A hydrogen fuel cell ultra-low platinum catalytic layer structure optimization control device and method

Abstract

The present application relates to a kind of hydrogen fuel cell ultra-low platinum catalytic layer structure optimization control device and method, by setting catalytic layer structure controllable synthesis module to carry out discrete dimension reduction and condition generation modeling, can efficiently synthesize a large amount of catalytic layer three-dimensional structure satisfying specified condition information, can fully explore high-dimensional structure design space, greatly reduce the computing burden.Not only this, in the cooperation setting of electrochemical property prediction module and statistical analysis device, the optimization control of ultra-low platinum catalytic layer can be decoupled into two sequentially executed steps, i.e., first synthesizing structures satisfying ultra-low platinum load, then quickly evaluating the electrochemical performance of the structures, and statistically analyzing the common characteristics of excellent structures, avoiding the dilemma of simultaneously reducing platinum load and optimizing performance, facilitating structure regulation engineering, and overcoming the technical defects of traditional methods for optimizing and regulating catalytic layer structures, which are costly and dependent on expert experience.

Description

Technical Field

[0001] This invention relates to the field of computer modeling and systems technology, and more specifically, to a device and method for optimizing and controlling the structure of an ultra-low platinum catalyst layer in a hydrogen fuel cell. Background Technology

[0002] Proton exchange membrane fuel cells (PEMFCs) have become the most widely used type of hydrogen fuel cell due to their high power density, fast start-up, and low operating temperature. They are also an ideal choice for the next generation of energy supply in the transportation sector and have broad application potential in portable power supplies, small power plants, and drones.

[0003] Although proton exchange membrane fuel cells (PEMFCs) have many advantages and broad application prospects, they are not yet fully commercialized and widely used in daily life because they still face bottlenecks in three areas: cost, performance, and lifespan. These bottlenecks primarily include the need to improve the performance of materials such as catalysts, reduce costs, and further optimize the structure of various components within the PEMFC.

[0004] In proton exchange membrane fuel cells (PEMFCs), the catalyst layer plays a crucial role as the site of electrochemical reactions, and its performance significantly impacts the overall performance of the fuel cell. A typical catalyst layer (CL) consists of randomly distributed catalyst (usually platinum) particles, a carbon support, ionomers, and pores. Generally, for catalysts with the same inherent activity, a larger electrochemical reaction area (ECSA) results in smaller activation losses. Meanwhile, proton conduction in the ionomer and electron conduction in the carbon-supported platinum determine the magnitude of ion and electron ohmic losses in the fuel cell, respectively, while oxygen transport within the catalyst layer has a decisive influence on the concentration loss. The highly coupled multi-physics charge-mass transfer and electrochemical reaction processes within the catalyst layer lead to mutual influences on activation losses, ohmic losses, and concentration losses in fuel cells with different catalyst layer structures. Furthermore, the disordered distribution of substances within the catalyst layer due to traditional catalyst layer fabrication processes makes it difficult to simultaneously reduce these electrochemical losses, resulting in lower catalyst utilization and higher mass transfer resistance. Furthermore, reducing the amount of the precious metal platinum used in the catalyst layer is key to reducing the cost of fuel cells (research indicates that catalysts account for 40% of the total fuel cell manufacturing cost). However, the significant increase in mass transfer resistance and the reduction in electrochemical reaction area in low-platinum catalyst layers severely hinder the improvement of cell performance. Therefore, optimizing and controlling the material distribution in the catalyst layer, and thus designing a high-performance low-platinum catalyst layer, is a crucial challenge that urgently needs to be overcome in the development of next-generation hydrogen fuel cells.

[0005] Current methods for optimizing and controlling catalyst layer structures generally employ a trial-and-error approach. This involves first fabricating the catalyst layer based on expert experience, then conducting experimental tests, and finally optimizing and improving it based on performance results. This approach demands a high level of expert experience, and due to its high cost, it can only conduct limited exploration, resulting in a very limited degree of optimization. Another trial-and-error method is based on computer simulation. First, numerical reconstruction methods are used to generate the catalyst layer structure. Then, pore-scale simulations are used to evaluate performance, and the reconstruction scheme is adjusted based on expert experience. This iterative process optimizes the structure. Although it uses computer simulation, the high dimensionality of the three-dimensional, multi-scale structure of the catalyst layer places a significant computational burden on numerical methods, again resulting in a very limited degree of optimization.

[0006] Therefore, existing trial-and-error methods, whether experimental or numerical simulation-based, are extremely costly and heavily reliant on expert experience for control. On the other hand, the catalyst layer structure is very complex and has a high design dimension. As a result, traditional methods, constrained by both high cost and high reliance on expert knowledge, have almost no chance of fully exploring the catalyst layer design space, thus failing to obtain the truly optimal structure and formulate optimal production and manufacturing recommendations. Summary of the Invention

[0007] The technical problem this invention aims to solve is how to overcome the shortcomings of existing technologies for optimizing and controlling the catalyst layer structure, which are characterized by high cost and reliance on expert experience. To overcome these shortcomings, this invention provides a device and method for optimizing and controlling the structure of an ultra-low platinum catalyst layer in a hydrogen fuel cell, specifically comprising a device for optimizing and controlling the structure of an ultra-low platinum catalyst layer in a hydrogen fuel cell and a method for optimizing and controlling the structure of an ultra-low platinum catalyst layer in a hydrogen fuel cell.

[0008] This invention provides a device for optimizing and controlling the structure of an ultra-low platinum catalyst layer in a hydrogen fuel cell, comprising:

[0009] The controllable synthesis module of the catalyst layer structure is set to sequentially perform discrete dimensionality reduction and conditional generation modeling on the three-dimensional scanning structural data of the fuel cell catalyst layer to obtain several sets of three-dimensional structure data of the catalyst layer that meet the specified condition information. A set of three-dimensional structure data of the catalyst layer depicts the three-dimensional structure of the catalyst layer with the specified condition information.

[0010] The electrochemical property prediction module communicates with the controllable synthesis module of the catalyst layer structure. It is set to predict the performance of the three-dimensional structure of the catalyst layer depicted by each group of three-dimensional structure data of the catalyst layer to obtain their respective electrochemical properties. Based on the electrochemical properties, all the depicted three-dimensional structures of the catalyst layer are classified into three categories: excellent performance, medium performance and poor performance. An equal number of random samples are retained in all categories for statistical analysis.

[0011] The statistical analysis device, communicating with the electrochemical property prediction module, is configured to perform numerical simulation and geometric statistical analysis on an equal number of random samples from the three categories obtained by the electrochemical property prediction module, in order to obtain structural statistical indicators reflecting the structural differences of each category; the structural statistical indicators include carbon content, average particle size, average ionomer thickness, local oxygen mass transfer resistance, and electrochemical reaction area.

[0012] The hydrogen fuel cell ultra-low platinum catalyst layer structure optimization and control device disclosed in this invention, by setting a controllable synthesis module for catalyst layer structure to perform discrete dimensionality reduction and conditional generation modeling, can efficiently synthesize a large amount of three-dimensional structure data of the catalyst layer that meets specified conditions and depicts the distribution characteristics of the catalyst layer. This allows for full exploration of the high-dimensional structure design space and provides a solid data foundation for discovering the common characteristics of high-performance structures and forming optimization and control suggestions. Simultaneously, the controllable synthesis module can also synthesize several sets of three-dimensional structure data of the catalyst layer based on the set initial slicing conditions, using three-dimensional structure conditions much larger than the training data size, achieving scalable structure synthesis and greatly reducing the computational burden. Furthermore, with the coordinated setup of the electrochemical property prediction module and the statistical analysis device, the optimization and control of the ultra-low platinum catalyst layer can be decoupled into two sequential steps: first, synthesizing structures that meet ultra-low platinum loading and rapidly evaluating the structures; second, statistically analyzing the common characteristics of excellent structures. This concept of decoupling platinum loading constraints from performance optimization reduces the difficulty of structural optimization and avoids the dilemma of simultaneously reducing platinum loading and optimizing performance. It is conducive to the engineering of structural regulation and overcomes the technical shortcomings of existing optimization and regulation of catalyst layer structure, which are high cost and rely on expert experience.

[0013] In one possible implementation, the controllable synthesis module for the catalyst layer structure includes:

[0014] The vector quantization variational autoencoder is configured to compress the three-dimensional scanned structural data to obtain continuous latent variables, and obtain the discrete variables corresponding to the continuous latent variables in the codebook. The discrete variables and their class probabilities are used to obtain several sets of three-dimensional structural data of the catalyst layer by upscaling the continuous latent variable space to the pixel space.

[0015] The conditional generation modeling device communicates with the vector quantization variational autoencoder and is configured to perform category modeling on the discrete variables based on the specified conditional information to obtain the category probabilities of the discrete variables.

[0016] The controllable synthesis module for catalytic layer structures, equipped with the aforementioned structure and functions, can divide the process of reconstructing continuous slices from scanned structural data into two stages: discrete dimensionality reduction performed by a vector quantization variational autoencoder and conditional generative modeling performed by a conditional generative modeling device. Since the three-dimensional catalytic layer structure has a very high dimensionality, directly performing generative modeling requires substantial computational resources. Therefore, this technical solution performs dimensionality reduction on the three-dimensional data and then performs generative modeling on the reduced space, which improves computational efficiency.

[0017] In one possible implementation, the vector quantization variational autoencoder includes:

[0018] The encoder is configured to compress the three-dimensional scanned structural data to obtain continuous latent variables;

[0019] The codebook unit, which communicates with both the encoder and the conditional generation modeling device, is configured to find the closest vector to the continuous latent variable in the codebook and obtain the position index of this closest vector, which is denoted as the discrete variable corresponding to the continuous latent variable.

[0020] The decoder, which communicates with both the encoder and the conditional generation modeling device, is configured to decode the discrete variables and their class probabilities into several sets of three-dimensional structural data of the catalytic layer by increasing the dimensionality from the continuous latent variable space to the pixel space.

[0021] The vector quantization variational autoencoder with the above structure can compress three-dimensional input data into continuous latent variables in a continuous latent variable space using an encoder, encode the continuous latent variables using codebook units, and then use the position numbers of the continuous latent variables in the codebook as discrete codes. These codes preserve the structural relative position information of the original high-dimensional data. Finally, the decoder decodes the original data from the latent space to achieve data reconstruction.

[0022] In one possible implementation, the specified condition information is divided into two categories: the first category is initial slice information in vector form, and the second category is property information in scalar form, including platinum loading, electrochemical reaction area, and local oxygen transport resistance. Using initial slice information as condition information can help generate three-dimensional structures from two-dimensional slices, and continuously use the last slice as the initial slice to carry out sequential stitching, synthesizing structures much larger than the training data size, thus achieving scalable generation.

[0023] In one possible implementation, the condition generation modeling apparatus includes:

[0024] The condition information processing module is configured to generate the corresponding initial slice information and property information based on the three-dimensional structure input by the model; the three-dimensional structure input by the model is composed of multiple consecutive two-dimensional slices cut from the three-dimensional scan structure data;

[0025] The Transformer decoder, which communicates with the condition information processing module, the codebook unit, and the decoder, is configured to perform category modeling on the discrete variable based on the specified condition information to obtain the category probability of the discrete variable.

[0026] In one possible implementation, the condition information processing module includes a first condition information unit and a second condition information unit. The first condition information unit is configured to generate the initial slice information corresponding to the three-dimensional structure input by the model, and the second condition information unit is configured to generate the property information based on the three-dimensional structure input by the model.

[0027] In one possible implementation, the Transformer decoder is composed of an input unit, a backbone network, and an output unit connected in series. The input unit communicates with the codebook unit to receive the discrete variables. The backbone network is a network structure composed of several attention mechanism modules connected in series. Each attention mechanism module is a network structure composed of three attention mechanism units connected in series. All attention mechanism units communicate with the second conditional information unit, and all attention mechanism units in the middle position communicate with the first conditional information unit. The output unit is configured to obtain the class probability of the discrete variables by using the output of the attention mechanism unit at the end through a fusion of layer normalization mapping, linear layer mapping, and softmax activation function.

[0028] Generative modeling using a Transformer decoder can transform discrete codes into one-dimensional sequences and estimate their joint probability density through autoregression for modeling.

[0029] In one possible implementation, in the attention mechanism module, the first attention mechanism unit is configured to sequentially perform layer normalization operation, multi-head attention operation, and random deactivation operation on the property information to obtain a first operation result, and then sum the first operation result with the input of the attention mechanism module to obtain the output of the first attention mechanism unit.

[0030] The attention mechanism unit in the middle position is configured to perform layer normalization operation, multi-head attention operation and random deactivation operation sequentially on the sum of the initial slice information and the property information to obtain a second operation result, and then sum the second operation result with the output of the attention mechanism unit in the first position to obtain the output of the attention mechanism unit in the middle position.

[0031] The attention mechanism unit at the end is configured to perform layer normalization, linear layer operation and random deactivation operation on the property information in sequence to obtain a third operation result, and then sum the third operation result with the output of the attention mechanism unit in the middle position to obtain the output result of the attention mechanism unit at the end.

[0032] The conditional generation modeling device with the above structure and functions can generate initial slice information and property information respectively through the conditional information processing module. With the support of the Transformer decoder, for the high-dimensional initial slice information, a cross-attention approach is used to integrate it. The information is extracted through the included convolutional neural network model, and then the information is fused using the sequence formed by cross-attention and discrete encoding. For the scalar property information, a layer normalization approach is used to map the conditional information into transformation coefficients. During layer normalization, the transformation coefficients are used to perform a linear transformation on the normalization result, thereby realizing the fusion of conditional information into discrete variables and finally obtaining the class probability of discrete variables, thus improving the reconstruction efficiency and accuracy of the three-dimensional structure of the catalytic layer.

[0033] In one possible implementation, the statistical analysis device is configured to perform the following steps:

[0034] A1: Numerical simulation and geometric statistical analysis are performed on an equal number of random samples from the three categories obtained by the electrochemical property prediction module to obtain the structural statistical indicators corresponding to each category;

[0035] A2: Obtain the correlation between the structural statistical indicators and performance indicators obtained in step A1 through correlation analysis;

[0036] A3: Using the correlation obtained in step A2, by comparing the correlation differences between different categories, we can identify the factors that have the greatest impact on performance under different platinum loadings and obtain the common characteristics of excellent structures.

[0037] The statistical analysis device that performs the above steps conducts numerical simulation and geometric statistical analysis on the three groups obtained by the electrochemical property prediction module. By comparing the correlation differences between different groups, it can identify the factors that have the greatest impact on performance under different platinum loadings, thereby obtaining the common characteristics of excellent structures. Furthermore, it can also use the full-scale model of the fuel cell to obtain the upper limit of optimal structural performance by utilizing the common characteristics of excellent structures.

[0038] Another technical solution of the present invention is to provide a method for optimizing and controlling the structure of an ultra-low platinum catalyst layer in a hydrogen fuel cell, comprising the following steps:

[0039] S1: Using the collected 3D structure data of the catalytic layer, the parameters of the vector quantization variational autoencoder to be trained are optimized through the autoencoder loss function to obtain the trained vector quantization variational autoencoder. The parameters of the conditional generation modeling device to be trained are optimized through the cross-entropy loss function to obtain the trained conditional generation modeling device.

[0040] S2: The trained vector quantization variational autoencoder is communicated with the conditional generation modeling device to obtain a controllable synthesis module for the catalytic layer structure;

[0041] S3: The controllable synthesis module for the catalyst layer structure sequentially performs discrete dimensionality reduction and conditional generation modeling on the three-dimensional scanning structure data of the fuel cell catalyst layer to obtain several sets of three-dimensional structure data of the catalyst layer that meet the specified condition information. A set of three-dimensional structure data of the catalyst layer depicts a three-dimensional structure of the catalyst layer.

[0042] S4: The electrochemical property prediction module predicts the performance of the three-dimensional structure of the catalyst layer depicted by the three-dimensional structure data of each group of catalyst layers to obtain their respective electrochemical properties. Based on the electrochemical properties, all the depicted three-dimensional structures of the catalyst layer are classified into three categories: excellent performance, medium performance and poor performance. An equal number of random samples are retained in all categories for statistical analysis.

[0043] S5: Numerical simulation and geometric statistical analysis are performed on the three categories of random samples obtained by the electrochemical property prediction module using a statistical analysis device to obtain structural statistical indicators that reflect the structural differences of each category.

[0044] The method disclosed in this application first optimizes the parameters of the vector quantization variational autoencoder to be trained using an autoencoder loss function to obtain a vector quantization variational autoencoder. Then, it optimizes the parameters of the conditional generation modeling device to be trained using a cross-entropy loss function to obtain a trained conditional generation modeling device. Next, the vector quantization variational autoencoder and the conditional generation modeling device communicate to obtain a controllable synthesis module for the catalytic layer structure. Subsequently, the controllable synthesis module for the catalytic layer structure performs discrete dimensionality reduction and conditional generation modeling, enabling the efficient synthesis of a large number of three-dimensional catalytic layer structures that meet specified conditional information. This allows for full exploration of the high-dimensional structure design space, providing a solid data foundation for uncovering the common characteristics of high-performance structures and forming optimization and control suggestions. Simultaneously, the controllable synthesis module for the catalytic layer structure can synthesize structures much larger than the training data based on the set conditional information, achieving scalable structure synthesis. This approach, which allows for smaller training data sizes, greatly reduces the computational burden. Furthermore, with the cooperation of the electrochemical property prediction module and the statistical analysis device, the optimization and control of the ultra-low platinum catalytic layer can be decoupled into two sequential steps: synthesizing structures that meet ultra-low platinum loading requirements and rapidly evaluating the structures, as well as statistically analyzing the common characteristics of excellent structures. This concept of decoupling platinum loading constraints from performance optimization reduces the difficulty of structural optimization and avoids the dilemma of simultaneously reducing platinum loading and optimizing performance. It is conducive to the engineering of structural regulation and overcomes the technical defects of optimizing and regulating the catalyst layer structure, such as high cost and reliance on expert experience. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of a hydrogen fuel cell ultra-low platinum catalyst layer structure optimization and control device disclosed in the embodiments of this application;

[0046] Figure 2 This is a schematic diagram illustrating the operational principle of the controllable synthesis module for the catalyst layer structure disclosed in the embodiments of this application.

[0047] Figure 3 This is a schematic diagram of the condition generation modeling device disclosed in the embodiments of this application;

[0048] Figure 4 This is a schematic diagram of the Transformer decoder operation disclosed in the embodiments of this application;

[0049] Figure 5 This is a schematic diagram of the operation of the ultra-low platinum catalyst layer structure optimization and control device for hydrogen fuel cells disclosed in the embodiments of this application. Detailed Implementation

[0050] First, those skilled in the art should understand that these embodiments are merely used to explain the technical principles of the embodiments of this application and are not intended to limit the scope of protection of the embodiments of this application. Those skilled in the art can make adjustments as needed to adapt to specific application scenarios.

[0051] In the embodiments of this application, unless otherwise explicitly specified and limited, communication or communication connection between the first feature and the second feature means that there is information transmission between the first feature and the second feature. This information transmission can be unidirectional or bidirectional, and the communication connection can be achieved by means of electrical connection of wires, wireless connection, electrical connection of electromagnetic medium (such as semiconductor), communication realized by channel, etc.

[0052] In the embodiments of this application, unless otherwise expressly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "on top of," and "over" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0053] The present application will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0054] See Figures 1-5 This application discloses a device for optimizing and controlling the structure of an ultra-low platinum catalyst layer in a hydrogen fuel cell. A schematic diagram of the device is shown below. Figure 1 As shown in the diagram, the operation is illustrated below. Figure 5 As shown, the control device includes a controllable synthesis module for the catalyst layer structure, an electrochemical property prediction module, and a statistical analysis device. The electrochemical property prediction module communicates with the controllable synthesis module for the catalyst layer structure, and the statistical analysis device communicates with the electrochemical property prediction module.

[0055] See Figure 2In this controllable device, the catalyst layer structure controllable synthesis module is configured to sequentially perform discrete dimensionality reduction and conditional generation modeling on the three-dimensional scanned structural data of the fuel cell catalyst layer, obtaining several sets of three-dimensional structure data of the catalyst layer that meet specified condition information. Furthermore, one set of three-dimensional structure data depicts a three-dimensional structure of the catalyst layer with specified condition information (i.e., a random three-dimensional structure of the catalyst layer). Specifically, in this embodiment, the catalyst layer structure controllable synthesis module includes a vector quantized variational autoencoder (VQVAE) and a conditional generation modeling device, wherein the conditional generation modeling device communicates simultaneously with both the vector quantized variational autoencoder and the electrochemical property prediction module.

[0056] Please continue reading Figure 2 In the controllable synthesis module of the catalyst layer structure, the vector quantization variational autoencoder is configured to compress the three-dimensional scanned structural data to obtain continuous latent variables, and then retrieve the discrete variables corresponding to the continuous latent variables from the codebook. Several sets of three-dimensional structure data of the catalyst layer are obtained using the discrete variables and their class probabilities, following the method of upscaling the continuous latent variable space to the pixel space (this discrete variable, after data upscaling, can obtain the reconstructed three-dimensional structure). Please continue to see... Figure 2 In this embodiment, the vector quantization variational autoencoder includes an encoder, a codebook unit, and a decoder. The codebook unit communicates with both the encoder and the conditional generation modeling device, and the decoder also communicates with both the encoder and the conditional generation modeling device. The encoder is configured to compress the 3D scanned structural data to obtain continuous latent variables. The codebook unit is configured to find the closest vector to the continuous latent variable in the codebook and obtain the position index of this closest vector, which is denoted as the discrete variable corresponding to the continuous latent variable. The decoder is configured to decode the discrete variable and its class probability into several sets of catalytic layer 3D structural data by increasing the dimensionality from the continuous latent variable space to the pixel space.

[0057] See Figure 2 , Figure 3 and Figure 4 In the controllable synthesis module of the catalyst layer structure, the conditional generation modeling device is configured to perform category modeling on discrete variables based on specified conditional information to obtain the category probabilities of the discrete variables. In this embodiment, the conditional information is divided into two categories: the first category is initial slice information in vector form, and the second category is property information in scalar form. The property information includes platinum loading (i.e.,... Figure 4 In ), electrochemical reaction area (i.e. Figure 4 In ) and local oxygen transport resistance (i.e. Figure 4 In See also Figure 3 and Figure 4 In this embodiment, the condition generation modeling device includes a condition information processing module and a Transformer decoder, wherein the Transformer decoder communicates with the condition information processing module, the codebook unit and the decoder simultaneously.

[0058] In the conditional generation modeling apparatus, the conditional information processing module is configured to generate the corresponding initial slice information and property information based on the 3D structure input to the model; the 3D structure input to the model is composed of multiple consecutive 2D slices extracted from 3D scanned structural data. See also Figure 3 In this embodiment, the condition information processing module includes a first condition information unit and a second condition information unit. The first condition information unit is configured to generate the corresponding initial slice information based on the three-dimensional structure input by the model, and the second condition information unit is configured to generate the property information based on the three-dimensional structure input by the model.

[0059] In the conditional generative modeling apparatus, the Transformer decoder is configured to perform class modeling on discrete variables based on specified conditional information to obtain the class probabilities of the discrete variables. In this embodiment, the Transformer decoder consists of an input unit, a backbone network, and an output unit connected in series. The input unit communicates with the codebook unit to receive discrete variables. The backbone network is a network structure composed of several attention mechanism modules connected in series. Each attention mechanism module is a network structure composed of three attention mechanism units connected in series. All attention mechanism units communicate with the second conditional information unit, and all attention mechanism units in the middle positions communicate with the first conditional information unit. The output unit is configured to obtain the class probabilities of the discrete variables using the output of the attention mechanism unit at the end, through a fusion of layer normalization mapping, linear layer mapping, and the softmax activation function. Using the Transformer decoder for generative modeling can transform discrete codes into a one-dimensional sequence. The joint probability density is estimated by autoregression. To perform modeling, i.e.

[0060] ,

[0061] in This represents the trainable parameters.

[0062] See Figure 4In the attention mechanism module of this embodiment, the first attention mechanism unit is configured to sequentially perform layer normalization, multi-head attention, and random deactivation operations on the property information to obtain a first operation result. Then, the first operation result is summed with the input of the attention mechanism module to obtain the output of the first attention mechanism unit. The input of the attention mechanism module at the head of the backbone network is the output of the input unit. In this embodiment, the output of the input unit is the result obtained by performing linear layer operations on discrete variables and incorporating a preset embedding vector and positional encoding. The attention mechanism unit in the middle position is configured to sequentially perform layer normalization, multi-head attention, and random deactivation operations on the summation result of the initial slice information and property information to obtain a second operation result. Then, the second operation result is summed with the output of the first attention mechanism unit to obtain the output of the middle attention mechanism unit. The attention mechanism unit at the end is configured to perform layer normalization, linear layer operation and random deactivation operation on the property information in sequence to obtain the third operation result. Then, the third operation result is summed with the output of the attention mechanism unit in the middle position to obtain the output result of the attention mechanism unit at the end, that is, the output of the attention mechanism module.

[0063] Specifically, in the attention mechanism module of this embodiment, the first attention mechanism unit is configured to execute the following calculation:

[0064] ;

[0065] The attention mechanism unit in the middle position is configured to perform the following calculation:

[0066] ;

[0067] The attention mechanism units at the beginning and end are configured to perform the following calculation:

[0068] ;

[0069] In the formula,

[0070] The output of the attention mechanism unit;

[0071] A positive integer, representing the execution order of the attention mechanism modules, i.e., according to... Figure 4 The order of the arrow directions;

[0072] This represents the summation operation on an image;

[0073] Represents random deactivation operation;

[0074] Represents multi-head attention computation;

[0075] Representative layer normalization operation;

[0076] Represents linear layer operations;

[0077] Represents the initial slice information, i.e. Figure 4 ① in the middle;

[0078] Representative nature information, that is Figure 4 ② in the middle.

[0079] See Figure 1 and Figure 5 In this control device, the electrochemical property prediction module is set to predict the performance of the three-dimensional structure of the catalyst layer (i.e. the synthesized three-dimensional structure of the catalyst layer) depicted by each group of three-dimensional structure data of the catalyst layer, so as to obtain their respective electrochemical properties. Based on the electrochemical properties, all the depicted three-dimensional structures of the catalyst layer are classified into three categories: excellent performance, medium performance and poor performance. In addition, an equal number of random samples are retained in all categories for statistical analysis.

[0080] See Figure 1 and Figure 4 In this control device, the statistical analysis device is set to perform numerical simulation and geometric statistical analysis on an equal number of random samples from the three categories obtained by the electrochemical property prediction module to obtain structural statistical indicators that reflect the structural differences of each category; the structural statistical indicators include carbon content, average particle size, average ionomer thickness, local oxygen mass transfer resistance, and electrochemical reaction area.

[0081] In this embodiment, the statistical analysis device is configured to perform the following steps: A1: Numerical simulation and geometric statistical analysis are performed on an equal number of random samples from the three categories obtained by the electrochemical property prediction module to obtain the structural statistical indicators corresponding to each category; A2: The correlation between the structural statistical indicators and performance indicators obtained in step A1 is obtained through correlation analysis; A3: Using the correlation obtained in step A2, the factors that have the greatest impact on performance under different platinum loadings are identified by comparing the correlation differences between different categories, and the common characteristics of excellent structures are obtained.

[0082] See Figure 2 and Figure 5 The following will further disclose the method of using the hydrogen fuel cell ultra-low platinum catalyst layer structure optimization and control device described in this embodiment. The method includes the following steps:

[0083] S1: Using the collected 3D structure data of the catalytic layer, the parameters of the vector quantization variational autoencoder to be trained are optimized through the autoencoder loss function to obtain the trained vector quantization variational autoencoder.

[0084] The autoencoder loss function in this embodiment consists of three parts, and its calculation formula is as follows:

[0085] ,

[0086] in, It is the output of the vector quantization variational autoencoder to be trained. It is the output of the encoder. It is a codebook vector. It is the closest discrete vector in the codebook. This indicates that the gradient has stopped. This involves a trade-off with hyperparameters. In this loss function, the reconstruction loss (the first term) uses negative log-likelihood (such as cross-entropy or MSE) to measure the decoder's reconstructed data. The quality is improved to ensure that the generated result closely approximates the original input. Codebook loss (the second term) is used. L 2-distance, pushing codebook vector Output to encoder Near (due to) The existence of this (only the codebook can be trained). Commitment loss (the third term) is also based on... L 2 distance, but reverse optimize the encoder to make its output More stable matching of discrete codebook vectors ( (Freezing the codebook) avoids frequent jumps in the latent representation between codebook vectors. The three work together to ensure that the controllable synthesis module of the obtained catalytic layer structure maintains high reconstruction accuracy while learning in the discrete latent space.

[0087] S2: The vector quantization variational autoencoder communicates with the conditional generation modeling device to obtain a controllable synthesis module for the catalytic layer structure; wherein, the conditional generation modeling device can be optimized for parameters using the cross-entropy loss function, the calculation formula of which is as follows:

[0088] ,

[0089] in, This represents the mean.

[0090] S3: The three-dimensional scanning structural data of the fuel cell catalyst layer are sequentially discretely reduced and conditionally generated and modeled through the controllable synthesis module of the catalyst layer structure to obtain several sets of three-dimensional structure data of the catalyst layer that meet the specified condition information. One set of three-dimensional structure data of the catalyst layer depicts the three-dimensional structure of the catalyst layer with the specified condition information.

[0091] S4: The electrochemical property prediction module is used to predict the performance of the three-dimensional structure of the catalyst layer depicted by the three-dimensional structure data of each group of catalyst layers to obtain their respective electrochemical properties. Based on the electrochemical properties, all the depicted three-dimensional structures of the catalyst layers are classified into three categories: excellent performance, medium performance and poor performance. An equal number of random samples are retained in all categories for statistical analysis.

[0092] S5: Numerical simulation and geometric statistical analysis are performed on the three categories of random samples obtained by the electrochemical property prediction module using a statistical analysis device to obtain structural statistical indicators that reflect the structural differences of each category.

[0093] The hydrogen fuel cell ultra-low platinum catalyst layer structure optimization and control device disclosed in this embodiment, by setting a controllable synthesis module for catalyst layer structure to perform discrete dimensionality reduction and conditional generation modeling, can efficiently synthesize a large number of three-dimensional catalyst layer structures that meet specified condition information. This allows for full exploration of the high-dimensional structure design space, providing a solid data foundation for discovering the common characteristics of high-performance structures and forming optimization and control suggestions. Simultaneously, the controllable synthesis module can also synthesize structures much larger than the training data through sequential splicing based on the set condition information, achieving scalable structure synthesis. This approach, which allows for smaller training data sizes, greatly reduces the computational burden. Furthermore, with the coordinated setup of the electrochemical property prediction module and the statistical analysis device, the optimization and control of the ultra-low platinum catalyst layer can be decoupled into two sequential steps: synthesizing structures that meet ultra-low platinum loading and rapidly evaluating the structures, as well as statistically analyzing the common characteristics of excellent structures. This concept of decoupling platinum loading constraints from performance optimization reduces the difficulty of structure optimization and control, avoids the dilemma of simultaneously reducing platinum loading and optimizing performance, facilitates the engineering of structure control, and overcomes the technical shortcomings of existing catalyst layer structure optimization and control methods, such as high cost and reliance on expert experience.

[0094] In the description of the embodiments of this application, it should be noted that the terms "inner" and "outer" and other terms indicating direction or positional relationship are based on the direction or positional relationship shown in the drawings. This is only for the convenience of description and does not indicate or imply that the device or component must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this application.

[0095] In the description of this application, the references to terms such as "one embodiment," "some embodiments," "in this embodiment," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0096] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A hydrogen fuel cell ultra-low platinum catalytic layer structure optimization and control device, characterized in that, The application relates to a method for synthesizing a fuel cell catalyst layer structure, comprising the following steps: a catalyst layer structure controllable synthesis module is arranged to sequentially perform discrete dimension reduction and conditional generation modeling on three-dimensional scanning structure data of a fuel cell catalyst layer, so as to obtain a plurality of groups of catalyst layer three-dimensional structure data meeting specified condition information, and one group of catalyst layer three-dimensional structure data describes a catalyst layer three-dimensional structure with specified condition information; an electrochemical property prediction module is in communication with the catalyst layer structure controllable synthesis module and is arranged to perform performance prediction on the catalyst layer three-dimensional structure described by each group of catalyst layer three-dimensional structure data respectively, so as to obtain respective electrochemical properties, and to perform category division on all the described catalyst layer three-dimensional structures according to the electrochemical properties, so as to divide into three categories of excellent performance, medium performance and poor performance, and all categories keep equal numbers of random samples for statistical analysis; a statistical analysis device is in communication with the electrochemical property prediction module and is arranged to perform numerical simulation and geometric statistical analysis on the equal numbers of random samples kept in the three categories obtained by the electrochemical property prediction module, so as to obtain structure statistical indexes reflecting structure differences of the categories; the structure statistical indexes include carbon proportion, average particle size, average ionomer thickness, oxygen local mass transfer resistance and electrochemical reaction area; the catalyst layer structure controllable synthesis module comprises: a vector quantization variational autoencoder is arranged to perform data compression on the three-dimensional scanning structure data to obtain continuous hidden variables, and to obtain discrete variables corresponding to the continuous hidden variables in a codebook, and to obtain a plurality of groups of the catalyst layer three-dimensional structure data by using the discrete variables and category probabilities in a manner of dimension increasing from a continuous hidden variable space to a pixel space; a conditional generation modeling device is in communication with the vector quantization variational autoencoder and is arranged to perform category modeling on the discrete variables according to the specified condition information, so as to obtain category probabilities of the discrete variables; the specified condition information is divided into two categories, the first category is initial slice information in a vector form, and the second category is property information in a scalar form; the property information includes platinum loading, electrochemical reaction area and oxygen local transmission resistance.

2. The hydrogen fuel cell ultra-low platinum catalytic layer structure optimization and regulation device according to claim 1, characterized in that, the vector quantization variational autoencoder comprises: an encoder is arranged to perform data compression on the three-dimensional scanning structure data to obtain continuous hidden variables; a codebook unit is in communication with the encoder and the conditional generation modeling device simultaneously and is arranged to find a closest vector to the continuous hidden variables in the codebook, and to obtain a position index of the closest vector, which is recorded as a discrete variable corresponding to the continuous hidden variable; a decoder is in communication with the encoder and the conditional generation modeling device simultaneously and is arranged to decode the discrete variables and category probabilities into a plurality of groups of the catalyst layer three-dimensional structure data in a manner of dimension increasing from a continuous hidden variable space to a pixel space.

3. The hydrogen fuel cell ultra-low platinum catalytic layer structure optimization and regulation device of claim 2, wherein, the conditional generation modeling device comprises: a condition information processing module is arranged to generate the initial slice information and the property information corresponding to a three-dimensional structure input by a model according to the three-dimensional structure; the three-dimensional structure input by the model is composed of a plurality of continuous two-dimensional slices cut from the three-dimensional scanning structure data; The Transformer decoder, in communication with the conditional information processing module, the codebook unit, and the decoder in the vector quantization variational autoencoder, is configured to perform category modeling on the discrete variable based on the specified conditional information to obtain a category probability of the discrete variable.

4. The hydrogen fuel cell ultra-low platinum catalytic layer structure optimization and regulation device of claim 3, wherein, The conditional information processing module includes a first conditional information unit and a second conditional information unit, the first conditional information unit is configured to generate the initial slice information corresponding to the three-dimensional structure of the model input, and the second conditional information unit is configured to generate the property information based on the three-dimensional structure of the model input.

5. The hydrogen fuel cell ultra-low platinum catalytic layer structure optimization and regulation device of claim 4, wherein, The Transformer decoder is composed of an input unit, a backbone network, and an output unit in series; the input unit is in communication with the codebook unit to receive the discrete variable, the backbone network is a network structure composed of a plurality of attention mechanism modules in series, the attention mechanism module is a network structure composed of three attention mechanism units in series, all the attention mechanism units are in communication with the second conditional information unit, all the attention mechanism units in the middle position are in communication with the first conditional information unit, and the output unit is configured to obtain the category probability of the discrete variable by fusing a layer normalization mapping, a linear layer mapping, and a softmax activation function using the output result of the attention mechanism unit at the end.

6. The hydrogen fuel cell ultra-low platinum catalytic layer structure optimization and regulation device of claim 5, wherein, In the attention mechanism module, the attention mechanism unit at the first position is configured to sequentially perform layer normalization operation, multi-head attention operation, and random inactivation operation on the property information to obtain a first operation result, and then sum the first operation result and the input of the attention mechanism module to obtain the output of the attention mechanism unit at the first position; The attention mechanism unit at the middle position is configured to sequentially perform layer normalization operation, multi-head attention operation, and random inactivation operation on the sum result of the initial slice information and the property information to obtain a second operation result, and then sum the second operation result and the output of the attention mechanism unit at the first position to obtain the output of the attention mechanism unit at the middle position; The attention mechanism unit at the end is configured to sequentially perform layer normalization operation, linear layer operation, and random inactivation operation on the property information to obtain a third operation result, and then sum the third operation result and the output of the attention mechanism unit at the middle position to obtain the output result of the attention mechanism unit at the end.

7. The hydrogen fuel cell ultra-low platinum catalyst layer structure optimization and regulation device according to any one of claims 1-6, characterized in that, The statistical analysis device is configured to perform the following steps: A1: performing numerical simulation and geometric statistical analysis on random samples of the same number reserved in the three categories obtained by the electrochemical property prediction module to obtain the structure statistical indicators corresponding to each category; A2: obtaining the correlation between the structure statistical indicators and the performance indicators obtained in step A1 by correlation analysis. A3: Using the correlation obtained in step A2, by comparing the correlation differences between different categories, the factors that most affect the performance under different platinum loadings are mined, and the excellent structural common features are obtained.

8. A method for optimizing and regulating the structure of a hydrogen fuel cell ultra-low platinum catalytic layer, characterized in that, The hydrogen fuel cell ultra-low platinum catalyst layer structure optimization and regulation device suitable for any one of claims 1-7, comprising the following steps: S1: using the collected three-dimensional structure data of the catalyst layer, the vector quantization variational autoencoder to be trained is parameterized by the self-encoder loss function to obtain the trained vector quantization variational autoencoder, and the conditional generation modeling device to be trained is parameterized by the cross-entropy loss function to obtain the trained conditional generation modeling device; S2: the trained vector quantization variational autoencoder and the conditional generation modeling device are communicated to obtain a catalyst layer structure controllable synthesis module; S3: the three-dimensional scanning structure data of the fuel cell catalyst layer is sequentially subjected to discrete dimension reduction and conditional generation modeling by the catalyst layer structure controllable synthesis module to obtain a plurality of groups of catalyst layer three-dimensional structure data meeting the specified condition information; S4: the performance of the catalyst layer three-dimensional structure described by each group of catalyst layer three-dimensional structure data is predicted by the electrochemical property prediction module to obtain the respective electrochemical properties, and all the catalyst layer three-dimensional structures described are classified according to the electrochemical properties into three categories of excellent performance, medium performance and poor performance, and the same number of random samples are retained in all categories for statistical analysis; S5: the same number of random samples retained in the three categories obtained by the electrochemical property prediction module are subjected to numerical simulation and geometric statistical analysis by the statistical analysis device to obtain structural statistical indicators reflecting the structural differences of each category.

Images