Applications of Data-Driven Learning Models in Fluid Mechanics: Solid-Fluid Multiphase Systems and Bat Flight

Abstract

Fluid-solid interactions are central to numerous industrial, environmental, and biological systems, encompassing particle-laden flows in chemical processing, urban airflow, and biological locomotion. Understanding and studying these interactions is essential for optimizing engineering designs and advancing scientific knowledge. This thesis explores three distinct yet interconnected topics within this domain: (1) deep learning-based drag force modeling for particulate suspensions, (2) reduced-order modeling (ROM) for flow predictions in randomly arranged solid arrays, and (3) data-driven analysis of bat flight kinematics. First, we address the challenge of accurately predicting individual particle drag forces in dense suspensions. Meso-scale models such as CFD-DEM often rely on empirical drag correlations that neglect microscale particle-particle interactions, leading to inaccuracies. To bridge this gap, we develop deep learning models—including convolutional neural networks (CNNs) and graph neural networks (GNNs)—that incorporate both local particle neighborhood information and global suspension parameters. We demonstrate the effectiveness of these models in predicting particle-scale drag forces more accurately than existing empirical methods, paving the way for improved meso- and macro-scale simulations of particulate flows. Additionally, we employ a physics-based learning paradigm and see that incorporating physical constraints to traditionally "black box" models improves performance, especially in cases of data paucity. Next, we investigate the application of deep learning and data-driven approaches for reduced-order modeling of complex flow fields. Traditional CFD methods, while highly accurate, are computationally expensive and unsuitable for real-time control or parameter sweeping for optimization tasks. Here, we compare convolutional autoencoder (CAE) models combined with recurrent architectures such as long short-term memory (LSTM) networks and temporal convolutional networks (TCNNs) against dynamic mode decomposition (DMD) for predicting the evolution of flow fields over time. Our analysis focuses on two-dimensional flow across randomly arranged infinite cylinders, highlighting the strengths and limitations of each method in predicting the long-term flow dynamics. Finally, in the domain of solid and fluid phase interactions, we shift our focus to biological fluid-structure interactions, specifically relating to bat flight. Unlike birds, bats possess highly flexible membrane wings, enabling enhanced maneuverability and complex aerodynamic adaptations. We analyze high-resolution bat flight data using proper orthogonal decomposition (POD) to identify key motion patterns and independent marker points on the bat's wing surface. Additionally, we develop data-driven models to map bat wing kinematics in the body-fixed or local frame to its global trajectory and vice versa, offering insights that could inform the design of bio-inspired flapping-wing unmanned aerial vehicles (UAVs). Together, these studies contribute to advancing computational methods for fluid-solid interactions across multiple domains, from industrial flow systems to biological locomotion. By integrating deep learning with physics-based modeling, this work provides novel frameworks for understanding and predicting complex flow phenomena, with broad applications in engineering and bio-inspired robotics.