Ph.D. Proposal

Sean P. Engelstad

(Advisor: Graeme J. Kennedy)

Computational Methods and Machine Learning
for High-Fidelity Aerostructural Optimization



Thursday, March 13

12 p.m.
Weber 200

Abstract

High-fidelity aerostructural optimization is important to the design of high-aspect ratio wings and high-speed vehicles. Due to high computational cost, high-fidelity design tools are often limited by mesh size, only consider part of the structure, and have narrower optimization scope. The goal of this work is to use computational methods and machine learning to reduce the cost of high-fidelity design optimizations. Namely, log-scale methods will be used to speedup sizing optimizations, using a novel combination of geometric programming and the fully-stressed design method. Machine learning is used to develop surrogate models of failure modes while maintaining computational efficiency for optimization.  In-house FEA and CFD tools are converted to GPU parallelism to speed up the baseline high-fidelity analysis, with scalable sparse linear solvers. Finally, parametric ROMs are built from the high-fidelity aerostructural analyses with a POD reduced basis to speedup solve time. These tools and strategies lead to faster high-fidelity optimizations and increase the scope of capabilities. Optimizations with geometric nonlinear structures, aeroelastic and aerothermoelastic physics are performed to demonstrate the effectiveness of these methods.

Committee
•    Prof. Graeme Kennedy – GT, School of Aerospace Engineering (advisor)
•    Prof. Christos Athanasiou – GT, School of Aerospace Engineering
•    Prof. Elizabeth Qjan – GT, School of Aerospace Engineering
•    Prof. Kai James – GT, School of Aerospace Engineering
•    Dr. Vinay Goyal – UCLA, School of Aerospace Engineering