Evaluation of Prediction Performances of Deep Learning Models for the Aerodynamic Characteristics of Flettner Rotors

Abstract This study investigates the prediction of the aerodynamic characteristics of Flettner rotors through three deep learning models. Various numbers of Flettner rotors, arrangements, and spin ratios are employed to consider these effects in the dataset. For the training of deep learning models, a dataset of aerodynamic force coefficients and flow fields is generated using Computational Fluid Dynamics (CFD). Three deep learning architectures (U-net, Encoder-Decoder, and Decoder models) are employed and trained to predict the aerodynamic characteristics of Flettner rotors. Three deep learning models are established through a training stage with a hyperparameter study and by altering the number of layers. The aerodynamic force coefficients and flow fields are predicted by established deep learning models and show small absolute errors compared to those from the CFD analysis. Moreover, predicted flow fields reflect the flow characteristics according to the difference of spin ratio and arrangement of Flettner rotors. In conclusion, the established deep learning models demonstrate rapid and robust predictions of aerodynamic force coefficients and flow fields for Flettner rotors under varying arrangements and spin ratios. Furthermore, a significant reduction in computational time is measured when comparing the analysis time of CFD simulations to the training and testing time of the deep learning models.