A DNN-based data-driven modeling employing coarse sample data for real-time flexible multibody dynamics simulations

To achieve real-time simulations for flexible multibody dynamics (FMBD) systems, we suggest data-driven modeling based on deep neural networks (DNNs). While a DNN can represent system dynamics accurately, two main factors of FMBD systems require demanding computational costs for training a DNN. One is a fine discretization of flexible bodies, which produces a large number of training data. The other is the nonlinearity of FMBD, which requires train DNN models to have numerous weight and bias parameters. To overcome these difficulties, we propose a data-driven modeling algorithm for training a DNN efficiently that consists of two steps. First, sets of randomly chosen coarse data sequentially train a DNN model. This helps speed up the training process, even for highly parametrized DNNs. At some point, the model no longer improves, and introducing an error correction step increases the performance of the model. The proposed algorithm is easy to employ and utilizes an efficient size of training data while achieving high performance of the DNN as demonstrated by numerical examples.