As the core component of direct methanol fuel cells, the catalyst layer plays the key role as a species, proton and electron transport channel. However, due to the complexity of the system, optimizing its performance involves a large number of experiments and high costs. In this study, finite element simulation combined with machine learning model was constructed to accelerate power density prediction and evaluate the influence of catalyst layer parameters on the maximum power density of direct methanol fuel cells. We built a fuel cell simulation model corresponding to different parameters, obtaining a database of more than 200 sets of 19 eigenvalues, and then used different machine learning models for training and prediction. Finally, three tree-integration methods were selected to rank the importance of 19 characteristic parameters. In addition, we performed a high-throughput screening of 200?000 different parameter combinations based on sequential model-based algorithm configuration. We selected the top 10 parameter combinations with high expected improvement scores and employed them into a numerical simulation model. The results show that a majority of the polarization curves obtained from the top combinations exceed the maximum power density of the original database. This method greatly saves the time of collecting fuel cell data for experiments and speeds up the parameter optimization process.