Ph.D. Thesis Proposal by Petro Junior Milan (Advisor: Prof. Vigor Yang)

Friday, Nov. 20 @ 3:30-5:30 PM (EST)

Deep-Learning Enhanced Multiphysics Flow Computations for Engineering Applications

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. However, even with the most advanced high-performance computing facility, high-fidelity numerical simulations are extremely costly in time and resources. Thus, for example, to survey the design parameter space for a single-element injector for a propulsion application (such as the RD-170 rocket engine) using a Large Eddy Simulation technique might take several months of computing time on a major cluster computer. The reason is that to characterize the flowfields, we need to resolve a multitude of coupled thermodynamic, fluid dynamic, 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 mixtures at supercritical conditions. If such models are used for statistical analysis or optimization, the total computation time grows too large to be economically feasible.

Recent developments in deep learning techniques offer the possibility of significant advances in dealing with these challenges while accelerating 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 modeling approaches and software using deep neural networks for large-scale simulation problems. The approaches will include (1) regression models for function approximation and solver acceleration, (2) deep autoencoders for nonlinear dimensionality reduction, (3) multi-level surrogate models for simulator approximation and rapid exploration of design spaces, and (4) probabilistic models for uncertainty quantification.

In this thesis proposal, we outline the databases generated from high-fidelity simulations for deep learning and reduced-basis model development, along with results obtained to date. Emphasis is placed on two model problems with near-realistic geometries and operating conditions relevant to automotive and aerospace applications, namely the multiphase cavitating flow in a diesel engine injector, and the supercritical turbulent flow in a swirl rocket injector.

Committee members:

• 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. Timothy C. Lieuwen, School of Aerospace Engineering, Georgia Institute of Technology
• Dr. Gina M. Magnotti, Energy Systems Division, Argonne National Laboratory