UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Deep learning-based reduced order modeling for unsteady flow dynamics and fluid-structure interaction Gupta, Rachit


This work presents data-driven predictions of nonlinear dynamical systems involving unsteady flow and fluid-structure interaction. Of particular interest is to develop a new simulation framework integrating high-fidelity models with deep learning towards Digital Twin. The final goal is to learn and predict the coupled dynamics via the digital twin of ship vessels and propellers. End-to-end deep learning-based reduced order models (DL-ROMs) are presented for digital twin development. The first part of this study develops an overall framework for DL-ROMs. The emphasis is to investigate the predictive performance of the hybrid DL-ROMs, which vary in obtaining the low-dimensional features, i.e., proper orthogonal decomposition (POD) and convolutional autoencoders. The low-dimensional features are evolved in time using recurrent neural networks (RNNs). This leads to the formulation of two DL-ROM frameworks: the POD-RNN and the convolutional recurrent autoencoder network (CRAN). To assess data-driven predictions, POD-RNN and CRAN are applied to predict unsteady flows and instantaneous forces for flow past static bluff bodies. We perform flow prediction analysis for a configuration of side-by-side cylinders with wake interference. For systems with moving interfaces and three-dimensional (3D) geometries, we develop modular DL-ROM techniques. The second part of this study includes model reduction strategies to predict vortex-induced vibration and 3D unsteady flows. The knowledge gained in the previous parts is utilized to develop partitioned and scalable DL-ROMs for unsteady flows with moving interfaces and parametric effects. We first develop a partitioned DL-ROM framework for fluid-structure interaction. The novel multi-level DL-ROM combines the effect of POD-RNN and CRAN by modular learning of two physical fields independently. While POD-RNN provides extraction of the fluid-structure interface, the CRAN enables the prediction of flow fields. For time series prediction of 3D flows, we present a 3D CRAN-based framework for predicting the fluid forces and vortex shedding patterns. We provide an assessment of improving learning capabilities using transfer learning for complex 3D flows with variable Reynolds numbers. The simplicity and computational efficiency of the proposed DL-ROMs allow investigation for various geometries and physical parameters. This research opens ways for digital twin development for near real-time prediction of unsteady flows and fluid-structure interaction.

Item Media

Item Citations and Data


Attribution-NonCommercial-NoDerivatives 4.0 International