Topology Optimisation under uncertainty for gas turbine internal flows
File(s)
Author(s)
Gaymann, Audrey Eeva
Type
Thesis
Abstract
The concept of Topology Optimization (TO) aims at designing optimum structures for
given boundary conditions and for a specific goal. In fluid dynamics, TO optimizes geometries based on a target set of performances required for the flow paths. This thesis covers the development and implementation of two innovative TO algorithms in the context of a solid surface in contact with a fluid domain.
The first mathematical approach developed relies on a continuous adjoint based method (ABM) combined with a sedimentation process. A scalar variable is introduced, the impermeability,
whose values are updated according to its local impact on the cost function.
Impermeability yields the optimum design inside the domain considered. The ABM is
further developed to include the presence of two type of flows inside the domain: a forward
flow and a reverse flow. The final goal of the optimization is the design of a diode
without moving parts whose properties are low pressure losses for the forward flow while
the reverse flow is faced with high pressure losses. The last attribute of the developed
ABM takes into account the impact of stochastic variations as opposed to the deterministic
optimization run on the diode. A Polynomial Chaos Expansion is used to compute
the effective gradient of the cost function and update the impermeability. For the different
optimization in the ABM, 2D and 3D geometries were obtained and validated on
commercial software.
The second mathematical approach developed in this thesis breaks from traditional
TO methods and relies on machine learning, specifically deep neural networks (DNNs).
Cells are updated inside the domain following the rule of a binary state; they are either
in a solid or in a fluid state. This enables an analogy of the domain to a binary image
and to be process with convolution neural networks. The system does not require human
knowledge and learns through an algorithm based on a Deep Neural Network and a Monte
Carlo Tree Search. The network is combined with an incompressible fluid solver, though
the method is flexible to accept any solver. ABM and DNN results are compared and
validated.
given boundary conditions and for a specific goal. In fluid dynamics, TO optimizes geometries based on a target set of performances required for the flow paths. This thesis covers the development and implementation of two innovative TO algorithms in the context of a solid surface in contact with a fluid domain.
The first mathematical approach developed relies on a continuous adjoint based method (ABM) combined with a sedimentation process. A scalar variable is introduced, the impermeability,
whose values are updated according to its local impact on the cost function.
Impermeability yields the optimum design inside the domain considered. The ABM is
further developed to include the presence of two type of flows inside the domain: a forward
flow and a reverse flow. The final goal of the optimization is the design of a diode
without moving parts whose properties are low pressure losses for the forward flow while
the reverse flow is faced with high pressure losses. The last attribute of the developed
ABM takes into account the impact of stochastic variations as opposed to the deterministic
optimization run on the diode. A Polynomial Chaos Expansion is used to compute
the effective gradient of the cost function and update the impermeability. For the different
optimization in the ABM, 2D and 3D geometries were obtained and validated on
commercial software.
The second mathematical approach developed in this thesis breaks from traditional
TO methods and relies on machine learning, specifically deep neural networks (DNNs).
Cells are updated inside the domain following the rule of a binary state; they are either
in a solid or in a fluid state. This enables an analogy of the domain to a binary image
and to be process with convolution neural networks. The system does not require human
knowledge and learns through an algorithm based on a Deep Neural Network and a Monte
Carlo Tree Search. The network is combined with an incompressible fluid solver, though
the method is flexible to accept any solver. ABM and DNN results are compared and
validated.
Version
Open Access
Date Issued
2019-09
Date Awarded
2020-04
Copyright Statement
Creative Commons Attribution NonCommercial Licence
License URL
Advisor
Montomoli, Francesco
Sponsor
Engineering and Physical Sciences Research Council (EPSRC)
Grant Number
1662246
Publisher Department
Aeronautics
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)