Generative adversarial network-based fuel cell catalyst layer three-dimensional structure generation method

Abstract

The invention discloses a fuel cell catalyst layer three-dimensional structure generation method based on a generative adversarial network. The method comprises the steps of catalyst layer random reconstruction, generative adversarial network model training and catalyst layer generation. The three-dimensional catalyst layer is constructed step by step by adjusting the volume fractions of platinum, carbon and electrolyte components in the catalyst layer and the generation probability of the components on each structural lattice point. The generator generates a three-dimensional catalyst layer structure according to the input random noise vector and volume fraction information, and the discriminator evaluates the difference between the generated catalyst layer and the actual catalyst layer. And the generator selects the component with the maximum probability in the catalyst layer structure as the component of the grid point through multi-layer convolutional neural network processing, and finally a single-channel three-dimensional structure is formed. Compared with a traditional random reconstruction method, the method has the advantages that tedious multiple traversal processes are avoided, the calculation efficiency is greatly improved, the volume fractions of platinum, carbon and electrolyte can be accurately controlled, it is ensured that the catalyst layer meets actual requirements, and the method has wide application prospects and engineering applicability.

Description

Technical Field

[0001] The present invention belongs to the field of electrochemical fuel cells, and in particular relates to a method for generating a three-dimensional structure of a fuel cell catalyst layer based on a generative adversarial network in a proton exchange membrane fuel cell. Background Art

[0002] Proton exchange membrane fuel cells, as a clean energy technology with efficient energy conversion and zero emission characteristics, have been widely used in the field of sustainable energy. One of the core components of proton exchange membrane fuel cells is the catalyst layer, which is composed of platinum, carbon and electrolyte. It is the core of electrochemical reaction and its structure directly affects the overall performance of the battery. However, the design and optimization of the catalyst layer faces many challenges, especially how to efficiently generate three-dimensional catalyst layers with different structures to meet the needs of various working conditions.

[0003] The microstructure and component distribution of the catalytic layer directly affect its electrochemical performance. The current method of generating the catalytic layer mainly relies on random reconstruction technology, that is, by adjusting the volume fraction of each component and the generation probability at each grid point, and traversing multiple times to generate different catalytic layer structures. Although this process can obtain new structures, it often consumes a lot of time and computing resources and is inefficient.

[0004] In this context, the present invention proposes a method for generating a three-dimensional structure of a fuel cell catalyst layer, which, by utilizing the advantages of generative adversarial networks, can efficiently and flexibly batch generate a three-dimensional catalyst layer that meets the preset requirements of the volume fractions of platinum, carbon, and electrolyte. This method not only effectively shortens the generation time, but also provides more accurate and diversified catalyst layer design solutions for a variety of applications. Summary of the invention

[0005] The purpose of the present invention is to provide a method for generating a three-dimensional structure of a fuel cell catalyst layer based on a generative adversarial network, which can efficiently generate catalyst layers of different three-dimensional structures through a generative network and meet the volume fraction requirements of components such as platinum, carbon and electrolyte. Compared with the current random reconstruction method, it not only improves the generation efficiency and avoids the process of multiple traversals and adjustments, but also can generate different catalyst layers in batches, which is suitable for the design and optimization of large-scale fuel cell catalyst layers.

[0006] The method for generating the three-dimensional structure of the fuel cell catalyst layer based on the generative adversarial network includes three parts: random reconstruction of the catalyst layer, training of the generative adversarial network model and generation of the catalyst layer. The specific steps are as follows:

[0007] (1) Random reconstruction of the catalytic layer

[0008] By adjusting the volume fractions of the three components of platinum, carbon and electrolyte in the catalytic layer, as well as the generation probability of the components at each structural lattice point of platinum, carbon and electrolyte, a three-dimensional catalytic layer that meets the design requirements is gradually constructed.

[0009] (1.1) Determine the volume fraction range of the three components of platinum, carbon and electrolyte in the catalyst layer according to design requirements, and perform batch random reconstruction.

[0010] (1.2) Based on the set volume fraction, the generation probability of the three components, platinum, carbon and electrolyte, at different lattice points is controlled to generate the structure point by point.

[0011] (1.3) Through multiple iterations and adjustments, ensure that the proportions of the three components in the catalyst layer structure meet the volume fraction requirements, and construct catalyst layer structures with different volume fractions by cycling steps 1.1-1.3.

[0012] (2) Conduct adversarial network model training

[0013] The generator of the adversarial network model generates a three-dimensional catalytic layer structure according to the input random noise vector and volume fraction information, and the discriminator of the adversarial network model is used to evaluate the difference between the generated catalytic layer and the actual catalytic layer. The generator of the adversarial network model and the discriminator of the adversarial network model can be referred to as the generator and the discriminator.

[0014] (2.1) Set the input features of the generator. The features consist of two parts: the first part is the noise vector, which represents the randomness of the generator in generating diverse structures; the second part is the volume fraction of platinum, carbon and electrolyte in the catalytic layer, which serves as a control parameter to guide the generation process of the adversarial network model.

[0015] (2.2) The generator processes the input data through a multi-layer convolutional neural network, and generates a four-channel tensor through a network layer with a residual module, corresponding to the distribution of pores, electrolytes, carbon, and platinum. The value of each channel is 0 or 1, indicating whether the component exists at the grid point.

[0016] (2.3) The loss function of the generator consists of two parts: data loss and physical information loss. The data loss optimizes the generator by calculating the difference between the generated result and the target structure, and sets a smoothing factor to improve the generalization ability of the adversarial network model. The physical information loss calculates the volume fraction of each channel in the generated result by taking the average value, and compares it with the target volume fraction to ensure that the generated result meets the specified ratio requirements of platinum, carbon and electrolyte.

[0017] (2.4) The generated four-channel tensor is used with a function that determines the maximum value, and the component with the highest probability at each grid point in the catalyst layer structure is selected as the component of the grid point, ultimately forming a single-channel three-dimensional structure.

[0018] (2.5) The input tensor of the discriminator is to select a single-channel three-dimensional structure from the four-channel tensor output by the generator. Through learning, the three-dimensional structure generated by each generator is evaluated to see whether it is close to the real catalytic layer, thereby optimizing the generation process of the generator.

[0019] (3) Generation of catalytic layer

[0020] The trained generator is used to generate the catalytic layer, and the structure is generated in batches according to the volume fractions of platinum, carbon and electrolyte in the catalytic layer set during generator training.

[0021] The characteristics and benefits of the present invention are:

[0022] (1) The characteristics of the present invention are its high efficiency, accuracy, structural diversity and rapid convergence ability. By adopting a generative adversarial network, the catalytic layer can be generated in large quantities, fundamentally avoiding the tedious traversal process in the traditional method, which not only greatly improves the calculation efficiency, but also saves time cost. Secondly, the generator introduces a smoothing factor and a physical information loss function to ensure that the generated catalytic layer meets the actual requirements at the physical level, especially in terms of controlling the volume fraction. The generator can accurately adjust the volume fractions of platinum, carbon and electrolyte, thereby ensuring that the final generated catalytic layer meets the needs of different applications.

[0023] (2) It provides higher diversity in the catalyst layer structure, has stronger flexibility, and can adjust the component distribution according to different needs. Compared with the traditional method, the present invention can more conveniently realize the generation of diversified structures and meet the optimization design of different performance requirements of fuel cells.

[0024] (3) By introducing the residual module, the generator can accelerate the training process, making the model converge faster, while improving the stability and reliability of the generated results. In addition, the generator can flexibly control the volume fraction ratio of platinum, carbon and electrolyte to meet the needs of different experiments and engineering applications, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 A diagram of the training process of generating an adversarial network in the present invention.

[0026] Figure 2 The present invention generates a final result graph of the adversarial network structure. Figure 2 The three-dimensional numerical simulation effect is expressed using color images. If black and white images are used, it is impossible to distinguish the structural features and implementation effects that each microscopic physical model wants to express. DETAILED DESCRIPTION

[0027] The specific modeling method of the present invention is further described below with reference to the accompanying drawings and examples. It should be noted that this example is a narrative description for the purpose of clearly explaining the simulation steps, and does not limit the protection scope of the method of the present invention.

[0028] The three-dimensional structure generation method of the fuel cell catalyst layer based on the generative adversarial network includes three parts. The specific method is as follows:

[0029] (1) Random reconstruction of the catalytic layer

[0030] During the catalytic layer generation process, the volume fraction ranges of platinum, carbon and electrolyte in the catalytic layer are determined by specifying the platinum loading, platinum to carbon mass ratio and electrolyte to carbon mass ratio of the catalytic layer, where the volume fraction of platinum is set between 0.035-0.045, the volume fraction of carbon is set between 0.3-0.4, and the volume fraction of the electrolyte is set to 0.17-0.37.

[0031] In the process of generating the catalytic layer, the volume fraction of each component is set according to the requirements, and the structure of the three-dimensional catalytic layer is generated point by point by controlling the generation probability of each component at different grid points in the catalytic layer structure. The size of the generated catalytic layer is set to 400 nanometers in length, 400 nanometers in width and 5120 microns in height, respectively, and the structural resolution is 5 nanometers. Such a resolution can ensure that the details of the generated catalytic layer are accurate and meet the design requirements. The amount of data generated is 10,000 different catalytic layers to meet the generation requirements of different volume fractions and structural sizes.

[0032] (2) Generative Adversarial Network Model Training

[0033] (2.1) The model structure of the generative adversarial network adopts a multi-layer convolutional neural network combined with a residual module. The training process is as follows: Figure 1 The input of the generator consists of two parts: one part is a random noise vector of length 100, which provides the randomness for generating diversified catalyst layer structures; the other part is the volume fraction information including platinum, carbon and electrolyte, which guides the generator to generate the components of the catalyst layer according to the set volume fraction information.

[0034] The generator of the generative adversarial network adopts five deconvolution layers. Except for the first deconvolution layer, the bilinear interpolation upsampling method is used before each deconvolution layer to reduce the number of parameters. Except for the last deconvolution layer, the residual module is used after each deconvolution layer. Finally, a four-channel three-dimensional tensor with the same size as the catalytic layer is generated through a normalized exponential function. Each channel represents the distribution of pores, electrolytes, carbon and platinum.

[0035] The loss function of the generator consists of two parts: data loss and physical information loss. The data loss optimizes the difference between the generated result and the real catalytic layer structure generated by random reconstruction through the cross entropy loss function; while the physical information loss ensures that the volume fraction of the generated catalytic layer meets the set requirements for the ratio of platinum, carbon and electrolyte through the mean square error function. The task of the discriminator is to judge whether the generated catalytic layer is similar to the real catalytic layer. Its model structure is similar to that of the generator. It extracts the features of the input three-dimensional catalytic layer structure through five convolutional layers, and finally outputs a probability value, which indicates the possibility that the discriminator judges that the three-dimensional catalytic layer structure is the real catalytic layer structure. The loss function of the discriminator uses binary cross entropy loss to evaluate the authenticity of the catalytic layer structure generated by the generator.

[0036] (2.2) During the training process, the generator and discriminator are optimized using the Adam optimizer, with a learning rate of 0.0002, a batch size of 64, and a total number of training iterations of 10,000. With this setting, the generator can efficiently and stably learn to generate a catalyst layer structure that meets the volume fraction requirements. The discriminator helps the generator to continuously optimize the generation process to ensure that the final generated catalyst layer can approximate the real catalyst layer structure. The training process and final results are shown in Figure 2. Figure 2 shown.

[0037] (3) Generation of catalytic layer

[0038] After training, the generator can batch generate catalytic layer structures that meet the volume fraction requirements of platinum, carbon and electrolyte based on the volume fraction inputs of platinum, carbon and electrolyte in the catalytic layer and a random noise vector. Figure 2 This is the final result graph of the generative adversarial network structure. These generated catalytic layers can be used in a variety of applications such as catalytic layer optimization and performance improvement in fuel cell design, providing efficient solutions for experiments and engineering design.

[0039] The present invention generates a three-dimensional catalytic layer structure based on a method of generating adversarial networks. Compared with the traditional random reconstruction method, the catalytic layer can be generated efficiently and in large quantities, avoiding the cumbersome multiple traversal process and greatly improving the calculation efficiency. At the same time, by introducing a smoothing factor and a physical information loss function, the generator can accurately control the volume fractions of platinum, carbon and electrolyte to ensure that the generated catalytic layer meets actual needs. In addition, the generated catalytic layer structure has a high diversity, and the proportions of different components can be flexibly adjusted to meet various application requirements. The introduction of the residual module not only accelerates the training process of the model, but also improves the stability and reliability of the generated results, and has broad application prospects and strong engineering applicability.

Claims

1. A method for generating a three-dimensional structure of a fuel cell catalyst layer based on a generative adversarial network, characterized by: The generation method includes three parts: random reconstruction of the catalytic layer, training of the generative adversarial network model, and generation of the catalytic layer. The specific steps are as follows: (1) Random reconstruction of the catalytic layer By adjusting the volume fractions of the three components of platinum, carbon and electrolyte in the catalyst layer, as well as the generation probability of the components at each structural lattice point of platinum, carbon and electrolyte, a three-dimensional catalyst layer that meets the design requirements is gradually constructed. (1.1) Determine the volume fraction range of the three components of platinum, carbon and electrolyte in the catalyst layer according to design requirements, and perform batch random reconstruction; (1.2) Based on the set volume fraction, the generation probability of the three components of platinum, carbon and electrolyte at different grid points is controlled to generate the structure point by point; (1.3) Through multiple iterations and adjustments, ensure that the proportions of the three components in the catalyst layer structure meet the volume fraction requirements. By looping steps (1.1)-(1.3), catalyst layer structures with different volume fractions are constructed. (2) Conduct adversarial network model training The generator of the adversarial network model generates a three-dimensional catalytic layer structure based on the input random noise vector and volume fraction information, and the discriminator of the adversarial network model evaluates the gap between the generated catalytic layer and the actual catalytic layer. (2.1) Setting the input features of the generator. The features consist of two parts: the first part is the noise vector, which represents the randomness of the generator in generating diverse structures; the second part is the volume fraction of platinum, carbon and electrolyte in the catalyst layer, which serves as a control parameter to guide the generation process of the adversarial network model; (2.2) The generator of the adversarial network model processes the input data through a multi-layer convolutional neural network, and generates a four-channel tensor through a network layer with a residual module, which corresponds to the distribution of pores, electrolytes, carbon, and platinum. The value of each channel is 0 or 1, indicating whether the component exists at the grid point; (2.3) The loss function of the adversarial network model generator consists of two parts: data loss and physical information loss. The data loss optimizes the generator of the adversarial network model by calculating the difference between the generated result and the target structure, and sets a smoothing factor to improve the generalization ability of the adversarial network model. The physical information loss calculates the volume fraction of each channel in the generated result by taking the average value method and compares it with the target volume fraction to ensure that the generated result meets the specified ratio requirements of platinum, carbon and electrolyte. (2.4) Using the function that determines the maximum value of the generated four-channel tensor, the component with the highest probability in each grid point in the catalyst layer structure is selected as the component of the grid point, and finally a single-channel three-dimensional structure is formed; (2.5) The input tensor of the discriminator of the adversarial network model is to select a single-channel three-dimensional structure from the four-channel tensor output by the adversarial network model generator. Through learning, the three-dimensional structure generated by each adversarial network model generator is evaluated to see whether it is close to the real catalytic layer, thereby optimizing the generation process of the generator. (3) Generation of catalytic layer The trained anti-network model generator is used to generate the catalytic layer, and the structure is generated in batches according to the volume fractions of platinum, carbon and electrolyte in the catalytic layer set during the training of the anti-network model generator.

Images