“Predictive Simulation of Turbulent Atomization and Spray Dynamics: High-Fidelity Simulations and Machine Learning-Driven Models”

CBE Seminar by Yue (Stanley) Ling, Ph.D.

Abstract

Accurate simulation of sprays is crucial for various industrial and natural processes, such as liquid fuel injection and sea sprays. The ability to predict spray characteristics, such as droplet size distribution, is essential for improving spray system performance. Despite advances in interface-capturing methods, direct numerical simulations (DNS) that fully resolve atomization and spray evolution remain infeasible, even with foreseeable computational power. By conducting DNS of canonical atomization configurations and droplet swarm evolution, we aim to better understand the fundamental physics and provide high-quality data, which are difficult to obtain experimentally, for data-driven modeling. This study presents DNS results for spray formation in a two-phase mixing layer between parallel gas and liquid streams, accompanied by a linear stability analysis of the shear longitudinal instability. Particular attention is given to the effect of inlet gas turbulence on the modulation of shear instability features, such as the transition from convective to absolute instability and the variation in the dominant frequency of breaking interfacial waves. We will also present high-fidelity simulation results for the secondary breakup of spray droplets, the data will serve as the basis for developing novel Lagrangian models for droplet deformation and drag. Interface-resolved DNS of single and cloud droplets are key to the rigorous development of next-generation Euler-Lagrange (EL) closure models, enabling accurate tracking of billions of droplets. A fundamental challenge lies in droplet shape characterization, for which spherical harmonics are employed. The temporal data of the spherical harmonic modal coefficients, along with the drag coefficient, are extracted from the simulations and used to train a Nonlinear Autoregressive Network with Exogenous Inputs (NARX) recurrent neural network. The machine learning (ML) model predictions show excellent agreement with the high-fidelity simulation results. Additionally, SHAP (Shapley Additive Explanations) analysis is performed to identify important features for the predictions and to enhance understanding of the physics involved in the ML models.

Career Bio

Dr. Yue (Stanley) Ling is an Associate Professor in the Department of Mechanical Engineering at the University of South Carolina, Columbia, South Carolina, USA. He obtained his B.S. from Beihang University, Beijing, China, and received his Ph.D. from the University of Florida, Gainesville, Florida, USA. Before joining USC, he was an assistant professor at Baylor University and a postdoctoral researcher at Sorbonne University in Paris, France. His research focuses on high-fidelity simulation and modeling of interfacial multiphase flows with heat and mass transfer, including sprays and atomization, droplet vaporization, turbulence-interface interaction, and shock interaction with particles and droplets. He received the NSF CAREER Award in 2020 and his current research is supported by NSF, NASA, ARO and others.