Proceedings of the ASME 2015 International Design Engineering Technical Conferences &

Computers and Information in Engineering Conference

IDETC/CIE 2015
August 2-5, 2015, Boston, Massachusetts, USA

DETC2015-46832

ON USING ADAPTIVE SURROGATE MODELING IN DESIGN FOR EFFICIENT FLUID

POWER

Lakshmi Gururaja Rao, Jonathon Schuh, Randy H. Ewoldt, James T. Allison

University of Illinois at Urbana-Champaign, Urbana, IL, 61801
Email: {lrao2,schuh4,ewoldt,jtalliso}@illinois.edu

ABSTRACT 1 INTRODUCTION
In the last several decades fluid power has been used ex- Many engineering applications involve multiple design ob-
tensively in diverse industries such as agriculture, construction, jectives, typically centered on enhancing performance and reduc-
marine, offshore resource extraction, and even entertainment. ing cost. The trade-off between conflicting objectives can be
With a vast and ever-increasing spectrum of potential applica- investigated using Multi-Objective Optimization (MOO) tools.
tions, the design of efficient and leak-free components in fluid The solution of a MOO problem involves a set of solutions,
power systems has become essential. Previous experiments and known as the Pareto-optimal set (or Pareto set), that consists of
studies have shown that the use of microtextured surfaces in hy- non-dominated optimal points and quantifies design trade-offs in
draulic components achieves performance enhancement by re- the objective function space. Many strategies exist for addressing
ducing friction and leakage. This article aims to build on this MOO design problems [1, 2], some of which include condensing
recent work through a systematic optimization-based study of the multiple objectives into a single objective function by using a
performance improvement through microtexture surface design. weighted sum or a geometric mean. An alternative treats all but
These studies evaluate the potential of Newtonian fluid proper- one objective as a constraint. In either case multiple optimiza-
ties, coupled with varying surface features, to achieve design ob- tion solutions are required to build up the Pareto set, each time
jectives for efficiency. This early-stage design strategy aims to varying objective weights or constraint bounds. Another class
find optimal surface features that minimize apparent fluid viscos- of MOO solution methods are evolutionary algorithms such as
ity (low friction) and the area of the microtexture. The resulting the Multi-Objective Genetic Algorithm (MOGA) and the Non-
multi-objective optimization (MOO) problem involves a compu- dominated Sorting Genetic Algorithm (NSGA-II) [3, 4], both of
tationally intensive simulation of the system based on computa- which produce multiple Pareto-optimal solutions in a single op-
tional fluid dynamics (CFD). As a strategy to reduce overall com- timization execution.
putational expense, this paper describes the development of a Evaluating design performance of engineering systems as
new adaptive surrogate modeling strategy for multi-objective op- part of design optimization studies often involves computation-
timization. Two case studies are presented: a simple analytical ally intensive simulations. Despite advances in computing power
case study illustrating the details of the method and a more so- in the recent past, computationally intensive analysis methods
phisticated case study involving the two-dimensional CFD sim- (e.g., Computational Fluid Dynamics (CFD)) can be impractical
ulation of Newtonian fluids on symmetric surface textures. This to use with optimization directly [5]. Important fluid power sys-
design approach embraces the potential of using rheologically tem design objectives, such as friction across sliding contacts,
complex fluids in engineering system design and optimization. often require an expensive CFD simulation of the Navier-Stokes
This study can be further extended to a more generalized prob- equation for each new design candidate. One effective strat-
lem by coupling both fluid and geometrical design decisions. egy for reducing the total number of high-fidelity simulations

1 Copyright © 2015 by ASME

required to solve a design optimization problem is optimize us- is computationally inexpensive, allowing many function evalua-
ing an approximate surrogate model (SM)—sometimes called a tions required for identifying an optimal design. The resulting
meta-model—of the high-fidelity model [6,7]. Accurate solution design solution is only optimal with respect to the SM, so SM
requires acceptable SM accuracy in the neighborhood of the op- accuracy at the predicted optimum must be assessed by checking
timum. Adaptive surrogate modeling methods iteratively assess high-fidelity model outputs at this point. One of three strategies
SM accuracy, and evaluate the high-fidelity model at new sam- typically are used for ensuring accurate solution (Fig. 1).
ple points to improve SM accuracy if needed in strategic regions
of the design space [8, 9]. The cost of evaluating sample points
using the high-fidelity model must be accounted for when com-
paring SM to direct optimization.
SM has been investigated for specific application domains,
such as structural design [10] and injection molding [11]. While
most studies have focused on single-objective problems (e.g.,
Ref. [12]), extensions of SM to MOO have been made [11]. This
article presents advancements of adaptive SM for MOO that aim
to reduce the number of required simulations further. New fea-
tures of this method include enforcement of constraints during
high-fidelity model sampling. The algorithmic approach is elu-
cidated using a mathematical example, followed by a physical
engineering application that employs a CFD simulation of New-
tonian fluids on microtextures.
2 ADAPTIVE SM METHODOLOGY
In this section the adaptive surrogate modeling (ASM) FIGURE 1: Strategies for SM-based design optimization: (a) Se-
method is explained in detail, including steps for formulating quential, (b) Adaptive, (c) Direct Sampling [6]
the multi-objective optimization problem and implementing the
adaptive strategy for incrementally improving surrogate model
accuracy in strategic design space regions. The first (Fig. 1(a)) is a sequential approach with a single
2.1 MOO Problem Formulation validation check after SM construction but before optimization.
This requires error assessment across the entire modeling domain
Formulating an underlying engineering design problem as a
to ensure accurate solution, involving a large number of valida-
mathematical optimization problem involves identifying design
tion points that must be evaluated using the high-fidelity model.
objectives, variables, and constraints based on a firm understand-
Sequential sampling methods that treat the model development
ing of the physical system and design requirements and intent. A
and optimization as independent tasks. Efficiency can be im-
multi-objective optimization problem can be formulated as fol-
proved by adopting an iterative method where additional training
lows:
data is obtained via additional sampling to improve SM accuracy
if model validation fails (Fig. 1(b)). Wang et al. studied adap-
min f(x) = [ f1 (x), . . . , fm (x)]T ,
x=[x1 ,x2 ,...,xn ]T tive sampling as a means to improve a surrogate model prior to
(single objective) optimization [13]. In this case, both the opti-
subject to g(x) = [g1 (x), . . . , g p (x)]T ≤ 0 (1) mization and model validation points (a more global approach)
T
h(x) = [h1 (x), . . . , hq (x)] = 0 are used to form a new sample set. High-fidelity model evalua-
xLB ≤ x ≤ xUB tions can be reduced further by using a direct sampling approach
where new sample points are concentrated near the optimum pre-
dicted via SM optimization (Fig. 1(c)).
where f(x) is a vector-valued objective function, x is the vector
Additional challenges are introduced for MOO problems.
of design variables, g(·) and h(·) are constraint functions, and
Instead of needing to ensure SM accuracy in the neighborhood
xLB and xUB are design variable bounds.
of a single predicted optimal point, the SM must be accurate in
2.2 SM-Based Multi-Objective Optimization the neighborhood of the set of Pareto-optimal designs. Wilson
The next step in solving Prob. (1) using a surrogate model- et al. introduced an approach that ensured model validation and
ing strategy is to construct a model based on training data, i.e., accuracy prior to identifying the Pareto-set [14]. Li et al. approx-
a set of design points and the corresponding objective and con- imated Pareto sets using a hyper-ellipse [15]. Shan and Wang
straint function values based on high-fidelity simulation. The SM developed a new method—Pareto-Set Pursuing (PSP)—that ad-

2 Copyright © 2015 by ASME

dresses the challenges of solving MOO problems based on com- be the same as the modeling domain, as no information is avail-
putationally expensive models [16]. able yet to help focus sampling efforts near the predicted Pareto-
The Adaptive Surrogate Modeling method for Multi- optimal points. It is desirable to avoid using sample points that
Objective Optimization (ASM-MOO) methodology presented violate design constraints for two reasons: 1) Often infeasible
here is a direct sampling approach. The case study involves a points cannot be simulated [17], and 2) even if simulation is pos-
challenging fluid-flow system design problem, but the methodol- sible, simulation of infeasible points is wasted effort as SM ac-
ogy can be extended to a general class of design problems. We curacy is unimportant in infeasible regions. Here a set of pre-
introduce design constraint enforcement at the sampling stage, liminary samples are generated using one of several appropriate
and study the resulting efficiency of this modified sampling space-filling sampling techniques (such as Latin Hypercube) in
method. the preliminary sampling domain D̃s . This domain may include
3 ASM-MOO Algorithm Framework infeasible points. This set of sample points is then filtered to re-
move infeasible points, i.e., retained sample points belong to the
A more detailed illustration of the ASM-MOO methodology set Ds = D̃s ∈ F, where F is the feasible domain (i.e., the set of
is shown in Fig. 2. The modeling domain Dm is the region of points that satisfy design constraints). In the studies presented
the design space over which the SM is constructed. This may here design constraints are analytical, so SMs for design con-
be fixed, or may contract and shift to conserve modeling efforts. straints are not needed. Future work will address the case where
The superscript k is the iteration counter. The sampling domain design constraints depend on simulation outputs.
Ds is the design space region over which new sample points are T is then formed by evaluating each retained sample point
defined; design sample points and their corresponding simulation using the high-fidelity model, and organizing these data pairs.
outputs are used for the training dataset T . The SM is then constructed using T . Here the specific approxi-
mation Here radial basis functions are used to create the model
[18]. The MOO problem is then solved using an appropriate al-
Initialize k = 0 gorithm (MOGA here) and function evaluations based on the SM
Dmk = D̃ k
s Dmto obtain a set Modeling domain
of approximately Pareto-optimal points x∗ .
D̃s Model Initial Sampling domainby assessing SM error at
validation is performed
Generate Samples 2 D̃sk points in the Pareto set (analogous to assessing SM error at the
F Feasible domain
predicted optimum for single-objective problems). If the selected
Derror
s Feasible
metric does notSampling domain error tolerance, the pro-
satisfy a specified
Identify feasible sample domain F k
Find feasible samples Dsk = D̃sk 2 F k cess
T is repeatedTraining Dataset sample points near the current
by generating new
Pareto set, running the high-fidelity model at these new points,
x∗ Optimal design variable
Obtain T k using hi-fi model and adding these points to T . A new (more accurate) SM is
Dformed,
v Validationisdomain
and optimization repeated. Additional details for these
steps
D̃u are presented below.
Updated Sampling domain
Construct SM based on T k
High Fidelity
using RBF
Model D3.1 Sampling
u Updated Feasible Sampling domain
D̃uk = octree sample(Dvk ) The modeling domain D̃m is defined here by the design vari-
Duk = D̃uk 2 F k Using MOGA, obtain Pareto-set able bounds. Due to the SM type and the nature of the case
T k+1 = T k [ Duk [ Dvk and identify xk⇤ corresponding
Shrink/Shift Dmk to the Pareto-set studies, D̃m was not adjusted during the the ASM-MOO process.
k =k+1 Here D̃s = Dm , and a suitable design of experiments (DOE) tech-
nique is used to generate sample points that lie in D̃s . The appro-
Define Dvk = regression
P surf ace(x⇤ k )
Error SN E = ||f (xi ) f̂ (xi )|| priate choice of sampling technique depend on factors such as (i)
8xi 2Dvk
design space dimension, (ii) estimated level of response noise,
and (iii) SM type. For noisy data (e.g., probabilistic models),
DOEs that reduce the sensitivity to noise are desirable, such as
No?
Tolerance met? central composite designs, face-centered cubic designs, factorial
designs, and Box-Behnken designs [19]. For systems with less
Yes?
noise (such as deterministic models), Latin-Hypercube Sampling
End
(LHS), minimum bias designs, and Orthogonal Arrays (OAs) are
FIGURE 2: ASM-MOO algorithm preferred. Here LHS is used, which is an example of a space-
filling design that is often effective for computer experiments
without variance [20–22], including adaptive surrogate model-
During initialization the sampling domain Ds is defined to ing [23] and in conjunction with OAs [24].

3 Copyright © 2015 by ASME

3.2 Filtering Infeasible Sampling Points pared polynomial response and radial basis functions and found
Engineering design problems often involve inequality con- that RBFs are more effective at handling non-linearities in some
straints g(·) ≤ 0. The sampling procedure described above en- cases [36]. RBF models have been selected for use in this arti-
sures that only feasible design points are evaluated using the cle. Comparison with other model types is an important topic for
high-fidelity model, improving overall efficiency. In some cases future work.
this strategy offers another important advantage: sample points When using an RBF approximation for multi-objective
that violate constraints may not be physically meaningful, or for problems, an approximate function is defined for each of the m
other reasons would cause simulation failure. Removing infeasi- objective function values. As part of this definition, the interpo-
ble preemptively avoids these simulation problems that are other- lation condition for the jth objective function is formulated using
wise difficult to manage in an optimization implementation [17]. nt training points:
In this article we investigate how filtering out infeasible sample
nt
points influences solution efficiency, which is a new concept in
ASM.
f j (xi ) = ∑ wkj ψ(||xi − ck ||), i = 1, 2 . . . nt
(2)
k=1
While this article addresses only analytically defined linear j = 1, 2 . . . m
and non-linear constraints, we acknowledge the need to extend
this study to constraints that require the simulation of the high-
fidelity model. Filtering out infeasible points based on approxi- where ψ(.) is the radial basis function, wkj are unknown weight-
mated constraint functions is an important topic for future work. ing coefficients, xi is the ith training point, and ck is the kth ba-
It is expected that additional challenges will be encountered as sis function center. The RBF used here is the thin plate spline
there is uncertainty in whether sample points are actually infea- function [37]: ψ(r) = r2 ln r, where r is the Euclidean distance
sible during the ASM solution process. between the training point and function center: r = (||xi − ck ||).
During SM construction we seek to find the coefficients wij
3.3 Surrogate Model Construction
that solve ψ w j = f j , where ψi,k = ψ(||xi − ck ||), i, k = 1, 2 . . . nt
Once the set of feasible sample points is defined, these de- is the Gram matrix, and the vector of training outputs for the
sign points are evaluated using the high-fidelity model. This pro- jth objective function is f j = [ f j (x1 ), f j (x2 ), . . . , f j (xnt )]T [37].
duces input-output pairs that are added to the set of training data This equation has a unique solution since the Gram matrix is
T to be used for SM construction. A wide variety of approxi- square. To reduce problem complexity here we assume that
mation function types may be used for SM construction. Sev- the RBF centers coincide with training points, i.e., ci = xi .
eral general-purpose options are described below. Application- Once the coefficients wkj are identified for each objective func-
specific SMs may help improve solution efficiency further. tion the approximate objective functions can be defined: f̂(x) =
General purpose approximation functions for SMs help to [ fˆ1 (x), . . . , fˆm (x)]T . For a given point in the design space x, we
estimate simulation output at design points other than those in can compute the row vector ψ̂ ψ (x) = [ψ(||x − c1 ||), . . . , ψ(||x −
T . Well-known options include Kriging [22, 25], artificial neu- cnt ||)]. The surrogate model for the jth objective function is
ral networks (ANNs) [26], and radial basis functions [27, 28]. fˆj (x) = ψ̂ ψ (x)w j .
Other methods include multivariate adaptive regression splines
(MARS) [29], least interpolating polynomials [30], and induc- 3.4 Surrogate Model Optimization
tive learning [31]. Jin et al. compared several different surro- The multi-objective design optimization problem is then
gate models and concluded that the appropriate model choice solved by replacing the original nonlinear functions in Prob. (1)
depends on the degree of system nonlinearity [32]. Stander et with approximate functions obtained via surrogate modeling.
al. compared polynomial response surface approximation, Krig- This produces an approximate Pareto set that is only accepted
ing, and neural networks [33], and discovered that while neural as the solution if the validation step is passed. If all objective
networks and Kriging models require a larger number of initial and constraint functions are approximated using surrogates, the
points, their algorithmic efficiencies are comparable to a polyno- updated problem formulation is:
mial model. The general conclusion from the literature demon-
strates that there is no single best SM type that satisfies all prob- min f̂(x) = [ fˆ1 (x), . . . , fˆm (x)]T ,
lems. Haftka et al. address this issue by exploring the use of an x=[x1 ,x2 ,...,xn ]T
ensemble or a weighted average SM in place of individual surro- subject to ĝ(x) = [ĝ1 (x), . . . , ĝ p (x)]T ≤ 0 (3)
gates [34]. The best model choice depends on the nature of the
ĥ(x) = [ĥ1 (x), . . . , ĥq (x)]T = 0
problem and the sampled data set.
Radial basis functions have been used extensively as SMs xLB ≤ x ≤ xUB
for engineering design optimization [18, 35]. Fang et al. com-
This MOO problem was solved here using a Multi-Objective

4 Copyright © 2015 by ASME

Genetic Algorithm (MOGA), a specific class of evolutionary al- we developed an algorithm that generated random points on the
gorithms [38]. The primary advantage of evolutionary algo- hypersurface that were within a convex hull of a subset of the
rithms for MOO is that an approximate Pareto set can be gen- known Pareto-optimal points. This provided an efficient means
erated via a single MOGA solution. NSGA is another class of of generating validation points that were approximately in the
evolutionary algorithms for MOO [4, 39]. true Dv .
3.5 Surrogate Model Validation Once validation points are chosen, the high-fidelity model
is evaluated at these points for comparison to the SM to evalu-
Surrogate models used in design optimization need to be ate accuracy. Several options for SM validation exist, but stan-
accurate in the region of the optimum. Guaranteeing accuracy dard methods are intended primarily for cases where accuracy
within a tolerance at other points in the design points is wasteful is being assessed across Dm . Cross-validation is a well-known
of resources. Validation is straightforward for single objective model validation method, but is computationally cumbersome.
problems where SM accuracy must be verified near the single Meckesheimer suggested a leave-one out cross-validation strat-
predicted optimal point. This can be accomplished by evaluating egy, applicable to radial basis functions [41, 42]. The method
the high-fidelity model at the predicted optimum, and comparing used here is to evaluate the standard normal error (SNE) using
the result to the SM output at this point. SM validation is more points in Dv . For a set of nv validation points, the norm of the dif-
involved for MOO problems. We need to ensure that the SM is ference between the actual function value and the approximated
accurate for points in the design space that are Pareto-optimal. function value is defined as the SNE:
Instead of validating the SM at a single point, it must be vali-
dated across a bounded hypersurface. For this purpose we define nv
the validation domain Dv to be the set of points over which the SNE = ∑ ||(f(xi ) − f̂(xi ))|| (4)
SM must be validated. If the design space is in Rn , the hypersur- i=1
face that defines Dv is a manifold of dimension n − 1.
Validating over bounded hypersurface introduces two core where xi is the design vector value for the ith validation point.
challenges: 1) Instead of validating a single point in the design
space, validation must be performed across an infinite number of 3.6 Sampling Domain Update
points on a hypersurface, and 2) The boundaries of this hypersur- If the SNE ≥ εv , where εv is a specified validation tolerance,
face may be complicated (possibly non-convex or even discon- model accuracy must be improved in the neighborhood of the
nected), and we should focus validation efforts on points within Pareto set. This is done by generating new sample points that
the hypersurface bounds. The first challenge is addressed by us- are in D̃u , which is the updated sampling domain that includes
ing regression to fit a surface through Pareto-optimal points in a small neighborhood of the design space near Dv . The high-
the design space to define an approximate (unbounded) valida- fidelity model is evaluated at these new sample points, and these
tion domain. A uniform sampling technique is then used to select data pairs (along with the validation points from the previous set)
a finite set of validation points from Dv for validation. Generat- are appended to T for use in constructing the updated SM with
ing these new validation points in Dv is required because sim- improved accuracy near the estimated Pareto set.
ply using the Pareto-optimal points identified by the MOGA as Sample points need to be defined that lie within D̃u , which
validation points may not be adequate due to non-uniform distri- is a hypervolume. One approach would be to define the bound-
bution in the design space. The error between the SM and the ary of this hypervolume directly (using SVDD, level set meth-
high-fidelity model is evaluated at these validation points, and ods, etc.) and filter out sample points that are not within this
these error quantities are combined into a single error metric for boundary. A classification method such as a Support Vector Ma-
SM validation. chine (SVM) could be used to determine whether candidate sam-
In addressing the second challenge we recognize that not ple points are within D̃u . Alternatively, force-directed layout or
every point on an unbounded regression surface that is fit through field functions could be used to increase the density of sample
Pareto-optimal points is in the Pareto set. We must identify and points near Dv .
stay within hypersurface bounds to ensure that validation points The method used here is an Octree-based sampling method
are (approximately) in the Pareto set. [43] that generates uniformly distributed points in the neighbor-
Several options exist for defining the validation hypersur- hood of validation points as shown in Fig. 3. Sample point den-
face Dv , including simple regression, non-uniform rational basis sity can be adjusted as the ASM-MOO algorithm progresses.
splines (NURBS) [40], or T-splines. Quadratic regression per- Given a specified distance tolerance, points lying far from Dv
formed well for the small-dimension case studies used here. Sup- are eliminated to produce D̃u . In addition, only feasible sample
port Vector Data Description (SVDD) could be used to define a points are retained. Du = D̃u ∈ F defines the updated feasible
precise hypersurface boundary that includes all known Pareto- sampling domain that contains new sample points to be simu-
optimal points [17]. Rather than defining a precise boundary, lated.

5 Copyright © 2015 by ASME

 ∆k+1
i = β ∆ki , where ∆k+1
i is the range of xi at the kth ASM-MOO
iteration, and β < 1 is a fixed contraction parameter. One strat-
egy for re-centering the modeling domain would be to align its
center with the centroid of Dv . Rapid changes in Dm center may
cause numerical instabilities, which could be ameliorated with
numerical damping (e.g., the heavy ball method). Here Dm re-
mains fixed. Variable bounds are determined by experimentally
appropriate limits.
FIGURE 3: Octree-based sampling algorithm (two design vari- 3.8 Training Data Set Update
able example: x1 and x2 ). With an initial set of points to define
the domain to sample from as shown in (a), the octree-based al- Before construction of the updated SM, the training data
gorithm uniformly samples points within these bounds sparsely set is augmented with both the validation data and new sample
in (b) and more densely in (c) depending on sampling require- points: T k+1 = T k ∪ Dku ∪ Dkv , where k denotes the kth iteration of
ments. the ASM-MOO algorithm.
4 NUMERICAL RESULTS AND DISCUSSION
The ASM-MOO algorithm is demonstrated first using a sim-
Figure 4 illustrates conceptually the validation and sam-
ple analytical multi-objective constrained optimization problem,
ple point generation strategies described above using a two-
and then using a more sophisticated case study involving design
dimensional design space. Yellow markers are MOGA Pareto-
of efficient fluid power systems.
points. The solid black line is the validation domain Dv obtained
via regression. Dark blue markers are the uniformly distributed 4.1 Analytical Example
validation points. The red curve is the design constraint boundary Consider the following MOO problem:
(up and to the right in infeasible). Red markers are new (infeasi-
ble) Octree sample points. Magenta markers are feasible sample
points that are too far away from Dv to use. Cyan markers are the min f(x) = [ f1 (x), f2 (x)]T ,
x=[x1 ,x2 ]T
feasible sample points in Du .
where f1 (x) = (x1 − 2)2 + (x2 − 1)2
f2 (x) = (x1 )2 + (x2 − 6)2 (5)
subject to − 2.5x1 + x2 ≤ 1
Dv
Constraint (x2 − 2)2 ≤ 2
Du
Pareto Point x⇤
0.4 ≤ x1 ≤ 1.6, 2 ≤ x2 ≤ 5
Infeasible point 2
/F

Point in Dv
This problem was solved in two ways: 1) using the MOGA al-
Feasible point 2 Du
gorithm available in M ATLAB R
with direct objective function
x2 Feasible Point 2
/ Du
evaluation, and 2) using the ASM-MOO with the same MOGA
algorithm. The result of the first solution approach is the baseline
or ‘true’ solution. Both Pareto sets are shown in Fig. 5, and the
error values are presented in Table 1.
If we allow the same number of total original function eval-
uations for the ASM-MOO method, using sample constraints to
x1 eliminate infeasible sample points results in slightly improved
FIGURE 4: Conceptual illustration of validation point and sam- solution accuracy based on SNE. The total number of function
ple point generation evaluations was made approximately equal by increasing the
number of initial sample points for the case with sample con-
straints. Additional benefits of sampling constraints may be real-
ized with simulation-based problems (e.g., preventing simulation
3.7 Modeling Domain Update failure by avoiding infeasible designs).
The modeling domain Dm can either remain static or can be 4.2 Design for Efficient Fluid Power
contracted and re-centered to ease model construction. Dm con-
traction can be accompanied through a simple iterative formula: A hydraulic system design study motivated the development
of the ASM-MOO method introduced in this article. This study

6 Copyright © 2015 by ASME

 19 TABLE 1: Results from using sampling constraints on the analyt-
True ical example. In each case the methods used approximately the
ASM-MOO
18.5 same number of original function evaluations.

18
Method No. of iterations SNE
17.5
f2

ASM-MOO 8 8.26e-04
17 w/o sampling
constraints
16.5
ASM-MOO 6 6.20e-04
16 with sampling
constraints
15.5
1 1.5 2 2.5 3 3.5 4
f1
lubricants. For the experiments, a DHR-3 gap controlled rota-
FIGURE 5: Comparison between Pareto sets obtained via direct tional rheometer was used with the textured plates mounted to a
optimization and via ASM-MOO temperature controlled Peltier plate with Crystalbond, a thermo-
reversible adhesive. A schematic of the experimental set up is
2.035
given in Fig. 7, and the measured geometric properties of the
Pareto points textures tested are given in Fig. 8.
Validation Points
2.03 Updated Sampling Points
Validation domain
2.025

2.02
x2

2.015

2.01

2.005

2
0.4 0.6 0.8 1 1.2 1.4 1.6
x1

FIGURE 6: Pareto-values and the update domain in the design

space.

involves the investigation of different surface texture designs for

components used in fluid power systems where metal parts are in
sliding contact separated by hydraulic fluid. The competing de- FIGURE 7: Experimental setup. Design variables are defined in
sign objectives investigated here include reducing effective vis- Eqn. (6).
cosity (related to friction) and the cross-sectional area of the tex-
ture (related to cost).
Previous studies have shown that the use of surface tex- In their experiments, Schuh and Ewoldt found that symmet-
tures aids in decreasing friction in lubricated sliding contact [44]. ric textures decreased the viscous friction more than the asym-
Schuh and Ewoldt [45] have experimentally examined the effects metric textures, but the asymmetric textures produced larger nor-
of using symmetric and asymmetric textures to decrease viscous mal forces than the symmetric ones (important for sealing).
friction and increase normal force production with Newtonian Since each texture profile met both of the given design ob-

7 Copyright © 2015 by ASME

 the viscous force was converted to an apparent viscosity using:

Fv
ηa = (7)
W L Uh

where U is the velocity of the top plate, which for these simula-
tions was set as U = 1.425m/s.
FIGURE 8: Measured geometric properties
The design constraint x1 ≤ 0.95x3 ensures that texture width
is no greater than 95% of the length of the moving plate. The
jectives, but met them to differing degrees, this presents an inter- simulation would be infeasible (or produce insensible output) if
esting opportunity to study optimal tradeoffs for texture profile the constraint is violated. RBFs were used as the surrogate model
design. A computer experiment of the physical setup provides here. ASM-MOO was then applied to the design system to iden-
a basis for evaluating texture design candidates. The simulation tify the optimal texture features. In this case, the validation do-
of the fluid-texture system requires a finite volume based CFD main Dv is a hypersurface as we have a 3-dimensional design
simulation, which is a computationally-expensive process, par- space (Fig. 9).
ticularly when scaled to 3D. ASM-MOO was used to mitigate
this expense. The MOO problem for this case study is defined as
follows:

min f(x) = [ f1 (x), f2 (x)]T , (6a)

x=[x1 ,x2 ,x3 ]T

where f1 = ηa /η0 (6b)

f2 = x1 x2 (6c)
x1 = Gap width: W (mm) (6d)
x2 = Texture depth: D (mm) (6e)
x3 = Plate length: L (mm) (6f)
subject to x2 ≤ 0.95x4 , (6g)
0 ≤ x1 ≤ 6 (6i)
0 ≤ x2 ≤ 2 (6j)
0.8953 ≤ x3 ≤ 8.953 (6k)

FIGURE 9: The Pareto-points, validation points, and updated

sample points for the case study. The surface shown represents
Training data was obtained via FLUENT, a finite volume-based the validation domain, Dv .
CFD solver. Symmetric and asymmetric texture profiles were
simulated at steady state in FLUENT using a Newtonian lubri-
cant with a viscosity of 0.140 Pa s and a density of 866.8 kg/m3 . Results presented here are preliminary and limited to the
The simulations were performed using double precision and par- case of symmetric textures. The result is visualized as a Pareto
allel computing with two processes. The pressure was solved set in Fig. 10. The Pareto-set obtained is compared with the
using the SIMPLE algorithm and the momentum was solved us- CFD points (obtained while building the training data set). We
ing the second order upwind technique. A hybrid initialization can see that the Pareto-set is approaching the true Pareto-frontier
scheme was used, and the simulation was run until all the resid- of the simulated points in Fig. 10. It is insightful to see that
uals for the system were less than 1 × 10−5 . h was fixed at 0.25 ηa /η0 < 1 can be achieved with a very small texture area. The
mm. Several different values of W , D, and L were simulated extreme Pareto-point (corresponding to ηa /η0 > 1) is not phys-
in order to examine different texture profiles. The results from ically meaningful, but it is expected that the ASM-MOO algo-
FLUENT were given as a normal force and viscous force, and rithm will allow for its correction as the iterations proceed and
the surrogate model accuracy is improved. Table 2 reports the
Pareto-set values for both CFD data points and the ASM-MOO

8 Copyright © 2015 by ASME

algorithm. This tradeoff information is valuable for designers 5 CONCLUSIONS AND FUTURE WORK
who are interested in determining a texture design that best meets Engineering design optimization problems are often char-
overall system design requirements. acterized by multiple objectives and expensive simulations. As
a strategy for reducing computational expense, surrogate mod-
eling strategies have been developed and applied across a vari-
ety of engineering domains. In this article, one such strategy,
Adaptive Surrogate Model coupled with Multi-Objective Opti-
mization (ASM-MOO), was presented. A systematic formula-
tion of the multi-objective problem and a detailed framework for
the adaptive surrogate modeling was outlined. This algorithm
also incorporated the use of design constraints in the sampling
phase, which is an essential feature particularly in modeling sys-
tems that cannot be simulated in analytical example and is further
extended to a fluid power component problem.
Future investigations will involve nonlinear sample con-
straints and design constraints evaluated using a SM. Alterna-
tive strategies for generating feasible samples should be explored
(e.g., mapping a hypercube to F in the neighborhood of Dv ). The
FIGURE 10: The CFD simulation outputs (gray markers) are case study provides great insights into plausible mictrotexture
shown along with the Pareto-set that is identified by the ASM- geometries for different design needs. However, it is limited cur-
MOO algorithm (orange circles). By plotting all CFD simula- rently to symmetric textures; ongoing work is addressing the case
tion results we can qualitatively identify the location of the true of asymmetric textures as well. Aside from texture geometry,
Pareto set at the lower-left boundary of the attainable set. The we envision studying a larger optimization problem by including
Pareto-set shown here is the result of an initial set of six ASM- fluid properties such as viscosity, density etc., in the design space
MOO iterations, and does not represent the converged optimal (i.e., simultaneous texture and fluid design). Other improvements
solution. Nonetheless, it is clear that the ASM-MOO algorithm include extension to 3D problems and non-Newtonian hydraulic
is moving toward the true Pareto-set. fluids. The ASM-MOO method defined here is a local search
algorithm; samples are taken from the neighborhood of the ap-
proximate Pareto set. As an extension, we would like to expand
it to a global method by also sampling in regions of low informa-
tion to improve global search properties.
TABLE 2: The labeled points in Fig. 10 are shown with numeric
values of both the design variable and the objective functions. 6 ACKNOWLEDGEMENTS
Rows 2 and 3 are values obtained from the ASM-MOO algo- This work was supported by the Engineering Research Cen-
rithm. Rows 1 and 4 are the Pareto-values that one can iden- ter for Compact and Efficient Fluid Power (CCEFP), supported
tify by plotting the CFD data points(shown as grey markers in by the National Science Foundation under Grant no. EEC-
Fig. 10). 0540834. The co-authors thank Mr. Yong Hoon Lee for his as-
sistance with numerical studies, and Mr. Nathaniel Bristow and
Ms. Madeline Popelka for their assistance with the experimental
ηa /η0 Area (mm2 ) W (mm) D (mm) L (mm) setup and data collection.

0.518 10.719 5.797 1.849 6.260 REFERENCES

[1] Zadeh, L., 1963. “Optimality and Non-scalar-valued Per-
0.527 8.000 4.202 1.978 4.432
formance Criteria”. IEEE Transactions on Automatic Con-
0.775 0.578 3.453 0.090 3.772 trol, 8(1), Jan, pp. 59–60.
[2] Farquhar, P., 1980. “Advances in Multiattribute Utility The-
1.00 1.4e-5 0.0039 0.0036 3.312
ory”. Theory and Decision, 12(4), pp. 381–394.
[3] Kalyanmoy, D., 2001. Multi-Objective Optimization Using
Evolutionary Algorithms. John Wiley & Sons, Inc., New
York, NY, USA.
[4] Deb, K., Pratap, A., Agarwal, S., and Meyarivan, T.,
2002. “A Fast and Elitist Multiobjective Genetic Algo-

9 Copyright © 2015 by ASME

 rithm: NSGA-II”. IEEE Transactions on Evolutionary p. 101014.
Computation, 6(2), Apr., pp. 182–197. [18] Kitayama, S., Arakawa, M., and Yamazaki, K., 2011.
[5] Koch, P. N., Simpson, T. W., Allen, J. K., and Mistree, “Sequential Approximate Optimization Using Radial Ba-
F., 1999. “Statistical Approximations for Multidisciplinary sis Function Network for Engineering Optimization”. Op-
Design Optimization: the Problem of Size”. Journal of Air- timization and Engineering, 12(4), pp. 535–557.
craft, 36(1), pp. 275–286. [19] Ferreira, S., Bruns, R., Ferreira, H., Matos, G., David, J.,
[6] Wang, G. G., and Shan, S., 2007. “Review of Meta- Brando, G., da Silva, E., Portugal, L., dos Reis, P., Souza,
modeling Techniques in Support of Engineering Design A., and dos Santos, W., 2007. “Box-behnken Design: an
Optimization”. Journal of Mechanical Design, 129(4), Alternative for the Optimization of Analytical Methods”.
pp. 370–380. Analytica Chimica Acta, 597(2), pp. 179 – 186.
[7] Queipo, N. V., Haftka, R. T., Shyy, W., Goel, T., [20] Simpson, T. W., Allen, J. K., Colton, J. S., Heikes, R. G.,
Vaidyanathan, R., and Tucker, P. K., 2005. “Surrogate- Rosen, W., and Schrage, D. P., 1998. A Concept Explo-
based Analysis and Optimization”. Progress in Aerospace ration Method for Product Family Design. Tech. rep., in
Sciences, 41(1), pp. 1 – 28. Mechanical Engineering. Atlanta, GA: Georgia Institute of
[8] Liu, G., Han, X., and Jiang, C., 2008. “A Novel Multi- Technology.
objective Optimization Method Based on an Approxima- [21] Park, J.-S., 1994. “Optimal Latin-hypercube Designs for
tion Model Management Technique”. Computer Meth- Computer Experiments”. Journal of Statistical Planning
ods in Applied Mechanics and Engineering, 197(3340), and Inference, 39(1), pp. 95 – 111.
pp. 2719 – 2731. [22] Sacks, J., Welch, W. J., Mitchell, T. J., and Wynn, H. P.,
[9] Chen, G., Han, X., Liu, G., Jiang, C., and Zhao, Z., 1989. “Design and Analysis of Computer Experiments”.
2012. “An Efficient Multi-objective Optimization Method Statistical Science, 4(4), pp. 409–423.
for Black-box Functions Using Sequential Approximate [23] Wang, G. G., 2003. “Adaptive Response Surface Method
Technique”. Appl. Soft Comput., 12(1), Jan., pp. 14–27. Using Inherited Latin Hypercube Design Points”. Journal
[10] Barthelemy, J.-F., and Haftka, R., 1993. “Approxima- of Mechanical Design, 125(2), pp. 210–220.
tion Concepts for Optimum Structural Design - a Review”. [24] Tang, B., 1993. “Orthogonal Array-Based Latin Hyper-
Structural optimization, 5(3), pp. 129–144. cubes”. Journal of the American Statistical Association,
[11] Zhou, J., and Turng, L.-S., 2007. “Adaptive Multiobjective 88(424), pp. 1392–1397.
Optimization of Process Conditions for Injection Molding [25] Cressie, N., 1988. “Spatial Prediction and Ordinary Krig-
Using a Gaussian Process Approach”. Advances in Polymer ing”. Mathematical Geology, 20(4), pp. 405–421.
Technology, 26(2), pp. 71–85. [26] Suykens, J., Vandewalle, J., and de Moor, B., 1996. Arti-
[12] Booker, A., Dennis, J.E., J., Frank, P., Serafini, D., Torc- ficial Neural Networks for Modelling and Control of Non-
zon, V., and Trosset, M., 1999. “A Rigorous Framework Linear Systems. Springer.
for Optimization of Expensive Functions by Surrogates”. [27] Dyn, N., Levin, D., and Rippa, S., 1986. “Numerical Proce-
Structural Optimization, 17(1), pp. 1–13. dures for Surface Fitting of Scattered Data by Radial Func-
[13] Wang, C., Duan, Q., Gong, W., Ye, A., Di, Z., and Miao, tions”. SIAM Journal on Scientific and Statistical Comput-
C., 2014. “An Evaluation of Adaptive Surrogate Model- ing, 7(2), pp. 639–659.
ing Based Optimization with Two Benchmark Problems”. [28] Fang, H., and Horstemeyer, M. F., 2006. “Global Response
Environmental Modelling Software, 60(0), pp. 167 – 179. Approximation with Radial Basis Functions”. Engineering
[14] Wilson, B., Cappelleri, D., Simpson, T., and Frecker, M., Optimization, 38(4), pp. 407–424.
2001. “Efficient Pareto Frontier Exploration using Sur- [29] Friedman, J. H., 1991. “Multivariate Adaptive Regression
rogate Approximations”. Optimization and Engineering, Splines”. Ann. Statist, 19, pp. 1–67.
2(1), pp. 31–50. [30] de Boor, C., and Ron, A., 1990. “On Multivariate Poly-
[15] Li, Y., Fadel, G., and Wiecek, M., 1998. “Approximating nomial Interpolation”. Constructive Approximation, 6(3),
Pareto Curves Using the Hyper-Ellipse”. American Insti- pp. 287–302.
tute of Aeronautics and Astronautics. [31] Langley, P., and Simon, H. A., 1995. “Applications of
[16] Shan, S., and Wang, G. G., 2004. “An Efficient Pareto Set Machine Learning and Rule Induction”. Commun. ACM,
Identification Approach for Multiobjective Optimization on 38(11), Nov., pp. 54–64.
Black-Box Functions”. Journal of Mechanical Design. [32] Jin, R., Chen, W., and Simpson, T., 2001. “Comparative
[17] Alexander, M. J., Allison, J. T., Papalambros, P. Y., and Studies of Metamodelling Techniques under Multiple Mod-
Gorsich, D. J., 2011. “Constraint Management of Reduced elling Criteria”. Structural and Multidisciplinary Optimiza-
Representation Variables in Decomposition-based Design tion, 23(1), pp. 1–13.
Optimization”. Journal of Mechanical Design, 133(10), [33] Stander, N., Roux, W., Giger, M., Redhe, M., Fedorova,

10 Copyright © 2015 by ASME

 N., and J, H., 2004. “A Comparison of Meta-modeling
Techniques for Crashworthiness Optimization”. 10th
AIAA/ISSMO Multidisciplinary Analysis and Optimization
Conference, pp. AIAA–2004–4489.
[34] Goel, T., Haftka, R., Shyy, W., and Queipo, N., 2007. “En-
semble of Surrogates”. Structural and Multidisciplinary
Optimization, 33(3), pp. 199–216.
[35] Deshmukh, A. P., and Allison, J. T., 2013. “Design of
Nonlinear Dynamic Systems Using Surrogate Models of
Derivative Functions”. Proceedings of IDETC/CIE 2013,
51, p. 61801.
[36] Fang, H., Rais-Rohani, M., Liu, Z., and Horstemeyer, M. F.,
2005. “A Comparative Study of Metamodeling Methods
for Multiobjective Crashworthiness Optimization”. Com-
put. Struct., 83(25-26), Sept., pp. 2121–2136.
[37] Forrester, A., Sobester, A., and Keane, A., 2008. Engineer-
ing Design via Surrogate Modelling: a Practical Guide.
John Wiley & Sons.
[38] Fonseca, C. M., and Fleming, P. J., 1993. “Genetic Algo-
rithms for Multiobjective Optimization: Formulation, Dis-
cussion and Generalization”. In Proceedings of the 5th
International Conference on Genetic Algorithms, Morgan
Kaufmann Publishers Inc., pp. 416–423.
[39] Srinivas, N., and Deb, K., 1994. “Muiltiobjective Optimiza-
tion Using Nondominated Sorting in Genetic Algorithms”.
Evol. Comput., 2(3), Sept., pp. 221–248.
[40] Foley, J. D., van Dam, A., Feiner, S. K., and Hughes,
J. F., 1990. Computer Graphics: Principles and Practice.
Addison-Wesley Longman Publishing Co., Inc., Boston,
MA, USA.
[41] Meckesheimer, M., Booker, A. J., Barton, R. R., and
Simpson, T. W., 2002. “Computationally Inexpensive
Metamodel Assessment Strategies”. AIAA Journal, 40,
pp. 2053–2060.
[42] Meckesheimer, M., 2001. “A Framework For Metamodel-
Based Design: Subsystem Metamodel Assessment and Im-
plementation Issues”. PhD thesis, The Pennsylvania State
University, University Park, PA.
[43] Laboratory, R. P. I. I. P., and Meagher, D., 1980. Octree En-
coding: a New Technique for the Representation, Manipu-
lation and Display of Arbitrary 3-D Objects by Computer.
Electrical and Systems Engineering Department Rensseiaer
Polytechnic Institute Image Processing Laboratory.
[44] Johnston, M. T., King, W. P., and Ewoldt, R. H., 2015.
“Shear Stress Characteristics of Microtextured Surfaces in
Gap-controlled Hydrodynamic Lubrication”. Tribology In-
ternational, 82, Part A(0), pp. 123 – 132.
[45] J.K. Schuh, R. H. E. “Effects of Surface Texture Depth Pro-
file on Decreasing Friction in Hydrodynamic Lubricated
Sliding Friction”. In preparation.

11 Copyright © 2015 by ASME