizes a full-process solution from the initial flow field guidance to the final structure optimization through key technical links such as constructing a parametric array structure, training an E-Net reduced-order surrogate model, an ETO intelligent optimization algorithm, and conjugate heat transfer topology optimization, providing an effective method framework and implementation plan for the multi-physics collaborative optimization under complex heat dissipation conditions.

[0037] Example 2: Application Example To verify the effectiveness and practicality of the above topological optimization improvement method, in this Example 2, a typical liquid-cooled plate radiator design scenario is selected as an application case. By constructing a parametric array in an asymmetric design domain and combining the SE-Net network with the ETO algorithm, the initial layout design and topological optimization iteration are completed to verify the superior thermal-fluid collaborative performance of the present invention.

[0038] Based on the topological optimization improvement method framework proposed in Example 1, in this example, an asymmetric liquid-cooled radiator is taken as the specific application object. Its design domain Ω is asymmetric, with one inlet and one outlet, as Figure 5 shown. The inlet temperature is known to be 293.15 K, the outlet pressure P0 = 0 Pa, the inlet width L is set to 6 mm, the heat source is 10^7 W / m 3 , and the inlet Reynolds number is 14000.

[0039] First, a 5×10 cylindrical array is constructed in the geometric region. The radius of each cylinder ranges from 0.001 m to 0.005 m, and the interval between cylinders is 0.02 m. Figure 6 The flow field diagram at an inlet Reynolds number of 14000 when the radius of each cylinder in the cylindrical array is 0.005 m is given. Then, using the Latin hypercube sampling method, the radius of each cylinder is randomly set between 0.001 m and 0.005 m to construct 300 sets of data as the training set, and 50 sets of data are constructed as the test set in the same way.

[0040] Subsequently, a neural network with SE attention mechanism (SE-Net) is constructed. The neural network with SE attention mechanism consists of squeezing, excitation, and impulse calibration. The algorithm flow chart is as Figure 2 shown. By training the SE neural network, the errors of the training set and test set of the average temperature and the pressure drop at the inlet and outlet are obtained respectively. The following figures are the errors of the training set and test set of the average temperature ( Figure 7 ) and the pressure drop at the inlet and outlet ( Figure 8 ). As Figure 7 shown, Figure (a) shows the average temperature prediction results of 300 samples in the training set, and Figure (b) shows the average temperature prediction results of 50 samples in the test set. The RMSE of the average temperature is 0.032 and 0.264 respectively, and the determination coefficient R 2 is 0.9977 and 0.902 respectively. The closer R² is to 1, the better the model fitting effect. As Figure 8 shown, Figure (a) shows the inlet and outlet pressure drop prediction results of 300 samples in the training set, and Figure (b) shows the inlet and outlet pressure drop prediction results of 50 samples in the test set. The RMSE of the inlet and outlet pressure drop is 2.67 and 12.21 respectively, and the determination coefficient R 2They are 0.998 and 0.976 respectively. The above results show that the SE-Net reduced-order surrogate model has achieved high prediction accuracy on both the training set and the test set, and can effectively replace the CFD simulation with high computational cost for rapid performance evaluation.

[0041] After the model training is completed, hundreds of thousands of potential samples are randomly generated by the Latin hypercube sampling method to construct the initial layout, and the optimal solution in the evaluation function is obtained using the ETO algorithm. The flowchart of the ETO algorithm is as Figure 3 shown. The mathematical expression of its modified objective function Π is: Among them, are the average or maximum temperature and the pressure drop at the inlet and outlet under the current initial layout sample respectively, are the temperature and pressure drop reference values of the reference sample respectively, w i 、 w j are the weight coefficients of the temperature and pressure drop targets respectively and w i + w j = 1, b 1 is the temperature difference correction factor and ; M is the normalized standard deviation of the current sample structure parameters, defined as In this embodiment, L max is set to 0.005 m, are the average temperature of the solid and the average temperature of the liquid respectively, where b 1 the estimated value is adopted, which is 0.08. S i 、 S all are the current total array area and the maximum array area respectively. are the temperature and pressure drop reference values of the reference sample respectively. In this embodiment, respectively represent the average temperature of the design domain and the pressure drop at the inlet and outlet when the array cylinder radius is 0.008 m. w i 、 w j represent the weight factors, which are 0.6 and 0.4 respectively here.

[0042] Through the optimization results of the SE-Net network and the ETO algorithm under the modified objective function, two target array layouts are obtained, as Figure 9As shown. The two target arrays exhibit obvious characteristics of non-uniform cylindrical radius distribution: in the inlet region of the flow field, the cylindrical radius is relatively small, which is beneficial to reducing the flow resistance and inlet pressure drop; in the middle region, the cylindrical size changes in a gradient manner, ensuring both sufficient heat transfer area and avoiding excessive flow separation; near the outlet, the cylindrical layout is carefully optimized to improve the flow uniformity. Both array layouts avoid the problems of large-scale flow separation and pressure drop concentration commonly found in traditional uniform arrays, and achieve the collaborative optimization of heat dissipation performance and flow performance through the non-uniform cylindrical radius distribution, providing an initial layout basis with good mainstream feature guidance for subsequent topology optimization.

[0043] Using the above two layouts as the initial layouts, topology optimization is carried out under the comprehensive performance objective evaluation function, and the results are comprehensively compared with the benchmark structure. The comparison objects include six types of typical channel layouts such as a bionic topology structure with an initial density of 0.5, a regular array type structure (including cylindrical and square column array channels), two groups of topology channels guided by mainstream features, and a regular rectangular channel. A uniform volumetric heat source (Q = 10^7 W / m 3 ) is applied in the design domain, the inlet Reynolds number is 14000, and all are adiabatic walls. The boundary conditions are the same among different models. The k-ε turbulence physical model and the conjugate heat transfer physical model are used to evaluate the hydraulic and heat transfer performance of the final design. The flow field characteristics are as Figure 10 shown, and the temperature field characteristics are as Figure 11 shown.

[0044] When the Reynolds number is as high as 14000, the pressure drops of the two groups of topology channels optimized by this method are 1467 Pa and 1530 Pa, and the relative pressure drop is reduced by nearly 80% compared with the square array channel and by nearly 73% compared with the circular channel, indicating that the topology channels play a very good role in improving the pressure drop. It can be seen from Figure 10 that many small vortices are generated at the walls of the rectangular array channel and the circular array channel, increasing the flow dissipation. While for the flow channels of the three topologies and the rectangular channel, they do not cause frequent flow separation and reattachment of the fluid like the circular array and square array channels, significantly reducing the pressure loss. At the same time, the topology channels have relatively smooth boundaries, making the fluid flow smoother during the turning process, further reducing the generation of vortices and better reducing the pressure drop.

[0045] When the Reynolds number is as high as 14000, the highest temperatures in the two groups of topology regions optimized by this method are 325 K and 330 K, and the lowest highest temperature of the other channels is 341 K, indicating that the topology channels optimized by this method play an obvious role in improving hot spots. It can be seen from Figure 11It can be significantly observed that the hot spot distributions of different flow channels are different. The hot spots of the topological flow channels optimized by this method are all in the upper left corner, which is caused by the forked flow that attenuates from the upper left. The hot spots of the square array and cylindrical array channels are both on the right, which is caused by the complex flow separation and the flow non-uniformity caused by the attachment. The hot spots of the rectangular channels are in the upper right, which is caused by the flow non-uniformity. The hot spots of the flow channels of the traditional topological optimization are in the lower right, which is caused by the refraction of the flow to the middle end.

[0046] In summary, the topological optimization model proposed by the present invention, which combines the SE-Net reduced-order surrogate model and the ETO optimization algorithm, can well solve the topological ill-conditioning problem of the topological optimization method of liquid-cooled plate radiators in the case of high Reynolds number turbulence with complex geometries. By constructing a small-area array field layout different from the traditional bionic mode, the topological results can be significantly improved, providing an important reference for synergistically and significantly improving the heat transfer performance and flow performance of liquid-cooled plate radiators under turbulence, and providing a feasible idea in the case of more complex flows. Without a large increase in the number of grids and the trial and error of continuous changes in cumbersome topological parameters, a better result can be obtained by guiding the optimization direction of the gradient algorithm while saving a large amount of time.

[0047] Through the above embodiments, the object of the present invention is fully and effectively achieved. Those skilled in the art can understand that the present invention includes but is not limited to the content described in the drawings and the above specific embodiments. Although the present invention has been described with respect to the currently considered most practical and preferred embodiments, it should be understood that the present invention is not limited to the disclosed embodiments, and any modification that does not deviate from the functional and structural principles of the present invention will be included in the scope of the claims.

Claims

1. A topological optimization improvement method for the initial layout of extracting mainstream features under complex heat dissipation conditions, characterized in that, At least include the following steps: S100. In the liquid cooling cavity design domain of the liquid cooling plate radiator, construct a two-dimensional array-type initial layout formed by arranging multiple structural units along the mainstream direction and the lateral direction. Set one or more of the radius, spacing, and arrangement method of the structural units as design variables, and establish an adjustable parameter space; S200. Conduct random sampling within the adjustable parameter space to generate a predetermined number of initial layout samples with different geometric parameter combinations; perform CFD numerical simulations for each sample under high Reynolds number conditions, and extract macroscopic performance indicators to construct a high-fidelity training data set containing geometric parameter combinations and their corresponding macroscopic performance indicators; S300. Based on the high-fidelity training data set, construct and train an SE-Net reduced-order surrogate model to establish a mapping prediction relationship between the geometric parameter combinations of the initial layout and the output of macroscopic performance indicators under high Reynolds number conditions, enabling the model to quickly evaluate the macroscopic performance of the initial layout without explicitly analyzing the flow field; S400. Define a modified objective function under the flow field freezing assumption for comprehensively evaluating the heat dissipation performance and pressure drop characteristics of the initial layout, at least including average or maximum temperature, inlet and outlet pressure drops, and the proportion of the structural unit area index, and construct it in combination with an empirical correction factor and a normalized weight; S500. Sample and generate a large number of potential initial layout samples within the adjustable parameter space, use the ETO algorithm to iteratively optimize the potential initial layout sample space, and call the SE-Net reduced-order surrogate model for rapid prediction when iteratively evaluating candidate solutions, and finally select the initial layout that satisfies the optimal solution of the objective function; S600. Take the selected initial layout as the initial field of topology optimization, construct a topology optimization model based on density field interpolation, introduce the k-ε turbulence model, the inverse magnetic permeability corrected momentum equation, and the conjugate heat transfer control equation, and perform density function evolution to achieve the optimal configuration of the materials in the design domain.

2. The topological optimization improvement method for the initialization layout of extracting mainstream features under complex heat dissipation conditions according to claim 1, characterized in that In step S100, the design domain is a symmetric or asymmetric structure, the structural unit is a cylinder or an equivalent geometric body, and the two-dimensional array-type initial layout is an orthogonal or non-orthogonal array composed of multiple rows and columns of structural units. By adjusting the design variables of the structural units, the velocity distribution and temperature gradient of the initial flow field are affected.

3. The topological optimization improvement method for the initialization layout of extracting mainstream features under complex heat dissipation conditions according to claim 1, characterized in that, In step S200, the generation of the high-fidelity training sample data set at least includes the following sub-steps: S201. Use the Latin hypercube sampling method to conduct random sampling within the adjustable parameter space to generate several groups of training samples and test samples, and each group of samples corresponds to a configuration of an initial layout geometric parameter combination; S202. For each initial layout sample, construct its corresponding two-dimensional or three-dimensional computational domain model, and configure unified high Reynolds number turbulence and conjugate heat transfer boundary conditions, at least including inlet high Reynolds number, inlet temperature, outlet pressure, heat source intensity, and characteristic length; S203. Conduct CFD numerical simulations at high Reynolds numbers for each initial layout sample under preset boundary conditions. The k-ε turbulence model is used to describe the turbulent flow characteristics at high Reynolds numbers, and the conjugate heat transfer coupling effect between the solid domain and the fluid domain is considered simultaneously. At the same time, calculate the velocity field and pressure field distributions in the fluid domain, as well as the temperature field distributions of the solid domain and the fluid domain; S204. Extract macroscopic performance indicators from the numerical solution results, including at least the average or maximum temperature indicator characterizing the heat dissipation performance, and the inlet and outlet pressure drop indicator characterizing the flow performance, to form a structured data record containing the one-to-one correspondence between the initial layout geometric parameter combinations and the macroscopic performance indicator outputs; S205. Use the initial layout geometric parameter combinations as input feature vectors and the corresponding macroscopic performance indicators as output label vectors to construct a high-fidelity training dataset containing the input-output mapping relationship, and divide part of the data into training data and the rest into test data.

4. The topological optimization improvement method for the initialization layout of extracting mainstream features under complex heat dissipation conditions according to claim 1, characterized in that, In step S300, the establishment and training of the SE-Net reduced-order surrogate model at least include the following sub-steps: S301. Construct a convolutional neural network structure with SE attention mechanism, including an input layer, a convolutional operation layer, a feature channel SE attention branch, a backbone activation branch, and an output layer. The input layer receives the feature input tensor encoded by the initial layout geometric parameters, extracts local spatial features through the first-layer convolutional operation, and sends them to the parallel SE attention path and the backbone feature path; S302. In the SE attention path, perform squeezing and excitation operations in sequence. The squeezing operation squeezes the features of the initial layout geometric parameters through global average pooling to generate a compact feature representation with a channel dimension descriptor; the excitation operation uses a gating mechanism composed of two fully connected layers to perform excitation operations on the squeezed feature descriptor. The first fully connected layer performs dimensionality reduction through the ReLU activation function, and the second fully connected layer generates channel weight coefficients between 0 and 1 through the sigmoid activation function; S303. After the convolutional operation in the backbone path, connect the ReLU activation function and repeat the stacking several times to extract deep features; Subsequently, perform element-wise weighted fusion with the output of the attention branch to achieve feature calibration based on importance weights, highlight the feature channels with high contribution to the prediction task, suppress unimportant feature information, and send the fused feature map into a multi-layer fully connected network; S304. The fused output passes through a multi-layer fully connected network in sequence, uses the ReLU activation function between layers, and adds a Dropout layer in the middle to prevent overfitting; the output layer is a continuous value output node for predicting the response of the initial layout macroscopic performance indicators; use the backpropagation algorithm to perform end-to-end training on the SE-Net network, with minimizing the mean square error between the predicted value and the true value as the optimization goal, and continuously update the network parameters through the gradient descent method until convergence; S305. Use an independent test dataset to verify the performance of the trained SE-Net reduced-order surrogate model, and statistically calculate the root mean square error RMSE and the coefficient of determination R of relevant macroscopic performance indicators 2 , and evaluate the prediction stability of the model in each region of the sample space to ensure that it has good generalization ability.

5. The topological optimization improvement method for the initialization layout of extracting mainstream features under complex heat dissipation conditions according to claim 1, characterized in that In step S400, the correction objective function is the objective function corrected relative to the sample standard under the frozen flow field freezing assumption, and its mathematical expression is: where, Π is the comprehensive performance objective function, are respectively the average or maximum temperature and the pressure drop at the inlet and outlet under the current initial layout sample, are respectively the temperature and pressure drop reference values of the reference sample, w i 、 w j are respectively the weight coefficients of the temperature and pressure drop targets and w i + w j = 1, S i is the area of the i th structural unit, n is the number of structural units, S all is the maximum total area allowed for the structural units; b 1 is the temperature difference correction factor and , are respectively the average temperatures of the solid region and the fluid region; M is the normalized standard deviation of the structural parameters of the current sample and , x i is the i th dimension parameter of the structural unit, is the average value of the dimensions of all structural units, L max is the maximum allowable dimension of the structural unit.

6. The topological optimization improvement method for the initialization layout of extracting mainstream features under complex heat dissipation conditions according to claim 1, characterized in that In step S500, the ETO algorithm adopts a global search strategy based on the combination of exponential function and trigonometric function. By iteratively optimizing in the large-scale potential initial layout sample space, it uses an exponential decay factor to control the dynamic adjustment of the search range, and combines the periodic characteristics of the trigonometric function to achieve a comprehensive exploration of the solution space. During the iteration process, the SE-Net reduction surrogate model is called for rapid performance evaluation.

7. The topological optimization improvement method for the initialization layout of extracting mainstream features under complex heat dissipation conditions according to claim 1 or 6, characterized in that In step S500, when performing global optimization based on the ETO algorithm to screen the optimal initial layout, it at least includes the following sub-steps: S501. Generate a large number of potential initial layout samples in the parameter space in a uniform or random manner as the initial candidate solution population, where each candidate solution corresponds to a set of initial layout geometric parameter combinations; meanwhile, initialize the key control parameters of the ETO algorithm, including at least the maximum number of iterations T max , the current number of iterations t, the exploration factor CM, the convergence control parameter CEi, and the stage switching parameter Ti; S502. Call the SE-Net reduced-order surrogate model to quickly predict the performance of each candidate solution, obtain the corresponding macroscopic performance indicators, calculate the fitness value of each candidate solution using the modified objective function, and identify the solution with the optimal fitness value in the current population as the global optimal solution F best , initialize the position update control parameter α 1、 α 2、 α 3 is used for subsequent exponential trigonometric transformation operations; S503. Check whether the current iteration number t has reached the maximum iteration number T max Restriction: if t≥T max then jump to step S507 to return the best solution; if t<T max then dynamically update the exploration factor CM, and achieve a smooth transition of the algorithm from global exploration to local development through the adaptive adjustment of the CM value; S504. First, it is judged whether the current iteration number t is equal to the convergence control parameter CEi. If t = CEi, the search space range is updated, and then the subsequent judgment is continued; then it is judged whether the exploration factor CM is greater than 1. If CM > 1, it enters the exploration stage and calculates the position update parameter α 1 and α 2. If CM ≤ 1, it enters the exploitation stage and calculates the update parameter α3; S505. Perform position updates for the corresponding stage according to the judgment result of step S504: In the exploration stage, if t < Ti, use the parameters α 1 and α 2 to combine with the current best solution F best Update the individual positions in the first and second exploration stages in sequence to achieve large-scale solution space exploration through the combination of exponential function and trigonometric function; In the development stage, if t < Ti, the parameters α 3 are used to update the individual positions in the first and second development stages in sequence, and the depth optimization of the solution is realized by local fine search around F best . S506. Re - call the SE - Net reduced - order surrogate model for performance prediction and fitness evaluation for all candidate solutions after position update, and compare the fitness value of the new solution with the fitness of the current global optimal solution F best If a better solution is found, update F best as the new optimal solution; Iteratively repeat the loop of steps S503 to S506 until a preset convergence criterion is met, and finally return F with the optimal objective function value. best The corresponding geometric parameter combination configuration.

8. The topological optimization improvement method of the initialization layout for extracting mainstream features under complex heat dissipation conditions according to claim 1, characterized in that In step S600, the initial layout topology optimization at least includes the following sub-steps: S601. Taking the optimal initial layout obtained by the ETO algorithm as the input of the initial density field distribution of topology optimization, mapping the geometric parameter combination configuration of the structural unit to the initial material density value of the corresponding grid unit in the topology design domain, and constructing a continuous material density distribution function ρ(x), where ρ ∈ [0, 1]; S602. Establish a multi-physics field analysis framework for complex thermal-fluid coupling at high Reynolds numbers, including: the Navier-Stokes momentum conservation equation for solving the velocity distribution and pressure distribution of the fluid in the design domain; the turbulence transport equation set based on the standard k-ε model; the convection-conduction coupling equation set based on energy conservation; S603. Aiming to achieve the collaborative optimum of heat dissipation performance and flow performance, construct a thermal-fluid collaborative topology optimization mathematical model with the material density function ρ(x) as the design variable. The constraint conditions at least include material volume fraction limitation, fluid domain connectivity constraint, and fluid domain boundary integrity constraint; and establish a mapping relationship between physical properties and design variables based on the SIMP and RAMP interpolation models; S604. Set unified conjugate boundary conditions, at least including inlet velocity or pressure boundary, outlet free pressure boundary, adiabatic boundary or temperature boundary of the structural wall, and apply a uniform volume heat source term Q in the specified area inside the structure to simulate the heating element under actual working conditions; S605. Define the comprehensive performance objective function , where are the average temperature and the inlet and outlet pressure drop of the current topology, respectively, a , b are the corresponding weight coefficients and a + b = 1, are the average temperature and the inlet and outlet pressure drop at the first step when the initial density field is 0.5; Considering the influence of the penalty factor in the SIMP and RAMP interpolation models on the sensitivity calculation, the adjoint method is used for topological sensitivity analysis. The gradient information of topological optimization is obtained by solving the adjoint temperature field equation and the adjoint flow field equation, and the gradient of the objective function with respect to the material density ρ(x) is calculated using the chain rule of differentiation; S606. Perform density filtering operation on the design variables, use the Helmholtz filter to smooth the material density distribution, apply the hyperbolic tangent projection function to project the filtered density value into the 0-1 interval, and achieve a gradual transition from a fuzzy boundary to a clear boundary by adjusting the projection parameter and threshold; S607. Based on the sensitivity analysis results, adopt a gradient-based mathematical programming algorithm to update the design variable of the material density distribution. Check the satisfaction of the constraint conditions in each iteration, and correct the design variables that violate the constraints, and use the penalty function method to handle the inequality constraints; S608. Set the convergence judgment criterion. If the convergence condition is met, terminate the optimization process and output the final topology optimization result. If not converged, return to step S602 to continue iterative optimization, and finally obtain the collaborative optimum material distribution topology structure of heat dissipation performance and flow performance under the given constraint conditions; S609. Perform threshold segmentation processing on the final density distribution result to form a clear topology boundary, and extract the manufacturable cooling channel geometric structure.

9. The topology optimization improvement method of the initialized layout for extracting mainstream features under complex heat dissipation conditions according to claim 8, characterized in that, In step S602, an artificial porous medium Darcy-type penalty term related to the material density is introduced into the momentum conservation equation and corrected by an inverse magnetic permeability function; in the coupled convection-heat conduction equations, the heat conduction and convective heat transfer mechanisms are defined in the solid domain and the fluid domain respectively, and the heat capacity, density, and thermal conductivity are interpolated through a material density function.

10. The topological optimization improvement method for the initialization layout of extracting mainstream features under complex heat dissipation conditions according to claim 1, 8 or 9, characterized in that, In step S600, the direction of topology optimization is guided by using the initial layout of the optimization search. By jointly optimizing the mainstream velocity field distribution and the temperature field distribution, the target value continuously decreases in the weighted objective function, so as to continuously update the topology until it is stable.

11. A computer program product, comprising computer instructions, characterized in that, The computer instructions are used to execute the topology optimization improvement method for the initial layout of extracting the mainstream features described in any one of claims 1 to 10 under complex heat dissipation conditions.

12. A computer-readable storage medium, on which a computer program is stored, characterized in that, When the computer program is executed by a processor, it implements the topology optimization improvement method for the initial layout of extracting the mainstream features described in any one of claims 1 to 10 under complex heat dissipation conditions.

Images