Proceedings of ASME Turbo Expo 2018

Turbomachinery Technical Conference and Exposition

GT2018
June 11-15, 2018, Oslo, Norway

GT2018-77098

Downloaded from https://asmedigitalcollection.asme.org/GT/proceedings-pdf/GT2018/51012/V02CT42A050/2788836/v02ct42a050-gt2018-77098.pdf by Milan Petrovic on 24 February 2020

MACHINE LEARNING ENABLED ADAPTIVE OPTIMIZATION
OF A TRANSONIC COMPRESSOR ROTOR WITH PRE-COMPRESSION

Michael Joly Soumalya Sarkar Dhagash Mehta

Thermal Fluid Science Department Systems Department Systems Department
United Technologies Research Center United Technologies Research Center United Technologies Research Center
East Hartford, CT, USA East Hartford, CT, USA East Hartford, CT, USA
Email: [email protected] Email: [email protected] Email: [email protected]

ABSTRACT Unorm Normalized predictive uncertainty of current surrogate

D(xi , x j ) Pairwise Mahalanobis distance between xi and x j
In aerodynamic design, accurate and robust surrogate C Covariance matrix
models are important to accelerate computationally expensive
CFD-based optimization. In this paper, a machine learning SAO Surrogate Assisted Optimization
framework is presented to speed-up the design optimization of EMO Evolutionary Multi-objective Optimization
a highly-loaded transonic compressor rotor. The approach is NEC No Evoluation Control
three-fold: (1) dynamic selection and self-tuning among several FEC Fixed Evolution Control
surrogate models; (2) classification to anticipate failure of the AEC Adaptive Evolution Control
performance evaluation; and (3) adaptive selection of new can- ML4SAO Machine Learning for Surrogate Assisted Optimization
didates to perform CFD evaluation for updating the surrogate, GP Gaussian Process
which facilitates design space exploration and reduces surrogate RBF Radial Basis Function
uncertainty. The framework is demonstrated with a multi-point NN Neural Network
optimization of the transonic NASA rotor 37, yielding increased RR Random Forest
compressor efficiency in less than 48 hours on 100 CPU cores. MAE Mean Absolute Error
The optimized rotor geometry features pre-compression that SMBO Sequential model-based Bayesian Optimization
relocates and attenuates the shock, without the stability penalty ROC Receiver Operating Characheristic
or undesired reacceleration usually observed in the literature. TPR True Positive Rate
FRP False Positive Rate
AUC Area Under the Curve
NOMENCLATURE DOE Design of Experiments

yi True performance output for the ith sample

ŷi Surrogate estimate for the ith sample INTRODUCTION
n Number of samples Aircraft engine manufacturers are subject to ever-increasing de-
nt Number of decision trees mands in terms of fuel efficiency and emissions. Consequently,
R Metric for ranking new design candidates modern compressor stages are required to deliver higher pressure
α Weight parameter ratios in more compact environments, with the maximum possi-
Dnorm Normalized Mahalanobis distance ble efficiency and stability margin.

1 Copyright © 2018 United Technologies Corporation

 The design space for highly-loaded transonic compressor Mesh. Two grid refinements are considered when generat-
stages is complex, with several loss mechanisms such as tip ing the grids with Numeca Autogrid5: a grid of 2.5 x 106 cells
clearance vortical structures, end-wall separation, and shock used in the optimization and a grid of 5.0 x 106 cells to guarantee
wave-boundary layer interaction to consider [1]. Beyond legacy the grid independency, as illustrated in Figure 1. The y+ is about
meanline loss models, computational fluid dynamics (CFD) 1. The finest grid comprises 25 cells in the tip gap, for a total of
solvers have been used in conjunction with optimization algo- 128 radially.

Downloaded from https://asmedigitalcollection.asme.org/GT/proceedings-pdf/GT2018/51012/V02CT42A050/2788836/v02ct42a050-gt2018-77098.pdf by Milan Petrovic on 24 February 2020

rithms in the past two decades to guide engineers in the under- CFD and validation UTCFD, a Reynolds-averaged Navier-
standing of these flow patterns and their impact on compressor Stokes (RANS) solver using the k − ω turbulence model, is val-
design [2]. idated against experimental data in Figure 1. The observed dis-
To accelerate aerodynamic design, simplified yet accurate crepancies with the measurements are consistent with previous
and robust models can perform as surrogates to computationally RANS-based comparisons [1, 6].
expensive CFD models. Traditionally, an evolutionary optimiza-
The present work aims to accelerate the multi-point opti-
tion methodology utilizes a surrogate model developed from a
mization of the NASA rotor 37. Two flow conditions are con-
CFD-based design of experiments.
sidered, illustrated as diamonds in Figure 1. Improvements in
In this paper, recent advancements in machine learning are
efficiency are targeted while maintaining pressure ratio and op-
incorporated into a surrogate-assisted optimization (SAO) im-
erating range across the nominal speedline.
plementation, acting in three ways: (1) dynamic selection and
tuning among several surrogate models throughout the optimiza-
tion; (2) prediction to anticipate failure of new CFD-based per-
formance evaluation; and (3) identification of new candidates for
CFD evaluation to refine the training of the surrogate models.
The framework is shown to outperform typical surrogate models
(such as Kriging) used in the turbomachinery community.
The proposed framework has significantly accelerated
the high-dimensional design optimization of a highly-loaded
transonic compressor rotor. The framework is also able to
identify detailed geometric features, such as pre-compression to
mitigate the shock wave-boundary layer interaction losses in a
transonic blade passage, and optimize such features for design
and off-design performance.

Problem formulation
The baseline selected to evaluate the proposed framework is
NASA rotor 37, a thoroughly investigated compressor rotor
tested in the 70s [3] and used as a validation reference for sev-
eral computational studies [1]. A survey of previous optimiza-
tion campaigns on this rotor has recently been published [4]. Al-
though a few studies consider off-design performance in post-
optimization analysis, the previous works focus on optimizing
performance at design point, often resulting in stability margin
deficit or choke point reduction. Meanwhile, multi-point aerody-
namic optimization is well documented [5], but hardly applied to
the highly-loaded rotor 37.
Parametrization. In the present effort, conventional
engineering parameters are considered for blade geometry
parametrization, including camber and thickness, lean and
sweep. Besides convenience for the design engineers, such
parametrization allows for detailed control of the section pro- FIGURE 1. Performance of NASA rotor 37, including CFD valida-
files, which is difficult to achieve with the alternative approach of tion, grid indenpency study, and baseline for optimization
free-form deformation. This is particularly important for highly-
loaded transonic compressors, where localized features are de-
sired to control the shock system.

2 Copyright © 2018 United Technologies Corporation

Methods to accelerate turbomachinery aero-design to allow the use of traditional optimization algorithms [13], and
Aerodynamic engineers have devoted considerable efforts in ac- in turn further enhancing the efficiency of the EMO methods.
celerating the CFD-based design optimization of turbomachin- There are various methods [14] to generate surrogate mod-
ery stages [7]. Meanwhile, the expectation of flow solver fi- els. One can use polynomial approximation models in which
delity somehow relentlessly follows and balances the continual one employs a combination of regression techniques assuming
increase of computational resources. Techniques for acceler- that the underlying functions are polynomials [15]. More

Downloaded from https://asmedigitalcollection.asme.org/GT/proceedings-pdf/GT2018/51012/V02CT42A050/2788836/v02ct42a050-gt2018-77098.pdf by Milan Petrovic on 24 February 2020

ated design therefore remain a critical need in industrial envi- sophisticated response surface methods such as Kriging and
ronments. radial basis function (RBF) have also been used in the past [16].
Multi-objective evolutionary optimization. Turbomachin- More recently, machine learning based methods such as support
ery design problems usually involve several objectives from dif- vector machines [17] and artificial neural networks [18] have
ferent disciplines of engineering and are known to be multi- gained traction due to their superior performance on a wide
objective optimization problems. Within the discipline of aero- range of problems.
dynamics, typical objectives include pressure ratio, efficiency,
and flow capacity. In principle, the target is to find a point in Contributions to surrogate-assisted optimization
the high-dimensional space at which all the objectives are simul- Surrogate-assisted optimization is used extensively for blade de-
taneously optimized. However, unlike the single-objective opti- sign, typically with Kriging or Neural Network surrogate mod-
mization problems, there may not exist such a point. Hence, the els [19]. Previous studies focused on case-specific model com-
goal is then to find Pareto optimal solutions; i.e., the points in parisons or on ensemble or stacking of surrogate models [20].
the space at which none of objectives can be further optimized The state-of-the-art of SAO, which involves direct replace-
without degrading at least one of the other objectives. The set ment of real function by surrogate, has been divided into three
of Pareto optimal solutions is called the Pareto front. One pop- categories according to the integration approach of surrogate in
ular approach to solve multi-objective optimization problems is SAO [14]: (i) No Evolution Control (NEC); (ii) Fixed Evolu-
evolutionary algorithms such as a genetic algorithm or particle tion Control (FEC); and (iii) Adaptive Evolution Control (AEC).
swarm optimization. NEC performs entire SAO with the surrogate modeled on initial
In recent years, the evolutionary multi-objective optimiza- DOE, but it is limited by erroneous convergence due to low sur-
tion (EMO) methods [8] have proven to be immensely success- rogate accuracy at high dimensionality. FEC adapts during high-
ful in applications such as aerodynamic design optimization and dimensional SAO via switching from surrogate to real function
other structural optimization problems [9]. However, as the and vice versa, but it is challenging to define the switching pa-
methods find more complex industrial applications such as tur- rameter a priori. Although state-of-the-art AECs have some au-
bomachinery design, the EMO tasks have become ever more dif- tomatic alternation mechanism between surrogate and real func-
ficult. In particular, the design space becomes high-dimensional tion, the techniques have not scaled up to large scale industrial
and the number of objectives grows rapidly due to requirements problems yet. Other than these shortcomings, several other chal-
from multiple disciplines. In turn, either the computational time lenges arise from the application of SAO to the multi-point opti-
or the experimental efforts to obtain each point in the high- mization of highly-loaded compressor blades such as CFD fail-
dimensional space becomes prohibitively expensive. Hence, it ure at off-design, high cost of CFD, dynamic selection of the op-
is essential to invent novel methods to enhance the efficiency of timal surrogate and online hyperparameter tuning of surrogate,
EMOs. and sparsity of the training patterns [21].
Surrogate modeling. A popular way to enhance the effi- The proposed framework complements a SAO implemen-
ciency of EMO algorithms for optimization is surrogate model- tation with machine learning techniques to efficiently train and
ing [10]. Here, based on a handful number of data-points, one effectively exploit the surrogate and to formalize automatic
constructs an approximation model of the simulation or exper- switching between surrogate and CFD in order to accelerate the
imental model which behaves as closely to the original model high-dimensional design cycle of turbomachinery stages. The
as possible while being computationally less expensive. Sur- following section describes the methods implemented within
rogate modeling is broadly classified in three types: (i) data- the framework, including algorithms for (i) constrained multi-
driven surrogate models which rely purely on input and output objective optimization; (ii) dynamic selection and tuning of
data-points; (ii) projection-based models in which the governing surrogate models; (iii) infeasibility detection; and (iv) adaptive
equations are projected onto a lower dimensional subspace; and down-selection of new CFD candidates for surrogate update. A
(iii) multifidelity models which are built using physics-based un- results section then presents the outcome of the framework when
derstanding. Such surrogate models have been shown to reduce applied to the aerodynamic design optimization of the NASA
numerical instabilities and in turn help perform uncertainty anal- rotor 37.
ysis [11], simplify complex models by weeding out irrelevant
parameters [12], to smooth the landscape of objective functions

3 Copyright © 2018 United Technologies Corporation

COMPUTATIONAL METHODS Dynamic selection and tuning of surrogate models
This section describes the proposed framework named ’Machine It is important to build a surrogate that is both accurate and gen-
Learning for Surrogate Assisted Optimization (ML4SAO)’ eralizable over the design domain to some extent. But man-
which enhances the accuracy and effectiveness of SAO in ually exploring for the right surrogate and then optimizing its
accelerating the design of turbomachinery stages. hyperparameter can be tedious and computationally expensive.
Also this type of exploration requires significant amount of do-

Downloaded from https://asmedigitalcollection.asme.org/GT/proceedings-pdf/GT2018/51012/V02CT42A050/2788836/v02ct42a050-gt2018-77098.pdf by Milan Petrovic on 24 February 2020

Constrained Multi-Objective Optimization (CMOO) main expertise in machine learning, which may not be accessible
A differential evolution algorithm [22] is implemented and ex- to other engineering domains. To automate and accelerate the
tended to multi-objective problems using a Pareto-based ap- process of obtaining the best surrogate candidate with its opti-
proach [23], using a similar non-dominated sorting and rank- mal hyperparameter set, a framework named ’Machine Learning
ing selection procedure as in NSGA-II [24]. The crowding dis- for Surrogate Assisted Optimization (ML4SAO)’ is devised and
tance [25] is used to ensure uniformly distributed selection to the schematic is shown in Figure 3. ML4SAO invokes a suite
build the Pareto fronts. Constraints are handled with the con- of surrogate candidates to obtain an optimal surrogate for each
strained tournament method [24], where penalties are prescribed performance output. The suite of surrogate candidates includes
directly on the Pareto ranks based on the severity of the con-
straints violation. This contributes to easing the handling of
severely constrained cases, where a well performing candidate
with a constraint slightly violated could be selected before a less
performing candidate whose constraints are all satisfied. This de-
scribes the non-assisted optimization algorithm, so called ”direct
optimization” as shown in Figure 2.
SAO is implemented by substituting the accurate CFD eval-
uations with the predictive surrogate, as illustrated in Figure 2.
Subsequently to a CFD-based design of experiment using Latin
hypercube sampling, the proposed ML4SAO machine learning
framework selects, trains and tunes a model among several sur-
rogate models. The evolutionary algorithm uses the predictive
surrogate to rapidly produce a feasible Pareto front, revealing the
design trade-offs between the different objectives. Candidates
are then adaptively selected along the Pareto front for further
CFD evaluation and added to the training set to refine the sur-
rogate model training.
FIGURE 3. Schematic of ML4SAO framework

conventional surrogate types such as linear regression models,

polynomial approximation models (e.g., polynomial lasso re-
gression), Gaussian process (GP) or Kriging, multi-layer Percep-
tron based neural network regression models (NN), RBF network
etc. Another unconventional surrogate model, according to the
surrogate based optimization literature, based on random forest
(RF) [26, 27] is added to the suite. Random forest regression is
an ensemble method that averages the predictions from nt num-
ber of decision regression trees. Those decision trees are trained
on nt number of bootstrap samples drawn from the data and each
tree grows without pruning by selecting best splits among a ran-
dom fraction of design variables at each step.
FIGURE 2. Surrogate Assisted Optimization (SAO)
The framework searches for the best surrogate by minimiz-
ing an error metric, e.g. mean absolute error, via cross validation
approach. Mean absolute error (MAE) is a measure of accuracy

4 Copyright © 2018 United Technologies Corporation

for the surrogate model and is defined as data which has been sequentially incremented on initial DOE.

Infeasibility Detection
Σn |yi − ŷi |
MAE = i=1 , (1) In a typical CFD computation, it is often the case for certain pa-
n rameter values that the computation breaks down resulting in an
undefined output, usually ‘not a number’ (NaN), representing in-

Downloaded from https://asmedigitalcollection.asme.org/GT/proceedings-pdf/GT2018/51012/V02CT42A050/2788836/v02ct42a050-gt2018-77098.pdf by Milan Petrovic on 24 February 2020

where yi and ŷi are the true performance outputs and correspond- feasibility of the parameters. NaNs yield difficulties while train-
ing surrogate estimates respectively for ith sample out of n total ing the corresponding response. As a common practice, a value
samples. (e.g., average of some of the nearest data-points, or a random
Cross validation is a robust model validation technique in value from some chosen distribution, etc.) is imputed in place of
machine learning that is applied to data during the training phase NaNs [32], although in many cases such an imputation may not
to increase generalizability to unseen data and to reduce overfit- represent the physics behind the phenomenon under considera-
ting. A k-fold cross validation [28] is used in this work to observe tion.
and select the best surrogate model or its optimal set of hyperpa- In this work, a different approach that learns the pattern of
rameters that has the least average MAE over k folds. k-fold cross appearance of NaNs is employed. When NaNs are present, a ma-
validation partitions the training data into k non-overlapping and chine learning model learns the likelihood of a new design candi-
exhaustive sets, trains the surrogate on k − 1 sets and tests on kth date to yield either a valid performance value or a NaN. For such
set and repeats this process k times to get an average of the model binary classification problems, a receiver-operating characteris-
performance. In this paper, k = 5 to 10 is mostly used, depending tic (ROC) curve for different cross-validations is used to evalu-
on the data size. ate the performance of the classification model. The x-axis of
A hyperparameter in machine learning is a parameter that is the ROC curve is true positive rate (TPR), i.e., the fraction of the
tuned before the training begins, as opposed to the parameters predictions which are correctly classified as NaNs. The y-axis
that are adjusted during the training via an optimization algo- of the ROC curve is false positive rate (FPR), i.e., the fraction
rithm. Some of the examples of hyperparameters are: regulariza- of the predictions which are incorrectly classified as NaNs. Both
tion hyperparameter for polynomial lasso, number of hidden lay- of these quantities are plotted for varying threshold on prediction
ers or hidden unites on a layer for NN, number of trees for a RF, probabilities, hence the ROC curve provides a way to evaluate
and type of kernel or a specific kernel’s length scale for GP based the performance of different models as well as to select the opti-
surrogates. A thorough search in the multi-dimensional hyperpa- mal value for the threshold. An optimal threshold usually max-
rameter spaces for surrogates based on RF, NN or GP with large imizes the TPR and minimizes the FPR. The optimal threshold
data can be prohibitively expensive. ML4SAO performs a quick value is found by computing the value of the threshold for which
and coarse grid search with 3-5 evaluation per hyperparameter |(1 − T PR)| + |FPR| is minimized. From an ROC curve, a sin-
dimension and finds, via cross validation for each of the sur- gle metric calls the area under the curve (AUC) of the ROC and
rogates, the model with lowest MAE. Then, ML4SAO applies ranges from 0 to 1, with 0 being the worst performance and 1
a sequential model-based Bayesian optimization (SMBO) [29] being a perfect model.
approach on that model to quickly obtain an optimal set of hy- The ROC curves for the response related to the prediction of
perparameter [30]. efficiency with 150 (a) and 300 (b) data are plotted in Figure 4.
SMBO [29, 31] is a global optimization technique for ROC curves for 5-fold cross validations are plotted separately
expensive black-box functions. When used in the context of in each of the two plots to demonstrate the variations among
hyperparameter (θ ) optimization, SMBO creates a smooth the curves over cross validations. A curve over mean values of
meta model (preferably GP) for mapping the hyperparameter the 5 curves is also shown in the figure. The dashed-line plot
to the cross validation error metric, e.g. MAE, based on a few on the diagonal shows a hypothetical worst case scenario where
initial data points and then sequentially invokes the expensive the corresponding model would correctly predict NaNs only by
surrogate (S) based on an acquisition function such as expected chance. Hence, the closer the mean curve to this diagonal curve
improvement. SMBO iteratively picks hyperparameters to run is the less accurate the corresponding model is. The closer a
the expensive surrogate while maximizing acquisition function curve follows the left-hand and then top top borders, the more
that trades off exploration for reducing GP uncertainty and accurate the model is. Hence, the plots in Figure 4 yield that the
exploitation for ‘better’ hyperparameter. As the SAO marches model for 300 data-points is clearly significantly better than that
forward, the initial data points are gradually replaced by the 150 data-points. This conclusion is reinforced by the mean AUC
optimal hyperparameters obtained in past SAO iterations. This values, 0.8 for 300 data-points and 0.59 for 150 data-points.
particular step helps in achieving a faster hyperparameter In other words, to use this infeasibility detection approach, the
optimization in the latter SAO iterations as it uses knowledge availability of larger DOE is crucial.
from the past hyperparameter optimizations on quasi-similar

5 Copyright © 2018 United Technologies Corporation

 where R is constructed as a weighted (α) sum of normalized pre-
dictive uncertainty (Unorm ) of current surrogate and normalized
Mahalanobis distance (Dnorm ) of design candidates from the ex-
isting data.
The pairwise Mahalanobis distance between vectors xi and
x j is defined as:

Downloaded from https://asmedigitalcollection.asme.org/GT/proceedings-pdf/GT2018/51012/V02CT42A050/2788836/v02ct42a050-gt2018-77098.pdf by Milan Petrovic on 24 February 2020

q
D(xi , x j ) = (xi − x j )T C−1 (xi − x j ), (3)

where C is the covariance matrix of the data. The average

pairwise Mahalanobis distance of a new design candidate
captures a dissimilarity metric with respect to the existing
database. On the other hand, predictive uncertainty expresses
the prediction interval on the new performance estimate by the
existing surrogate. Whereas GP or linear regression has closed
form analytical solution for prediction interval, it is calculated
via bagging approach for surrogates like RF.

Validation and Comparison of ML4SAO

The proposed machine learning-improved SAO framework is

benchmarked against typical Gaussian Process surrogate based
SAO on the mathematical Kursawe function, which consists of
the two objectives below subject to −5 ≤ xi ≤ 5 and 1 ≤ i ≤ 3:

 h  q i
 f1 (xx) = ∑2 −10 exp −0.2 x2 + x2
i=1 i i+1
Minimize 
0.8
 (4)
 f2 (xx) = ∑3 |xi | 3
+ 5 sin xi
i=1

This optimization has a discontinuous Pareto front and no

FIGURE 4. ROC curves of the response related to the prediction of analytical solution. Figure 5 compares the behavior of the SAO
efficiency on design of experiments of 150 (a) and 300 (b) candidates process with different surrogate models ( Matern kernel GP, RBF
intended for rotor 37 optimization. kernel GP, ML4SAO, ML4SAO with downselect) using metrics
of distance and spread of the current Pareto front to the “ideal”
Pareto front obtained with direct optimization. The distance met-
Adaptive down-selection of new CFD candidates for ric computes the sum of distances from each point of the current
surrogate update front to the nearest point of the ideal front. The spread metric
As the SAO marches forward, it is necessary to update and retrain is the sum of distances from the extreme solutions of the current
the surrogate to eventually curb the effects of error propagation. front to the extremes of the ideal front. Lines show the statisti-
A set of non-dominated design points are obtained from each it- cal mean of the metrics over 10 optimization runs, while the bars
eration of SAO. If all of them are utilized to update the surrogate, illustrate the variance.
the amount of required CFD evaluation would be large. It is observed that ML4SAO and ML4SAO with downselect
ML4SAO selects only a few design candidates for full CFD variants outperform other conventional SAO approaches with
evaluations. This is one of the keys towards SAO acceleration in respect to both metrics. Figures 6(a-c) presents how ML4SAO
this work. SAO selects those design candidates that score highest converges to the true Pareto earlier and closer than other conven-
average ranking over all the performance outputs on a metric R tional SAO approaches. Also Figure 6(d) shows that with 100
given as: generations (instead of 20) at each iteration of ML4SAO, the
surrogate based Pareto closely overlaps with the true Pareto.
R = αUnorm + (1 − α)Dnorm , (2)

6 Copyright © 2018 United Technologies Corporation

 Downloaded from https://asmedigitalcollection.asme.org/GT/proceedings-pdf/GT2018/51012/V02CT42A050/2788836/v02ct42a050-gt2018-77098.pdf by Milan Petrovic on 24 February 2020
FIGURE 5. SAO on Kursawe function (a, mean and variance of distance over 10 optimizations; b, mean and variance of spread over 10 optimizations)

FIGURE 6. SAO results on Kursawe function with most representative case of mean over 10 optimization with different surrogate models (a,Matern
kernel ; b, RBF kernel GP; c, ML4SAO); d, Best Pareto front with 100 generations on ML4SAO

7 Copyright © 2018 United Technologies Corporation

RESULTS As a result of the single-point optimization, the off-design
The target for demonstration of the framework is to improve the performance of the most efficient candidate at design condition
efficiency of the NASA rotor 37 over its operating range, while is shown in Figure 8. While the choke flow is not reduced, the
maintaining the pressure ratio, stability margin and choke flow. stability margin shows a deficit with the pressure characteristic
Assuming the availability of a 100 CPUs cluster, which is rep- rolling over sooner. This result motivates the following multi-
resentative of current industrial engineering day-to-day practice, point optimization to maintain stability margin.

Downloaded from https://asmedigitalcollection.asme.org/GT/proceedings-pdf/GT2018/51012/V02CT42A050/2788836/v02ct42a050-gt2018-77098.pdf by Milan Petrovic on 24 February 2020

a single-point point optimization is shown to deliver 1.5 point of
efficiency increase in less than 16 hours, while a multi-point op-
timization achieves about 1.0 point of efficiency increase while
maintaining off-design performance in less than 48 hours.
The design space comprises 38 parameters that are defined
along several span locations. They consist of spline offsets to the
fitted baseline geometry for inlet and outlet metal angle, camber
angle, and thickness, as well as lean and sweep.

Single-point optimization. The flow condition at maximum

efficiency is considered for design, as illustrated by a diamond
on Figure 1. The pressure ratio is constrained to be within 1%
of the baseline and the adiabatic efficiency is set as an objective
to be maximized. A second objective is selected to minimize
the residual of the momentum equation in the CFD in order to
leverage the benefits of the multi-objective algorithm.
The history of the single-optimization is shown in Figure 7.
Subsequently to a DOE with 50 individuals, 10 promising can-
didates are selected at each iteration for CFD evaluation using
ML4SAO’s down-selection approach. Only 6 iterations are nec-
essary to reach a significant efficiency improvement. A final
round of training and optimization is performed with selection
solely based on the crowding distance to produce a uniformly
distributed front of optimized solutions. All included, a total of
120 CFD evaluations are performed, lasting less than 16 hours
on the 100 CPUs cluster.

FIGURE 8. Performance of baseline and optima of single-point and

multi-point optimizations

Multi-point optimization. Besides the design flow condi-

tion at maximum efficiency, a second point representative of off-
design condition is considered for optimization to maintain the
stability margin, as illustrated by the two diamonds on Figure 8.
The preceding single-point optimization indicated that an addi-
tional flow condition was not required to maintain choke flow in
the present case. The two objectives are to maximize the effi-
ciency at both design and off-design flow conditions. While a
FIGURE 7. History of single-point optimization similar dual-bound constraint is imposed for the pressure ratio at

8 Copyright © 2018 United Technologies Corporation

design to be within 1% of the baseline, the pressure ratio at off-
design is only constrained to be larger than the one of the baseline
to allow potential increase of the slope for the pressure charac-
teristic. To prevent oscillating solutions at the off-design condi-
tion, the performance of any case not converging with a residual
of the momentum equation below 10−6 is discarded. An addi-

Downloaded from https://asmedigitalcollection.asme.org/GT/proceedings-pdf/GT2018/51012/V02CT42A050/2788836/v02ct42a050-gt2018-77098.pdf by Milan Petrovic on 24 February 2020

tional constraint is implemented to guide the optimizer towards
region of the design where the momentum equation residuals re-
main below 10−6 . The history of the multi-point optimization
is shown in Figure 9(a). The constrained tournament approach
with rank-based penalties is observed to allow the objectives to
be optimized even though not all the constraints are satisfied. FIGURE 10. History of surrogate models selection
Alleviating the upper bound of the pressure ratio constraint at
design condition (or, in other words, considering higher pressure
ratios at design speed as valid candidates), Figure 9(b) illustrates Pre-compression. The most efficient candidate is further
points satisfying all other constraints. Subsequently to a DOE analyzed. Figure 11 shows the loss contributions along the span,
with 100 individuals, 6 iterations are performed with 10 candi- with turbulent kinetic energy downstream of the blade signifi-
dates selected at each iteration for CFD evaluation. With two cantly reduced, especially around 70% of the span.
flow conditions, a total of 340 CFD evaluations are performed,
lasting less than 48 hours on the 100 CPUs cluster. The perfor-
mance characteristic of the candidate with highest efficiency at
design point is shown on Figure 8. While the pressure ratio is
increased by 2%, violating the upper bound on the design point
pressure ratio constraint, the choke flow is not reduced, and the
stability margin is maintained.
Figure 10 lists the models selected during the multi-point
optimization, where the unconventional Random Forest model is
recurrently identified by the proposed framework as the preferred
candidate surrogate over NN or GP (or Kriging) models typically
used in the turbomachinery community. It is observed that RF
scales better than GP as the DOE grows for a high dimensional FIGURE 11. Turbulent kinetic energy downstream of the rotor; base-
design optimization problem like the one discussed here. line (a) and optimized (b)

FIGURE 9. History of multi-point optimization

9 Copyright © 2018 United Technologies Corporation

 As a result of the optimization, the pre-shock Mach num- CONCLUSION
ber is reduced (Figure 13). The proposed framework identifies
a pre-compression shock system, where compression waves de- This paper presents a machine learning framework to achieve
celerate the flow upstream of the shock, resulting in a mitiga- a faster and accurate surrogate-assisted optimization. The ap-
tion of the shock wave-boundary layer interaction (or shock- proach is three-fold: (1) dynamic selection and tuning among
induced separation) in Figure 12. In the literature, typical S- several surrogate models throughout the optimization; (2) predic-

Downloaded from https://asmedigitalcollection.asme.org/GT/proceedings-pdf/GT2018/51012/V02CT42A050/2788836/v02ct42a050-gt2018-77098.pdf by Milan Petrovic on 24 February 2020

shape profiles exhibit undesired reacceleration downstream of tion to anticipate failure of new CFD-based performance evalua-
the pre-compression region, as well as on the pressure side [4]. tion; and (3) identification of new candidates to refine the training
The present result achieves the desired pre-compression with no of the surrogate models.
reacceleation upstream of the shock (Figure 13).
The proposed framework is first benchmarked on the math-
ematical Kursawe function, highlighting a complex and discon-
tinuous Pareto front. It is shown to outperform typical surrogate
models often used in the turbomachinery community.

During compressor rotor optimization, the proposed frame-

work identifies the RF-based surrogate as the preferred candidate
compared to more conventional surrogates like Neural Network
and Gaussian Process (or Kriging) models. Dynamic selection
from the surrogate suite also facilitates the SAO performance at
the high dimensionality required for the compressor rotor op-
timization. Assuming the availability of a 100 CPUs cluster,
which is representative of current industrial engineering day-to-
day practice, the framework delivers significant improvement of
the efficiency of a highly-loaded compressor rotor in less than 48
hours, while maintaining the stability margin. It is also capable
of identifying pre-compression features in order to mitigate the
FIGURE 13. Isentropic Mach number loading at 70% span shock wave-boundary layer interaction without any reaccelera-
tion upstream of the shock.

FIGURE 12. Relative Mach number filed at 70% span; baseline (a) and optimized (b)

10 Copyright © 2018 United Technologies Corporation

ACKNOWLEDGMENT Evolutionary Computation (CEC), 2013 IEEE Congress on, IEEE,
We acknowledge internal funding from UTRC. We thank Georgi pp. 2548–2555.
Kalitzin and Gorazd Medic of UTRC for useful discussions. [18] Schalkoff, R. J., 1997. Artificial neural networks, Vol. 1. McGraw-
Hill New York.
[19] Verstraete, T., Coletti, F., Bulle, J., Vanderwielen, T., and Arts,
REFERENCES
T., 2013. “Optimization of a u-bend for minimal pressure loss in
[1] Denton, J., 1997. “Lessons from rotor 37”. Journal of Thermal internal cooling channelspart i: Numerical method”. Journal of

Downloaded from https://asmedigitalcollection.asme.org/GT/proceedings-pdf/GT2018/51012/V02CT42A050/2788836/v02ct42a050-gt2018-77098.pdf by Milan Petrovic on 24 February 2020

Science, 6(1), pp. 1–13. Turbomachinery, 135(5), p. 051015.
[2] Oyama, A., Liou, M.-S., and Obayashi, S., 2004. “Transonic [20] Goel, T., Haftka, R. T., Shyy, W., and Queipo, N. V., 2007. “En-
axial-flow blade optimization: Evolutionary algorithms/three- semble of surrogates”. Structural and Multidisciplinary Optimiza-
dimensional navier-stokes solver”. Journal of Propulsion and tion, 33(3), pp. 199–216.
Power, 20(4), pp. 612–619. [21] Karakasis, M. K., and Giannakoglou, K. C., 2006. “On the use
[3] Reid, L., and Moore, R. D., 1978. “Design and overall perfor- of metamodel-assisted, multi-objective evolutionary algorithms”.
mance of four highly loaded, high speed inlet stages for an ad- Engineering Optimization, 38(8), pp. 941–957.
vanced high-pressure-ratio core compressor”. [22] Storn, R., and Price, K., 1997. “Differential evolution–a sim-
[4] John, A., Shahpar, S., and Qin, N., 2017. “Novel compressor blade ple and efficient heuristic for global optimization over continuous
shaping through a free-form method”. Journal of Turbomachinery, spaces”. Journal of global optimization, 11(4), pp. 341–359.
139(8), p. 081002. [23] Madavan, N. K., 2002. “Multiobjective optimization using a
[5] Luo, J., Zhou, C., and Liu, F., 2014. “Multipoint design optimiza- pareto differential evolution approach”. In Evolutionary Computa-
tion of a transonic compressor blade by using an adjoint method”. tion, 2002. CEC’02. Proceedings of the 2002 Congress on, Vol. 2,
Journal of Turbomachinery, 136(5), p. 051005. IEEE, pp. 1145–1150.
[6] Chima, R. V., 2009. “Swift code assessment for two similar tran- [24] Deb, K., Agrawal, S., Pratap, A., and Meyarivan, T., 2000. “A fast
sonic compressors”. AIAA paper, 1058, p. 2009. elitist non-dominated sorting genetic algorithm for multi-objective
[7] Wang, G. G., and Shan, S., 2007. “Review of metamodeling tech- optimization: Nsga-ii”. In International Conference on Parallel
niques in support of engineering design optimization”. Journal of Problem Solving From Nature, Springer, pp. 849–858.
Mechanical design, 129(4), pp. 370–380. [25] Konak, A., Coit, D. W., and Smith, A. E., 2006. “Multi-objective
[8] Abraham, A., and Jain, L., 2005. “Evolutionary multiobjective optimization using genetic algorithms: A tutorial”. Reliability En-
optimization”. Evolutionary Multiobjective Optimization, pp. 1– gineering & System Safety, 91(9), pp. 992–1007.
6. [26] Breiman, L., 2001. “Random forests”. Machine Learning, 45(1),
[9] Jin, Y., and Sendhoff, B., 2009. “A systems approach to evolu- Oct, pp. 5–32.
tionary multiobjective structural optimization and beyond”. IEEE [27] Liaw, A., and Wiener, M., 2002. “Classification and Regression
Computational Intelligence Magazine, 4(3). by randomForest”. R News, 2(3), pp. 18–22.
[10] Jin, Y., 2011. “Surrogate-assisted evolutionary computation: Re- [28] Kohavi, R., 1995. “A study of cross-validation and bootstrap for
cent advances and future challenges”. Swarm and Evolutionary accuracy estimation and model selection”. In Proceedings of the
Computation, 1(2), pp. 61–70. 14th International Joint Conference on Artificial Intelligence - Vol-
[11] Doherty, J., and Christensen, S., 2011. “Use of paired simple ume 2, IJCAI’95, Morgan Kaufmann Publishers Inc., pp. 1137–
and complex models to reduce predictive bias and quantify un- 1143.
certainty”. Water Resources Research, 47(12). [29] Feurer, M., Springenberg, J. T., and Hutter, F., 2015. “Initializ-
[12] Young, P. C., and Ratto, M., 2011. “Statistical emulation of large ing bayesian hyperparameter optimization via meta-learning”. In
linear dynamic models”. Technometrics, 53(1), pp. 29–43. Proceedings of the Twenty-Ninth AAAI Conference on Artificial
[13] Vuik, C., Segal, A., and Meijerink, J., 1999. “An efficient precon- Intelligence, AAAI’15, AAAI Press, pp. 1128–1135.
ditioned cg method for the solution of a class of layered problems [30] Jones, D. R., Schonlau, M., and Welch, W. J., 1998. “Efficient
with extreme contrasts in the coefficients”. Journal of Computa- global optimization of expensive black-box functions”. Journal of
tional Physics, 152(1), pp. 385–403. Global Optimization, 13(4), Dec, pp. 455–492.
[14] Dı́az-Manrı́quez, A., Toscano, G., Barron-Zambrano, J. H., and [31] Brochu, E., M. Cora, V., and De Freitas, N., 2010. “A tutorial on
Tello-Leal, E., 2016. “A review of surrogate assisted multiob- bayesian optimization of expensive cost functions, with applica-
jective evolutionary algorithms”. Computational intelligence and tion to active user modeling and hierarchical reinforcement learn-
neuroscience, 2016. ing”.
[15] Myers, R. H., Montgomery, D. C., and Anderson-Cook, C. M., [32] Saar-Tsechansky, M., and Provost, F., 2007. “Handling missing
2016. Response surface methodology: process and product opti- values when applying classification models”. Journal of machine
mization using designed experiments. John Wiley & Sons. learning research, 8(Jul), pp. 1623–1657.
[16] Hardy, R. L., 1971. “Multiquadric equations of topography and
other irregular surfaces”. Journal of geophysical research, 76(8),
pp. 1905–1915.
[17] Rosales-Pérez, A., Coello, C. A. C., Gonzalez, J. A., Reyes-
Garcia, C. A., and Escalante, H. J., 2013. “A hybrid surrogate-
based approach for evolutionary multi-objective optimization”. In

11 Copyright © 2018 United Technologies Corporation