Implementation and Evaluation of Machine Learning Assisted Adjoint Sensitivities Applied to Turbomachinery Design Optimization

Implementation and Evaluation of Machine Learning Assisted Adjoint Sensitivities Applied to Turbomachinery Design Optimization

Bibliographic Details
Main Author: Ugolotti, Matteo
Language:English
Published: University of Cincinnati / OhioLINK 2020
Subjects:
Aerospace Materials
Aerodynamic optimization
Adjoint Method
Turbomachinery Design
CFD
Machine Learning
Aerodyanamics
Online Access:http://rave.ohiolink.edu/etdc/view?acc_num=ucin1593267985073912
id ndltd-OhioLink-oai-etd.ohiolink.edu-ucin1593267985073912
record_format oai_dc
collection NDLTD
language English
sources NDLTD
topic Aerospace Materials
Aerodynamic optimization
Adjoint Method
Turbomachinery Design
CFD
Machine Learning
Aerodyanamics
spellingShingle Aerospace Materials
Aerodynamic optimization
Adjoint Method
Turbomachinery Design
CFD
Machine Learning
Aerodyanamics
Ugolotti, Matteo
Implementation and Evaluation of Machine Learning Assisted Adjoint Sensitivities Applied to Turbomachinery Design Optimization
author Ugolotti, Matteo
author_facet Ugolotti, Matteo
author_sort Ugolotti, Matteo
title Implementation and Evaluation of Machine Learning Assisted Adjoint Sensitivities Applied to Turbomachinery Design Optimization
title_short Implementation and Evaluation of Machine Learning Assisted Adjoint Sensitivities Applied to Turbomachinery Design Optimization
title_full Implementation and Evaluation of Machine Learning Assisted Adjoint Sensitivities Applied to Turbomachinery Design Optimization
title_fullStr Implementation and Evaluation of Machine Learning Assisted Adjoint Sensitivities Applied to Turbomachinery Design Optimization
title_full_unstemmed Implementation and Evaluation of Machine Learning Assisted Adjoint Sensitivities Applied to Turbomachinery Design Optimization
title_sort implementation and evaluation of machine learning assisted adjoint sensitivities applied to turbomachinery design optimization
publisher University of Cincinnati / OhioLINK
publishDate 2020
url http://rave.ohiolink.edu/etdc/view?acc_num=ucin1593267985073912
work_keys_str_mv AT ugolottimatteo implementationandevaluationofmachinelearningassistedadjointsensitivitiesappliedtoturbomachinerydesignoptimization
_version_ 1719457704262500352
spelling ndltd-OhioLink-oai-etd.ohiolink.edu-ucin15932679850739122021-08-03T07:15:31Z Implementation and Evaluation of Machine Learning Assisted Adjoint Sensitivities Applied to Turbomachinery Design Optimization Ugolotti, Matteo Aerospace Materials Aerodynamic optimization Adjoint Method Turbomachinery Design CFD Machine Learning Aerodyanamics The need for quick and low-cost numerical aerodynamic shape optimization capabilities has become of primary importance in recent years for aircraft and jet engine manufacturers to improve their product performance and reduce design costs. This performance enhancement process is very complex and might involve a few objective functions but, it could readily consider hundreds of design parameters. The success of the fast gradient-based approaches for optimization is due to the introduction of the Adjoint Method, which allowed the computation of the objective function gradient in a very computationally efficient fashion compared to alternative existing methods. The Adjoint Method made it possible to quickly calculate the derivatives of the objective function with respect to the volume grid nodes thanks to a particular differentiation of the Computational Fluid Dynamic solver, whereas several techniques could be used to obtain the design parameter derivatives of volume grid nodes. Despite its popularity, the Adjoint Method for design optimization is still not a standard tool among designers. Possible explanations include the lack of the functional gradient computation capability in some numerical codes for Computational Fluid Dynamics, the need of a well-converged flow solution (sometimes difficult or impossible to achieve), the quite common instability of the adjoint solver, the inability to easily link the derivatives to the actual design variables (especially for turbomachinery cases) and the inability to properly and robustly morph the geometry and the computational domain during the optimization phase.This dissertation details the implementation of the discrete Adjoint Method and its application in the design optimization of turbomachinery blades. The intent is to tackle some of the aforementioned drawbacks of the Adjoint Method and advance the applicability of this approach for automatic design optimization. Some of the contributions of the present work to this area of research include the use of a robust implicit discrete adjoint solver in a discontinuous Galerkin framework, an alternative implementation approach to compute the derivatives respect to the volume grid nodes that utilized both hardcoded and automatic differentiation, and the complete differentiation of the SA turbulence model with PDE-based- computed wall distance field for accurate objective function derivatives.Furthermore, this work successfully demonstrates the use of finite-differences to approximate the design parameter derivatives of volume grid nodes for structured meshes and proposes an original method that utilizes Machine Learning (ML). This innovative ML approach is an alternative to the inaccurate finite-difference and the popular mesh deformation techniques and requires minimum implementation effort while guaranteeing sufficient accuracy of the derivatives and satisfying the important link between the objective function and the actual design variables.The adjoint infrastructure is also tested and its performance assessed using several fluid dynamic cases. Eventually, the entire gradient-based infrastructure is applied to concrete optimization problems typical of turbomachinery design, such as the design of a turbine vane and an outlet guide vane. A multi-objective function optimization case is also studied to prove that the adjoint capability can be successfully utilized to describe Pareto Fronts and carry out multi-objective and/or multi-point shape optimizations. 2020-10-22 English text University of Cincinnati / OhioLINK http://rave.ohiolink.edu/etdc/view?acc_num=ucin1593267985073912 http://rave.ohiolink.edu/etdc/view?acc_num=ucin1593267985073912 unrestricted This thesis or dissertation is protected by copyright: all rights reserved. It may not be copied or redistributed beyond the terms of applicable copyright laws.

Similar Items