Ph.D. Defense: Petro Junior Milan

Deep-Learning Enhanced Multiphysics Flow Computations for Propulsion Applications

(Advisor: Prof. Vigor Yang)

Wednesday, November 17

4 - 6 PM (EST)

MK317, Bluejeans: https://bluejeans.com/774079904/3931

Abstract

Numerical simulation is a critical part of research into and development of engineering systems. Engineers often use simulation to explore design settings both analytically and numerically before prototypes are built and tested. Even with the most advanced high performance computing facility, however, high-fidelity numerical simulations are extremely costly in time and resources. For example, a survey of the design parameter space for a single-element injector for a propulsion application (such as the RD-170 rocket engine) using the large eddy simulation technique may require several tens of millions of CPU hours on a major computer cluster. This is because the flowfields can only be fully characterized by resolving a multitude of strongly coupled fluid dynamic, thermodynamic, transport, multiphase, and combustion processes. The cost is further increased by grid resolution requirements and by the effects of turbulence and high-pressure phenomena, which require treatment of real-fluid physics at supercritical conditions. If such models are used for statistical analysis or design optimization, the total computation time and resource requirements may render the work unfeasible.

Recent developments in deep learning techniques offer the possibility of significant advances in dealing with these challenges and significant shortening of the time-to-solution. The general scope of this thesis research is to set the foundations for new paradigms in modeling, simulation and design by applying deep learning techniques to recent developments in computational science. More specifically, the research aims at developing an integrated suite of data-driven surrogate modeling approaches and software for large-scale simulation problems. The techniques to be put into practice include: (1) deep neural networks for function approximation and solver acceleration, (2) deep autoencoders for nonlinear dimensionality reduction, and (3) spatiotemporal emulators based on multi-level neural networks for simulator approximation and rapid exploration of design spaces.

A hierarchy of benchmark cases has been studied to generate databases to enable and support the development and verification of the proposed approaches. Emphasis is placed on canonical examples, as well as on engineering problems for aerospace and automotive applications, including supercritical turbulent flows in a rocket-engine swirl injector, and  multiphase cavitating flows in a diesel engine injector.

Committee:

• Prof. Vigor Yang (Advisor), School of Aerospace Engineering, Georgia Institute of Technology
• Prof. Joseph C. Oefelein, School of Aerospace Engineering, Georgia Institute of Technology
• Prof. Edmond Chow, School of Computational Science and Engineering, Georgia Institute of Technology
• Prof. Jean-Pierre Hickey, Department of Mechanical and Mechatronics Engineering, University of Waterloo
• Dr. Gina M. Magnotti, Energy Systems Division, Argonne National Laboratory