Simulation-based aerodynamic shape optimization has been greatly
pushed forward during the past several decades, largely due to the
developments of CFD, geometry parameterization methods, mesh deformation
techniques, sensitivity computation, and numerical optimization algorithms.
Effective integration of these components has made aerodynamic shape
optimization a highly automated process, requiring less and less human
interference. Mesh generation, on the other hand, has become the main
overhead of setting up the optimization problem. Obtaining a good
computational mesh is essential in CFD simulations for accurate output
predictions, which as a result significantly affects the reliability of
optimization results. However, this is in general a nontrivial task,
heavily relying on the user’s experience, and it can be worse with the
emerging high-fidelity requirements or in the design of novel configurations.
On the other hand, mesh quality and the associated numerical errors are
typically only studied before and after the optimization, leaving the design
search path unveiled to numerical errors. This work tackles these issues
by integrating an additional component, output-based mesh adaptation, within
traditional aerodynamic shape optimizations.
First, we develop a more suitable error estimator for optimization
problems by taking into account errors in both the objective and
constraint outputs. The localized output errors are then used to drive
mesh adaptation to achieve the desired accuracy on both the objective
and constraint outputs. With the variable fidelity offered by the adaptive
meshes, multi-fidelity optimization frameworks are developed to tightly
couple mesh adaptation and shape optimization.
The objective functional and its sensitivity are first evaluated on an
initial coarse mesh, which is then subsequently adapted as the shape
optimization proceeds. The effort to set up the optimization is minimal
since the initial mesh can be fairly coarse and easy to generate.
Meanwhile, the proposed framework saves computational costs by reducing the
mesh size at the early stages of the optimization, when the design is far from
optimal, and avoiding exhaustive search on low-fidelity meshes when the outputs
are inaccurate.
To further improve the computational efficiency, we also introduce new
methods to accelerate the error estimation and mesh adaptation using
machine learning techniques. Surrogate models are developed to predict the
localized output error and optimal mesh anisotropy to guide the adaptation.
The proposed machine learning approaches demonstrate good performance in
two-dimensional test problems, encouraging more study and developments to
incorporate them within aerodynamic optimization techniques.
Although CFD has been extensively used in aircraft design and
optimization, the design automation, reliability, and efficiency are
largely limited by the mesh generation process and the fixed-mesh
optimization paradigm. With the emerging high-fidelity requirements and the
further developments of unconventional configurations, CFD-based optimization
has to be made more accurate and more efficient to achieve higher design
reliability and lower computational cost. Furthermore, future aerodynamic
optimization needs to avoid unnecessary overhead in mesh generation and
optimization setup to further automate the design process.
The author expects the methods developed in this work to be the keys to
enable more automated, reliable, and efficient aerodynamic shape optimization,
making CFD-based optimization a more powerful tool in aircraft design.
@phdthesis{Chen_2020_Thesis,
title = {Enabling Automated, Reliable, and Efficient Aerodynamic Shape Optimization With Output-Based Adapted Meshes},
author = {Chen, Guodong},
school = {University of Michigan},
address = {Ann Arbor, Michigan, USA},
year = {2020},
url = {http://hdl.handle.net/2027.42/163034}
}
