Beruflich Dokumente
Kultur Dokumente
ABSTRACT
Target design load setting is one of the critical aspects in preliminary design phase of
aircraft design process. Overestimation or underestimation of design loads lead to the air-
craft’s performance and reliability implications, risk of costly re-desisgn, and specific design
lead time. Selection of reliable design loads under uncertainties require a large number of
analyses to evaluate loads for numerous flight conditions like maneuver loads, gust loads,
ground loads, time dependent response under control deflection and so on leading to huge
computational burden. To overcome this issue, Principal Component Analysis assisted
Surrogate Modeling (PCA-SM) is developed in the current work for rapid reconstruction
of correlated loads with relatively smaller number of analyses using Design of Experiments
(DoE). The advantages of PCA-SM method over other surrogate modeling techniques have
been discussed and demonstrated on a twin-aisle commercial transport aircraft case.
Nomenclature
λi ith eigenvalue of Σ
Lg Data matrix of gust load data
Lm Data matrix of static maneuver load data
Λ Diagonal matrix containing eigenvalues of Σ
φi ith eigenvectors of Σ
Φ Matrix containing all the eigenvectors of Σ
Φk Matrix containing k eigenvectors of Σ with largest k eigenvalues
F Flight condition
lm , lg Load vector for static maneuver and gust load respectively
U Uncertain variable
w Vector of weight or scores of x associated with eigenvectors stored in Φ
X Data matrix consisting of all obeservation of x
x Random vector
Σ Correlation Matrix of X
θ Polar angles
Cg Convex hull for gust loads
Cm Convex hull for static maneuver loads
Fx , Fy , Fz Shear forces along x, y and z axis
∗ PhD Candidate, School of Aerospace Engineering, Aerospace Systems Design Laboratory, Student Member.
† Boeing Regents Professor of Advanced Aerospace Systems Analysis, School of Aerospace Engineering, AIAA Fellow
‡ Loads technical skill leader for uncertainty management
1 of 15
I. Introduction
Note: This paper is intended to be the first part of another paper (Title: A Methodology to Create
Approximate Models of Load Envelopes Under Uncertainty) 1 presented in the same conference and session.
Both papers present a different methodology to model the equivalent load envelopes of airplane wings.
Downloaded by UNIVERSITY OF COLORADO on April 20, 2017 | http://arc.aiaa.org | DOI: 10.2514/6.2015-3092
Target design loads2 setting is one of the critical aspects of aircraft design and development process. The
target design loads are required not only for the the design and analysis of an aircraft’s basic structure but
also for the detailed design of its components. Therefore, evaluation and selection of design loads have direct
impact on aircraft’s structural design, weight, performance and other aspects of aircraft’s characteristics.
In general, target design load setting is not critical at conceptual design phase. For a conventional design
in conceptual design phase, the loads are not explicitly evaluated.3, 4 The variation in weights are directly
evaluated using historical data or empirical relationships as function of basic shape parameters. The selection
of target design loads in preliminary design phase is obtained by carrying out an iterative process between
load analysis and structural design.5 The design loads are evaluated under different flight conditions like
gust loads, time dependent responses due to control deflections, maneuver loads, ground loads and so on6 .7
The critical loads can be estimated by analysis of loads under a large number of flight conditions driven
by regulations CS-258 and FAR25.9 Generally, engineering estimates are used to select subsets of flight
conditions to carry out analysis for critical load evaluation. These design loads are then used to carry out
detailed structural design for better estimation of structural weight and elastic properties, which in turn is
used in the re-evaluation of design loads. If the newly obtained loads are higher than the loads selected
initially, then redesign or mitigation actions are required. On the other hand, if the newly obtained loads
are lower than the load selected initially, it can be used for structural modifications to reduce the aircraft’s
weight or to increase structural reserve factor.
In aircraft design process, the design loads are defined by the critical loads under various conditions
of dynamic and static maneuvers, mass distributions, flight conditions and so on. These are generally
expressed by internal loads, i.e. shears (Fx , Fy , Fz ) and moments (Mx , My , Mz ). For conventional aircraft
configurations, only a few of these internal loads are of interest, for example shear, bending moment and
torque at different span locations. These quantities or loads at different locations of aircraft are also referred
here as Interesting Quantities (IQs).
Load envelope approach is one of the methods used to identify design loads.5, 10–13 The design load
envelopes are made up of a number of correlated load cases such as Shear, Moment and Torque (SMT)
distributions along the wing for either a single steady balanced maneuver case or a specific time instant from
a dynamic case (e.g. gust case). The full range of load maneuvers are run for a range of mass distribution
cases, flight points, etc. Typically the critical load cases are identified in a two dimensional (2D) envelope
for each station using a Convex hull method (i.e. putting an elastic band around the cases). An example of
a 2D load envelope of bending moment and torque near the wing root for a commercial transport aircraft
under 797 maneuver loadsa is shown in Figure 1a. The loads associated with the red dots are considered
critical for that spanwise location. Similarly, for time-dependent cases like gust response, the load envelope
can be obtained by performing the convex hull algorithm on the time history of the loads. An example of
load envelope for gust response is shown in Figure 1b. The global load envelope is identified by evaluating
the convex hull of the loads for all the static and time-dependent scenarios.
During the preliminary design phase, some of the variables affecting the loads such as parameters as-
sociated with shape, aerodynamics, structures, etc. are uncertain. The selection of design loads without
considering uncertainty may lead to overdesign or underdesign of the structure which may cause significant
a The data is normalized to avoid the disclosure of sensitive data
2 of 15
Torque
Bending moment Bending moment
Downloaded by UNIVERSITY OF COLORADO on April 20, 2017 | http://arc.aiaa.org | DOI: 10.2514/6.2015-3092
(a) Example of the 2D load envelope based on (b) Example of the 2D load envelope for a sin-
maneuver loads for a commercial transport air- gle gust response for a commercial transport
craft near the wing root. In this example, a aircraft near the wing root.
total of 797 loading scenarios are considered.
The red dots are considered as critical loads
Figure 1: Example 2D load envelope on maneuver loads and gust loads for a commercial transport aircraft
increase in engineering cost and time.14, 15 Therefore, the consideration of the uncertainty in the definition
of the design loads is necessary.
Uncertainty analysis can be carried out by propagating uncertainties of design parameters to loads and
then to load envelopes using methods like Monte Carlo simulations, importance sampling method, local
expansion method, and so on. A large number of aero-elastic analyses are required under numerous flight
conditions, maneuver loads, ground loads, gust case, mass cases, etc. to propagate uncertainty which in turn
leads to a huge computational burden. A computationally efficient approach is to build surrogate models16
of each IQs as a function of uncertain parameters under each flight condition, maneuver load, ground load,
gust case, etc, by running relatively smaller number of analyses using Design of Experiments (DoE). These
surrogates can used to propagate uncertainty to loads and then to load envelopes.
Although, surrogate modeling of each IQs for various load cases is a computationally feasible option,
some critical issues need to be answered.
• Data management: Typically the number of IQs which are monitored can be in the order of hundreds.
For example, in wing, if the loads (SMT) are monitored at 50 different spanwise locations at 1, 000
different flight points, then number of surrogate models required are 150, 000. Typically IQs are also
monitored at other locations of aircraft like pylon, horizontal tail, vertical tail, fuselage etc. at early
preliminary design phase, requiring more number of surrogates. Building and handling of such a large
number of surrogate models can quickly become a data management issue. Although, automation is
possible, manual work might be required to check for the accuracy of the surrogate models.
• Accurate estimation of correlated loads: A case that forms one point of 2D envelope at a given spanwise
location will have associated and correlated values at other spanwise locations. For example, 2D
envelope at three different spanwise locations is shown in Figure 2a, 2b and 2c. The dots represented
by the same color (red, blue and magenta) are spanwise correlated loads for a certain flight conditions.
A correlated load generated by a particular case which is critical at one spanwise location may not be
critical at other locations. For example, the critical loads shown by the red dots at inner wing location
(Figure 2a) are not critical at the other two locations. Similarly, critical loads at mid and outer wings
(blue and magenta dots) may not be part of load envelope at other spanwise locations.
It is important to extract all the correlated loads which are present on the envelope of at least one
spanwise location. This step is critical because the structural and component analysis requires loading
of individual correlated load cases to correctly define the associated internal loads within the structure.
3 of 15
Mx
Mx
Fz Fz Fz
(a) 2D load envelope at inner wing (b) 2D load envelope at mid wing (c) 2D load envelope at outer wing
span span span
Downloaded by UNIVERSITY OF COLORADO on April 20, 2017 | http://arc.aiaa.org | DOI: 10.2514/6.2015-3092
Figure 2: 2D load envelope at three different wing span and correlated loads (dots with same color)
The drawback of using different surrogates of IQs at each spanwise location is that it leads to generation
of non-physical correlated loads due to error associated with surrogate models.
Mx
My
My
True
Reconst.
0 1 2 3 4 0 1 2 3 4 Mx
time (s) time (s)
(a) Time history of bending moment (b) True and reconstructed time his- (c) True and reconstructed 2D load
(Mx ) tory of torque (My ) envelope
Figure 3: Large error associated with 2D envelope reconstruction at certain spanwise location due to small
error in time history reconstruction of My only
• Time-dependent surrogates: Similar to static loads, it is important to accurately reconstruct the time
history of dynamic load cases. This may require non-linear or temopral-spatial surrogate modeling
techniques.17 Due to the complexity of these techniques, the relative error in reconstruction can be
high. A very small error in reconstructing time history leads to a large error in 2D envelope which is
shown in Figure 3.
One option to handle unsteady load is to build surrogate models of only trough and crest point of
time history.18 This may be appropriate to generate 1D load envelope but can be erroneous for 2D
envelopes as the vertex of 2D envelopes can be formed by points which are not trough and crest points
as shown in Figure 4.
One way to handle the above issues is to build surrogate models of load vectors (or vector of IQs) as
function of uncertain variables.
A set of n IQs for static load cases is given by n dimensional load vector lm . The load vector lm is a
function of p uncertain variable U and flight condition F , and is represented as:
4 of 15
My
My
Mx
Time history
Load Envelope vertices Load Envelope
0 1 2 3 4 0 1 2 3 4 Mx
time(s) time(s)
(a) Time history of bending moment (b) Time history of torque with the (c) Load envelope and the vertices
with the vertices of load envelope vertices of load envelope
Downloaded by UNIVERSITY OF COLORADO on April 20, 2017 | http://arc.aiaa.org | DOI: 10.2514/6.2015-3092
Figure 4: The vertices of load envelopes generated by points other than crest and trough
U,F )
lm = lm (U (1)
Similarly, a set of n IQs for dynamic load cases at a given instance of time t is given by n dimensional load
vector lg and has functional relationship with uncertain variable U and flight condition F and is represented
as:
U , F , t)
lg = lg (U (2)
Let Conv() be the convex hull function. Let q and r be the indices of the IQs in the load vector lm or lg
for which 2D envelope needs to be evaluated. For example lm [q] = Mx and lm [r] = Fz at mid span location
respectively. Then, a critical load envelope of two IQs with indices q and r in the load vector lm or lg for a
given instance of uncertain variable is given as
!
[
Cm (UU |q, r) = Conv hllm [q](F
F , U ), lm [r](F
F , U )i (3)
∀F
F
!!
[ [
U |q, r) = Conv
Cg (U hllg [q](F
F , U , t), lg [r](F
F , U , t)i (4)
∀F
F ∀t
where h·, ·i represents sample points in two dimensional Cartesian coordinate system, Cm is the load envelope
for static cases, and Cg is the load envelope of dynamic load cases.
In the current work a Principal Component Analysis assisted Surrogate Modeling (PCA-SM) has been
developed to build accurate models of correlated load vectors lm and lg as a function of uncertain variables
(and time for dynamic load cases). Firstly, Principal Component Analysis (PCA) is carried out to transform
correlated loads into uncorrelated principal components. Then, surrogate models of scores associated with
principal components are built. Since, the number of principal components are generally independent of the
number of IQs in the correlated load vector, the total number of surrogate models required is independent
of the number of IQs that are need to be tracked.
5 of 15
35
25
0.6
Mx
20
15
2 0.4
10
1.5 5
0 1 0.2
10 20 30 40
0.5 0.5
Mx Index along span
Downloaded by UNIVERSITY OF COLORADO on April 20, 2017 | http://arc.aiaa.org | DOI: 10.2514/6.2015-3092
0 time (s)
1
x/b
(b) Correlation among bending moment at different
(a) Bending moment as a function of span and time spanwise location, where color represents the magni-
tude of correlation
The current methodology (PCA-SM) takes advantage of this correlation to reduce the dimensions of
load vector by transforming the correlated loads into uncorrelated variables. This is done using Principal
Component Analysis (PCA),19 which is a method for orthogonal transformation of correlated variable into
a set of linearly uncorrelated variables called principal components. Generally, more than 99% variability of
correlated loads can be represented by only few principal components. Once the principal components are
selected, surrogate models of scores associated with principal components are built.
Σ = ΦΛΦT (7)
6 of 15
w =ΦTk x
x =(x1 , x2 , . . . , xp )T ∈ Rp
w =(w1 , w2 , . . . , wk )T ∈ Rk (8)
The vector w are the weights of x associated with the eigenvectors stored in Φk , and has been referred
to as scores in the current research. In geometrical sense, w is the projection of x in a lower dimensional
orthogonal space given by the eigenvectors stored in Φk . The lower dimensional (k ≤ p) vector w can
then be used to carry out further study. To extract the original variables from the latent space, an inverse
Downloaded by UNIVERSITY OF COLORADO on April 20, 2017 | http://arc.aiaa.org | DOI: 10.2514/6.2015-3092
A notional example is shown in Figure 6. The scatter plot of a correlated data is shown in Figure 6a.
The figure also shows the two eigenvectors or the principal components of the data. As observed, the first
principal component is the direction along which the maximum variance of the data is captured while the
second principal component is the orthogonal to the first one and captures the second largest variance.
Transforming the original data onto the space defined by principal components generate uncorrelated set
of data as shown in Figure 6b. Since, the variance along the second principal component is small, it can
be neglected. Reconstruction of the data in original space using only first principal component is shown in
Figure 6c.
2 2
3
1 2 1
1
w2
0 0
x2
x2
0
−1 −1
−1
−2 −2
−2
−2 −1 0 1 2 −2 −1 0 1 2 3 −2 −1 0 1 2
x1 w1 x1
(a) Scatter plot of data in original (b) Scatter data in latent (w-space) (c) Data reconstruction using only
space (x-space) and eigenvectors space after PCA scores of 1st principal component
7 of 15
• Number of principal components: A classical approach to select the number of principal component
(k) is given as
Pk
λ
Ppi=1 i ≥ α (12)
λ
j=1 j
where usually α ≥ 0.8. Higher value of α reduces the reconstruction error of PCA.
In this section a step by step process for building surrogate model of correlated load vectors using PCA
scores is discussed for both static and dynamic cases. The process is defined for a single flight condition and
needs to be repeated for all the other flight conditions.
Analysis is carried out for each set of uncertain variables. The loads for static cases are stored in a single
matrix Lm where ith row corresponds to load vector (llm T ) associated with ui . The loads associated with
dynamic scenarios are stacked in a single matrix Lg , where each row corresponds to correlated dynamic load
vector of lg T generated by a particular uncertain variable ui and time tj .
ŵkm = ŵkm (u
ui ) (15)
Various methods of surrogate model techniques such as Response Surface Method (RSM), Artificial
Neural Network (ANN), Support Vector Regression (SVR), etc can be used to build the surrogate models.
Once the surrogate models of principal component’s scores are built, the static load vector for any given
uncertainty ui can be estimated as
Km
X
ˆlm (u
ui ) = ŵkm (u φm
ui )φ (16)
k
k=1
8 of 15
wkg (u φgk )T lg (u
ui ) = (φ ui , tj ) (17)
Due to the complexity associated with above mentioned surrogate modeling techniques for temporal data,
a relativly simple approach has been used in the current work. Typically, the loads which are associated
with the critical load envelopes are observed in the first few time cycles of the time history. A subset of
discrete time location (tj ) are selected from the first few time period and principal component scores are
evaluated only for these discrete time. Separate surrogate models of the principal component scores are built
separately at each of these discrete time steps.
The load vectors at the any other time is interpolated using interpolation methods like polynomial, spline,
etc.
2. Centroid (given as (xc , yc )) of the all the vertices is evaluated and a polar coordinate system is consid-
ered with (xc , yc ) as origin. It is assumed that the uncertainty is small enough that (xc , yc ) lies inside
all the samples of load envelopes.
3. Polar angle (θ) is discretized between [0, 2π].
4. For each discretized value of (θ), the intersection points of ray with all the samples of envelopes are
evaluated and the cumulative distribution function (Fr (r|θ)) of radial coordinates (r) is estimated for
all the intersection points as shown in Figure 7.
5. For a given confidence α, radial coordinates with Fr (r(1−α)/2 )|θ) = (1 − α)/2 and Fr (r(1+α)/2 |θ) =
(1 + α)/2 are evaluated.
6. The α confidence interval envelopes is represented by polygon formed by all points given by r(1−α)/2
and r(1+α)/2 for all θ.
9 of 15
Intersection points of
radial vector with
polar angle
( , )
Downloaded by UNIVERSITY OF COLORADO on April 20, 2017 | http://arc.aiaa.org | DOI: 10.2514/6.2015-3092
III. Results
This section presents the preliminary results of PCA-SM approach to predict maneuver loads and time
history of gust loads as a function of uncertain variables of a twin-aisle commercial transport aircraft in
wing. The analysis for design of experiments are carried out using tools used for aircraft certification at
Airbus. The data is shown after normalization to remove the commercial sensitivity.
0.8
Normalized Load
0.6
0.4
0.2
−0.2
0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1
x/s x/s x/s
Principal Component Analysis is carried out on Lm . Contribution of each component in terms of variance
of load vector is shown in Figure 9a. For the current case, more than 99% variability are contributed by
10 of 15
Normalized RMSE
Cumulative %
0.015
Variance 100 57%
80 46%
0.01
60 34%
40 23%
0.005
20 11%
0 0%
1 2 3 4 5 6 0
0 0.2 0.4 0.6 0.8 1
Principal Components
x/s
(a) Contribution and variance captured by each
(b) Spanwise root mean squared error for nor-
Downloaded by UNIVERSITY OF COLORADO on April 20, 2017 | http://arc.aiaa.org | DOI: 10.2514/6.2015-3092
Figure 9: Variance captured by principal components and RMS error of reconstruction using PCA-SM for
maneuver loads
only first six principal components. Next, the surrogate models using ANN with two-layer feedforward
network with five to six sigmoid hidden neurons are built for each principal component’s score. The process
is repeated for all the flight conditions. A validation process is carried with 100 reserved data. Spanwise
root mean squared error for normalized data for all the flight condition is shown in Figure 9b.
Reconstruction of Mx distribution along span for three randomly selected set of uncertain variables at a
randomly selected flight condition is shown in Figure 10a. For the same set of random variables, loads are
evaluated for all 700 flight conditions. Reconstruction of 2D load envelope of Mx and My at wing mid span
location is shown in Figure 10b and is compared with true envelope.
True
PCA−SM
True
My
PCA−SM
Mx
Figure 10: Reconstruction of bending moment distribution and load envelopes for three randomly selected
set of uncertain variables
For uncertainty quantification Gaussian probability distribution is assumed on all the uncertain variables
with standard deviations equal to 10% of baseline values. Uncertainty propagation is carried out using
11 of 15
My
Fz
95% CI 95% CI 95% CI
Median Median Median
Downloaded by UNIVERSITY OF COLORADO on April 20, 2017 | http://arc.aiaa.org | DOI: 10.2514/6.2015-3092
Fz My Mx
(a) Confidence interval of 2D load en- (b) Confidence interval of 2D load en- (c) Confidence interval of 2D load en-
velope for Mx and Fz velope for Fz and My velope for My and Mx
Figure 11: Confidence interval of 2D load envelopes at mid span location for maneuver load
30 100%
Cumulative %
20 67%
Variance
Mx
My
10 33%
0 2 4 6 0 2 4 6 0 0%
1 2 3 4 5 6
time(s) time(s) Principal Components
(a) Time history of bending moment (b) Time history of torque at mid (c) Contribution and variance cap-
at mid span location for different set span location for different set of un- tured by each principal component for
of uncertain variables certain variables gust loads
Figure 12: Time history for bending moment and torque at midspan location from all the experiments and
variance of principal components
12 of 15
PC Score 2
PC Score 3
10
5 0 0
0
−5 −1
−5
−10 −10 −2
0 0.5 1 1.5 2 0 0.5 1 1.5 2 0 0.5 1 1.5 2
time(s) time(s) time(s)
(a) Time history of scores associated (b) Time history of scores associated (c) Time history of scores associated
with first principal component with second principal component with third principal component
Figure 13: Time history of scores associated with first three principal component
Downloaded by UNIVERSITY OF COLORADO on April 20, 2017 | http://arc.aiaa.org | DOI: 10.2514/6.2015-3092
which have been used for training the models. The root mean squared error of normalized torque along the
span and as a function of time is shown in Figure 14.
Figure 14: Root mean squared error for normalized torque as a function of span and time
The reconstruction of time history for inner wing Mx and inner wing My is shown in Figure 15a and 15b
for three randomly selected uncertain variable sets. The reconstruction of the 2D envelope for same set of
uncertain variables is shown in Figure 15c.
To demonstrate the uncertainty quantification of load envelope, Guassian distribution is assumed for
uncertainty variables with one standard deviation equal to 10% of baseline. Uncertainty propagation is
carried out using Monte Carlo simulation, to estimate the 95% confidence interval of load envelope for a
given gust length. The 95% confidence interval of load envelope at mid span location for the current case is
shown in Figure 16.
IV. Conclusion
This paper presents a Principal Component Analysis assisted Surrogate Modeling (PCA-SM) approach
to reconstruct the load vectors for static and dynamic scenarios to analyze load envelope under uncertainty.
The number of surrogate models required by PCA-SM to reconstruct the load vectors is independent of
number of IQs to be monitored. Therefore, the number of surrogate models required for all load scenarios
13 of 15
True
Inner Wing My
PCA−SM
True
PCA−SM
0 0.5 1 1.5 2 True
time (s) 0 0.5 1 1.5 2
PCA−SM
time (s)
Inner Wing Mx
Figure 15: Reconstruction of time history of Mx and My and their load envelopes for three randomly selected
uncertain variables
My
Fz
Fz My Mx
(a) 95% confidence interval of Mx and (b) 95% confidence interval of Fz and (c) 95% confidence interval of My and
Fz My Mx
Figure 16: 95% confidence interval of load envelope for a fixed gust length at mid span location
are manageable, making it computationally efficient for uncertainty analysis. Since PCA-SM generates entire
vector of correlated IQs for given set of uncertain variables, the reconstructed load vector is coherent with
the underlying physics. Using PCA-SM methods for dynamic cases, time history of only few principal
component’s scores are need to be modeled. This reduces the reconstruction error of time history of loads
and load envelopes for dynamic load cases. Results show good agreement with the reconstruction of static
loads, load history of dynamic scenarios and load envelopes. The PCA-SM enables computationally efficient
uncertainty quantification of loads and load envelopes. In the current work a simplified approach has been
used to carry out local uncertainty quantification of load envelope. In future, method for “global” confidence
level and reliability analysis will be explored. Future work will also focus on sensitivity analysis and resource
allocation study for uncertainty management of load envelopes.
References
1 Rancourt, D., Ghosh, S., Mavris, D. N., and Coggon, S., “A Methodology to Create Approximate Models of Load
Envelopes Under Uncertainty,” 16th AIAA Multidisciplinary Analysis and Optimization Conference, 2015.
2 Coggon, S., “Uncertainty Management Application to Target Design Load Setting, An AIRBUS Industrial Use Case,”
14 of 15
American Institute of Aeronautics and Astronautics, Reston, VA, 2nd ed., 2002.
5 Stauffer, W. A., Lewolt, J. G., and Hoblit, F. M., “Application of advanced methods to design loads determination for
the L-1011 transport,” Journal of Aircraft, Vol. 10, No. 8, 1973, pp. 449–458.
6 Howe, D., Aircraft loading and structural layout, Professional Engineering Publishing, 2004.
7 Luber, W. and Fullhas, K., “Design Loads for Future Fighter Aircraft,” IMAC-XXI: Conference & Exposition on Struc-
September 2014.
10 Appa, K., “Recent advances in maneuver loads analysis,” Computer methods in applied mechanics and engineering,
Post-Processing,” Proceedings of the 28th International Congress of the Aeronautical Sciences, 2012.
12 Werner-Westphal, C., Heinze, W., and Horst, P., “Multidisciplinary integrated preliminary design applied to unconven-
tional aircraft configurations,” Journal of aircraft, Vol. 45, No. 2, 2008, pp. 581–590.
Downloaded by UNIVERSITY OF COLORADO on April 20, 2017 | http://arc.aiaa.org | DOI: 10.2514/6.2015-3092
13 Lee, I. S., “System, method and computer-readable storage medium for defining a load envelope that envelopes a plurality
2014.
16 Forrester, A. I. J., Engineering design via surrogate modelling: a practical guide, J. Wiley, 2008.
17 Kyriakidis, P. and Journel, A., “Geostatistical SpaceTime Models: A Review,” Mathematical Geology, Vol. 31, No. 6,
Theory,” Journal of Sound and vibration, Vol. 252, No. 3, 2002, pp. 527–544.
22 Strang, G., Linear Algebra and Its Applications, Brooks Cole, 4th ed., July 2005.
23 Vershynin, R., Compressed Sensing, Theory and Applications, chap. 5, Cambridge Univ. Press, Cambridge, England,
U.K, 2012.
15 of 15