Multi-fidelity and multi-objective optimization under uncertainty in aerospace design

Lead Research Organisation: University of Cambridge
Department Name: Engineering

Abstract

In the past, design optimization was used to finding the best design at a single operating point which would usually have degraded performance at off-design conditions. Nowadays, the focus has shifted to guiding the optimization process towards finding designs such that they achieve as good performance as possible over a range of uncertain operating conditions. This project will focus on improving recent techniques in optimization under uncertainty which find stochastically non-dominated designs such as horsetail matching. Since CFD computations are expensive, uncertainty propagation is done by first building a surrogate model over the uncertainty space and then sampling it many times. This project aims at reducing the computational cost associated with building surrogate models, such as improving current sampling techniques and using multiple fidelity models. In addition, improving efficiency of multiple stage machine (i.e. jet engine) optimization under uncertain inputs and targets will be looked at. Suitable applications include turbomachinery and hypersonic vehicle designs.

Publications

10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/N509620/1 01/10/2016 30/09/2022
1961585 Studentship EP/N509620/1 01/10/2017 30/09/2020 Daumantas Kavolis
 
Description Real-life engineering designs, be it aircraft or cars, are subject to a multitude of variable operating conditions like speed over its lifetime. Thus it is important to take those uncertain conditions into account during the design stage while changes are still cheap to make. However, this uncertainty quantification requires potential designs to be evaluated multiple times each with different operating conditions which can quickly make some design problems infeasible. Especially in aerospace, where simulations can be extremely expensive, reducing the number of those simulations required can save a lot of time and money. The current issue with a vast majority of existing uncertainty quantification methods is that they are developed for analysis purposes, that is they do not rely on any prior knowledge of the design. In contrast, as optimization converges towards solution or solutions new designs are increasingly similar to previous ones. Using this knowledge, the uncertainty quantification can be sped up since an approximate evaluation has already been done. This can be used to achieve the same optimal solutions faster, find better ones in the same period of time or take into account more variables at the design stage.
Exploitation Route Reduced optimization time can be used by businesses to take their product to market faster which would result in higher market share. Academics can tackle more difficult problems on the same computational budget, such as hypersonic vehicle design.
Sectors Aerospace, Defence and Marine
 
Title Adaptive sampling for surrogates in optimization 
Description We can use the discrepancies at each unused sample between the previous design surrogate and partial surrogate constructed using only the same sample points as the current design to guide sampling towards less accurate regions. 
Type Of Material Computer model/algorithm 
Year Produced 2021 
Provided To Others? No  
Impact For accurate and smooth surrogates, this showed significant computational savings with improved statistical moments as computed from the constructed surrogates compared to constructed surrogates based on a space-filling sampling sequence. It can also be combined with surrogate information reuse for additional computational savings. Computational savings are achieved by shifting larger changes to surrogates earlier in surrogate construction so without adjustments to tolerances some accuracy loss can be expected. Non-smooth model responses did not show any improvements as the noise from the residuals interfered with accurate estimates of sample accuracies and further research is needed. 
 
Title Partial interpolator construction for integrated path optimization 
Description In aerospace, the performance of some designs may be affected by the flight path such as hypersonic reentry gliders. However, constructing interpolators for path optimization based on fixed bounds is unnecessarily wasteful - some regions of the interpolator will simply be never used. Using the information obtained from already evaluated similar designs, current design interpolator bounds can be estimated to reduce the cost of constructing an interpolator with no accuracy loss in the integrated paths. 
Type Of Material Computer model/algorithm 
Year Produced 2021 
Provided To Others? No  
Impact 50% computational savings were observed in a hypersonic reentry glider shape optimization with a Cartesian grid-based tri-linear interpolator with negligible differences in optimal paths, with most differences due to stochastic path optimization algorithm. The method can also be used to reduce the designer's burden of specifying appropriate interpolator bounds without overconstraining the path optimization for every design. 
 
Title Surrogate information reuse 
Description By simply starting from a similar design surrogate, significant savings can be achieved without a large impact to its accuracy. In optimization, there is a plethora of similar designs to choose. Compared to regular surrogate construction, zero baseline is replaced by an already constructed similar design surrogate. 
Type Of Material Computer model/algorithm 
Year Produced 2021 
Provided To Others? No  
Impact Depending on model responses, the computational savings achieved can vary between 10-80% or possibly even more without affecting accuracy. 
 
Title Hypersonic vehicle analysis tool 
Description A reduced-order model for quickly quantifying hypersonic vehicle parameters, such as lift and drag, using ideal/real gas models, similar to CBAERO used by NASA. Implemented in C++ for rapid evaluations but Python bindings are included for use with scientific packages only available on Python. Also contains a simple waverider generator. 
Type Of Technology Physical Model/Kit 
Year Produced 2020 
Impact The model can evaluate hundreds of designs per second on a desktop which allows testing ideas very quickly. Genetic algorithms require hundreds and thousands of evaluations to find the solution which quickly becomes infeasible as a test tool for many problems. Previously, ideas were tested on algebraic problems and later on 2D aerofoils but this allows testing more interesting problems.