Experimental and numerical optimizations of an upwind

mainsail trimming
Matthieu Sacher, Frédéric Hauville, Régis Duvigneau, Olivier Le Maître,
Nicolas Aubin

To cite this version:

Matthieu Sacher, Frédéric Hauville, Régis Duvigneau, Olivier Le Maître, Nicolas Aubin. Experimental
and numerical optimizations of an upwind mainsail trimming. THE 22nd CHESAPEAKE SAILING
YACHT SYMPOSIUM, Mar 2016, Annapolis, United States. �hal-01387783�

HAL Id: hal-01387783

https://inria.hal.science/hal-01387783
Submitted on 26 Oct 2016

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 THE 22nd CHESAPEAKE SAILING YACHT SYMPOSIUM
ANNAPOLIS, MARYLAND, MARCH 2016

Experimental and numerical optimizations of an upwind mainsail trimming

Matthieu Sacher1 , Naval Academy Research Institute – IRENav, France
Frédéric Hauville, Naval Academy Research Institute – IRENav, France
Régis Duvigneau, INRIA – Sophia Antipolis, France
Olivier Le Maı̂tre, LIMSI – CNRS, France
Nicolas Aubin, Naval Academy Research Institute – IRENav, France
Mathieu Durand, K-EPSILON, France
(BS) or Velocity Made Good (VMG). The different loads
accounted by the VPPs can be based on empirical formu-
ABSTRACT las, experimental data or numerical simulations (Hansen
et al., 2003, Korpus, 2007). However, due to the complex-
This paper investigates the use of meta-models for optimiz- ity and multi-physic characters of the yachts dynamics, per-
ing sails trimming. A Gaussian process is used to robustly formance studies often consider the hydrodynamic (Huetz
approximate the dependence of the performance with the and Guillerm, 2014) and aerodynamic (Augier et al., 2012,
trimming parameters to be optimized. The Gaussian pro- Menotti et al., 2013, Trimarchi, 2012) aspects separately.
cess construction uses a limited number of performance ob- Here, we focus on the aerodynamics, optimizing the per-
servations at carefully selected trimming points, potentially formance of a sail system, but the numerical procedure de-
enabling the optimization of complex sail systems with mul- veloped below can be used to perform hydrodynamic opti-
tiple trimming parameters. We test the optimization proce- mization or even fully coupled yacht performance optimiza-
dure on the (two parameters) trimming of a scaled IMOCA tion. Sail systems are subjected to very complex phenom-
mainsail in upwind conditions. To assess the robustness of ena, such as nonlinear Fluid-Structure Interaction (FSI) ef-
the Gaussian process approach, in particular its sensitivity to fects and instabilities. Moreover, the modeling of real sail-
error and noise in the performance estimation, we contrast ing conditions is still an open research problem due to the
the direct optimization of the physical system with the op- large uncertainties in wind and sea states. To our knowledge,
timization of its numerical model. For the physical system, the optimization of sails has thus been limited so far to ideal-
the optimization procedure was fed with wind tunnel mea- ized situations. For instance, sail shape optimizations (with-
surements, while the numerical modeling relied on a fully out accounting for the FSI problem) are performed in (Rous-
non-linear Fluid-Structure Interaction solver. The results selon, 2008), while the trimming of two-dimensional sails
show a correct agreement of the optimized trimming param- is numerically considered in (Chapin et al., 2008). Regard-
eters for the physical and numerical models, despite the in- ing FSI in three spatial dimensions, the authors in (Ranzen-
herent errors in the numerical model and the measurement bach et al., 2013) mention an optimization of the trimming
uncertainties. In addition, the number of performance esti- of sails, but within an inviscid flow approximation and not
mations was found to be affordable and comparable in the much details are provided on the optimization procedure
two cases, demonstrating the effectiveness of the approach. used.
The present work aims at pursuing these efforts toward
the development of efficient numerical optimization proce-
INTRODUCTION
dures capable of dealing with complex sail systems, with
realistic physical models (e.g. nonlinear FSI and turbulent
Researches on sailing yachts have fostered the development
flows) and a large number of optimization variables (i.e.
of advanced methods dedicated to the prediction and im-
trimming parameters). Denoting x 2 ⌦ the optimization
provement of racing yacht performance. The performance
variables, the optimization problem can be written as
is usually analyzed using so-called Velocity Prediction Pro-
grams (VPPs) (Oossanen, 1993), which solve equilibrium xopt = arg min P(x),
equations (balancing hull, appendages and sails loads) to de- x2⌦
termine several performance indicators, such as Boat Speed
where xopt 2 ⌦ are the sought optimal parameters and
1 matthieu.sacher@ecole-navale.fr P : ⌦ 7! R is a measure of performance. The main dif-
ficulty preventing the straightforward application of stan- timization, this experiment is used as a reference to assess
dard optimization procedures to sail systems is related to the relevance of an optimization relying on numerical res-
the cost of estimating of the performance P at tentative val- olutions of the FSI problem to compute the performance.
ues x of the parameters. Indeed, the estimation of P(x) in- To this end, the experimental sail system was measured
volves the resolution of a nonlinear FSI problem, which re- (dimensions, mechanical characteristics of mast and boom,
quires several convergence iterations between the nonlinear . . . ), and wind tunnel inflow conditions recorded, to create
elastic and flow solvers. Also, adjoint based techniques are a numerical model of the experiment. State of the art FSI
hardly amenable to FSI problems involving coupled nonlin- solvers is then used for the resolution of the resulting numer-
ear solvers; this fact precludes the use of efficient gradient- ical model at the sequence of trimming points requested by
based iterative methods in favor of optimization algorithms the GP-based optimizer. The numerical resolution involves
such as the simplex based (Nelder and Mead, 1965) or evo- a nonlinear structural solver with a mesh deformation util-
lutionary (Bäck and Schwefel, 1993, Hansen, 2006) meth- ity (K-FSI tools) developed by K-EPSILON, coupled with
ods. However, depending on the considered problem, these the finite volume turbulent flow solver FINETM /Marine
so called gradient-free algorithms are known to require a from Numeca Software. The Unsteady Reynolds-Average
large number of evaluations of P(x), making applications to Navier-Stokes Equations (URANS) turbulence model is
sail systems very costly as a single evaluation may routinely used in these numerical experiments.
require several hours of CPU on modern parallel computers. The paper is organized as follows. The first Section
Based on these observations, we advocate the use of briefly reviews the construction of the Gaussian Process ap-
meta-model approaches to mitigate the large computational proximation and the selection of the new parameters in the
cost of optimizing the trimming parameters of sail systems. iterative optimization procedure. We then detail the experi-
Specifically we rely below on Gaussian Process (GP) ap- mental set-up and the results of the corresponding optimiza-
proximations for the mapping P : ⌦ 7! R. This statisti- tion procedure in the second Section. The numerical model-
cal approach uses a coarse set of performance evaluations at ing of the experiment and optimization results are reported
some parameters values x 2 ⌦ to infer a GP G(x) ⇡ P(x). in third Section. Finally, conclusions of this work and di-
One can then apply its favorite optimization procedure to rection for future developments are provided in the fourth
G(x) to obtain the corresponding approximation of xopt . Section.
The surrogate-based optimization procedure is embedded in
an iterative scheme, where new evaluations of the perfor-
GP MODEL BASED OPTIMIZATION
mance at carefully selected new points x are introduced in
order to refine the GP approximation in regions of ⌦ suscep- In this Section we start by briefly summarizing the construc-
tible to include the optimum. The GP approach is then ex- tion of a Gaussian Process to model a function from noisy
pected to improve the optimization by a) requiring an over- measurements. More details on GP models can be found for
all lower number of performance evaluations, compared to instance in (Gibbs, 1997, Rasmussen and Williams, 2006).
direct gradient-free approaches, and b) enabling the use of We then describe the GP model optimization procedure (Du-
efficient global optimization tools. In addition, the GP con- vigneau and Chandrashekar, 2012), detailing the selection
struction provides a natural way to estimate convergence on of successive optimal candidates.
the approximation of P and then to characterize the accu-
racy on the retrieved optimum.
Another interest in considering an optimization based on Gaussian Process Model
GP meta-model is that it naturally accommodates for errors T
Consider a dataset Xn = (x1 · · · xn ) of n training inputs
and noise in the performance evaluation. This specificity vectors xi 2 ⌦ ⇢ Rd . Each element xi 2 Xn is associ-
is exploited in the present work to perform the optimiza- ated to an observation (or measurement) yi 2 R which is
tion of an actual physical sail system, consisting in a scaled assumed to be dependent on a latent function f (x) through
IMOCA mainsail in upwind conditions. The objective is
to find the optimal trimming of the sail, for a performance yi = f (xi ) + "i , i = 1, . . . , n, (1)
criterion combining the drive and side aerodynamic force
coefficients. Here, the GP-based optimizer used values of where "i is a random measurement error (i.e. the measure-
P(x) measured in the wind tunnel of the Yacht Research ment noise). In this work, the "i are assumed independent
Unit (Auckland), for the sequence of trimming points re- and to follow the same (centered) Gaussian distribution:
quested by the iterative optimization procedure. Because
of the imperfections in the experimental apparatus and in- "i ⇠ N 0, ✏
2
, (2)
herent noise in the measurements, the estimates of P(x)
were subjected to significant errors, that would have com- where N µ, ⌃2 denotes the Gaussian distribution with
promised the convergence of descent methods (Saul’ev and mean µ and variance ⌃2 . Thus, ✏ 2 is referred to as noise
Samoilova, 1975) without using a GP reconstruction. variance. The objective is therefore to model the latent func-
In addition to evidence the robustness of GP based op- tion f (x) on the basis on the noisy observations yi .
 The latent function is considered as a realization of a zero- The best prediction of y(x) is the mean ŷ(x) of the distri-
mean multivariate Gaussian process F , with unknown co- bution; the prediction variance ˆy2 (x) quantifies the uncer-
variance function CF , that is F ⇠ N (0, CF ), with tainty in the prediction. The second order properties of the
. prediction y(x) can be explicitly expressed as
CF (x, x0 ) = E {F (x), F (x0 )}, (3)
1
where E {·} denotes the expectation operator. ŷ(x) = kT (x) C(⇥) + ✏
2
I Yn , (10)
The covariance function of F must be specified. In this ˆy2 (x) = (x) + 2 T
k (x) C(⇥) + 2
I
1
k(x).
✏ ✏
work, we consider the Matérn class (Stein, 2012) of station-
(11)
ary covariance functions having for one-dimensional gener-
ator
!⌫ ! The hyper-parameters ⇥ and noise variance ✏ 2 are un-
p p
21 ⌫ 2⌫r 2⌫r known a priori and need to be learned from the data. They
M⌫ (r, l) = K⌫ . (4) can be determined by maximizing the log-marginal likeli-
(⌫) l l
hood (Rasmussen and Williams, 2006) given by
Here r = |x x0 |, ⌫ and l are two positive parameters, and
K⌫ is the modified Bessel function of the second kind. We 2 n 1 2
L(⇥, ✏ )= log (2⇡) log C(⇥) + ✏ I
shall further restrict ourselves to covariances with ⌫ ! 1, 2 2
leading to the squared exponential covariance family with 1 T 1
Y C(⇥) + ✏ 2 I Yn . (12)
generator 2 n
✓ 2◆ The optimal hyper-parameters and noise variance are then
r
M1 (r, l) = exp . (5) found by minimizing L with respect to its arguments. An
2l2
evolution strategy algorithm (Hansen, 2006) is used for this
The multidimensional counterpart is obtained by tensor purpose.
product of the one-dimensional generator. The final expres- Once the optimal hyper-parameters are determined, the
sion of the covariance function for the GP approximation GP model can be used to predict values at new points us-
is ing (10) and (11). The most computationally consuming
Yd ✓ ◆ part of the GP construction is the assembly of the (full) ma-
(xi x0i )2
c(x, x0 ; ⇥) = ✓1 exp + ✓2 . (6) trix C + ✏ 2 I , and the evaluation of its determinant and
i=1
2li2 inverse required in the definition of the log-marginal likeli-
hood and for new predictions. This can be done efficiently
In the expression (6) of c, ⇥ = {✓1 , ✓2 , l1 , l2 , . . . , ld } is
by LU decompositions.
a vector of hyper-parameters. The first hyper-parameter, ✓1 ,
Figure 1 illustrates for a one-dimensional function the
scales the distance-dependent correlation, while ✓2 is an off-
effect of the observation noise ✏ on the constructed GP
set from zero. The other parameters li are the anisotropic
model. The constructions use 6 observations points de-
correlation lengths associated to the d directions of ⌦. From
picted with circles in the plots and the covariance hyper-
the parametrized covariance function c(x, x0 ; ⇥) we derive
parameters are determined by maximizing the log-marginal
the covariance matrix C(⇥) 2 Rn⇥n of the observation
likelihood. However, in the first case, shown in Figure 1(a)
points in Xn . The covariance matrix C(⇥) has for entries
a value ✏ = 0 is imposed, while in the second case in Fig-
.
Ci,j (⇥) = c(xi , xj ; ⇥), 1  i, j,  n. (7) ure 1(b) the noise level is also optimized. In addition to the
observations, the two plots report the mean of the GP mod-
Given the n noisy observations yi , collected into the vec- els with classical ±3ˆ uncertainty range. It is seen from
T
tor Yn = (y1 · · · yn ) , the predicted observation y(x) at a Figure 1(a) that when the measurements are assumed to be
new point x 2 ⌦ is given by the joint Gaussian distribution noise-free ( ✏ = 0), the resulting GP model is interpolating
✓ ◆ ✓  ◆
Yn C + ✏2I k(x) the data, i.e. the variance of the prediction is zero at the data
Xn , ⇥ ⇠ N 0, . points. However, the mean of the GP model exhibits signifi-
y(x) kT (x) (x) + ✏ 2
(8) cant oscillations such that over-fitting can be suspected. On
the contrary, optimizing the noise level ✏ results in a mean
In (8), the dependence of C on the hyper-parameters has process free of spurious oscillations but that is no more in-
been removed to simplify the notation, and terpolating, as it can be appreciated from Figure 1(b). The
. . T averaged distance of the best prediction to the observations
(x) = c(x, x; ⇥), k = (c(x, x1 ; ⇥) · · · c(x, xn ; ⇥)) ,
is ⇠ ✏ .
while I is the identity of Rn . Using the conditional rules of a
joint Gaussian distribution (Rasmussen and Williams, 2006, Optimization strategy using Gaussian Process models
Von Mises, 1964), it comes
2
The GP model (build, previously) can be used to determine
y(x)| Yn , Xn , ⇥, ✏ ⇠ N ŷ(x), ˆy2 (x) . (9) the next control parameters x⇤ to be included in the data
 nalizing areas where the variance ˆy2 is small. It writes as
1.0
0 1
Observations
Meta-model
0.5
✏
AEI(x) = EI(x) @1 A, (13)
+ 3σ
- 3σ
q
0.0
ˆy2 (x) + ✏
2

-0.5

-1.0
with the Expected Improvement defined by
-1.5 EI(x) = ˆy2 (x) [u(x) (u(x)) + (u(x))] ,
-2.0 ŷ(x⇤⇤ ) ŷ(x)
u(x) = , (14)
-2.5 ˆy2 (x)
-3.0
-1.50 -1.25 -1.00 -0.75
x
-0.50 -0.25 0.00
where and = 0 denote respectively the cumulative dis-
tributions (Erf-function) and density of the standard Gaus-
(a) Assuming ✏ = 0. sian distribution, and x⇤⇤ is the effective best solution:
.
x⇤⇤ = arg min [ŷ(x) + ˆ (x)] . (15)
1.0 x2⌦
Observations
0.5 Meta-model
+ 3σ
- 3σ
When the optimum x⇤ of the AEI is determined, the cor-
0.0 responding performance P(x⇤ ) is evaluated and is included
in the database. A new GP model can then be reconstructed
-0.5
using the extended data base, leading to a new maximizer
-1.0 x⇤ of the updated AEI, and so on. The iterations carry on
-1.5 until some convergence criterion is satisfied or the resources
allocated to the optimization procedure have been exhausted
-2.0
(e.g. reaching a prescribed number of performance evalua-
-2.5
tions). Classical convergence criteria compare the distance
-3.0
-1.50 -1.25 -1.00 -0.75 -0.50 -0.25 0.00
between two successive iterates, in terms of optimal param-
x eters x⇤ or/and performance prediction P(x⇤ ). Overall, we
remark that each iteration requires the resolution of two op-
(b) Optimal ✏ = 0.143.
timization problems (one for the covariance parameters and
one for the AEI) and one evaluation of the performance. In
Figure 1: Effect of ✏ on the GP model.
practice, all results reported in this work were obtained us-
ing the nonlinear non-convex black-box optimization library
based on the Covariance Matrix Adaptation Evolution Strat-
egy (Hansen, 2006, Hansen and Ostermeier, 2001).
base. Deterministic optimization approaches would classi-
cally choose x⇤ as the best control parameters, that is the EXPERIMENTAL OPTIMIZATION
minimizer of the meta-model over ⌦. However, GP mod-
els are random and not even bounded, so the definition of This Section concerns the sail trimming optimization per-
the best control parameters needs to be clarified in this con- formed in the wind tunnel of the Yacht Research Unit
text. This is classically achieved by introducing an appro- (YRU) (Flay, 1996) at the University of Auckland.
priate (deterministic) merit function, and combining the ex-
pected prediction ŷ and its variance ˆy2 , requiring x⇤ to be
Experimental setup
the maximizer of this merit function. In fact, the merit func-
tion should balance a selection of x⇤ yielding the minimal The sail model is inspired by an IMOCA 60-foot design
expected prediction ŷ (optimality) with a selection of x⇤ in mainsail at 1:13 scale. It was designed and produced by
areas of large variance ˆy2 to reduce the GP model uncer- the sail-makers of INCIDENCES SAILS company. The sail
tainty. A complete summary of various merit functions pro- has a surface area of 1 m2 , for an height of h = 2 m, and
posed in the literature is provided in (Jones, 2001). In this is supported by a rig consisting of a flexible circular section
work, we use the Augmented Expected Improvement (AEI) carbon mast (constant diameter 14 mm), clamped at its base
merit function (Huang et al., 2006) which is an extension of and without spreader, backstay or forestay. The sail and rig
the popular Expected Improvement (EI) (Jones et al., 1998) are set in the open jet test section of the YRU wind tunnel,
in the case of noisy estimations. The AEI merit function see Figure 2(a). The test section is 7.2 wide and 3.5 m high.
AEI(x) estimates the expected increase in the performance, Three stepper motors and a control card control remotely
taking into account the noise in the observed values and pe- the main sheet length (Lsheet ) and main car position (Lcar )
as shown in Figure 2(b). In the following, Lsheet and Lcar responding Reynolds number, based on the reference chord
are the only trimming parameters to be optimized. Note length c = S/h = 0.5 m, is Re = 1.2 ⇥ 105 . A multi-hole
that the remote system allows for changing these trimming pressure probe (Cobra Probe) was used to measure profiles
parameters without switching off the wind tunnel flow and of the flow velocity at several locations inside the wind tun-
making a new tare of the measurement instruments. A pre- nel. These measurements were repeated with and without
cision of ±2 mm on the imposed trimming parameters was the sail model in the test section to verify its effect on the
estimated through repeated measurements. flow field. Typical profiles are shown in Figure 3; it can be
seen that the inflow has no twist.

(a) Mainsail in the YRU wind tunnel.

Figure 3: Velocity profiles with sail model.

The optimization problem is finally defined as the max-

imization with respect to x = (Lsheet , Lcar ) of the perfor-
mance P(x) taken as a composite function of the thrust CX
and side CY aerodynamic coefficients:

P(x) = CX + 0.1CY . (16)

(b) Close view on the trimming system. The aerodynamic coefficients are deduced from the aerody-
namic forces through normalization by the reference force
Figure 2: Experimental setup. being q1 S, where q1 is the reference dynamic pressure
measured in the wind tunnel (precision ±1 Pa). In (16), the
A six-components force balance, located under the floor coefficient 0.1 penalizes the side force to account for the re-
of the wind tunnel, was used to measure the aerodynamic sulting hydrodynamic drag and leeway that would be detri-
forces. The X-direction corresponds to the model longitu- mental to the performance. The optimization of the trim-
dinal forward direction (i.e. thrust force direction), while ming parameters then follows the procedure illustrated in
the Y -direction is defined as the positive port-side (i.e. side Figure 4. The primary loop, “Sampling loop”, generates an
force direction) and the Z-direction is the vertical. Af- initial Latin Hypercube Sample (LHS) set of 10 trimming
ter careful calibration and testing, the balance precisions in parameters (McKay et al., 2000). For each initial sample,
the X, Y and Z directions were estimated to be ±0.09 N, the experimental model is remotely set to the corresponding
±0.11 N and ±0.27 N respectively. An additional load sen- trimming values of Lsheet and Lcar . After the transient flow
sor, with 5 daN range, was used to record the load in the is over, the aerodynamic loads reported by the balance are
sheet with a precision of ±0.02 N. Flying shapes were also averaged over an acquisition time of 30 s to smooth-out the
recorded with a V-SPARS acquisition system (Le Pelley and remaining noisy fluctuations in the signals. When the (time-
Modral, 2008), tracking the position of five dark red stripes averaged) loads are collected for all the initial samples (blue
across the sail (see Figure 2(a)). loop in Figure 4), the GP based optimization is carried out
The wind tunnel inflow velocity was measured and found on the initial data set (red block “Trimming optimization”).
to have an apparent wind speed (AWS) of 3.5 ± 0.15 m/s The optimization provides a new trimming point, x⇤ , maxi-
for an apparent wind angle (AWA) of 40 ± 2 deg. The cor- mizing the AEI merit function.
 In the “Optimization loop”, the experimental apparatus tion, contrasting two cases. In the first case, the measure-
is remotely tuned to the new trimming point x⇤ . The aero- ment noise is set to ✏ = 0 (noise-free situation) so the con-
dynamic loads are then averaged over 30 s and the new es- structed GP models are interpolating the experimental data.
timate of P(x⇤ ) is included in the database. The Gaussian In the second case, the actual measurement noise is deter-
Process model is updated consequently, generating a new mined experimentally and it is subsequently used in the GP
trimming points x⇤ . These steps (red loop in Figure 4) are model construction.
repeated until the convergence of the trimming parameters,
which is considered achieved when the algorithm proposes
two successive trimming points within a distance less than Experimental optimization with ✏ =0
1 mm (i.e the precision on the enforcement of the trimming
parameters in the experimental apparatus). The results of the experimental trimming optimization in the
case where ✏ is set to zero is now presented.
Start The GP model for the experimental P(x) function, af-
ter 36 iterations of the optimization algorithm, is reported
in Figure 5. Specifically, Figure 5(a) depicts the color con-
tours of the GP model mean as a function of the trimming
Dataset sampling parameters Lcar and Lsheet , while Figure 5(b) shows the
(10 points) standard deviation of the GP model. The black dots are the
data points where the performance was experimentally es-
timated. Regarding the location of the observation points,
Wind tunnel measurements we notice a large dispersion, highlighting the lack of con-
- Set trimming New trim vergence in the successive tentative optimal trimming can-
- Run and save 30 s didates x⇤ . In fact, the algorithm has explored the param-
eter domain ⌦ without discovering a particular sub-domain
of ⌦ likely to contain the global optimum. This can also be
Trimming optimization appreciated from the mean field of the GP model, in Fig-
- Build metamodel
ure 5(a), which, although smooth, presents at least 2 local
minima. The presence of multiple local minima is in fact
- Optimize AEI
Sampling Loop spurious and induced by the interpolating nature of the GP
model for ✏ = 0: the model is fitting the experimental
New trim
noise. This can also be appreciated from the standard devia-
tion field reported in Figure 5(b), which is zero at the obser-
vation points, denoting an inappropriate level of confidence
Wind tunnel measurements in the GP model approximation of P(x) at these locations.
- Set trimming Further, departing from the observation points, the variance
- Run and save 30 s of the GP model prediction quickly increases (observe, in
particular, the standard deviation field in the neighborhood
of isolated data points) and becomes large. As a result of the
Trimming optimization over-confidence in the model at measured points and high
- Build metamodel variance (low confidence) in unexplored areas, the optimiza-
- Optimize AEI tion process is led by the AEI merit function to propose new
candidates in relatively less populated areas.
New trim Figure 6 depicts the measured values of P(x⇤ ) at the suc-
cessive tentative optima as selected along the iterations of
the algorithm (the first 10 iterations correspond to the initial
Conv ? No Latin Hypercube Sampling of ⌦, and are not actual itera-
tions of the algorithm). The plot shows that the measured
Optimization Loop
performance is not converging and it sustains large fluctu-
Yes
ations having the same magnitude as for the initial random
sample: the complete absence of an improvement trend in
the successive measurement of P(x⇤ ) is characteristic of the
End
failure of the present approach. This unsuccessful test high-
lights the negative effect of not accounting for the noise in
Figure 4: Trimming optimization procedure. the estimates of P(x): it prevents the GP model to discover
trends in the actual performance function from the noisy ob-
Below, we present the results of the trimming optimiza- servations.
 marginal likelihood in (12). Instead, ✏ is directly estimated
from the experimental apparatus, using repeated measure-
ments at the same trimmings. The same 10 previous initial
LHS points are used to determined the first GP model, with
the prescribed value of ✏ . Then the optimization proceeds
and a different sequence of proposed optima is generated as
the GP models differ from the previous experiment.
In particular, the optimization now converges in 33 iter-
ations, the last two proposed optima being in sufficiently
close distance (less than 1 mm). The convergence of the se-
quence of proposed optima can be seens in Figure 7, where
the mean and standard deviation of the GP model of P(x) at
convergence are plotted. In contrast to the case with ✏ = 0,
the clustering of the successive proposed optima is clearly
(a) GP model expectation. visible. Also, the mean of the GP model in Figure 7(a) re-
mains smooth and now exhibits a single well-defined global
maximum. The converged optimal trimming is found for
Lsheet = 133 mm and Lcar = 138 mm corresponding to
a predicted performance P(xopt ) = 0.397. The standard
deviation of the GP model, depicted in Figure 7(b), is seen
to be minimal in the neighborhood of the optimum, though
assuming values & ✏ . Other regions of ⌦ far from the opti-
mum are not explored by the optimization process, although
the variance can be large.
The convergence of the optimization procedure can also
be appreciated from the plot of Figure 8 which should be
contrasted with the results shown Figure 6. It shows a clear
improvement of the measured performance (after the first
10 random points). In fact after iteration 25, the remain-
ing fluctuations in the measured performances can be essen-
(b) GP model standard deviation.
tially attributed to the measurement noise. These remaining
fluctuations, with amplitude ⇠ ✏ , are much less significant
Figure 5: GP model of the experimental performance P(x)
compared to the previous case.
( ✏ = 0).
In summary, the global optimal trimming parameters are
determined despite the noise in the measurements, thanks to
0.5
the non-interpolating nature of the GP approximation which
smooths out the noise. Further, the optimum is found in
0.4
few iterations only, owing to the AEI merit function which
0.3 is able to disregard non-interesting areas of ⌦, even if they
carry large prediction variance.
Measurement

0.2

0.1

0 NUMERICAL OPTIMIZATION
-0.1
A numerical model of the wind tunnel and sail model has
-0.2 been created to reproduce the previous experimental opti-
-0.3
0 5 10 15 20 25 30 35 40
mization problem. The objective is to assess the capabilities
Iteration of the optimization method, when applied to a coupled FSI
software, and compare the resulting optimum with the ex-
Figure 6: Experimental measurements of the performance perimental one.
for the sequence of proposed optima. Case of ✏ = 0.
Numerical model
Experimental optimization with noise
We briefly present the structural and fluid solvers used for
In a second experiment we set ✏ 2 = 0.0272 . This value the resolution of the FSI problem. Steady solution of the FSI
is not determined as part of the optimization of the log- problem are sought by means of a quasi-steady approach.
 cables and Constant Strain Triangles (CST) membrane el-
ements of various types) for the static or dynamic simula-
tion of sail boat rigs in large displacement regime (Augier,
2012). The structural model for the simulations presented
hereafter is illustrated in Figure 9; it uses dimensions and
mechanical characteristics (for the mast, boom, and sail fab-
rics) measured on the experimental model.

(a) GP model expectation.

(a) Structural mesh. (b) Mainsail stiffness.

Figure 9: Numerical model of the sail.

The solver ARA is coupled to the ISIS-CFD software

(from FINETM /Marine) which solves the Navier-Stokes
equations in the flow domain. ISIS-CFD is based on finite
volume methods accommodating both structured and un-
structured meshes; it also proposes several turbulence mod-
(b) GP model standard deviation.
els and boundary condition. For the present computations,
we consider a parallelepiped computational domain, with
Figure 7: GP model of the experimental performance P(x)
spatial extension 7.5h, 12h and 1.8h in the X, Y and Z di-
( ✏ = 0.027).
rections respectively. These dimensions were selected on
the basis of previous numerical experiments (Viola et al.,
2013). The boundary conditions, applied on the different
faces of the computational domain, are schematically illus-
0.5

0.4 trated in Figure 10.

0.3
Measurement

0.2

0.1

-0.1

-0.2

-0.3
0 5 10 15 20 25 30 35
Iteration

Figure 8: Experimental measurements of the performance

for the sequence of proposed optima. Case of ✏ = 0.027.
Figure 10: Boundary conditions for the flow solver.

For the structural model of the sail we rely on the ARA The fluid domain is meshed using HEXPRESSTM , a
software developed by K-EPSILON. The code ARA consid- semi-automated mesh generator. Note that the mast is not
ers different structural elements (e.g. Timoshenko beams, meshed in the fluid solver. Regarding the turbulence model,
the SST k ! model (Menter et al., 2003) was selected with of the numerical optimization procedure is identical to that
wall function boundary conditions (Kalitzin et al., 2005). of the experiment in Figure 4. However, for the numeri-
This choice requires a sufficiently fine mesh over the sail cal optimization, ✏ is directly inferred from the data when
surface and at the bottom of domain (sea level) to correctly minimizing (12). In addition, the parameter domain for the
capture the vertical profile of the inflow velocity. The later numerical case has been increased to encompass higher val-
is estimated from the experimental profiles previously re- ues of Lcar .
ported in Figure 3, and its shown in Figure 11.
Numerical optimization results
Figure 12 depicts the mean value of the GP model based on
2.5 mean_U
the numerical evaluation of the performance. The GP model
Interp_0.25
is reported at the end of the optimization procedure, which
2 has converged in 34 iterations. The black dots correspond
again to the sequence of optimization points x⇤ in the data
1.5
set. We first remark the smoothness of the mean GP model
Z [m]

which exhibits a single global optimum, as for the experi-

1
mental case (considering measurement noise). In fact, the
0.5
inferred ✏ = 0.022 has a value close to the experimental
one. Further, in the range Lcar 2 [ 210, 150] mm, the ter-
0 minal GP model is seen to be in good agreement with its
0 0.5 1 1.5 2 2.5 3 3.5 4 experimental counterpart shown in Figure 7(a). However,
U [m/s] the valley containing the numerical minimum is larger and
flatter, compared to the experimental one, and the numerical
Figure 11: Inflow velocity profile. optimum appears at a value of a Lcar value larger than for
the experimental case. The variance of the GP model predic-
For the FSI simulations, the mesh of the fluid domain tion, y2 , exhibits a structure similar in shape and magnitude
has to be deformed to adapt to the changes in the bound- to the experimental case in Figure 7(b) (not shown).
ary geometries. A mesh deformation propagation method,
proceeding from the deformable boundaries toward the in-
side of the fluid domain, was developed at K-EPSILON for
this purpose (Durand et al., 2014). The FSI problem is then
solved coupling ARA and ISIS-CFD solvers with a quasi-
monolithic algorithm (Durand, 2012), which is an implicit
coupling procedure adapted to a partitioned solver. Briefly,
the resolution of the structural problem is nested inside the
iterations on the nonlinear steady flow solution. This ap-
proach preserves the convergence and stability properties of
the monolithic approach. More details on the solvers and the
coupling algorithms can be found in (Durand, 2012, Roux
et al., 2002, 2008).
A convergence analysis was conducted in order to select
spatial discretization capturing correctly the physics of the
FSI problem, while maintaining a reasonable computational Figure 12: Mean GP model of the numerical performance
cost permitting the optimization of the trimming parameters. (inferred ✏ = 0.022).
In particular, different fluid meshes with up to 4.3 million fi-
nite volumes were considered. Eventually, a discretization To understand the differences between the numerical and
of the fluid domain with roughly 1.5 million finite volumes experimental optima, we first compare in Figure 13 and
and a sail discretization with 2 700 membrane elements was Figure 14 the flyings shapes for two trimming parameters
selected for the computations presented below. The com- (Lsheet , Lcar ) equal to (160, 0) and (133, 138) respectively
putations were carried out on a 64 CPUs cluster; an aver- (lengths in mm). For the case with centered car, Lcar = 0,
aged computational time of 5 h was reported for solving a small wrinkle is visible in the experimental flying shape
individual FSI problems. From the FSI solution, the asso- (see Figure 13(a)). At the optimal experimental trimming
ciated aerodynamic forces acting on the sail are computed point (Lsheet , Lcar ) = (133, 138), shown in Figure 14(a),
in the same reference frame as in the experimental setup, the wrinkle in the experimental flying shape is even more
and the performance in (16) is finally returned to the opti- pronounced. This is in contrast with the corresponding nu-
mizer. Except for the determination of P(x), the flowchart merical flying shapes, shown in Figures 13(b) and 14(b)
respectively, which present no such wrinkle. Modeling er-
rors and experimental uncertainties are deemed responsible
for this difference. In particular, the absence of wrinkle in
the numerical solution could be mostly due to an incorrect
prescription of the tensions in the two full battens of the
sail. Another important source of discrepancy between the
flying shapes are the boundary conditions of the numerical
wind tunnel and effects of confinement, which were shown
to have a significant impact on the computed aerodynamic
forces (Viola et al., 2013).
(a) Experimental.

(a) Experimental.

(b) Numerical.

Figure 14: Comparison of experimental and numerical fly-

ing shapes for trimming parameters Lsheet = 133 and
Lcar = 138.

Lsheet [mm] Lcar [mm] Pred. P(xopt )

Exp. 133 138 0.397 ± 0.027
Num. 122 226 0.413 ± 0.024

(b) Numerical. Table 1: Comparison of the experimental and numerical op-

tima.
Figure 13: Comparison of experimental and numerical fly-
ing shapes for trimming parameters Lsheet = 160 and
Lcar = 0. that the thrust coefficients are in fact equal up to the third
significant digit. On the contrary, the side force coefficients
To complete the comparison of the experimental and nu- and the sheet tensions exhibit larger discrepancies. The dif-
merical optimizations, we report in Table 1 the computed lo- ferences in the sheet tensions can be directly attributed to
cation of the two optima and the corresponding best predic- the different optimal Lcar . The higher disagreement in the
tion of the performance. We observe that the experimental side force coefficients is not surprising. In our experience
and numerical optimal trimming are significantly different the side forces are very sensitive to model errors.
for the optimal Lcar . This difference can be explained by
the numerical performance function that is particularly flat CX CY T [N]
along the Lcar direction around the optimal point: variation
of Lcar around the optimum weakly affects the predicted Exp. 0.497 ± 0.012 1.026 ± 0.015 14.9 ± 0.02
performance. This is consistent with the predicted perfor- Num. 0.497 0.803 20.1
mances reported in the last column of Table 1, that are in
close agreement despite the differing trimming parameters. Table 2: Comparison of aerodynamic coefficients and sheet
tension at the experimental and numerical optima.
A more detailed investigation of the measured and com-
puted fluid forces at the optima, reported in Table 2, reveals
CONCLUSION AND DISCUSSION der grant agreement No PIRSES-GA-2012-318924, and
from the Royal Society of New Zealand for the UK-
We have proposed to use a Gaussian Process model to en- France-NZ collaboration project SAILING FLUIDS (see
able the optimization of the trimming parameters of a com- www.sailingfluids.org). This work was also
plex nonlinear sail systems. The approach has been first supported by the “Laboratoire d’Excellence” LabexMER
tested on the trimming of an experimental model sail in (ANR-10-LABX-19) and co-funded by a grant from the
the Yacht Research Unit wind tunnel. The experiments French government under the program “Investissements
have validated the approach and have shown its robustness d’Avenir”.
against noisy estimates of the performance. For this two pa-
rameters problem, the experimental optimal point was found
REFERENCES
within few iterations of the algorithm. These tests have val-
idated the proposed optimization method and have demon-
B. Augier. Etudes expérimentales de l’interaction fluide-
strated its robustness against experimental variabilities, as
structure sur surface souple: application aux voiles de
well as the key role of the noise parameter.
bateaux. PhD Thesis, Universite de Bretagne Occiden-
A detailed numerical model of the wind tunnel exper-
tale, 2012.
iment has been established, considering the full Fluid-
Structure Interactions problems. The numerical model is B. Augier, P. Bot, F. Hauville, and M. Durand. Experimen-
based on a turbulent flow solver coupled with a nonlinear tal validation of unsteady models for fluid structure in-
elastic solver, using a mesh deformation method. The nu- teraction: Application to yacht sails and rigs. Journal of
merical and experimental optima were found to be consis- Wind Engineering and Industrial Aerodynamics, 101:53–
tent, given all the modeling and measurement errors. In par- 66, 2012.
ticular, it is found that the predicted performances and fluid
forces are in much better agreement than the optimal param- T. Bäck and H. P. Schwefel. An overview of evolutionary al-
eters. In fact, it can be reasonably claimed that the differ- gorithms for parameter optimization. Evolutionary com-
ences are consistent with the current predictive capabilities putation, 1(1):1–23, 1993.
of state of the art FSI solvers, and that the GP approximation
does not introduce noticeable errors in the optimization pro- V. G. Chapin, R. Neyhousser, G. Dulliand, and P. Chassaing.
cedure. Specifically, we have shown that, from a limited set Design optimization of interacting sails through viscous
of computations, the GP model is able to reconstruct accu- cfd. In INNOVSail, Innovation in high performance sail-
rately the numerical estimate of the performance. Therefore, ing Yacht, Lorient, 2008.
for the present tests, only a better numerical modeling of the
M. Durand. Interaction fluide-structure souple et légère,
experimental set-up would help reducing the observed dis-
application aux voiliers. PhD Thesis, Ecole Centrale de
crepancies. Possible avenues in this direction are: improved
Nantes, 2012.
battens modeling, better boundary conditions for the flow,
accounting for mast/flow interaction, more advanced turbu- M. Durand, A. Leroyer, C. Lothodé, F. Hauville, M. Vison-
lence model, . . . At a more fundamental level, the question neau, R. Floch, and L. Guillaume. Fsi investigation on
of the treatment of experimental uncertainties remains criti- stability of downwind sails with an automatic dynamic
cal. On this aspect, we are considering sensitivity and uncer- trimming. Ocean Engineering, 90:129–139, 2014.
tainty quantification (Le Maı̂tre and Knio, 2010) studies to
account for the experimental uncertainties. In particular, the R. Duvigneau and P. Chandrashekar. Kriging-based op-
results presented in this work point to the need for an appro- timization applied to flow control. International Jour-
priate characterization of the optima and the assessment of nal for Numerical Methods in Fluids, 69(11):1701–1714,
their robustness to uncertainties and modeling errors. Future 2012.
works will develop these aspects along with the deployment
of the GP-based optimization method to problems involving R. G. J. Flay. A twisted flow wind tunnel for testing yacht
large numbers of trimming parameters (full yacht rig). sails. Journal of Wind Engineering and Industrial Aero-
dynamics, 63(13):171 – 182, 1996. ISSN 0167-6105.
Special issue on sail aerodynamics.
ACKNOWLEDGEMENTS
M. N. Gibbs. Bayesian Gaussian Processes for Classifi-
The authors wish to acknowledge the people of the cation and Regression. PhD thesis, University of Cam-
Yacht Research Unit for their welcoming and help dur- bridge, 1997.
ing the experiments in the wind tunnel. Contribu-
tions of INCIDENCES SAILS and NUMECA compa- H. Hansen, P. S. Jackson, and K. Hochkirch. Real-time ve-
nies are also acknowledged. This work was partially locity prediction program for wind tunnel testing of sail-
funded by the European Union’s Seventh Program for re- ing yachts. Proc. The Modern Yacht, Southampton, UK,
search, technological development and demonstration un- 2003.
N. Hansen. The cma evolution strategy: a comparing review. P. V. Oossanen. Predicting the speed of sailing yachts.
In Towards a new evolutionary computation, pages 75– SNAME, 101:337–397, June 1993.
102. Springer, 2006.
R. Ranzenbach, D. Armitage, and A. Carrau. Mainsail plan-
N. Hansen and A. Ostermeier. Completely derandom- form optimization for irc 52 using fluid structure interac-
ized self-adaptation in evolution strategies. Evolutionary tion. In The 21st Chesapeake Sailing Yacht Symposium,
Computation, 9(2):159–195, 2001. SNAME, 2013.
D. Huang, T. T. Allen, W. I. Notz, and N. Zeng. Global op- C. E. Rasmussen and C. K. I. Williams. Gaussian processes
timization of stochastic black-box systems via sequential for machine learning. MIT Press, 2006.
kriging meta-models. Journal of global optimization, 34
(3):441–466, 2006. N. Rousselon. Optimization for sail design. In modeFRON-
TIER Conference, 2008.
L. Huetz and P. E. Guillerm. Database building and sta-
tistical methods to predict sailing yacht hydrodynamics. Y. Roux, S. Huberson, F. Hauville, J. P. Boin, M. Guilbaud,
Ocean Engineering, 90:21–33, 2014. and B. Malick. Yacht performance prediction : Towards
a numerical vpp. In 1st High Performance Yacht Design
D. R. Jones. A taxonomy of global optimization methods Conference Auckland, 4-6 December, Auckland, 2002.
based on response surfaces. Journal of global optimiza-
tion, 21(4):345–383, 2001. Y. Roux, M. Durand, A. Leroyer, P. Queutey, M. Visonneau,
J. Raymond, J. M. Finot, F. Hauville, and A. Purwanto.
D. R. Jones, M. Schonlau, and W. J. Welch. Efficient global Strongly coupled vpp and cfd ranse code for sailing yacht
optimization of expensive black-box functions. Journal performance prediction. In 3rd High Performance Yacht
of Global optimization, 13(4):455–492, 1998. Design Conference Auckland, 2-4 December, pages 215–
225, Auckland, 2008.
G. Kalitzin, G. Medic, G. Iaccarino, and P. Durbin. Near-
wall behavior of rans turbulence models and implications V. K. Saul’ev and I. I. Samoilova. Approximation methods
for wall functions. Journal of Computational Physics, for the unconstrained optimization of functions of several
204(1):265–291, 2005. variables. Journal of Soviet Mathematics, 4(6):681–705,
R. Korpus. Performance prediction without empiricism: A 1975.
rans-based vpp and design optimization capability. In M. L. Stein. Interpolation of spatial data: some theory for
The 18th Chesapeake Sailing Yacht Symposium, SNAME, kriging. Springer Science & Business Media, 2012.
2007.
D. Trimarchi. Analysis of downwind sail structures using
O. P. Le Maı̂tre and O. M. Knio. Spectral methods for un- non-linear shell finite elements. PhD Thesis, University
certainty quantification. Scientific Computation. Springer, of Southampton, 2012.
New York, 2010.
I. M. Viola, P. Bot, and M. Riotte. Upwind sail aerodynam-
D. J. Le Pelley and O. Modral. V-spars: A combined sail and ics: A rans numerical investigation validated with wind
rig shape recognition system using imaging techniques. tunnel pressure measurements. International Journal of
In Proc. 3rd High Performance Yacht Design Conference Heat and Fluid Flow, 39:90–101, 2013.
Auckland, New Zealand, Dec, pages 2–4, 2008.
R. Von Mises. Mathematical Theory of Probability and
M. D. McKay, R. J. Beckman, and W. J. Conover. A com- Statistics. Academic press, 1964.
parison of three methods for selecting values of input
variables in the analysis of output from a computer code.
Technometrics, 42(1):55–61, 2000.
W. Menotti, M. Durand, D. Gross, Y. Roux, D. Glehen, and
L. Dorez. An unsteady fsi investigation into the cause of
the dismating of the volvo 70 groupama 4. In INNOVSail,
Innovation in high performance sailing Yacht, page 197,
Lorient, 2013.
F. R. Menter, M. Kuntz, and R. Langtry. Ten years of in-
dustrial experience with the sst turbulence model. Turbu-
lence, heat and mass transfer, 4:625–632, 2003.
J. Nelder and R. Mead. A simplex method for function min-
imization. Computer Journal, 7(4):208–313, 1965.