A method and device for multi-scale intelligent collaborative optimization of a fuel cell cathode flow field

Abstract

This application relates to the field of fuel cell design optimization technology, and discloses a multi-scale intelligent collaborative optimization method and device for fuel cell cathode flow field, with the device corresponding to the method. Addressing the problems of simulation and experimental parameter mismatch and the inefficiency of manual parameter tuning, it achieves intelligent calibration of underlying micro-parameters through optimization algorithms and automated interactive interfaces. To overcome the limitations of traditional discrete optimization, which easily gets trapped in local optima and is computationally time-consuming, a neural network surrogate model integrating physical dual constraints is constructed, namely the fuel cell surrogate model. To address the deficiency of ignoring the dynamic characteristics of the air compressor, the fuel cell surrogate model and the air compressor model are coupled at the system level, avoiding the research and development pitfall of high parasitic power estimation errors caused by dynamic changes in air compressor efficiency. Relying on a multi-objective decision-making mechanism, it successfully achieves synergistic improvement based on system efficiency and internal indicators on a global scale, providing a scientific automated solution for the design of high-efficiency, high-durability bipolar plates for fuel cells.

Description

Technical Field

[0001] This application relates to the field of fuel cell design optimization technology, specifically a method and device for multi-scale intelligent collaborative optimization of fuel cell cathode flow field. Background Technology

[0002] Existing technologies have improved fuel cell performance to some extent and explored optimization methods, but in the multi-scale collaborative optimization design of cathode flow field for actual vehicle fuel cells, the following significant defects still exist: there is parameter mismatch between simulation models and actual physical processes, and a lack of accurate calibration; traditional methods are prone to getting trapped in local optima and have low design efficiency; multi-objective optimization fails to accurately reflect the dynamic changes in the isentropic efficiency of air compressors. Summary of the Invention

[0003] The purpose of this application is to provide a multi-scale intelligent collaborative optimization method and device for the cathode flow field of a fuel cell, in order to solve the technical problems in the prior art, such as the flow field structure design being highly dependent on manual experience for parameter tuning, the three-dimensional CFD fluid dynamics simulation being time-consuming and prone to getting trapped in local optima, and the failure to accurately reflect the dynamic changes in the isentropic efficiency of the air compressor.

[0004] To achieve the above objectives, this application provides a multi-scale intelligent cooperative optimization method for the cathode flow field of a fuel cell, comprising: Based on experimental test data of polarization curves of single fuel cells, automated scripts are used in conjunction with three-dimensional numerical simulation to quickly identify parameters and accurately calibrate a high-precision three-dimensional numerical model of fuel cells. Based on the three-dimensional numerical model of fuel cells, automated batch sampling and CFD joint calculation are performed within the preset geometric design parameter space to obtain various performance responses of fuel cells and generate a high-quality training dataset. A fuel cell surrogate model is constructed by combining multilayer perceptron algorithm and physical constraints. This fuel cell surrogate model is used to deeply learn the nonlinear mapping relationship between geometric and operating parameters and system performance. The fuel cell proxy model is coupled with the preset air compressor proxy model at the system level. Taking into account the system efficiency and preset internal indicators, the preset multi-objective optimization algorithm and decision algorithm are run to output the optimal combination of flow field geometric parameters to complete the multi-scale intelligent collaborative optimization of the fuel cell cathode flow field.

[0005] To achieve the above objectives, this application also provides a multi-scale intelligent cooperative optimization device for the flow field of a fuel cell cathode, which applies the multi-scale intelligent cooperative optimization method for the flow field of a fuel cell cathode as described above, including: The fuel cell 3D numerical model calibration module is used to quickly identify parameters and accurately calibrate a high-precision fuel cell 3D numerical model based on experimental test data of polarization curves of a single fuel cell using automated scripts in conjunction with 3D numerical simulation. The training dataset generation module is used to automatically perform batch sampling and CFD joint calculation based on the three-dimensional numerical model of the fuel cell within a preset geometric design parameter space to obtain various performance responses of the fuel cell and generate a high-quality training dataset. The fuel cell proxy model construction module is used to construct a fuel cell proxy model by combining multilayer perceptron algorithm and physical constraints. This fuel cell proxy model is used to deeply learn the nonlinear mapping relationship between geometric and operating parameters and system performance. The fuel cell and air compressor coupling optimization module is used to couple the fuel cell proxy model with the preset air compressor proxy model at the system level. It comprehensively considers the system efficiency and preset internal indicators, runs preset multi-objective optimization and decision algorithms, and outputs the optimal combination of flow field geometric parameters to complete the multi-scale intelligent collaborative optimization of the fuel cell cathode flow field.

[0006] Beneficial effects: The multi-scale intelligent collaborative optimization method and device for the cathode flow field of fuel cells in this application addresses the technical shortcomings of existing technologies, such as the high dependence of flow field structure design on manual experience parameter tuning, the long time consumption of three-dimensional CFD fluid dynamics simulation and the easy getting trapped in local optima, and the failure to accurately reflect the dynamic change characteristics of the isentropic efficiency of air compressors. It constructs a complete data-driven closed-loop framework for experimental calibration, automated simulation, surrogate model construction, and system collaborative optimization, thereby achieving a synergistic improvement in the performance and lifespan of automotive fuel cells. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 A flowchart illustrating the multi-scale intelligent collaborative optimization method for the cathode flow field of a fuel cell provided in this application embodiment; Figure 2 A flowchart illustrating the calibration process of the three-dimensional numerical model of the fuel cell provided in this application embodiment; Figure 3 A comparison chart of the calibrated simulation model output and experimental data provided for embodiments of this application; Figure 4 This is an architecture diagram of the fuel cell proxy model provided in the embodiments of this application; Figure 5 The algorithm regression diagram for the fuel cell proxy model provided in the embodiments of this application; in the diagram: (a) current density; (b) voltage drop; (c) ; Figure 6 A comparison chart of model output and experimental data provided for embodiments of this application; Figure 7 A flowchart illustrating system-level coupling and multi-objective global optimization decision-making provided in the embodiments of this application; Figure 8 The multi-objective optimization results provided in the embodiments of this application; Figure 9 A flowchart illustrating the overall process of the multi-scale intelligent collaborative optimization method for the cathode flow field of a fuel cell provided in this application embodiment; Figure 10 The diagram shows the structure of the multi-scale intelligent collaborative optimization device for the cathode flow field of a fuel cell provided in this application embodiment; in the diagram: 10, fuel cell three-dimensional numerical model calibration module; 20, training dataset generation module; 30, fuel cell proxy model construction module; 40, fuel cell and air compressor coupling optimization module.

[0009] The implementation, functional features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0010] The technical solutions in the embodiments of this application will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0011] In this document, the term "comprising" is intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0012] With the increasing importance of clean energy in energy systems, proton exchange membrane fuel cells (PEMFCs) have become a key development direction for automotive power systems due to their advantages such as high energy density, zero emissions, and high efficiency. Inside the fuel cell, the cathode flow field plays a crucial role in transporting reactant gases and draining liquid water. Its structural design directly determines the material distribution, current density uniformity, and water management capabilities within the cell. Optimizing the flow field structure can increase local gas velocity and change the flow direction, thereby significantly enhancing mass transfer and representing an effective means to improve the overall performance of the cathode flow field.

[0013] Currently, existing technologies for the design and optimization of the cathode flow field structure in fuel cells mainly rely on a combination of computational fluid dynamics (CFD) simulation and optimization algorithms. These existing technologies primarily include the following four aspects: Existing technology 1: Design of overall topology configuration and local auxiliary features of flow field.

[0014] For the design and optimization of the cathode flow field in proton exchange membrane fuel cells (PEMFCs), this type of approach mainly falls into two categories: research focusing on the overall flow field topology and research focusing on novel flow channel cross-sections or auxiliary structures. In the design focusing on the overall flow field configuration, researchers typically make hybrid improvements based on basic configurations such as radial, spiral, parallel, and serpentine shapes, or draw inspiration from biological structures to conduct biomimetic designs of the flow field, such as snowflake-shaped or lung-shaped flow fields. In the design focusing on novel flow channel cross-sections or auxiliary structures, this type of approach mainly improves the overall performance of the flow field by changing the shape of the flow channel (horizontal or vertical plane) or adding auxiliary structures such as baffles within the flow channel. These approaches usually utilize three-dimensional computational fluid dynamics (CFD) numerical models to evaluate the impact of different flow field structures on the battery's internal mass transfer and water management capabilities.

[0015] Existing technology 2: Simulation optimization based on empirical discrete point sampling.

[0016] Because three-dimensional CFD simulation calculations are time-consuming, when designing a flow field, engineers usually rely on their experience to select several discrete design points within the range of flow field geometric parameters, and then conduct simulations or experimental tests in sequence to select a set of relatively optimal design points as the final solution.

[0017] Existing technology 3: Optimization of some parameters based on the proxy model.

[0018] To overcome the time-consuming nature of CFD computation, some studies have introduced machine learning algorithms. For example, Zou, G.; Chen, K.; Chen, W.; Deng, Q.; Chen, B. Multi-objective optimization of proton exchange membrane fuel cell flow channel baffle based on artificial neural network and genetic algorithm. Fuel 2025, 380: 133205. DOI: 10.1016 / j.fuel.2024.133205. discloses a scheme for multi-objective optimization of the flow channel baffle based on an artificial neural network (ANN) to generate a surrogate model and combined with a genetic algorithm; Ghanbari, S.; Ghasabehi, M.; Asadi, MR; Shams, M. An inquiry into transport phenomenon and artificial intelligence-based optimization of a novel bio-inspired flow field for proton exchange membrane fuel cells. Applied Energy2024, 376: 124260. DOI: In 10.1016 / j.apenergy.2024.124260, the Response Surface Methodology (RSM) is used to establish a surrogate model, and the Non-dominated Sorting Genetic Algorithms (NSGA) are used to find the optimal operating conditions.

[0019] Existing technology four: Empirical parameter identification at the fuel cell stack level.

[0020] To improve the accuracy of simulations, some existing technologies, such as Sharma, P.; Raju, S.; Salgotra, R. An evolutionary multi-algorithm based framework for the parametric estimation of proton exchange membrane fuel cell. Knowledge-Based Systems 2024, 283: 111134. DOI: 10.1016 / j.knosys.2023.111134., employ improved intelligent optimization algorithms (such as the Rao algorithm and the Walrus optimization algorithm) to fit the polarization curves of the PEMFC stack, thereby identifying unknown parameters at the system level.

[0021] Although the aforementioned existing technologies have improved fuel cell performance to some extent and explored optimization methods, the following significant shortcomings still exist in the multi-scale collaborative optimization design of the cathode flow field for practical automotive fuel cells: First, there is a parameter mismatch between the simulation model and the actual physical process, and a lack of accurate calibration.

[0022] Whether designing for overall flow field configuration or local flow channel characteristics, there are often limitations in validating simulation models. Some works oversimplify the problem, especially for single-channel models, and experimental calibration of parameters is generally lacking, leading to discrepancies between model results and reality. Existing techniques typically only construct simple single-channel models and lack experimental data support. When calibrating parameters for three-dimensional numerical models, current techniques usually rely on experience to manually select and adjust unknown parameters to match experimental results. This is not only inefficient in the face of a large and complex parameter space but also makes it difficult to minimize the error between model output and experimental results.

[0023] Secondly, traditional methods are prone to getting stuck in local optima and are inefficient in design.

[0024] If the empirical discrete point sampling method is used, since the continuous parameter space is not traversed, the so-called optimal solution obtained is often only a local suboptimal solution within the parameter space, which cannot bring out the ultimate potential of the flow field structure.

[0025] Third, multi-objective optimization fails to accurately reflect the dynamic changes in the isentropic efficiency of air compressors.

[0026] In existing technologies for multi-objective flow field optimization, some studies have neglected the parasitic power loss caused by the air compressor. Even when some studies begin to focus on the system-level parasitic power loss due to increased flow resistance, they often oversimplify the energy consumption assessment of the air compressor. Specifically, existing technologies generally treat the isentropic efficiency of the air compressor as a fixed constant for theoretical calculations. However, in actual operating conditions, changing the geometric characteristics of the cathode flow field, such as the channel size and the shape of the turbulence elements, will significantly change the overall pressure drop and the flow rate demand of the reactant gas, thereby forcing the operating point of the air compressor, i.e., the pressure ratio and mass flow rate, to shift. According to the actual operating characteristics of the air compressor, its isentropic efficiency changes dynamically with highly nonlinearity and pressure ratio. Treating it as a static constant severely isolates the actual physical coupling mechanism, leading to significant errors in the estimation of the air compressor's parasitic power. This lack of consideration for nonlinear dynamic efficiency characteristics may result in the optimized bipolar plate structure parameters exhibiting a negative effect in the actual vehicle system, where the apparent performance of individual battery cells is improved, but the system efficiency decreases.

[0027] In response to the aforementioned technical deficiencies, this embodiment discloses a multi-scale intelligent collaborative optimization method and device for fuel cell cathode flow field. It is a multi-scale collaborative optimization framework for fuel cell cathode flow field, which addresses the technical deficiencies in the prior art, such as the flow field structure design being highly dependent on manual experience for parameter tuning, the three-dimensional CFD fluid dynamics simulation being time-consuming and prone to getting trapped in local optima, and the failure to accurately reflect the dynamic changes in the isentropic efficiency of the air compressor. This achieves a synergistic improvement in the performance and lifespan of automotive fuel cells.

[0028] The multi-scale intelligent collaborative optimization method for the fuel cell cathode flow field in this embodiment will now be described.

[0029] Before proceeding, an exemplary scenario and the hardware used in this embodiment will be described in detail. In this embodiment, the design and optimization of a baffle-type cathode flow field are used as an example to illustrate the core framework. The hardware devices and software dependencies used in this embodiment are as follows: In terms of hardware environment: The central processing unit (CPU) adopts an AMD EPYC 9J14 96-Core Processor, whose multi-core high-concurrency processing capability is mainly used to support the grid parallel computing and multi-task joint solution of the 3D CFD simulation model; the graphics processing unit (GPU) adopts an NVIDIA GeForce RTX 4090D, which, relying on its powerful tensor computing power, is mainly used to accelerate the tensor matrix operation and backpropagation training process of deep learning surrogate models, namely MLP networks.

[0030] Regarding the software and dependency environment: the operating system used is Windows 11 Pro; the simulation software used is Ansys 2024 R2, which serves as the basic solver in this embodiment and is responsible for solving the conservation equations of the fuel cell that involve the coupling of multiple physical fields such as mass, momentum, composition, and charge; the programming language used is Python 3.10; and the deep learning framework used is PyTorch, which is used to build, train, and validate the neural network surrogate model that fits the nonlinear mapping relationship of the fuel cell.

[0031] The core component libraries it relies on include: ansys-fluent-core (PyFluent), which enables low-level communication between Python scripts and the Ansys Fluent solver, automating the entire process of boundary condition setting, solving, and result export, and serving as a key interface for automatic parameter identification and batch data generation; scikit-optimize, which performs model parameter identification and hyperparameter optimization based on strategies such as differential evolution and Bayesian optimization; pandas, which cleans, structures, integrates, and preprocesses the massive amounts of experimental and sample data exported from simulations; and scikit-learn, which provides a set of general machine learning tools for data normalization, evaluation metric calculation, and other tasks.

[0032] Reference Figure 1 , Figure 1 This is a flowchart illustrating the multi-scale intelligent collaborative optimization method for the cathode flow field of a fuel cell provided in an embodiment of this application.

[0033] like Figure 1 As shown, in a first aspect, this embodiment discloses a multi-scale intelligent collaborative optimization method for the cathode flow field of a fuel cell, including: S10: Based on experimental test data of polarization curves of single fuel cells, automated scripts are used in conjunction with three-dimensional numerical simulation to quickly identify parameters and accurately calibrate a high-precision three-dimensional numerical model of fuel cells.

[0034] S10 of this embodiment aims to establish a three-dimensional numerical model of a single fuel cell with multi-physics coupling, i.e., a three-dimensional numerical model of the fuel cell. The three-dimensional numerical model of the fuel cell in this embodiment comprehensively considers the two-phase flow transport in the flow channel, the mass transfer behavior in the porous medium, and the electrochemical reaction, and follows the conservation equations of mass, momentum, energy, composition, and charge.

[0035] Figure 2 A flowchart illustrating the calibration process of the three-dimensional numerical model of the fuel cell provided in this application embodiment.

[0036] To address the mismatch between real microscopic physical quantities and empirical coefficients, this embodiment proposes using a set of key microscopic electrochemical and porous medium parameters as the objects to be identified. A closed-loop system is constructed by deeply integrating a CFD simulation engine with a global heuristic optimization algorithm through an automated interactive interface. At discrete operating voltage points, the simulated current density is extracted, and the mean square error between the simulated and experimental data is calculated to construct a normalized objective function. The algorithm iterates through parameter combinations within a given search space until the error converges to within the tolerance range, thus obtaining a precisely calibrated numerical model. The specific flowchart is as follows: Figure 2 As shown.

[0037] Specifically, the calibration of the three-dimensional numerical model of the fuel cell includes: Import experimental test data of polarization curves of single fuel cells; The physical boundaries of the microscopic key parameters to be identified are set; among them, the microscopic key parameters to be identified include cathode reference exchange current density, gas diffusion pore blockage index, GDL porosity, CL porosity, cathode ionomer resistance and water removal coefficient. The preset optimization algorithm generates the individual parameter vectors for the current iteration; Automated interface mapping parameters to the three-dimensional numerical model of the fuel cell; The solver performs multiphysics coupled simulation calculations; Extract steady-state simulation results for each operating condition; Calculate the value of the preset objective function; Determine whether the convergence condition or the maximum number of iterations is met; if so, output the optimal combination of simulation parameters that minimizes the error; otherwise, retain the better individual and proceed to the next iteration.

[0038] In a specific application, the following six key physical quantities are used as a combination of parameters to be identified: cathode reference exchange current density, gas diffusion pore blockage index, GDL porosity, CL porosity, cathode ionomer resistance, and water removal coefficient. To achieve accurate calibration of these six parameters, this embodiment uses the Python-based PyFluent interface to deeply integrate the AnsysFluent simulation engine with the Differential Evolutionary Algorithm (DE) to construct an automated parameter tuning closed loop. The DE parameters are set as follows: population size of 10, iteration cutoff condition of error loss less than 25, maximum number of iterations of 20, scaling factor of 0.8, and crossover probability of 0.9. The search range for the six model physical parameters is shown in Table 1.

[0039] Table 1 In a specific application, this embodiment selects 10 different discrete operating voltage points, extracts the corresponding steady-state current density, and calculates the mean square error. And normalization is performed to prevent numerical overflow. Objective function The mathematical expression is as follows: (1) (2) in, The simulated current density (A / m) 2 ), The experimental current density (A / m) 2 ).exist Divide by get This is to prevent the objective function from becoming too large.

[0040] Reference Figure 3 , Figure 3 A comparison chart of the output of the calibrated simulation model and the experimental data provided in the embodiments of this application.

[0041] The parameter identification results of this embodiment are shown in Table 2. Substituting the calibrated parameters into the three-dimensional numerical model, the output polarization curves and experimental data are shown below. Figure 3 As shown.

[0042] Table 2 S20: Based on the three-dimensional numerical model of the fuel cell, automated batch sampling and CFD joint calculation are performed within the preset geometric design parameter space to obtain various performance responses of the fuel cell and generate a high-quality training dataset.

[0043] In this embodiment, S20, based on the calibration model, sets the key geometric design parameters of the flow field as design variables and uses a space-filling design method, such as orthogonal experimental design or Latin hypercube sampling, to generate a geometric parameter combination matrix covering the typical characteristics of the design space. An automated script traverses the design schemes, automatically loads boundary conditions (i.e., operating parameters), and triggers the solution calculation. The simulation results are used to extract the overall pressure drop, polarization curve data, and other internal evaluation indicators, such as the relative standard deviation of the current density distribution. Structured data is used to build the underlying database required for system-level multi-objective optimization.

[0044] Specifically, the generation of the training dataset includes: Based on the three-dimensional numerical model of the fuel cell, the key geometric design parameters of the cathode flow field are set as design variables, and a geometric parameter combination matrix covering the typical features of the design space is generated by a preset space filling design method. Among them, the key geometric design parameters include the number of trapezoidal baffles, the bottom angle of the baffles, and the height of the baffles in the cathode flow field. The design scheme is traversed using automated scripts, which automatically load boundary conditions and trigger the solution calculation. Extract the overall pressure drop and polarization curve data of the simulated cathode flow field, as well as the preset internal evaluation indicators, to complete the construction of the underlying database required for system-level multi-objective optimization; among them, the internal evaluation indicators include the relative standard deviation of current density distribution.

[0045] In a specific application, this example uses the orthogonal experimental method to construct the sample. Three key geometric parameters of the trapezoidal baffle in the cathode flow field—the number of baffles, the base angle of the baffle, and the height of the baffle—are set as design variables, and an orthogonal array is selected. A geometric parameter combination matrix covering the typical characteristics of the design space is generated. Finally, a PyFluent automation script is used again to traverse each geometric design scheme in the orthogonal array. The script automatically loads the simulation model, applies boundary conditions, and triggers the solution calculation. After the simulation is complete, the script automatically extracts the overall pressure drop, polarization curve data, and current density distribution uniformity index of the flow field. The data is aligned with the corresponding geometric inputs and operating parameters to form a high-quality simulation dataset.

[0046] S30: A fuel cell surrogate model is constructed by combining a multilayer perceptron algorithm with physical constraints. This fuel cell surrogate model is used to deeply learn the nonlinear mapping relationship between geometric and operating parameters and system performance.

[0047] In this embodiment, S30 utilizes the generated high-quality simulation dataset to train a fuel cell proxy model that replaces CFD simulation. In practice, using a black-box model may result in poor model generalization ability. Therefore, this embodiment proposes a neural network architecture that integrates dual physical constraints.

[0048] Reference Figures 4 to 6 . Figure 4 This is an architecture diagram of the fuel cell proxy model provided in the embodiments of this application. Figure 5 This is the algorithm regression graph for the fuel cell proxy model provided in the embodiments of this application. Figure 6 A comparison chart of model output and experimental data provided in the embodiments of this application.

[0049] Specifically, the fuel cell proxy model takes preset key geometric design parameters and operating parameters as inputs and outputs current density, cathode flow field pressure drop and preset internal evaluation indicators.

[0050] In applications, the network input feature matrix contains key geometric design parameters of the cathode flow field. and operating parameters The output response matrix includes current density. Flow field pressure drop And other internal evaluation indicators. After normalizing the data by standard deviation, the open-circuit voltage is extracted. and anchor values ​​in normalized space Simultaneously, extract the normalized anchor point value of the output response at the physical zero point. and It should be noted that the internal indicator used in this embodiment is the relative standard deviation of the current density distribution, which characterizes the uniformity of the current density distribution. The following is based on This can be explained, but it can be replaced with other internal indicators.

[0051] Specifically, in the forward propagation of the network, the fuel cell proxy model incorporates a physical hard constraint mechanism to address the inherent physical boundary condition of the polarization curve at the open-circuit voltage, i.e., when the voltage equals the open-circuit voltage, the output current must be 0. First, calculate the non-negative value of the current voltage overpotential: (3) in, It is the open circuit voltage. The anchor value obtained after standard deviation normalization; It is the input voltage after standard deviation normalization.

[0052] The Softplus function is used to ensure monotonic activation. After multiplying by the overpotential, the result is superimposed to the zero anchor point to force a reconstructed current and uniform output. (4) (5) in, This is the normalized predicted current density value. It is the normalized physical zero-point value of the current density. This is the predicted current density value for the MLP network. For normalized Predicted value For normalized Physical zero point value This represents the predicted RSD value for the MLP network.

[0053] Since the flow field voltage drop is decoupled from the voltage, the original prediction is directly output, i.e., the unnormalized MLP network output value. .

[0054] The final model output value requires denormalization of the normalized variables.

[0055] Specifically, when updating network weights through backpropagation, the fuel cell proxy model introduces a weighted data fidelity loss, assigning a higher penalty weight to the mean square error (MSE) in the high potential range. .

[0056] (6) in, It is the model output value. It is the actual value. Indicates the penalty weight. Indicates mean square error. , The total number of data points in the training dataset generated for the three-dimensional numerical model of the fuel cell.

[0057] Simultaneously, a monotonic physical penalty loss is introduced, since the slope of the polarization curve must be less than or equal to 0 under constant geometry. By using automatic differentiation techniques, penalties are imposed on regions of positive slope that violate physical laws. (7) in, This represents the partial derivative of the predicted current density with respect to voltage. This represents the total number of unlabeled physical points, corresponding to a batch of points randomly sampled in the input space, i.e., between the upper and lower bounds of the voltage, in this embodiment. This represents the physical information regularization term.

[0058] Combining the two, a physical penalty coefficient is introduced. Balance the proportions of the two: (8) An optimizer is used to iteratively solve the network, ultimately outputting a high-frequency, high-precision PINN surrogate model. The neural network architecture described above is as follows: Figure 4 As shown. The results after training are as follows. Figure 5 As shown, high R 2 This demonstrates that the surrogate model is highly effective and can excellently replace the 3D numerical model. After constructing the surrogate model, this embodiment selected a set of parameter combinations and experimentally measured data, which was then compared with the output data of the surrogate model. Figure 6 As shown, the output data of the surrogate model matches the actual experimental data very well, proving the generalization performance of the Physically Constrained Surrogate Model (PINN) architecture used in the fuel cell surrogate model of this embodiment.

[0059] S40: Couples the fuel cell proxy model with the preset air compressor proxy model at the system level, comprehensively considers system efficiency and preset internal indicators, runs preset multi-objective optimization algorithm and decision algorithm, and outputs the optimal combination of flow field geometric parameters to complete the multi-scale intelligent collaborative optimization of the fuel cell cathode flow field.

[0060] Reference Figure 7 , Figure 7The flowchart of system-level coupling and multi-objective global optimization provided for the embodiments of this application is shown.

[0061] In this embodiment, S40 performs system-level multi-objective collaborative optimization by coupling the dynamic characteristics of the air compressor. This is achieved by deeply coupling a trained fuel cell physical information neural network (PINN) surrogate model with a pre-built air compressor surrogate model. Then, to avoid physically feasible parameter combinations in the geometric parameter space, geometric interference and physical boundary constraints are introduced into the multidimensional variable space. Based on this, the Pareto front algorithm is used to extract the optimal solution set, and subsequently, a multi-objective decision-making algorithm is introduced for the final engineering decision. The flowchart is as follows... Figure 7 As shown. Specifically, in this embodiment, firstly, in a given geometric parameter space, such as the number of baffles... baffle height Baffle bottom corner and target operating voltage Next, the established PINN model is invoked to quickly predict the corresponding current density. and cathode flow field pressure drop Other internal metrics, such as the relative standard deviation of current density distribution Subsequently, the above prediction results are substituted into the system-level physical formula and coupled in series with the air compressor proxy model. The specific calculation process of model coupling consists of three parts.

[0062] Specifically, the fuel cell proxy model is coupled with the pre-defined air compressor proxy model at the system level, including: The total output power of the fuel cell is calculated based on the operating voltage, predicted current density, and effective activation area; the theoretical cathode air mass flow rate required to maintain the predicted current density is calculated according to Faraday's law and stoichiometry; and the pressure ratio required by the air compressor is calculated based on the overall pressure drop of the cathode flow field.

[0063] In applications, the fuel cell output and cathode air demand are calculated. The total output power of the fuel cell is... From the working voltage Predicting current density and effective activation area The calculation shows that: (9) Calculate the theoretical cathode air mass flow rate required to maintain this current density based on Faraday's law and stoichiometry. Then, based on the cathode flow field pressure drop... Calculate the required pressure ratio of the air compressor (Pressure Ratio): (10) in, Standard atmospheric pressure This is the preset pipeline and back pressure loss coefficient.

[0064] The theoretical cathode air mass flow rate and pressure ratio are converted into equivalent flow rate and equivalent pressure ratio. Pre-calibrated compressor surge and blockage boundary functions are introduced. If the current operating point exceeds these pre-calibrations, the equivalent flow rate is forcibly truncated to the boundary value, resulting in a corrected compressor input operating condition. The corrected equivalent flow rate and equivalent pressure ratio are input into the compressor surrogate model to predict the compressor isentropic efficiency under the current operating condition. Based on the compressor isentropic efficiency, the system's true parasitic power and the fuel cell system's net power are calculated. Based on the fuel cell system's net power and the total input power calculated using the consumed hydrogen, the fuel cell system efficiency is obtained. Optimization decisions are made based on the fuel cell system efficiency and preset internal evaluation indicators.

[0065] In this application, the calculated airflow and pressure ratio are converted into equivalent flow and equivalent pressure ratio. To ensure the safe operation of the air compressor, this embodiment introduces pre-calibrated air compressor surge and blockage boundary functions. If the current operating point exceeds this safety envelope, the equivalent flow is forcibly truncated to the boundary value, thus obtaining the corrected air compressor input operating condition. The air compressor parasitic power and system efficiency are calculated, and the corrected equivalent flow and equivalent pressure ratio are input into a pre-trained air compressor machine learning surrogate model to predict the air compressor isentropic efficiency under the current operating condition. Furthermore, the true parasitic power of the calculation system is... Net power of fuel cell system : (11) (12) in, It is the inlet velocity of the flow channel. It is the inlet area of ​​the flow channel.

[0066] Total input power calculated based on hydrogen consumption To obtain the efficiency of the fuel cell system : (13) When generating a global evaluation mesh in a multidimensional parameter space, geometric interference and physical boundary constraints need to be set to avoid redundant calculations in invalid or physically unmanufacturable design regions.

[0067] Specifically, before making optimization decisions based on the fuel cell system efficiency and the comprehensive performance evaluation index of the internal flow field, effective design points are marked based on the preset design variable boundaries and the preset interference judgment formula, and optimization decisions are made only for the effective design points.

[0068] In this embodiment, the design variable boundaries are shown in Table 3. Geometric interference and physical boundary constraints need to take into account the total length of the flow channel. Limited by the fact that the axial span of a single baffle increases significantly when the baffle height increases or the base angle decreases, an interference criterion formula is constructed: (14) in: for The total length of the bottom edge of the baffle.

[0069] Table 3 Before performing optimization, the system automatically checks the grid space if and only if At that time, the set of geometric parameters Only then are they marked as valid design points for subsequent evaluation. For the baseline case, i.e., without baffles, The system uses special markings to prevent calculation errors.

[0070] After obtaining the predicted data for all valid design points, this embodiment conducts multi-objective optimization and decision-making to improve system efficiency while ensuring the uniformity of current density distribution. Building upon the Pareto optimal set obtained using the non-dominated sorting method, this embodiment further proposes a multi-objective decision-making algorithm to obtain a unique comprehensive optimal solution. Since system efficiency is a benefit-oriented indicator (maximizing expected value), while the relative standard deviation of current density is a cost-oriented indicator (minimizing expected value), the two often have a physically interdependent game-like relationship. To solve this multi-objective game problem, this embodiment divides the optimization process into two stages: solution and decision-making.

[0071] Specifically, optimization decisions based on fuel cell system efficiency and comprehensive performance evaluation indicators within the flow field include a solution phase and a decision phase.

[0072] The Pareto front algorithm is used in the solution phase, and the output consists of a set of Pareto optimal solutions that are not mutually dominant.

[0073] Reference Figure 8 , Figure 8 The multi-objective optimization results provided for the embodiments of this application.

[0074] In the application, the Pareto front algorithm is used to extract the Pareto optimal solution set. Multi-objective optimization algorithms, such as non-dominated sorting algorithms or the NSGA-II non-dominated sorting genetic algorithm with elitist retention, are used to perform global iterative optimization within the effective constraint space of geometric features and working conditions. By continuously eliminating inferior solutions, the algorithm ultimately outputs a Pareto optimal solution set consisting of a set of non-dominated solutions that are not mutually dominant. Figure 8As shown, this intuitively and comprehensively reveals the multi-objective game boundary curve between system efficiency and flow field uniformity.

[0075] In the decision-making stage, a multi-objective decision-making algorithm is introduced. Based on the Pareto optimal solution set, the high-dimensional discrete data points in the set are evaluated twice. The corresponding weights are assigned according to the actual focus of the engineering, and the relative closeness between each non-dominated solution and the theoretical positive and negative ideal solutions is calculated. Thus, the unique compromise solution with the best comprehensive performance is determined as the optimal combination of flow field geometric parameters as the output.

[0076] In this application, a multi-objective decision-making algorithm is introduced for the final engineering decision. Based on obtaining the Pareto optimal solution set, a comprehensive evaluation decision model is used to perform a secondary evaluation of the high-dimensional discrete data points within the set. For example, the TOPSIS algorithm, which approximates the ideal solution ranking method, is used in this embodiment to evaluate benefit-type indicators. Cost-related indicators Normalization is performed, and corresponding weights are assigned based on the actual engineering focus. The relative closeness between each non-dominated solution and the theoretical positive and negative ideal solutions is calculated. Finally, the unique compromise solution with the best overall performance is strictly determined from these solutions, as shown in Table 4. This provides clear quantitative guidance for the selection of flow field structure for vehicle fuel cells in different scenarios such as long driving range or high durability.

[0077] Table 4 In conclusion, the multi-scale intelligent collaborative optimization method for the fuel cell cathode flow field in this embodiment is summarized as follows.

[0078] Reference Figure 9 , Figure 9 The overall flowchart of the multi-scale intelligent collaborative optimization method for the cathode flow field of fuel cells provided in this application embodiment is shown.

[0079] like Figure 9As shown, the multi-scale intelligent collaborative optimization method for the fuel cell cathode flow field in this embodiment is implemented in four core stages. First, this embodiment innovatively introduces three-dimensional numerical model parameter calibration based on a small amount of experimental data. Relying on the experimental polarization curves of a small number of single fuel cells, automated scripts combined with numerical simulation software are used for rapid parameter identification to accurately calibrate the three-dimensional high-precision fuel cell numerical model. Subsequently, automated batch sampling and calculation are performed within the defined geometric design parameter space to generate a high-quality training dataset. Next, a novel multilayer perceptron algorithm combined with a physical constraint (PINN) framework is proposed, and a fuel cell performance prediction surrogate model is constructed based on this framework. Finally, this surrogate model is innovatively coupled with a pre-constructed air compressor surrogate model. With system efficiency and internal evaluation indicators (such as cathode current density distribution uniformity) as comprehensive optimization objectives, multi-objective optimization algorithms and decision algorithms are used to perform a global optimal search for the core geometric design parameters of the flow field, thereby achieving a synergistic improvement in the performance and lifespan of automotive fuel cells.

[0080] The multi-scale intelligent collaborative optimization device for the cathode flow field of the fuel cell in this embodiment will now be described.

[0081] Reference Figure 10 , Figure 10 The diagram shows the structure of the multi-scale intelligent collaborative optimization device for the cathode flow field of a fuel cell provided in this application embodiment; in the diagram: 10, fuel cell three-dimensional numerical model calibration module; 20, training dataset generation module; 30, fuel cell proxy model construction module; 40, fuel cell and air compressor coupling optimization module.

[0082] like Figure 10 As shown, in a second aspect, this embodiment also discloses a multi-scale intelligent cooperative optimization device for the cathode flow field of a fuel cell, which applies the multi-scale intelligent cooperative optimization method for the cathode flow field of a fuel cell as described above, including: The fuel cell three-dimensional numerical model calibration module 10 is used to quickly identify parameters and accurately calibrate a high-precision fuel cell three-dimensional numerical model based on experimental test data of polarization curves of a single fuel cell by using automated scripts in conjunction with three-dimensional numerical simulation. The training dataset generation module 20 is used to perform automated batch sampling and CFD joint calculation based on the three-dimensional numerical model of the fuel cell within the preset geometric design parameter space to obtain various performance responses of the fuel cell and generate a high-quality training dataset. The fuel cell proxy model construction module 30 is used to construct a fuel cell proxy model by combining the multilayer perceptron algorithm and physical constraints. This fuel cell proxy model is used to deeply learn the nonlinear mapping relationship between geometric and operating parameters and system performance. The fuel cell and air compressor coupling optimization module 40 is used to couple the fuel cell proxy model with the preset air compressor proxy model at the system level, comprehensively consider the system efficiency and preset internal indicators, run the preset multi-objective optimization algorithm and decision algorithm, and output the optimal combination of flow field geometric parameters to complete the multi-scale intelligent collaborative optimization of the fuel cell cathode flow field.

[0083] It should be noted that the fuel cell cathode flow field multi-scale intelligent collaborative optimization device of this embodiment corresponds to the aforementioned fuel cell cathode flow field multi-scale intelligent collaborative optimization method. Therefore, the contents not specifically described in the fuel cell cathode flow field multi-scale intelligent collaborative optimization device of this embodiment, including but not limited to functional definitions, working principles, and technical effects, can all be referred to the description in the aforementioned fuel cell cathode flow field multi-scale intelligent collaborative optimization method, and will not be repeated here.

[0084] The following describes the alternative technical contents of the multi-scale intelligent collaborative optimization method and device for the fuel cell cathode flow field in this embodiment, which can be replaced in practical applications.

[0085] The underlying physics solver and automated driving scripts are variable. When performing high-precision 3D numerical simulations of flow fields, parameter identification, and batch automated data generation, in addition to using the Ansys Fluent solver and the Python-based PyFluent interface as described in the embodiment, the underlying multiphysics solver engine can also be replaced by other fluid dynamics software with multi-field coupling calculation capabilities, such as COMSOLMultiphysics, OpenFOAM, or STAR-CCM+. The corresponding automated driving interactive scripts and environment can also be replaced by MATLAB (e.g., using the COMSOL with MATLAB interface to achieve co-simulation), C++, or other programming languages ​​with low-level API calls and data batch processing capabilities.

[0086] The combination of microscopic physical and electrochemical parameters to be identified is variable. During high-precision automated parameter calibration of the three-dimensional numerical model, the combination of unknown parameters used as the optimization target is not limited to the six specific microscopic parameters listed in the examples (i.e., cathode reference exchange current density, gas diffusion pore blockage index, GDL porosity, CL porosity, cathode ionomer resistance, and water removal coefficient). Depending on different flow field configurations, membrane electrode assembly (MEA) material properties, and the accuracy requirements of actual physical modeling, the set of parameters to be identified can be equivalently replaced by any combination of unknown parameters in the fuel cell conservation equations and porous media transport models, including but not limited to flexible combinations of microscopic electrochemical kinetics and porous media transport parameters such as bipolar plate and diffusion layer contact resistance, catalyst layer specific surface area, proton exchange membrane aqueous permeability, water condensation / evaporation rate constant, and other multi-component effective diffusion coefficients.

[0087] The automated parameter identification and optimization algorithm is variable. The global parameter identification and calibration algorithm used to minimize the error between the simulated polarization curve and the real experimental data, in addition to the differential evolution algorithm (DE) used in the embodiments, can also be equivalently replaced by other heuristic intelligent algorithms with global optimization capabilities or probabilistic graphical models, including but not limited to particle swarm optimization (PSO), genetic algorithm (GA), Rao optimization algorithm with chaotic characteristics, walrus optimization algorithm (WOA), grid search, or Bayesian optimization algorithm.

[0088] The objective function for parameter identification is variable. When performing automated parameter calibration, the objective function used to quantify the difference between the simulated polarization curve and the actual experimental data, besides using the mean squared error (MSE) or relative mean squared error in the examples, can also be equivalently replaced by other statistical error assessment metrics that can effectively evaluate the accuracy of data fitting, including but not limited to root mean squared error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), Huber loss function, or metrics based on the coefficient of determination (…). The comprehensive fitness function is constructed.

[0089] Experimental design and dataset sampling methods are variable. In order to generate high-quality training datasets for CFD models, in addition to the orthogonal experimental method in the examples, other space-filling design methods, such as Latin hypercube sampling and Sobol quasi-random sequences, can also be used to effectively sample in the parameter space when driving CFD models to generate high-quality training datasets.

[0090] The method for constructing the fuel cell surrogate model in system-level coupling is variable. When constructing a fast predictive surrogate model for system-level coupling evaluation, in addition to adopting the Physical Information Neural Network (PINN) and Multilayer Perceptron (MLP) architecture that integrates dual physical constraints as described in the embodiment, it can also be equivalently replaced by other machine learning or surrogate modeling techniques with multidimensional regression prediction capabilities, including but not limited to Support Vector Regression (SVR), Gaussian Process Regression (GPR / Kriging model), Random Forest Regression (RF), Gradient Boosting Decision Tree (such as XGBoost, LightGBM), Radial Basis Function (RBF) Neural Network, or standard Deep Neural Network (DNN) without imposed physical constraints.

[0091] The comprehensive performance evaluation index within the flow field is variable. Besides using the relative standard deviation of the cathode current density distribution in the embodiment, this also includes... In addition to serving as a cost-based internal evaluation metric for measuring the flow field, it can also be replaced by other physical quantities that can characterize flow field performance, hydrothermal management capabilities, and reaction uniformity, including but not limited to the relative standard deviation of oxygen concentration distribution, the uniformity of liquid water mass fraction distribution, the variance of flow field pressure drop distribution, or the relative standard deviation of bipolar plate temperature distribution.

[0092] The key technical optimizations of this embodiment are summarized below.

[0093] This embodiment discloses a fuel cell proxy model and its construction method that integrates dual physical constraints. The method is characterized by the design of a Physical Information Neural Network (PINN), in which hard constraints based on overpotential are embedded in the forward propagation of the network topology to strictly lock the zero-point extreme boundary at the open-circuit voltage. Furthermore, a soft penalty based on the monotonicity of the polarization curve is introduced into the backpropagation loss function to ensure that the model's high-frequency predictions strictly conform to the underlying electrochemical laws. Figure 3 As shown.

[0094] This embodiment discloses a system-level multi-objective collaborative optimization method considering the dynamic energy consumption of an air compressor. The method is characterized by deep series coupling of the fuel cell proxy model and the air compressor model, taking into account the problem of high error in parasitic power estimation caused by the dynamic changes of the air compressor, and comprehensively quantifying the surge in parasitic power consumption caused by changes in flow field geometry. With the goal of maximizing system efficiency and optimizing internal evaluation indicators (such as the uniformity of current density distribution), under the premise of establishing geometric constraints, the Pareto front algorithm is used to extract the Pareto optimal solution set, and a multi-objective decision algorithm is introduced to make the final engineering decision, so as to obtain the comprehensive optimal configuration solution.

[0095] This embodiment discloses an AI-driven multi-scale collaborative optimization method for the cathode flow field of a vehicle fuel cell. This method constructs a complete data-driven closed-loop framework encompassing experimental calibration, automated simulation, surrogate model construction, and system collaborative optimization. Figure 1 or Figure 9 The flowchart is shown.

[0096] In conclusion, the technical effects of this embodiment are summarized as follows.

[0097] First, addressing the issues of parameter mismatch between simulation and experiment, and the inefficiency of manual parameter tuning, this embodiment utilizes optimization algorithms and automated interactive interfaces to achieve intelligent calibration of underlying micro-parameters, completely eliminating reliance on experience and significantly improving the baseline accuracy of the numerical model. Building upon this, to overcome the limitations of traditional discrete optimization, which easily gets trapped in local optima and is computationally time-consuming, this embodiment constructs a neural network surrogate model that integrates physical constraints, achieving globally accurate mapping within the multi-dimensional parameter space at extremely low computational cost. Most importantly, addressing the serious deficiency of existing optimization methods that often ignore the dynamic characteristics of the air compressor, this embodiment innovatively performs deep system-level coupling between the fuel cell surrogate model and the air compressor model. This effectively avoids the development pitfall of high parasitic power estimation errors caused by dynamic changes in air compressor efficiency. Finally, relying on a multi-objective decision-making mechanism, it successfully achieves a synergistic improvement in system efficiency and current density distribution uniformity globally, providing a scientific and automated solution for the design of high-efficiency, high-durability bipolar plates for automotive fuel cells.

[0098] In the embodiments provided in this application, it should be understood that the embodiments described herein can be implemented in hardware, software, firmware, middleware, code, or any suitable combination thereof. For hardware implementation, the processor may be implemented in one or more of the following: application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, other electronic units designed to implement the functions described herein, or combinations thereof. For software implementation, some or all of the processes of the embodiments may be performed by a computer program instructing the associated hardware. During implementation, the program may be stored in a computer-readable storage medium or transmitted as one or more instructions or code on a computer-readable storage medium. Computer-readable storage media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of a computer program from one place to another. Storage media may be any available medium accessible to a computer. Computer-readable storage media may include, but are not limited to, RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code having the form of instructions or data structures and accessible to a computer.

[0099] Finally, it should be noted that the above description is only a preferred embodiment of this application and is not intended to limit this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A multi-scale intelligent collaborative optimization method for the cathode flow field of a fuel cell, characterized in that, include: Based on experimental test data of polarization curves of single fuel cells, automated scripts are used in conjunction with three-dimensional numerical simulation to quickly identify parameters and accurately calibrate a high-precision three-dimensional numerical model of fuel cells. Based on the three-dimensional numerical model of fuel cells, automated batch sampling and CFD joint calculation are performed within the preset geometric design parameter space to obtain various performance responses of fuel cells and generate a high-quality training dataset. A fuel cell surrogate model is constructed by combining multilayer perceptron algorithm and physical constraints. This fuel cell surrogate model is used to deeply learn the nonlinear mapping relationship between geometric and operating parameters and system performance. The fuel cell proxy model is coupled with the preset air compressor proxy model at the system level. Taking into account the system efficiency and preset internal indicators, the preset multi-objective optimization algorithm and decision algorithm are run to output the optimal combination of flow field geometric parameters to complete the multi-scale intelligent collaborative optimization of the fuel cell cathode flow field.

2. The multi-scale intelligent collaborative optimization method for the cathode flow field of a fuel cell according to claim 1, characterized in that, The calibration of the three-dimensional numerical model of the fuel cell includes: Import experimental test data of polarization curves of single fuel cells; The physical value boundaries of the micro-critical parameters to be identified are set; wherein, the micro-critical parameters to be identified include cathode reference exchange current density, gas diffusion pore blockage index, GDL porosity, CL porosity, cathode ionomer resistance, and water removal coefficient. The preset optimization algorithm generates the individual parameter vectors for the current iteration; Automated interface mapping parameters to the three-dimensional numerical model of the fuel cell; The solver performs multiphysics coupled simulation calculations; Extract steady-state simulation results for each operating condition; Calculate the value of a preset objective function; wherein the objective function is: in: The value of the objective function. Indicates the first Simulated current density at each point, in A / m² 2 ; Indicates the first The experimental current density at each point is expressed in A / m². 2 ; , The number of experimental test data points; Determine whether the convergence condition or the maximum number of iterations is met; if so, output the optimal combination of simulation parameters that minimizes the error; otherwise, retain the better individual and proceed to the next iteration.

3. The multi-scale intelligent collaborative optimization method for the cathode flow field of a fuel cell according to claim 1, characterized in that, The generation of the training dataset includes: Based on the three-dimensional numerical model of the fuel cell, the key geometric design parameters of the cathode flow field are set as design variables, and a geometric parameter combination matrix covering the typical features of the design space is generated by a preset space filling design method; wherein, the key geometric design parameters include the number of trapezoidal baffles, the bottom angle of the baffles, and the height of the baffles in the cathode flow field; The design scheme is traversed using automated scripts, which automatically load boundary conditions and trigger the solution calculation. Extract the overall pressure drop and polarization curve data of the simulated cathode flow field, as well as the preset internal evaluation indicators, to complete the construction of the underlying database required for system-level multi-objective optimization; wherein, the internal evaluation indicators include the relative standard deviation of current density distribution.

4. The multi-scale intelligent collaborative optimization method for the cathode flow field of a fuel cell according to claim 1, characterized in that, The fuel cell proxy model takes preset key geometric design parameters and operating parameters as inputs, and outputs current density, cathode flow field pressure drop and preset internal evaluation indicators.

5. The multi-scale intelligent collaborative optimization method for the cathode flow field of a fuel cell according to claim 4, characterized in that, In the forward propagation of the fuel cell proxy model, a physical hard constraint mechanism is incorporated, including: in: The current voltage overpotential is non-negative. , It is the open circuit voltage. The anchor value obtained after standard deviation normalization It is the input voltage after standard deviation normalization; This is the normalized predicted current density value. It is the normalized physical zero-point value of the current density. This represents the predicted current density value for a multilayer perceptron network. This is the Softplus function; For normalized Predicted value For normalized Physical zero point value For multilayer perceptron networks Predicted value; It is an internal evaluation index, representing the relative standard deviation of the current density distribution, used to evaluate the uniformity of the current density distribution; The original prediction of the overall pressure drop output of the cathode flow field.

6. The multi-scale intelligent collaborative optimization method for the cathode flow field of a fuel cell according to claim 4, characterized in that, When the fuel cell proxy model updates network weights via backpropagation, the overall objective function is... for: in: It is the mean square error. , This is the output value of the fuel cell proxy model. It is the actual value corresponding to the output value. This indicates the preset penalty weight. , The total number of data points in the training dataset generated for the three-dimensional numerical model of the fuel cell; The loss term is a monotonic physical penalty. , This represents the partial derivative of the predicted current density with respect to voltage. This represents the total number of unlabeled physical points. Preset for balancing and The coefficient.

7. The multi-scale intelligent collaborative optimization method for the cathode flow field of a fuel cell according to claim 1, characterized in that, The system-level coupling of the fuel cell proxy model with the preset air compressor proxy model includes: The total output power of the fuel cell is calculated based on the operating voltage, predicted current density, and effective activation area; the theoretical cathode air mass flow rate required to maintain the predicted current density is calculated according to Faraday's law and stoichiometry; and the pressure ratio required by the air compressor is calculated based on the overall pressure drop of the cathode flow field. The theoretical cathode air mass flow rate and pressure ratio are converted into equivalent flow rate and equivalent pressure ratio. Pre-calibrated compressor surge and blockage boundary functions are introduced. If the current operating point exceeds these pre-calibrations, the equivalent flow rate is forcibly truncated to the boundary value, resulting in a corrected compressor input operating condition. The corrected equivalent flow rate and equivalent pressure ratio are input into the compressor surrogate model to predict the compressor isentropic efficiency under the current operating condition. Based on the compressor isentropic efficiency, the system's true parasitic power and the fuel cell system's net power are calculated. Based on the fuel cell system's net power and the total input power calculated using the consumed hydrogen, the fuel cell system efficiency is obtained. Optimization decisions are made based on the fuel cell system efficiency and the comprehensive performance evaluation index within the flow field.

8. The multi-scale intelligent collaborative optimization method for the cathode flow field of a fuel cell according to claim 7, characterized in that, Before making optimization decisions based on the comprehensive performance evaluation indicators of fuel cell system efficiency and internal flow field, valid design points are marked based on preset design variable boundaries and preset interference judgment formulas, and optimization decisions are made only for valid design points; wherein, the interference judgment formula is: in: for The total length of the bottom edge of the baffle plate. For the number of baffles, For the height of the baffle, For the bottom corner of the baffle; if and only if At that time, geometric parameters The marked point is a valid design point.

9. The multi-scale intelligent collaborative optimization method for the cathode flow field of a fuel cell according to claim 7, characterized in that, The optimization decision based on the fuel cell system efficiency and the comprehensive performance evaluation index of the internal flow field includes a solution stage and a decision stage. The solution stage employs the Pareto front algorithm, outputting a Pareto optimal solution set consisting of a set of non-dominated solutions that are not mutually dominant. The decision stage introduces a multi-objective decision algorithm, which performs a secondary evaluation of the high-dimensional discrete data points within the Pareto optimal solution set, assigns appropriate weights based on the actual engineering focus, calculates the relative closeness between each non-dominated solution and the theoretical positive and negative ideal solutions, and thus determines the unique compromise solution with the best overall performance as the output optimal combination of flow field geometric parameters.

10. A multi-scale intelligent collaborative optimization device for the cathode flow field of a fuel cell, employing the multi-scale intelligent collaborative optimization method for the cathode flow field of a fuel cell as described in any one of claims 1 to 9, characterized in that, include: The fuel cell 3D numerical model calibration module is used to quickly identify parameters and accurately calibrate a high-precision fuel cell 3D numerical model based on experimental test data of polarization curves of a single fuel cell using automated scripts in conjunction with 3D numerical simulation. The training dataset generation module is used to automatically perform batch sampling and CFD joint calculation based on the three-dimensional numerical model of the fuel cell within a preset geometric design parameter space to obtain various performance responses of the fuel cell and generate a high-quality training dataset. The fuel cell proxy model construction module is used to construct a fuel cell proxy model by combining multilayer perceptron algorithm and physical constraints. This fuel cell proxy model is used to deeply learn the nonlinear mapping relationship between geometric and operating parameters and system performance. The fuel cell and air compressor coupling optimization module is used to couple the fuel cell proxy model with the preset air compressor proxy model at the system level. It comprehensively considers the system efficiency and preset internal indicators, runs preset multi-objective optimization and decision algorithms, and outputs the optimal combination of flow field geometric parameters to complete the multi-scale intelligent collaborative optimization of the fuel cell cathode flow field.

Images