Deep Learning for Spatio-Temporal Modeling of Flows

Abstract  Data-driven approaches are gaining popularity in studying fluid mechanics, thanks to the availability of abundant numerical and experimental data. The advancements in deep learning have enabled its effective application in fluid mechanics, where it is emerging as a powerful tool for spatio-temporal modeling, particularly in turbulent flows. The seminar will focus on the development of auto-regressive transformers integrated with deep convolutional neural networks for efficient spatio-temporal learning. These innovative methods enable accurate prediction of future steps in fluid flow evolution based on appropriate initial conditions. The presented results, demonstrated using canonical flow configurations, showcase agreement with reference flow evolution and expected variances for long-term predictions. This work highlights the potential of deep learning models in accurately capturing and predicting spatio-temporal behavior in fluid flows. The seminar will conclude with an open discussion on various avenues for deploying state-of-the-art deep learning models, including deep learning-assisted acceleration of fluid dynamics solvers and improvement of experimental measurements using deep learning techniques. The aim of the seminar is to foster collaborations and explore the extensive potential of deep learning-assisted modeling in the study of turbulent flows.

Bio  Dr. Aakash Patil is a postdoctoral fellow at the Center for Turbulence Research at Stanford University, working on the intersection of turbulent reactive flows and deep learning. He holds a Bachelor's degree in Mechanical Engineering from the University of Pune in India, and a joint Master's degree in Turbulence from Ecole Centrale de Lille and ISAE-ENSMA in France. He recently defended his Ph.D. in Computational Mathematics from Mines ParisTech - Paris Sciences & Lettres University. During doctoral research, he worked with Elie Hachem at CNRS-CEMEF in Sophia Antipolis, focusing on the development of deep learning-assisted modeling methods for turbulent flows. His research interests include understanding and predictive capabilities of deep learning for turbulence in fluids, as well as the HPC side of deploying deep learning models in exascale numerical simulations.