Recently, polymer electrolyte fuel cell (PEFC), which are expected to be applied to heavy-duty vehicles with high environmental load, are required to further improve their output power. the output characteristics of PEFC depend on the structure of the catalyst layer, including the aggregation form of the carbon support and the ionomer coating state. The structure of the catalyst layer is also important. The catalyst layer structure is also determined by the composition of the catalyst ink that forms the structure and the agglomeration form of the particles. However, the correlation between each ink condition and output characteristics is difficult to understand directly because it involves complex factors of structure formation and transport reaction distribution. In this study, we fabricated various catalyst layer structures under various ink conditions with different compositions and fabrication processes, analyzed their output characteristics, and evaluated the correlation using machine learning as a learning model.In the cathode catalyst layer, oxygen, protons, and electrons move through voids, ionomer (electrolyte polymer), and carbon black (CB), respectively, and oxygen reduction reaction occurs on the Pt surface. Since the anodic reaction is sufficiently rapid compared to the cathodic reaction, the anodic reaction was neglected, and the inside of the battery was assumed to be isothermal and steady state. A 300 nm substrate aggregate was formed and filled into the computational domain, and the catalyst layer structure was fabricated by arranging the Pt catalyst and coating it with ionomer. The structure was simulated by varying the spatial probability density of the support arrangement. The structure was subdivided into blocks of primary particle size and transport calculations 1) were performed between adjacent blocks.The I/C ratio (weight ratio of ionomer to CB), platinum loading, and spatial probability density related to the arrangement of the CB support were varied, and the battery characteristics were determined by transport calculations for each structure using 100 structures that reproduced the coarse and dense structure of the catalyst layer. From the results of the transport calculations, the voltage at a current density of 0 A/cm2 and the current density for each 0.05 V between 0.6 and 0.85 V were associated with the images of the catalyst layer, and a dataset consisting of 399 cross-sectional images of the catalyst layer, each voltage and current density was created.CNN 2) was used for the machine learning model.Four hundred images were prepared for the study, of which 240 were used for training, 80 for validation, and 80 for testing. Adam was used for the optimization method, RMSE for the loss function, and Keras for implementation.Transport calculations were performed under conditions similar to current fuel cell operating conditions: temperature 80 ℃, RH 80 %, and supply oxygen partial pressure 20 kPa. In order to reproduce differences in the fabrication process, the spatial probability density associated with the arrangement of CB carriers was varied to fabricate structures with different aggregation morphologies.This calculation confirmed the voltage drop associated with the formation of aggregates, especially in the high current density region where mass transport is the rate-limiting factor. This is because oxygen transport was inhibited as the aggregates became larger in the aggregation area, and the high current density region inside the catalyst layer became localized.The voltage at a current density of 0 A/cm2 and the coefficient of determination (R2) between the production value of the transport calculation and the machine learning prediction at the current density at each voltage were all better than 0.8, confirming that the machine learning model we constructed can predict battery characteristics. In addition, the machine learning model was able to predict differences in battery characteristics where the I/C ratio and platinum loading were equal and only the coarseness and density of the structure differed.As described above, it is suggested that differences in the fabrication process can cause differences in battery characteristics, and a machine learning model was constructed to predict battery characteristics based on catalyst layer structure images where process differences are assumed.AcknowledgmentThis study was supported by NEDO program, Japan.
Read full abstract
Jan 10, 2025
Subhasish Chaki + 2
Just Published
Open Access