Numerical simulations are powerful tools for understanding phenomena in Polymer Electrolyte Membrane Fuel Cells (PEMFCs) and developing PEMFCs with high performance and durability. We have been developing a three-dimensional (3D) PEMFC simulator P-Stack [1]. P-Stack can perform 3D cell/stack simulations for long-time operation about 10 ~ 100 hours at most. However, fuel cell degradation occurs on longer time scales. To simulate such very long time scales within acceptable computational time, the computational speed need to be improved innovatively. As one of the approaches, we have integrated machine learning models into the simulation.In the power generation calculation of P-Stack, an in-plane current density distribution is determined by calculating a one-dimensional (1D) electrochemical reaction model at each point. In the present study, the 1D electrochemical reaction model is replaced with a machine learning model. The learning data is the result of “pseudo-1D” power generation calculation of 500 cases generated by P-Stack. Here, “pseudo-1D” means the power generation calculation for sampling is performed for small cell geometry, guaranteeing that in-plane distribution is almost homogeneous. It also suitable for sampling calculation because computational cost is very low. In the data generation by P-Stack, the input parameters of P-Stack were determined by Latin hypercube sampling. The input parameters for the machine learning model are the potential, the concentration of each gas species in the anode and cathode catalyst layers, and the temperature. The prediction parameters are current density, cell resistance, overvoltages, molar flux of each gas species and heat flux.As shown in Figure 1, the machine learning model well predicts the target parameters, and the coefficient of determination was about 0.9 for various prediction targets of the verification data. As a result, machine learning could be used to predict power generation. We have implemented this machine learning model into P-Stack. We will also discuss the stability and accuracy of computation coupled with machine learning model and P-Stack, and how much the calculation can be accelerated by our approach.Reference[1] T. Tsukamoto, T. Aoki, H. Kanesaka, T. Taniguchi, T. Takayama, H. Motegi, R. Takayama, S. Tanaka, K. Komiyama, M. Yoneda, J. Power Sources 488 (2021) 229412 Figure 1