Predicting dynamics in parameter-time domain with deep learning and data augmentation

The poster for Model Reduction and Surrogate Modeling (MORe2024), presaved for a preprint paper. We propose a non-intrusive parametric MOR method named Data-Augmented Predictive Deep Neural Network (DAPredDNN) that employing CAE-FFNN and KDMD for dynamics prediction in both the time and the parameter space. The proposed method enriches original data by augmenting the KDMD-decoder-extrapolated data during the offline stage. Training the CAE-FFNN with this augmented data allows the network to correlates time-parameter pairs with their extrapolated solution directly, avoiding the time-marching behaviour during the online phase.