CW type heat exchanger performance optimization method based on PINN and functional variation method
By coupling PINN with CFD, the problems of local optima and numerical solution stability in traditional heat exchanger design are solved, enabling rapid optimization and efficient computation of heat exchanger performance, while reducing mesh requirements and computational costs.
Patent Information
- Application Number
- CN202511773844.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-28
- Publication Date
- 2026-03-17
AI Technical Summary
Traditional heat exchanger design optimization methods are prone to getting trapped in local optima and are difficult to achieve global optima. Furthermore, traditional numerical methods suffer from non-physical oscillations and stability degradation when solving for negative diffusion terms, making efficient convergence difficult.
A coupled approach of Physical Information Neural Network (PINN) and Computational Fluid Dynamics (CFD) is adopted. By reducing the mesh resolution requirement through transfer learning, and combining functional variational optimization methods, a set of functional equations is constructed. PINN is used to solve the negative diffusion coefficient PDE, thereby achieving coupled optimization of multiphysics fields of flow and heat transfer.
It achieves rapid optimization of heat exchanger performance, reduces the number of meshes required, computation time and cost, and achieves a temperature field error of less than 1%, which is less than 90% of the results of traditional high-resolution mesh numerical solutions.
Smart Images

Figure CN121683076A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for optimizing the performance of a CW-type heat exchanger, belonging to the technical field of heat exchange device performance optimization. Background Technology
[0002] Thermal energy utilization is a core component of energy conversion and power systems. Among the types of resources available to humankind, thermal energy constitutes a major portion, with 80%–90% of energy first converted into thermal energy before utilization. Improving heat transfer efficiency is crucial for energy systems to enhance thermal energy utilization and reduce energy loss, and it is also a prerequisite for ensuring the efficient and stable operation of power systems. For example, in the aerospace field, the extremely high heat flux generated by aircraft under high Mach numbers and combustion environments places stringent demands on thermal management technologies; in nuclear energy systems, structural expansion and cracking of fuel rods due to high temperatures can easily lead to safety hazards; in the fields of microelectronics and high-performance computing, the ever-increasing power density of chips makes traditional heat conduction mechanisms insufficient to meet heat dissipation requirements, severely restricting the improvement of device performance. All of these industrial scenarios urgently require novel heat exchange technologies based on enhanced heat transfer theory to achieve significant increases in heat flux density and temperature uniformity within limited spaces.
[0003] Micro gas turbines offer significant advantages in micro mobile power applications due to their high energy density, but their thermal efficiency is typically below 20% due to their relatively low temperature and pressure ratios. Introducing a regenerative heat exchanger can transfer the waste heat from the high-temperature gas at the turbine outlet to the high-pressure air at the compressor outlet, thus raising the temperature of the air at the combustion chamber inlet. CW-type heat exchangers, commonly used in regenerative processes, are crucial for improving gas turbine efficiency through optimized heat transfer performance. However, achieving a balance between heat transfer efficiency and total pressure loss in multi-objective optimization of heat exchanger performance is challenging; therefore, the industry requires a reasonable mathematical model and governing equations suitable for engineering design.
[0004] Traditional stochastic optimization methods are prone to getting trapped in local optima when solving heat exchanger design optimization problems, failing to guarantee the global optimality of the heat exchanger design. This can lead to excessively large heat exchanger structures and increased equipment investment costs. Moreover, traditional design methods often treat discrete variables such as plate length and width as continuous variables, obtaining a nonlinear programming model and then rounding it off. This approach does not take into account the discrete nature of the design variables during the optimization process, easily resulting in infeasible design solutions.
[0005] Functional variational optimization methods couple the multiphysics characteristics of flow and heat transfer by constructing the extremum conditions of the objective functional, thereby achieving multi-objective synergistic optimization. Variables such as entropy production, heat loss, and heat consumption can all be used as optimization objectives for heat transfer performance to characterize thermodynamic irreversibility. However, regardless of the variables chosen to derive the Euler-Lagrange equations for functional variation, they all exhibit strong nonlinear characteristics. For example, the presence of the negative diffusion term coefficient in the inverse temperature equation destroys the ellipticity of the equation, causing non-physical oscillations in the numerical solution and consequently degrading its stability. Traditional finite volume methods and Newton's iteration methods are limited by the time-step constraints of explicit schemes or the matrix ill-conditioned nature of implicit schemes, making efficient convergence difficult. To address these issues, this paper compares the results with those of functional variational optimization targeting heat transfer entropy production, analyzing the feasibility and advantages of functional variational optimization based on vorticity field synergy. To address the challenge of numerically solving partial differential equations with negative diffusion coefficients, a solution strategy coupling Computational Fluid Dynamics (CFD) and Physical Information Neural Network (PINN) is proposed. This strategy reduces the mesh resolution requirements of functional variational optimization and provides a novel solution for enhancing the design of passive convective heat transfer technologies and solving hybrid conventional-unconventional partial differential equation systems. Summary of the Invention
[0006] To address the problem of low efficiency in heat exchanger design in existing engineering projects, this invention proposes a performance optimization method for CW-type heat exchangers based on PINN and functional variational methods.
[0007] The technical solution adopted by the present invention to solve the above problems is as follows: The steps of the present invention include: Step 1: Determine the geometric parameters of the CW type heat exchanger, and determine the flow conditions and channels of the combustion gas and cooling gas based on the geometric parameters, and draw the flow field mesh. Step 2: Determine the basic equations describing the physical model, then construct the corresponding functional equations based on the variational method, and determine the weighting coefficients of the entropy production term according to the application requirements. Step 3: Integrate the inverse temperature equation with negative diffusion coefficient into the physical information neural network, and solve the conservation equation using the CFD method. Step 4: The CFD solver extracts the inverse temperature field information from the neural network calculation results and sends the temperature, pressure and velocity physical field information to the PINN solver to realize the coupled calculation method of the two.
[0008] Furthermore, in step 2, a set of generalized functional equations is constructed based on the variational method, characterized by minimizing entropy production and flow power consumption. The energy equation is used to characterize maximizing the gradient of the flow term and minimizing flow power consumption, and heat transfer entropy production is used to approximate the total entropy production. Additionally, the integral of the viscous dissipation function in the fluid domain represents the total mechanical energy required to maintain the flow. Based on this, a variational optimization equation based on heat transfer entropy production is constructed, and the EL equation is ultimately derived. Subsequently, based on the vorticity field synergy, a corresponding variational equation is constructed.
[0009] Furthermore, PINN is used to provide a solution for solving the negative diffusion coefficient PDE, and fundamental physical laws are integrated into the neural network loss to apply regularization to the physical and data-driven model parameters.
[0010] Furthermore, OpenFOAM is used to calculate all the equations once, and the resulting inverse temperature field A is used for data-supervised learning in the first PINN. Then, the training results are transferred to the second PINN through transfer learning; The second PINN works in conjunction with OpenFOAM, and the two exchange data during the solution process; during the training of the second PINN, PINN and OpenFOAM exchange information once every 2000 training iterations. Both PINNs use the same network structure and prediction task objective. The second PINN uses a point distribution consistent with the low-resolution grid CFD distribution, and uses more points than the first PINN, thus ensuring that the weight and bias information transferred by transfer learning is more accurate.
[0011] Furthermore, the CFD-PINN coupled solution method in step 3 includes two PINN modules and a transfer learning module, a transfer learning module, and a low-resolution grid CFD module.
[0012] Furthermore, the governing equations are combined with their initial and boundary conditions to form a physics-based loss function. A feedforward neural network is then used to optimize the model parameters through the loss function.
[0013] Furthermore, the PINN used is a fully connected neural network with one input layer, one output layer, and eight hidden layers. The input variable is the spatial coordinates (x, y), and the pressure P, temperature T, and velocity vector U at each coordinate point are incorporated into the loss function. The output variable is the inversion temperature A at each collocation point. During the forward navigation process of the neural network, the activation function y=tanh(x) is selected to normalize the input data and achieve nonlinear mapping. Automatic gradient descent is implemented using TensorFlow-Adam Optimizer to quickly update the weight matrix w and bias coefficient b of the neural network. The parameters of TensorFlow-Adam Optimizer are set to hyperparameter E=10. -6 β1=0.9, β2=0.999; the weight matrix w of the first PINN is randomly initialized, and the bias coefficient b is set to zero.
[0014] The beneficial effects of this invention are: 1. The physical information neural network and computational fluid dynamics coupling method provided by this invention can be applied to the performance optimization of heat exchangers, filling the gap in domestic and foreign research in this field; 2. The coupled information transmission method used in this invention can reduce the number of grids required and has a shorter computation time, saving a lot of time, resources and costs; 3. This invention can quickly predict the flow field of CW type heat exchangers with different parameters. Compared with the results of high-resolution grid numerical solutions, the average error of the temperature field is less than 1%, and the number of grids required for calculation is reduced by 90%. Attached Figure Description
[0015] Figure 1 It is a CFD-PINN coupled solution framework; Figure 2 It involves setting up the computational domain and related dimensions for a simple two-dimensional pipe flow field. Figure 3 This is a schematic diagram of the grid nodes of a two-dimensional flow field; Figure 4 This is a schematic diagram comparing the local temperature results of high-precision grid CFD calculations and low-precision grid CFD-PINN coupled calculations. Wherein, a) the temperature distribution obtained by different schemes, and b) the relative temperature error obtained by different schemes; Figure 5 These are schematic diagrams showing the inversion temperature A and its relative error obtained by different calculation methods; Figure 6 It is a stacked structure of heat transfer units of the original surface heat exchanger of type CW; Figure 7 It is a heat transfer unit grid for a CW type heat exchanger; Figure 8It refers to the temperature distribution across the cross-section of heat transfer channels of different sizes; Figure 9 This is a distribution diagram of the overall heat transfer coefficient for different core boards. Detailed Implementation
[0016] Specific implementation method one: as follows Figures 1 to 9 As shown, a performance optimization method for CW-type heat exchangers based on PINN and functional variational methods includes the following specific steps: Step 1: Determine the geometric parameters of the CW type heat exchanger, and determine the flow conditions and channels of the combustion gas and cooling gas based on the geometric parameters, and draw the flow field mesh. Step 2: Determine the fundamental equations describing the physical model, then construct the corresponding functional equations based on the variational method, and determine the weighting coefficients of the entropy production term according to application requirements; construct the corresponding functional equations based on the variational method, characterizing them by minimizing entropy production and flow power consumption, characterizing them by maximizing the gradient of the flow term and minimizing flow power consumption using the energy equation, and approximating the total entropy production using heat transfer entropy production; furthermore, use the integral of the viscous dissipation function in the fluid domain to represent the total mechanical energy required to maintain the flow. Based on this, construct the variational optimization equations based on heat transfer entropy production, and finally derive the EL equations based on heat transfer entropy production; subsequently, construct the corresponding variational equations based on vorticity field coordination. PINN is used to provide a solution for solving the negative diffusion coefficient PDE. The fundamental physical laws are integrated into the neural network loss to apply regularization to the physical and data-driven model parameters. All equations are calculated once using OpenFOAM, and the resulting inverse temperature field A is used for data-supervised learning in the first PINN. Then, the training results are transferred to the second PINN through transfer learning; The second PINN works in conjunction with OpenFOAM, and the two exchange data during the solution process; during the training of the second PINN, PINN and OpenFOAM exchange information once every 2000 training iterations. Both PINNs use the same network structure and prediction task objective. The second PINN uses a point distribution consistent with the low-resolution grid CFD distribution, and uses a larger number of points than the first PINN, thus ensuring that the weight and bias information transferred by transfer learning is more accurate. Step 3: Integrate the inverse temperature equation with negative diffusion coefficient into the physical information neural network, and solve the conservation equation using the CFD method; the CFD-PINN coupled solution method includes two PINN modules and one transfer learning module, one transfer learning module, and one low-resolution grid CFD module; Initially used to propagate physical field information to the first PINN module for approximate initialization, it was subsequently used to iteratively exchange solution information with the second PINN module. Given that PINN only requires a coarse initial approximation, the pressure equations were linearly discretized using a first-order upwind scheme during the first use of LCFD, with the residual convergence criterion set to 10. -4 The second time LCFD was used, the pressure equation employed a second-order upwind scheme, with a residual convergence criterion set at 10. -6 In summary, this invention delegates the task of solving the negative diffusion term coefficient equations for the Lagrange operator A in the EL equations of both functional variational optimizations to PINN, while other PDEs are solved in a CFD solver configured with a low-resolution grid (2.25 × 10⁵). The final result is similar to that of high-fidelity CFD (HCFD) using 2.5 × 10⁶ grid nodes in traditional CFD, but the required number of grid nodes is reduced to 10%. The governing equations, combined with their initial and boundary conditions, form a physics-based loss function;
[0017] In the formula, FNN represents the total loss of the neural network, and Θ represents the trainable parameters, consisting of weights ω and biases b, which are continuously optimized during training to reduce the loss term. The fully connected network depicts the relationship between spatial coordinates (x, y) and related physical quantities (temperature inversion A). This paper employs a feedforward neural network to optimize the model parameters Θ through the loss function.
[0018]
[0019] For the steady-state problem under study, LF, LB, and LD represent the losses caused by PDE error, boundary error, and data supervision, respectively. The first two loss terms are defined as follows:
[0020] ; The PINN used is a fully connected neural network with one input layer, one output layer, and eight hidden layers. The input variable is the spatial coordinates (x, y), and the pressure P, temperature T, and velocity vector U at each coordinate point are incorporated into the loss function. The output variable is the temperature inversion A at each collocation point. During the forward navigation process of the neural network, the activation function y=tanh(x) is selected to normalize the input data and achieve nonlinear mapping. Automatic gradient descent is implemented using TensorFlow-Adam Optimizer to quickly update the weight matrix w and bias coefficient b of the neural network. The parameters of TensorFlow-Adam Optimizer are set to hyperparameter E=10. -6β1=0.9, β2=0.999; the weight matrix w of the first PINN is randomly initialized, and the bias coefficient b is set to zero; since the two-stage PINN has similar network structure and PDE to be solved, transfer learning can be used to improve training efficiency. In transfer learning methods, the object of learning is usually called the source domain, and the target of learning is called the target domain. This paper uses this method to transfer the learning object (source domain D) to the target domain. S The existing "knowledge" of a dense collocational PINN supervised by low-resolution CFD data is reused in task B (target domain DT: a sparse collocational PINN coupled with the CFD solver for information exchange), which can significantly improve learning efficiency while maintaining the feasibility of the target task. Source domain D S Task A: Train the inverse temperature equation with data supervision using the P, T, U, and A distributions provided by LCFD. Target domain D T Task B: Exchange information with the CFD solver and solve the problem in a coupled manner, outputting the inversion temperature A to the CFD solver and receiving the P, T, and U distributions; Step 4: The CFD solver extracts the inverse temperature field information from the neural network calculation results and sends the temperature, pressure and velocity physical field information to the PINN solver to realize the coupled calculation method of the two.
[0021] Example First, the geometric parameters of the CW-type heat exchanger are determined, and based on this, the flow conditions and channels for the combustion gas and cooling gas are determined, and the flow field mesh is drawn. Further, the fundamental equations describing the physical model are determined, and then a corresponding set of functional equations is constructed based on the variational method. The weighting coefficients of the entropy production term are determined according to application requirements. Then, the inverse temperature equation with a negative diffusion coefficient is integrated into the physical information neural network, and the conservation equations are solved using the CFD method. The CFD solver extracts the inverse temperature field information from the neural network calculation results and inputs physical field information such as temperature, pressure, and velocity into the PINN solver, realizing the coupled calculation method. The coupled implementation flowchart is shown below. Figure 1 .
[0022] First, a simple two-dimensional flow field heat transfer simulation is performed: taking the convective heat transfer process within a two-dimensional channel as the research object, air is selected as the working fluid, and its density is... =1.177 kg / m3, specific heat at constant pressure Cp=1004.9 J / (kg·K), dynamic viscosity μ=1.802×10 -5 Pa·s, thermal conductivity =0.02624 W / (m·K). The computational domain size is 30 mm × 5 mm, and no-slip boundary conditions are applied at both the top and bottom boundaries, such as... Figure 2As shown. The bottom wall heat flux density is set to 3000 W / m². The left side is the inlet, with fully developed flow conditions set, and the outlet is set to 101325 Pa. Two different calculation methods are used with high and low precision meshes, as shown in the mesh diagram. Figure 3 .
[0023] Next, a set of corresponding functional equations is constructed based on the variational method. First, a variational optimization equation based on heat transfer entropy production is constructed:
[0024]
[0025]
[0026]
[0027] The above is the EL equation derived from heat transfer entropy production. Similarly, the variational equation based on vorticity field synergy can be obtained:
[0028]
[0029]
[0030] PINN provides an efficient solution for solving the negative diffusion coefficient PDE. Fundamental physical laws are integrated into the neural network loss to regularize the physics and data-driven model parameters, ensuring that the accuracy of PINN is unaffected by the PDE form. Since regularization can be naturally introduced into data-supervised PINN training, PINN has unique advantages in solving ill-posed problems. The OpenFOAM solver is selected to solve the remaining related equations. The final numerical solution scheme is as follows: First, OpenFOAM is used to calculate all equations once, and the generated inverse temperature field A is used for data-supervised learning in the first PINN. Then, the training results are transferred to the second PINN through transfer learning. The second PINN works in conjunction with OpenFOAM, exchanging data during the solution process. A schematic diagram of the coupling process is shown below. Figure 1 .
[0031] During the training of the second PINN, information is exchanged between PINN and OpenFOAM every 2000 training iterations (or when the loss function error is less than 1e-4). However, during the initial 20,000 training iterations, training is performed only by receiving the P, T, and U information output from the first PINN, ensuring that the inverse temperature field initially fed into OpenFOAM is roughly reasonable, thus improving the accuracy of OpenFOAM's solution. Both PINNs use the same network structure and prediction task objective, allowing for effective utilization of transfer learning methods to reduce training costs. The second PINN's point allocation is consistent with the low-resolution grid CFD distribution, using a larger number of points compared to the first PINN, ensuring more accurate weight and bias information transferred through transfer learning. Given the low training cost of the first PINN, the increased training cost due to point allocation is negligible. The CFD-PINN coupled solution method developed in this paper includes two PINN modules and one transfer learning module. Additionally, a low-resolution grid CFD (LCFD) module is included, which is used twice in total. Initially used to propagate physical field information to the first PINN module for approximate initialization, it was subsequently used to iteratively exchange solution information with the second PINN module. Given that PINN only requires a coarse initial approximation, the pressure equations were linearly discretized using a first-order upwind scheme during the first use of LCFD, with the residual convergence criterion set to 10. -4 The second time LCFD was used, the pressure equation employed a second-order upwind scheme, with a residual convergence criterion set at 10. -6 In summary, this invention delegates the task of solving the negative diffusion term coefficient equations for the Lagrange operator A in the EL equations of both functional variational optimizations to PINN, while other PDEs are solved in a CFD solver configured with a low-resolution grid (2.25 × 10⁵). The final result is similar to that of high-fidelity CFD (HCFD) using 2.5 × 10⁶ grid nodes in traditional CFD, but the required number of grid nodes is reduced to 10%.
[0032] The governing equations can be combined with their initial and boundary conditions to form a physics-based loss function:
[0033] In the formula, FNN represents the total loss of the neural network, and Θ represents the trainable parameters, consisting of weights ω and biases b, which are continuously optimized during training to reduce the loss term. The fully connected network depicts the relationship between spatial coordinates (x, y) and related physical quantities (temperature inversion A). This paper employs a feedforward neural network to optimize the model parameters Θ through the loss function.
[0034]
[0035] For the steady-state problem studied in this paper, LF, LB, and LD represent the losses caused by PDE error, boundary error, and data supervision, respectively. The first two loss terms are defined as follows:
[0036]
[0037] The first PINN loss function consists of PDE loss, boundary loss, and data loss. The data loss is derived from the LCFD solution set, aiming to enhance the prediction accuracy of the first PINN without significantly increasing time cost, thereby providing initial weights and biases for the second PINN through transfer learning. The second PINN loss function only includes PDE loss and boundary loss, with the PDE solved by coupling with CFD. At the start of its training, the prediction problem exhibits a high local error function, and training is difficult to converge in some cases. To address this, this paper introduces node-level adaptive weights, focusing optimization on high-error nodes rather than uniformly weighting the entire network. Finally, the loss for each collocation point in the computation domain Ω is calculated as follows:
[0038] Using ε as the pointwise error set in the domain Ω, the error weight at position (i,j) can be expressed as:
[0039] In the above formula θ and θ are the shift factor and scaling factor, respectively. The final loss function is:
[0040] The PINN used is a fully connected neural network with one input layer, one output layer, and eight hidden layers (each with 40 neurons). The input variables are spatial coordinates (x, y), and the pressure P, temperature T, and velocity vector U at each coordinate point are incorporated into the loss function. The output variable is the inversion temperature A at each collocation point. During the forward propagation of the neural network, the activation function y=tanh(x) is selected to normalize the input data and achieve nonlinear mapping. Automatic gradient descent is implemented using TensorFlow-Adam Optimizer to quickly update the weight matrix w and bias coefficients b of the neural network. The main parameter of TensorFlow-Adam Optimizer is set to the hyperparameter E=10. -6β1=0.9, β2=0.999. The weight matrix w of the first PINN is randomly initialized, and the bias coefficient b is set to zero. Since the two-stage PINN has similar network structure and PDE to be solved, transfer learning can be used to improve training efficiency. In transfer learning methods, the object being learned is usually called the source domain, and the target being learned is called the target domain. This paper uses this method to transfer the learning object (source domain D) to the target domain. S The existing "knowledge" of a dense collocational PINN supervised by low-resolution CFD data is reused in task B (target domain DT: a sparse collocational PINN coupled with the CFD solver for information exchange), which can significantly improve learning efficiency while maintaining the feasibility of the target task. Source domain D S Task A: Train the inverse temperature equation with data supervision using the P, T, U, and A distributions provided by LCFD. Target domain D T Task B: Exchange information with the CFD solver and solve the problem in a coupled manner, outputting the inversion temperature A to the CFD solver and receiving the P, T, and U distributions.
[0041] Compare the results obtained after solving, such as Figure 4 It can be observed that the CFD-PINN coupled solution method has high accuracy in numerical simulations involving inverse temperature fields. The CFD-PINN coupled solution results are relatively accurate, and the results are highly consistent with HCFD results in most watersheds. Therefore, the CFD-PINN coupled solution method can effectively improve the accuracy of LCFD results and enhance the convergence of PINN calculations.
[0042] Simulation of the flow field of the CW-type heat exchanger: The physical model of the CW-type original surface heat exchanger studied in this paper is as follows: Figure 6 As shown, this heat exchanger is commonly used in the regenerative structure of micro gas turbines. The key geometric parameters of the CW heat transfer channel include longitudinal amplitude A, flow wavelength L, channel spacing P, and channel height H. The key parameters of the CW original surface heat exchanger heat transfer unit are L / Ac=2.11, R1 / R2=1.77, H=1.1mm, H / P=1.25, and d... e (Air side) = 0.6 mm, d e (Gas side) = 0.8 mm. Mass flow rate and total temperature boundary conditions are applied to the air and gas side inlets, with total temperatures of 432 K and 840 K respectively, and Reynolds number of 3000 for both. Static pressure boundary conditions are applied to the air and gas side outlets, both set to 101325 Pa. The mesh diagram is shown below. Figure 7 After verifying the mesh independence, the feasibility of the numerical method is demonstrated.
[0043] By changing the H / P value, multiple heat transfer channels of different sizes were obtained, and the solutions were performed separately. The temperature distribution diagrams are shown below. Figure 8Through the various results obtained from CFD-PINN coupled calculations, we can analyze that as the H / P ratio increases, the coverage area of the high-temperature region in the air passage gradually increases, and the temperature distribution in the gas passage also exhibits a corresponding distribution pattern of intermediate high temperature and boundary low temperature, such as... Figure 8 As shown.
[0044] This method also allows for the analysis and study of the influence of curvature on the flow-heat transfer performance of the core plate. Heat transfer studies were conducted on an uncurved straight core plate (k=0), three types of arc-shaped core plates (k=0.167, 0.2, and 0.25 respectively), and two types of variable curvature core plates (involute and cycloidal forms respectively). The straight heat transfer core plate is used only as a reference, as uncurved core plates cannot form a complete annular CW-type original surface heat exchanger. All curved plates are manufactured using a straight plate integral deformation process, thus ensuring that the geometric surface area of the CW-type original surface heat exchanger core plates with different curvature profiles remains unchanged, thereby guaranteeing that each scheme has the same flow channel volume and heat transfer area. The curvature distribution of the cycloidal and involute forms exhibits a non-linear variation characteristic with the relative arc length (0≤s≤1, corresponding to the position from the inner diameter to the outer diameter of the annular CW-type original surface heat exchanger). Figure 9 As shown, the overall heat transfer coefficient UA of the CW-type original surface heat exchanger core plates with different curvature profiles generally follows the same trend with the Re number, showing a positive correlation where the larger the Re number, the higher the UA. However, it is worth noting that compared with the uncurved flat plate structure of the CW-type original surface heat exchanger core plate, the UA value of the curved core plate shows a significant decrease, and the heat transfer performance degradation is more pronounced for the two variable curvature core plates. Compared with the flat plate reference scheme, the maximum decrease in the overall heat transfer coefficient reaches 8%. Therefore, when designing based on a single-channel flat core plate, the negative impact of the curvature profile on the heat transfer performance should be carefully evaluated to avoid the degradation of heat transfer performance caused by the curvature of the core plate. The curvature of the involute and cycloidal core plates increases sharply near the inner diameter. This region is the air inlet and has the largest temperature difference. The drastic curvature change induces flow separation and reduces local heat transfer efficiency.
[0045] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.
Claims
1. A PINN and functional variational method-based performance optimization method for a CW-type heat exchanger, characterized in that, The specific steps include: Step 1, determine the geometric shape parameters of the CW heat exchanger, determine the flow conditions and channels of the gas and cooling gas according to the geometric shape parameters, and draw the flow field grid; Step 2, determine the basic equations describing the physical model, and then construct the corresponding functional equation set based on the variational method, and determine the weight coefficient of the entropy term according to the application requirements; Step 3, integrate the inverse temperature equation with negative diffusion coefficient into the physical information neural network, and solve the conservation equation by CFD method; Step 4, the CFD solver extracts the inverse temperature field information from the neural network calculation results, and delivers the temperature, pressure and velocity physical field information to the PINN solver to realize the coupling calculation method of the two.
2. The method of claim 1, wherein, In step 2, the corresponding functional equation set is constructed based on the variational method, which is characterized by the minimization of entropy production and flow power consumption, and the energy equation is used to maximize the flow power consumption and minimize the flow power consumption. The total mechanical energy required to maintain flow is represented by the integral of the viscous dissipation function in the fluid domain. Based on this, the variational optimization equation based on heat transfer entropy production is constructed, and the E-L equation is finally derived based on heat transfer entropy production; Subsequently, based on the vortex field coordination, the corresponding variational equation is constructed.
3. The method of claim 1, wherein the method is characterized by: PINN is used to provide a solution to the negative diffusion coefficient PDE, and the basic physical law is integrated into the neural network loss to regularize the physical and data-driven model parameters.
4. The method of claim 1, wherein the method is characterized by, All equations are calculated once using OpenFOAM, and the inverse temperature field A generated is used for data supervised learning in the first PINN; Then the training results are transferred to the second PINN through transfer learning; The second PINN works cooperatively with OpenFOAM, and the two exchange data during the solving process; every 2000 times of training in the training process of the second PINN, PINN exchanges information with OpenFOAM once; The two PINNs use the same network structure and prediction task target, and the second PINN uses more number of points than the first PINN, so as to ensure that the weight and bias information transferred by the transfer learning is more accurate.
5. The method of claim 1, wherein the method is characterized by: The CFD-PINN coupling solving method in step 3 includes two PINN modules and a transfer learning module, a transfer learning module, and a low-resolution grid CFD module.
6. The method of claim 1, wherein the method is characterized by: The control equation is combined with its initial condition and boundary condition to form a physics-based loss function, a feedforward neural network is used, and the model parameters are optimized through the loss function.
7. The method of claim 1, wherein the method is characterized by: The PINN adopted is a fully connected neural network containing 1 input layer, 1 output layer and 8 hidden layers, the input variable is spatial coordinates (x, y), the pressure P, temperature T and velocity vector U of each coordinate point are coded into the loss function, and the output variable is the inverse temperature A at each distribution point. In the forward process of the neural network, the activation function y = tanh(x) is selected, the input data is normalized to realize nonlinear mapping, and TensorFlow-Adam Optimizer is used to realize automatic gradient descent and quickly update the weight matrix w and bias coefficient b of the neural network; the parameter setting of TensorFlow-Adam Optimizer is super parameter E = 10 -6 , β1=0.9, β2=0.999; the weight matrix w of the first PINN is randomly initialized, and the bias coefficient b is set to zero.
