You are on page 1of 18

Journal of Ocean Engineering and Science 8 (2023) 15–32

Contents lists available at ScienceDirect

Journal of Ocean Engineering and Science


journal homepage: www.elsevier.com/locate/joes

An artificial intelligence-aided design (AIAD) of ship hull structures


Yu Ao a,b, Yunbo Li a,c, Jiaye Gong c, Shaofan Li b,∗
a
College Of Shipbuilding Engineering, Harbin Engineering University, China
b
Department of Civil and Environmental Engineering, University of California, Berkeley, CA 94720, USA
c
College of Ocean Science and Engineering, Shanghai Maritime University, China

a r t i c l e i n f o a b s t r a c t

Article history: Ship-hull design is a complex process because the any slight local alteration in ship hull structure may
Received 15 July 2021 significantly change the hydrostatic and hydrodynamic performances of a ship. To find the optimum hull
Revised 18 September 2021
shape under the design requirements, the state-of-art of ship hull design combines computational fluid
Accepted 9 November 2021
dynamics computation with geometric modeling. However, this process is very computationally intensive,
Available online 18 November 2021
which is only suitable at the final stage of the design process. To narrow down the design parameter
Keywords: space, in this work, we have developed an AI-based deep learning neural network to realize a real-time
Artificial intelligence prediction of the total resistance of the ship-hull structure in its initial design process. In this work,
Deep learning neural network we have demonstrated how to use the developed DNN model to carry out the initial ship hull design.
Hull deformation The validation results showed that the deep learning model could accurately predict the ship hull’s total
Machine learning resistance accurately after being trained, where the average error of all samples in the testing dataset is
Ship hull design
lower than 4%. Simultaneously, the trained deep learning model can predict the hip’s performances in
Total resistance
real-time by inputting geometric modification parameters without tedious preprocessing and calculation
processes. The machine learning approach in ship hull design proposed in this work is the first step
towards the artificial intelligence-aided design in naval architectures.
© 2021 Shanghai Jiaotong University. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/)

1. Introduction with the objective to achieve the best hydrodynamic performances.


The design process of the hull form is a try-and-error or cyclic
The most critical stage in shipbuilding is its design stage in ship process. Designers usually need to select an initial hull model to
engineering [1,2]. The ship designer must consider all technical re- avoid a heavy modeling process. Based on the initial hull model,
quirements from the shipowners as well as the performance re- designers can obtain the optimal solutions by looping over geome-
quirements comprehensively [3]. The whole ship design process try modification and performance evaluation [6]. The hull form de-
is a very complex process. When processing a ship design pro- sign process may require a large workload and cost much time and
cess, what designers often encounter are complex design steps and resources, which depends not only on the hull modification but
massive workloads. In detail, the whole design process can be di- also on the performance evaluation cost. After designers choose a
vided into five steps, including concept design, initial design, ba- set of geometric parameters for a possible new hull, the designer
sic design, detail design, and manufacturing design. Each design must evaluate the hydrodynamic/hydrostatic performance of the
stage has its purpose. For instance, in the conceptual design stage, ship hull. Take resistance calculation as an example: the Compu-
the ship’s specifications and purpose need to be determined, while tational Fluid Dynamics (CFD) calculations, e.g. [7], often may cost
in the preliminary design, the hull form needs to be determined. several hours or even a whole day, not to mention the modeling
Since the hull shape form has a decisive influence on the ship’s time cost. Thus, in order to shorten the time required for the hull
sailing performance and seakeeping performance, the initial design form design, the best choice is to reduce the workload of the geo-
is a very critical design stage e.g. [4]. metric modification process and the calculation process.
The initial ship hull design [5] is a process of designing and The objective of the ship performance evaluation process is to
evaluating hull forms to determine the final optimal hull structure, evaluate the performance of the designed hull form. Therefore, if
we can use a more effective method to evaluate the performances

of the hull form, the time required for hull design can be signifi-
Corresponding author.
cantly shortened and less dependent on the designers’ experiences.
E-mail address: shaofan@berkeley.edu (S. Li).

https://doi.org/10.1016/j.joes.2021.11.003
2468-0133/© 2021 Shanghai Jiaotong University. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
Y. Ao, Y. Li, J. Gong et al. Journal of Ocean Engineering and Science 8 (2023) 15–32

Often time, it is just impossible to go through all the possible de- be a design target, which is rarely considered in the current ap-
sign scenariosbecause of the time or the limitation of designers’ plication of DL technology in hull design. A DL model which can
experience. However, if we are able to find the mappings of all only conduct the relationship between the principal dimensions of
possible hull forms in the design space to the resistance value, we the hull and its performances cannot satisfy the ship designer’s re-
may further simplify and optimize the hull form design process in quirements.
the largest possible design space, and hence when designers de- With the rapid developments of artificial intelligence technol-
signed a new hull form, they can immediately find the relation- ogy, it is possible to use DL-based machine learning methods to ef-
ship between the ship performance and ship hull geometric shape, ficiently obtain the mapping relationship between any hull form in
saving much time, labor, experience, and cost. Then, the question is the design space and the performance results required in the per-
how to find the complete mapping between various hull forms and formance evaluation stage by revealing the inner relationships of a
their performances. The mapping relation between the hull form sample dataset. In detail, following the ideas of DL technology, we
and its performance is complicated, nonlinear, and non-intuitive need to prepare a dataset that contains a suitable amount of sam-
in general. The first reason is that more than hundreds and even ples located in the designers-required design space to show map-
thousands of parameters are required to describe a ship hull form, ping style between hull forms and corresponding evaluated per-
which leads to the complexity of the mapping. The second reason formance, which will be further used to train the DL model. Af-
is that this mapping relation is implicitly embedded in a super- ter establishing the dataset, we can establish an FCNN model with
manifold with thousands of dimensions, requiring many different correct hyperparameters to ensure the model is suitable for the
implicit representations to describe such mapping relation, if it current mission, analyze and reveal internal mapping relationships
is explainable and rationalizable. Even if it may be explainable, between hull form and corresponding performance, and store the
it may still be difficult for us to obtain an explicit mathematical relationships in the model for further use. Once the results pre-
model that can meet the design accuracy requirements. dicted by the DL model are close enough to the actual result to
Deep Learning (DL) technology, in particular the deep neural satisfy design requirements, then designers can adopt it to evalu-
network (DNN) [8], is a class of artificial intelligence-based ma- ate the performances in real-time, which will save the workload in
chine learning methods [9]. DNN is a powerful method, which has the design loop process.
complex network structures and multiple layers of artificial neu- Simply put, the application of DL technology in ship engineer-
ral networks. In some cases, DNN can outperform human expert ing to assist the ship design can be treated as transforming the
analysis and performance. Recently, not limited to the computer ship design process into a ship selection process, making sure de-
science where it was initially located, DL technology has success- signers can immediately finish the performance evaluation pro-
fully applied in other fields, including computer vision, audio pro- cess after they finish the geometry modification process to get the
cessing, natural language processing, and search engines. The main hull form. The main reason for using DL technology-based method
advantage of DL technology is that it enables machines to improve rather than other machine learning methods is that DL technology-
their decision process based on both data as well as machine ex- based machine learning method has high accuracy and good adapt-
perience. Furthermore, in the past few years, DL technology has ability to very complicated problems, especially for an object like
become one of the most versatile and efficient tools for handling a hull form, a very complex object, which will need many param-
classification and regression tasks [10,11]. eters to express a complete hull form. Based on DL technology, we
As a powerful machine learning tool, DL technology can extract can create a large enough model to discover the mapping relation-
the complicated relationship between input data and output data ships among all parameters used to describe a ship hull form and
by feeding a DL model with a sample dataset e.g. [12,13]. Through its hydrodynamic performance.
the training process, the trained DL model can use the uncovered In this work, we have developed a Deep Learning Neural Net-
hidden relationship to make predictions. Thus, DL technology is work model for the initial design of ship hulls. In specific, we con-
very suitable for dealing with the task that needs to deduce the struct a Fully Connected Neural Network (FCNN) model to predict
mapping between two vectors since it allows a DL model to learn and evaluate the total resistance for hull form based on geome-
from the sample dataset and deduce the mapping relationships. try modification parameters, which can forecast not only the per-
In some recent studies, based on DL technology, some re- formance of hulls with different principal dimensions but also the
searchers have used data-driven approaches to solve many prob- performance of hulls with different hull forms. We believe the DL
lems in naval architecture engineering. Representatively, Milaković model establishment process in this work provides an effective and
et al. [14] developed a machine learning-based method for pre- practical solution to assist the hull form at its initial design stage.
dicting ship speed profile in a complex ice field under the situ- It is noted that the data-driven design method proposed in this
ation that computational methods have difficulties capturing the work aims to save ship designers’ workload and time cost in the
entire complexity of the ship-ice interaction process. Wackers et al. initial design stage. Compared with other design-assistant meth-
[15] used a machine learning approach to reduce the computa- ods, such as structural optimization methods [20,21], by adopting
tional effort of CFD simulations. Meanwhile, there have been many the method proposed in this paper, ship designers can immediately
researches adopting DL technology to aid engineering design due find the suitable or optimal hull shape and hence the hydrody-
to its convenience. For instance, Gunpinar et al. [16] adopted re- namics/hydrostatic performance of hulls by adjusting the geometry
gression/neural network methods to establish a design support modification parameters. However, this method sacrifices the cal-
system for car side silhouette design. Shaeffer et al. [17] used a ma- culation accuracy and ignores the inner structure, making it only
chine learning approach to regress existing data to obtain a model suitable to be used in the initial design stage.
to assist the early-stage of the hull form design. The paper is organized as follows. In the rest of the paper, we
In current literature, most reported researches use the principal first briefly introduce the method of expressing the hull form as
dimensions of the hull as the input parameters of ANN to forecast geometry modification parameters. We then outline the process
its performance. For instance, Cepowski et al. [18] established a DL for gathering the required sample data to train the deep learn-
model to estimate added resistance of ships based on the principal ing model before demonstrating the trained DL model, followed
dimensions of the hull. Grabowska et al. [19] forecast realized the by the detailed process about the hyperparameters tuning process
prediction of hull’s resistance based on parameters with the high- and model training process. Next is followed by several examples
est influence by adopting DL models. However, in the initial design demonstrating how to use the proposed deep learning model to
stage, both the principal dimensions and the hull surface should predict the total resistance of hull structures. Specifically, based

16
Y. Ao, Y. Li, J. Gong et al. Journal of Ocean Engineering and Science 8 (2023) 15–32

Table 1
Basic geometric parameters of the initial hull form model.

Geometric parameters Scaled hull model

Length between perpendiculars 7.279 m


Breadth 0.703 m
Draft 0.342 m
Wetted surface area 9.438 m2
Design speed 2.196 m/s

show the relationship between the position after deformation and


the initial position in Eq. (3),
 

L 
M 
N
P=ψ −1
blmn (ψ (P 0 ))μlmn (3)
Fig. 1. The relationship between the global coordinate system and the lattice coor-
l=0 m=0 n=0
dinate system.
where:

on several forms of hull structures located in design space rather • ψ - the transformation from real space to lattice space,
than in the sample dataset, we compare the prediction result of • ψ −1 - the inverse transformation from lattice space to real
the deep learning model with the evaluation result of the potential space,
flow method. Through these cases, we seek to demonstrate that • blmn - Bernstein polynomials,
the deep learning technology-based prediction method proposed in • P 0 - the initial position of the point.
this work can accurately predict the hull form performance based
As shown in Eq. (3), due to the characteristics of the Bézier
on a set of geometry modification parameters, and hence assisting
polynomials, the influence weight of a control point to an object
the hull form design process.
point will change based on the distance, which leads that with
the increase of the distance, the influence weight of the control
2. Geometry modification expression method
point to the object point decreases quickly, making a control point
can only significant modify nearby part of the object. The charac-
2.1. Free-form deformation (FFD)
teristic of distance determines the influence weight of the control
point ensures that one can adjust the shape of the object more in-
Free-Form Deformation (FFD) is a geometric modification tech-
tuitively.
nique mainly used to deal with the deformation of rigid objects.
The application of FFD used in this work is based on the PyGem
Sederberg and Parry developed FFD in 1986 [22], where they used
framework [23], which is a python library using Free-Form Defor-
the ternary tensor product Bernstein polynomial control a Bézier
mation (FFD), Radial Basis Functions (RBF), and Inverse Distance
volume in R3 space. In FFD, all the deformation processes hap-
Weighting (IDW) to parametrize and morph complex geometries.
pened in a local parametric normalized coordinate system, which
follows Eq. (1), whose boundary is the lattice space.
2.2. Geometry modification process of hull form
X (s, t, u ) = L0 + sS + t T + uU (1)
where In the hull form design process, designers usually select an ini-
tial hull form model and modify the initial model to find a new
• L0 - the origin of the lattice coordinate system, hull form to avoid a massive modeling process. In this paper, we
• S, T , U - the edge vectors along the axes of the local coordinate select the hull form of the KRISO Container Ship (KCS), which
system (S,T,U), was developed by the Korean Maritime and Ocean Engineering Re-
• s, t, u - coordinate values in the local coordinate system, lies in search Institute (KMOERI), as the initial hull form. To facilitate the
(0,1) for any point located at the interior of the lattice space. accuracy verification of the calculation program used to generate
All control points are attached to the pre-defined lattice coordi- samples in the dataset, the hull form used in this paper is a 1:31.6
nate system, with an artificially determined distribution indepen- scale model. As one of the most commonly used ship models, KCS
dent of the shape of the rigid object. An evenly distributed control has been used as benchmarks for many calculation methods for its
point was adopted here, which is also the most commonly used structure, lines plan, and model test results are open to the public.
distribution. In the even distribution, the lattice space will be di- Table 1 shows the basic geometric parameters of the scaled KCS
vided into l plane, m planes, and n planes in the direction of s- hull form model used in this paper.
axis, t-axis, and u-axis, respectively, where the intersection point In this work, we use the right-handed coordinate system to de-
of every three planes constitutes a control point. The relationship scribe the hull form model and its deformation range, where the
between the global coordinate system and the lattice coordinate x-axis parallel with the hull length direction and the y-axis parallel
system can be seen in Fig. 1. Eq. (2) shows the position of control with the hull width direction. Fig. 1 shows the hull structure po-
point P in the lattice space, which is formed by the intersection of sition in a fixed coordinate system. Consider that a hull is a sym-
the ith plane on the s-axis, the jth plane on the t-axis, and the kth metrical object, we only consider half of the hull model in order
plane on the u-axis, to reduce the number of geometry modification parameters, which
will decrease half the number of geometry modification parame-
i j k ters. In this work, we selected the hull in the negative direction of
Li jk = L0 + S + T + U . (2)
l m n the y-axis as the geometry modification object, which can be done
The position of all rigid object points will be changed based on by doing a mirroring process to obtain the complete hull model.
the total influence from all the control points, where the influence In this work, the geometry modification process of hull form
weight is inversely proportional to the distance the object point has been transformed into a set of geometry modification param-
and the control point. Taking the object point P as an example, we eters by combining the linear transformation method and the FFD

17
Y. Ao, Y. Li, J. Gong et al. Journal of Ocean Engineering and Science 8 (2023) 15–32

technique. In detail, the linear transformation method was adopted structure in the xyz coordinates are x̄max , ȳmax , z̄max , and the min-
to adjust the principal dimensions and sailing speed of the hull imum coordinate value of all points in the hull structure are x̄min ,
form model, and the FFD technique was adopted to modify the ge- ȳmin , z̄min . We can then obtain the length Bl , breadth Bb , and height
ometry shape of the hull form, which leads to two kinds of geome- Bh of the parallelepiped lattice space as follows,
try modification parameters. The geometry modification process of Bl x̄max − x̄min (1 + 2λx )
hull form in this paper has been divided into three sub-stages for Bb = ȳmax − ȳmin  (1 + 2λy ) (6)
more effortless intuitive operation. Bh z̄max − z̄min (1 + 2λz )
We called the first sub-stage of the geometry modification pro-
cess the hull stretch stage, where only the length, width, and where:
height will be changed. Suppose that the initial hull model has λx , λy , λz - gap rates in the x, y, z directions,
been transformed to a hull point set So, whose density is dense  - the element multiplication.use
enough to make the result predicted by the calculation method
used satisfies the engineering requirements. Then the relationship As shown in Eq. (6), we added three different gap rates in
between the intermediate hull point set Se after the first sub-stage the x, y, and z directions to ensure the entire hull model can be
and the initial hull point set So is shown in Eq. (4). completely enclosed inside the lattice space. Besides, the position
⎡ ⎤ of the origin (Ox , Oy , Oz ) of the lattice space in the hull coordi-
x1 y1 z1 nate system is also related to the extreme coordinate value of hull
⎢ .. ⎥ points, which follows Eq. (7).
⎢ . ⎥ αx
⎢ ⎥ x̄min − λx (x̄max − x̄min ) αx
Se = So · α = ⎢ xi yi zi ⎥ · αy (4) Ox
⎢ .. ⎥ αz Oy = ȳmin − λy (ȳmax − ȳmin )  αy (7)
⎣ ⎦
. Oz z̄min − λz (z̄max − z̄min ) αz
xh yh zh
The control point layout adopts in this work is an even distri-
where: bution, which is the commonly used distribution and suitable for
the parallelepiped lattice. For the lattice volume that is constructed
xi , yi , zi - the position of the ith hull point of the initial model, based on the extreme coordinate value of hull points, the paral-
h - the total number of points used to present hull form model, lelepiped lattice should also be a slender body with a long length
αx , αy , αz - length change rate, breadth change rate, and height and a relatively short breadth and height. Due to the shape feature
change rate. of the lattice space, the number of control points in the length di-
After finishing the first sub-stage of the geometry modification rection should be much larger than the number of control points
process, we continue with the second sub-stage of the geometry in the other two directions so that the geometry modification can
modification process, which is responsible for changing the draft be more reasonable and the geometry modification results become
of the hull structure model. The relationship between the interme- abundant. In FFD, the movements of control points are measured
diate hull point set SI after the second sub-stage and the interme- based on the lattice coordinate system, whose actual movements
diate hull point set Se after the first sub-stage are given as follows, in the hull coordinate system are related to the size of the lat-
tice space, which gives us the advantage to set up the range of all
⎡ ⎤ control points without attention to the difference of the principal
0 0 Do
⎢ .. ⎥ dimensions of hull structures. We created a (l × m × n ) finite-size
⎢ . ⎥ lattice volume of control points based on the initial lattice space to
⎢ ⎥
SI = Se + αz · αd · ⎢0 0 Do ⎥ (5) do the geometry modification, whose current position of all control
⎢ .. ⎥ points in the hull coordinate system was represented by matrix .
⎣ .

Finally, as shown in Eq. (8), based on FFD, we can obtain the final
0 0 Do hull point set S f based on the lattice space and the position matrix
where:  of the lattice of control points,
S f = F F DSe , B, O, , (8)
αd - the draft change rate,
Do - the draft of the initial hull form model. where:

At the third sub-stage of the geometry modification process, we F F D - FFD operation.


adjust hull form shape based on the FFD technique on the interme- The entire geometry modification process of the hull form is
diate hull point set SI . Since FFD deforms the object based on the shown in Fig. 2. The geometry modification method adopted in
control points attached to the lattice coordinate system, and only this paper provides an excellent way to summarize the whole ge-
the object point inside the lattice space can be modified, we need ometry modification process with a smaller number of representa-
to determine the shape and position of the lattice space with the tive parameters which can represent the hull form modified from
intermediate hull point set SI to ensure every hull point located the KCS hull model. In this work, the geometry modification pa-
inside the lattice space. Meanwhile, all hull samples in the dataset rameters used to express the hull form include length change rate,
have different principal dimensions that lead to the position, and breadth change rate, height change rate, draft change rate, and the
the size of the lattice space for each sample is different and needs movement amount of all control points. By modifying the principal
to be determined in a fixed mode for data validity. The lattice form dimension modification parameters, we can change the principal
chosen in this work is a parallelepiped lattice space, which is ap- dimensions of the final hull model, and by modifying the move-
propriate for a slender object like ship hull model, and it is more ment amount of control points, we can change the final hull form.
convenient to be adopted in real design process, In this work, we
use the extreme coordinate value of the intermediate hull point 2.3. Various measures to prevent irrational hull models
set SI to get the position and the size of the lattice space for easy
operation, where we compare all hull points inside it to obtain the Since training a DL model requires a large dataset, it will take
maximum and minimum coordinate values on each axis separately. much time to generate the dataset by manually changing the de-
Assume that the maximum coordinate value of all points in a hull formation parameters. Although the new hull model generated

18
Y. Ao, Y. Li, J. Gong et al. Journal of Ocean Engineering and Science 8 (2023) 15–32

Fig. 2. Entire geometry modification process of the hull form.

by artificially changing the deformation parameters can make the When the hull point of a hull is too close to the XoZ plane or
dataset more intuitive, it is difficult for the artificially generated crosses the XoZ plane to the area where Y > 0, then that hull will
hull dataset to cover the entire deformation space. Moreover, when not meet the engineering requirements and become irrational. The
the designer artificially changes the deformation parameters to hull points being too close to or crossing the XoZ plane are caused
generate a new hull model, the designer will give more consid- by the preset control points’ offset range. The offset range of the
eration to the hull model that he or she considers reasonable due control point determines the hull’s deformation space, and a suffi-
to the designer’s subjective influence. Considering the above rea- ciently large offset range of the control point can ensure that the
sons, we use random distribution functions to randomly change hull also has a large enough deformation space. To reduce the DL
the deformation parameters to specify the ship hull’s deformation. model’s training time, when setting the control point offset range,
Randomly changing the deformation parameters will save time and we will compress the control point offset range as much as pos-
ensure that the generated dataset elements can be distributed as sible while ensuring that the deformation space is large enough.
evenly as possible in the entire deformation space. However, since In this paper, the offset range of the control point will change ac-
we will randomly change all deformation parameters and the gen- cording to its location. Specifically, the control points near the bow
erated dataset is large, irrational hull forms will inevitably ap- and stern will have a more extensive offset range, while the control
pear during the hull generation process. Moreover, if we ignore points near the parallel middle body of the hull will have a smaller
these irrational hulls, this will cause the finally trained DL model’s offset range. We give different offset ranges for the control points
prediction accuracy to decrease. To avoid the appearance of irra- in different regions to reflect our different degrees of emphasis on
tional hulls, we have taken several measures to avoid irrational different regions. By giving the control points close to the bow and
hulls. stern a more extensive offset range, we ensure that the hull’s bow
The possibility of generating irrational hulls can be reduced by and stern part have a more extensive deformation range. Since the
adopting different truncated normal distributions and setting dif- hull surfaces at the bow and stern of the KCS are more complex
ferent standard deviations. However, due to the need to generate and have large varying curvatures, the more extensive offset range
many hull models, there is no guarantee that irrational hulls will makes the hull points near the bow and stern more likely to be too
not appear after deformation. After analyzing the irrational hull close to or crossing the XoZ plane. When generating the dataset,
we generated, we found that three reasons mainly cause the ir- most of the irrational hulls are caused by the hull points being too
rational hulls. The first and the most important reason is that the close to or crossing the XoZ plane. Even if only one hull point ap-
hull points are very close to or even crossing the XoZ plane. When proaches or crosses the XoZ plane, the deformed hull is irrational.
the control point approaches the XoZ plane, the hull point will also If we do not take any measures to avoid such errors, there will be
approach the XoZ plane. When the hull point is very close to the many invalid elements in the dataset. To solve this kind of error,
XoZ plane or even crosses the XoZ plane to the area where Y > 0, we correct the Y coordinate value of the hull point to ensure that
the width of a particular hull area will be very small or even be- it will not be too close to or cross the XoZ plane. In this way, we
come a negative value. From an engineering perspective, the hull can avoid such errors and obtain a thinner hull without irrational
at this time is irrational. The second reason is that the distance hull conditions. Eq. (9) expresses the relationship between the hull
between the hull points after deformation is too large. When the point after correction and the hull point before correction.
distance between the hull points is too large, the panel area will be x̄i = xi
too large, which will cause the accuracy of the calculation result of
yi , yi ≤ ys
the panel method to decrease. The third reason is that the local ȳi = (9)
ys · e−a(yi −ys ) , yi > ys
deformation of the hull is too large. Specifically, when a particular
area of the hull has an immense change and other areas have small z̄i = zi on i = 1, . . . , h
deformations, the calculation core based on the potential flow the- where:
ory may not handle such a hull. Even if it can, the result will be
inaccurate. (xi , yi , zi ) are ith hull point coordinates before correction,
(x̄i , ȳi , z¯i ) are ith hull point coordinates after correction,

19
Y. Ao, Y. Li, J. Gong et al. Journal of Ocean Engineering and Science 8 (2023) 15–32

ys is small constant for decide boundary, Table 2


Two parameters used to pre-
a is correction rate.
vent the appearance of irra-
tional hulls.
As shown in Eq. (9), we use the constant ys to determine
whether a hull point being too close to or crossing the XoZ plane. Parameters Actual values
Since the Y-coordinate values of all hull points are negative, if the τ 2.0
Y-coordinate value of a hull point is less than ys , then the hull ψ 0.8
point has enough distance from the XoZ plane, which is accept-
able to us. On the contrary, when the Y-coordinate value of a hull
point is greater than ys , the hull point is too close to the XoZ plane, hull points is too large according to the matrix W . First, we stipu-
which will not meet the requirements we set in advance. When late that the maximum element in the matrix W , i.e. Wmax , is not
the Y-coordinate value of a hull point is greater than ys , we will greater than a constant τ that we set in advance. Second, we stipu-
correct the hull point. Eq. (9) is equivalent to setting a repelling late that the standard deviation sW of all matrix W elements is not
point for each hull point, which ensures that the closer the hull greater than the constant ψ we set in advance. Therefore, in gen-
point is to the repelling point, the greater the rejection. eral, judging whether the new hull is reasonable or not depends
The second reason for the appearance of an irrational hull is on whether or not the matrix W satisfies the conditions posted in
that the distance between the hull points after deformation is too Eq. (13),
large. When the distance between the hull points of the deformed
hull is too large, the panel’s size will be too large, making the cal- Wmax ≤ τ
(13)
culation accuracy decrease. Unlike when the hull points are too and sW ≤ ψ
close or crossing the XoZ plane, which violates the engineering re- Adjusting parameter τ can ensure that the distance between ad-
quirements, the hull with a too large distance between the hull jacent hull points will not increase excessively, therefore ensuring
points does not violate the engineering requirements. The exis- that the panel area will not become too large. While adjusting pa-
tence of hulls with too large distances between hull points will rameter ψ ensures that the entire hull will not have sizeable local
make the accuracy of the dataset elements inconsistent, which will deformation. If the new hull matrix W does not satisfy these two
affect the prediction accuracy of the final DL model. We need to conditions in Eq. (13), we will deem the new hull as an irregu-
ensure that the accuracy of the elements in the dataset is roughly lar or irrational hull form. After several attempts, we can choose a
the same. A DL model trained based on a dataset with roughly the control parameter group with smaller values, which leads to a high
same element accuracy can predict the correct change trend of the pass rate for generating ship hull form. Table 2 shows the two pre-
hydrodynamic results with the change of deformation parameters set parameters used to prevent the appearance of irrational hulls.
and guide the design. Simultaneously, excessive local deformation
of the hull may also lead to a decrease in calculation accuracy, so
3. Data collection
we also need to avoid this type of hull.
To prevent an irrational hull from appearing in the dataset, we
3.1. Hull performance acquisition method
need to determine that the deformation of a ship hull is reason-
able every time when we generated a new hull. In this work, we
One sample in the dataset needs to contain hull form and cor-
determine whether the deformation of a ship hull is too large or
responding performances to provide input and output to the DL
the distance between the hull points is too large by judging the
model. All methods that can acquire the performance of the hull
change in the distance between the hull points before and after
form can be adopted. The acquisition method can be a perfor-
the deformation. As shown in Eq. (10), suppose that a hull model
mance evaluation method to calculate the performance of each
contains total h hull points, then we first define a matrix U ∈ Rh×k
sample in the dataset or a collection method to collect exist-
to show the distance between each hull point and the k other hull
ing samples to avoid computational work. The sample collection
points closest to it in the hull before FFD deformation.
method is not suitable for the acquisition method used here since
⎡ ⎤
U11 ... U1k this paper’s unique geometry modification method. We can use
U= ⎣ .. ..
.
.. ⎦ (10)
many calculation methods to evaluate the performance of the hull
. . form, from the high-precision CFD method based on Navier–Stokes
Uh1 ... Uhk h×k to the potential flow method based on assumptions.
After performing FFD deformation, we can obtain a new matrix The choice of calculation method depends on the design stage
based on U, and then we can define the new matrix as V ∈ Rh×k . and design purpose assisted by the DL model. In detail, because
The matrix V shown in Eq. (11) reflects the distance between the designers need to determine the principal dimensions and hull
hull point and the other k hull points closest to it (based on the form of the ship in the initial design stage, the design space in the
hull before the deformation) after deformation. initial design stage will be larger than that in other design stages,
⎡ ⎤ which leads to the dataset for training the DL model who guides
V11 ... V1k the initial design should contain more hull samples. It will cause
V = ⎣ .. ..
.
.. ⎦ (11) colossal computing time when using the high-fidelity CFD method
. .
to establish a dataset containing many samples. In contrast, design-
Vh1 ... Vhk h×k ers need minor and precise modifications to the hull form in the
Based on matrix U and matrix V , we can base on Eq. (12) to get detail design stage, making the DL model that guides the detail de-
the deformed hull’s distance change rate matrix named W ∈ Rh×k . sign needs a dataset with low number requirement but high pre-
⎡ ⎤ cision samples.
V11 /U11 ... V1k /U1k The purpose of the DL model established in this paper is to
W =⎣ ..
.
..
.
..
.
⎦ (12) assist the hull design process, which belongs to the initial design
stage, and because of the need to predict the performance of hull
Vh1 /Uh1 ... Vhk /Uhk h×k models with different main sizes and hull forms, it has a larger
After we obtain a new hull form, we can determine whether the design space. Simultaneously, since the hull model needs to be
new hull’s local deformation is too large or the distance between rechecked and fined at several design stages after the initial de-

20
Y. Ao, Y. Li, J. Gong et al. Journal of Ocean Engineering and Science 8 (2023) 15–32

sign, the accuracy of the evaluation results in the initial design Table 3
The variation range of part of the deformation parame-
stage is lower compared with subsequent design stages, making
ter.
the potential flow method can satisfy the accuracy requirement.
To sum up, in this work, we choose the potential flow method as Geometry modification parameter Variable range
the hull performance acquisition method to calculate the perfor- αx [0.8, 1.2]
mance of each sample. The role of CFD simulations in this work αy [0.8, 1.2]
is generating data and establishing the dataset to tune and train αz [0.8, 1.2]
αd [−0.4, 0]
the neural network model. The hull model of each sample in the
αf [0.7, 1.3]
dataset is generated by randomly adjusting the geometry modifi-
cation parameters, and the total resistance of each sample in the
dataset is evaluated based on the CFD method.
1
−φzz (φx ϕx + φy ϕy ) = φzz (φx2 + φy2 − U 2 )
3.2. The total resistance calculation theory 2
1 1
− φx (φx2 + φy2 )x − φy (φx2 + φy2 )y
To reduce the computational workload when constructing the 2 2
dataset, we treat the resistance performance of the hull as the only on z = 0 free surface. (17)
evaluation indicator. The resistance performance of the hull is one
where:
of the most critical factors when designers evaluate the hull form,
which will directly or indirectly affect many performances of the • g - acceleration of gravity,
ship, including but not limited to design speed, engine horsepower, • U - constant advance speed of the ship.
and fuel consumption. Considering the verification purpose of the
proposed method, only the situation that the ship hull sailing in The total resistance computation program used in this work is
still water has been considered. an in-house computer code developed by the authors’ laboratory
The total resistance of a ship sailing in still water consists of [26], which is based on the classic ITTC friction resistance for-
many sources, including wind, hull surface, and ship appendages. mula and the Dawson method. The validation and the accuracy
According to the simplification guideline provided by the Interna- of the method can be found in the previous works, e.g. [26,27].
tional Towing Tank Conference (ITTC) [24], the total resistance Rt The in-house computer code can be run in parallel in a personal
was decomposed as frictional resistance R f and wave-making re- computer, further shortening the calculation time to establish the
sistance Rw . In this paper, we adopt the classical ITTC frictional dataset.
resistance formula [24] to calculate the frictional resistance and
adopt Dawson’s method [25] to calculate the wave-making resis- 3.3. Design space of samples in the dataset
tance. The classic ITTC friction resistance formula plays a vital role
in resistance estimation, with the accuracy that can satisfy engi- The final valid space for the prediction of the DL model de-
neering requirements when dealing with most ship types, which pends entirely on the design space of the samples in the dataset.
follows Eq. (15). The larger the design space of the sample, the wider the effective
space of the final DL model, which also means that more samples
1
Rf = ρU 2 SC f (14) are needed to train the DL model. Since in this paper, we have rep-
2 resented the geometry modification process of the hull model into
a set of geometry modification parameters, the design space of all
0.075 samples depends on the variation range of geometry modification
Cf = (15)
(lgRe − 2 )2 parameters.
where: Moreover, we expand the design space by adding speed param-
eters to the sample to help designers further reduce the workload
• ρ - density of water, in the initial design stage. Each sample in the dataset has a differ-
• U - constant speed,
ent sailing speed, which means the final DL model will also have
• S - wetted surface area,
the ability to predict the total resistance of the hull model under
• C f - friction coefficient,
different Froude numbers. As shown in Eq. (18), the sailing speed
• Re - Reynolds number.
also has been transformed to a speed parameter, just like the ge-
Since Dawson’s method has high calculation efficiency, the ometry modification parameters.
wave-making resistance calculation based on it usually only needs
v = vd · α f (18)
a few minutes to complete one calculation case. Meanwhile, due to
its lower CPU requirements, Dawson’s method-based computation where:
code can be run in parallel computations in a personal computer.
In Dawson’s method, the total flow ψ is assumed to be the com-
• v - actual Froude number,
position of double-body flow φ and wavy flow ϕ , as follows,
• vd - initial Froude number,
• α v - Froude number change rate.
ψ = φ + ϕ. (16)
As shown in Table 3, we first set the range of principal di-
The double-body potential flow and wavy flow satisfy differ- mension modification parameters, including length change rate,
ent boundary conditions: the double-body flow must satisfy the breadth change rate, height change rate, draft change rate, and
boundary conditions on the hull surface, while the wavy flow must Froude number change rate, which determines the design space
satisfy the boundary conditions on both the hull surface and the of the hull model specification. In this paper, we set the vari-
free surface. After twice expansions based on the Taylor series, the able range of length change rate, breadth change rate, and height
free surface condition can be written as follows, change rate can be changed from 0.8 to 1.2. Moreover, since the
φx ( φx ϕ x + φy ϕ y ) x + φy ( φx ϕ x + φy ϕ y ) y draft of the ship will be changed drastically considering the differ-
1 1 ent situations of fully loaded and unloaded, we have given a range
+ ϕx (φx2 + φy2 )x + ϕy (φx2 + φy2 )y + gϕz of the draft change rate to simulate the unloaded condition, which
2 2
21
Y. Ao, Y. Li, J. Gong et al. Journal of Ocean Engineering and Science 8 (2023) 15–32

Table 4 the control points, and Table 6 shows the location of different ar-
Parameters used to determine the lattice space.
eas and the movement range of the control points in the area.
Parameters x-axis y-axisn z-axis Based on the preset design space and the aforementioned mea-
Offset rate 0.05 0.05 0.05 sures to prevent unreasonable hull models, we can obtain as many
Rotation angle 0.0 0.0 0.0 hull model samples as possible to train the DL model. Benefitting
N control points 20 4 6 from our setting for the movement range of the control points ac-
cording to different hull areas, the randomly generated hull sam-
Table 5 ples will continue to maintain a long parallel mid-body, and the
Location of different areas and their fixed directions. middle part is basically flat instead of a lot of bumps and de-
Boundary area Area location Fixed coordinate
pressions. Moreover, due to the large movement range of the con-
trol points near the bow and stern, we can make drastic geome-
x ∈ [−0.236, 7.063]
try modifications to the bow and stern. Fig. 5 shows a compari-
Boundary-1 y ∈ [−0.572, 0.063] X
z ∈ [−0.400, 0.455] son between a randomly deformed hull model and the initial hull
x ∈ [−0.236, 0.266] model, where we can notice the drastic geometry modifications at
Boundary-2 y ∈ [−0.572, 0.063] X, Z the bow area and a relatively minor modification at the mid-body.
z ∈ [−0.400, −0.105]
x ∈ [7.0625, 7.588]
3.4. Data collection
Boundary-3 y ∈ [−0.572, 0.063] X, Z
z ∈ [−0.012, 0.455]
Training a DL model requires a dataset contains with large
enough samples. In order to save the data collection process, we
can be changed from −0.4 to 0. Further, the Froude number change randomly generate geometric modification parameters according
rate can be changed from 0.7 to 1.3. to the preset variable range to generate each hull model in the
The number of control points determines the fineness of the dataset. For the principal dimension modification parameters, we
geometry modification of the hull form. The more control points, used continuous uniform distributions to generate them to make
the finer the geometric modification of the hull form. In this work, sure all samples can evenly distribute in the principal dimension
we created a (20 × 4 × 6 ) lattice of control points, where exist 20 design space. Besides, we used the truncated normal distribution
control points on the x-axis, 4 control points on the y-axis, and 6 [28], which is a normal distribution with upper and lower limits in
control points on the z-axis. As shown in Table 4, we pre-defined sample space, to generate the movement amount of control points.
the parameters used to determine the lattice space. The truncated normal distribution makes the movement amount
All control points are evenly distributed in the lattice space, of control points concentrated near the mean value, ensuring the
moving freely in three axis directions. To avoid unreasonable hull hull form after geometry modification will be more likely similar
forms and reduce the number of the geometric modification pa- to the initial KCS hull. Besides, since the more similar the hull form
rameters, we select different fixed directions for different control is to the initial hull form, the easier it is for the hull form to meet
points, which is determined by the area where it is located. Fig. 3 engineering requirements, and using a truncated normal distribu-
shows the area division depends on different fixed directions, and tion also means the hull form after geometry modification is more
Table 5 shows the location of different areas and their fixed di- likely in line with engineering requirements.
rections. As shown in Fig. 3 and Table 5, we do not make any re- The degree of similarity between the hull form after geometry
strictions to the bulb bow part, which ensures the fineness of the modification and the initial hull form can be adjusted by changing
geometry modification of the bulb bow. We make sure the control the parameters in the truncated normal distribution. In detail, the
point in Boundary-1 cannot move in the x-axis direction. Further- smaller the standard deviation of the truncated normal distribu-
more, the control points located in Boundary-2 and Boundary-3, tion, the larger possibility the hull form keeps similar to the initial
which are close to the upper structure and propeller, are restricted hull form. The larger the standard deviation of the truncated nor-
to be can only move in the y-axis direction. mal distribution, the greater the possibility of significant geome-
Similar to the principal dimension modification parameters, the try modification. Eq. (19) provides the probability density function
amount of the movement of the control points determines the ge- f (x ) of the truncated normal distribution, which is determined by
ometry modification space of the hull form. Different areas of the the mean μ, the standard deviation σ , low boundary value a, the
ship have different degrees of influence on the ship hull perfor- upper boundary value b, and the probability density function φ
mance, where the geometry modification of the bow and stern has and the cumulative distribution function of the standard normal
a more significant impact on the performance of the ship, and the distribution,
midship has a relatively low influence on the performance of the 1 φ ( x−σμ )
ship. To reflect the degree of influence for different hull areas, we f (x; μ, σ , a, b) = . (19)
σ ( b−σμ ) − ( a−σμ )
set up different control point movement ranges according to differ-
ent areas, making sure we can reduce the difficulty of training the Using continuous uniform distributions and truncated normal
DL model while ensuring sufficient geometry modification space. distributions to generate hull model samples can significantly re-
Fig. 4 shows the area division depends on the movement range of duce the time required for the data collection process. Meanwhile,

Fig. 3. Area division depends on different fixed directions.

22
Y. Ao, Y. Li, J. Gong et al. Journal of Ocean Engineering and Science 8 (2023) 15–32

Fig. 4. Area division depends on the movement range of the control points.

Table 6 ples. The performance evaluation process of the dataset was con-
Location of different areas and the movement range of the con-
ducted on a personal laptop computer (Alienware Area-51 m, CPU:
trol points in the area.
Intel Core I9-9900 K, 3.60 GHz, RAM: 32.0 GB, and GPU: Nvidia
Area Location Direction Scope Geforce RTX 2070). After performing evaluation and sample valid-
x ∈ [7.063, 7.588] ity tests, we generated a dataset that contains 31,058 hull samples
SABED y ∈ [−0.572, 0.063] X [−0.1, 0.1] inside it.
z ∈ [−0.400, −0.012]
x ∈ [−0.236, 7.588]
SACQM y ∈ [−0.572, 0.063] Y [−0.1, 0.1]
z ∈ [−0.400, 0.455] 4. Construction and training of the deep learning model
x ∈ [−0.236, 7.588]
SABPM y ∈ [−0.572, 0.063] Y [−0.2, 0.2] 4.1. Fully connected neural network (FCNN)
z ∈ [−0.400, −0.012]
x ∈ [−0.236, 0.266]
SKLQP y ∈ [−0.572, 0.063] Y [−0.3, 0.3] DL technology has a unique ability to extract internal relation-
z ∈ [−0.105, 0.455] ships from complex or even inaccurate data, thereby finding a
x ∈ [6.477, 7.588] mathematical model that is too complex for the human brain or
SABHG y ∈ [−0.572, 0.063] Y [−0.3, 0.3] other computer technologies [8]. In recent years, DL technology
z ∈ [−0.400, −0.012]
x ∈ [7.063, 7.588]
has developed rapidly and has been found to have many successful
SABED y ∈ [−0.572, 0.063] Y [−0.4, 0.4] applications in many other fields not limited to computer science
z ∈ [−0.400, −0.012] [29].
x ∈ [−0.236, 0.266] As one of the most representative DL methods, Fully Connected
SKLQP y ∈ [−0.572, 0.063] Z [−0.2, 0.2]
Neural Network (FCNN), also known as Artificial Neural Network
z ∈ [−0.105, 0.455]
x ∈ [6.477, 7.588] (ANN), is a DL technology-based parallel computational machine
SABHG y ∈ [−0.572, 0.063] Z [−0.2, 0.2] learning method, whose feature is that neurons will receive the
z ∈ [−0.400, −0.012] processed information from all neurons in the previous layer. Fig. 6
x ∈ [7.0625, 7.58811] shows a schematic diagram of the topology structure of FCNN and
SABED y ∈ [−0.572, 0.063] Z [−0.3, 0.3]
z ∈ [−0.400, −0.012]
a single neuron processing unit.
Different layers perform different information transformation
purposes in the FCNN model. Specifically, the input layer respon-
sible for receiving data information from the outside world, which
does not contain any calculation work. The hidden layer is equiva-
lent to the information processing system, which is responsible for
the main calculation work. Relying on the hidden layer, an FCNN
model can extract internal rules from the dataset and store “mem-
ory” and “experience” in it. The most distinguishing feature of the
FCNN model is that every hidden layer is fully connected with two
adjacent layers, where each neuron in the layer l receives all neu-
ron processed information in the layer (l − 1 ) and then distributes
its processed information to all neuron located in the layer (l + 1 ).
The output layer located at the end of the FCNN model is respon-
sible for exporting predicted results to the outside world.
In this work, the input layer accepts the geometry modification
parameters used to describe different hull forms, where each neu-
ron in this layer represents an individual geometry modification
parameter from a given sample in the dataset. The output layer,
Fig. 5. Comparison of deformed hull model and initial hull model. which coalesces the influence caused by the replacement of ge-
ometry modification parameters, exports the current inputted hull
form performances. Due to the structure topology characteristics,
not like manually generating hull model samples, randomly gener- when the FCNN model faces a large number of input parameters,
ating samples can make sure all samples can be as much as pos- it will become computationally intensive and prone to overfitting,
sible evenly distributed in the entire design space and also avoid leading to poor performance in specific tasks. However, many prac-
the subjective influence of designers. tices have proved that the FCNN model has shown stable perfor-
Before and after calculating the performance of hull samples, mance in dealing with non-image-related tasks. Hence, consider-
we checked the hull samples in the dataset twice to ensure the ing that we need to build a model to simulate the function be-
validity of the samples. Specifically, before calculating the perfor- tween the geometry modification parameters and the total resis-
mance of hull samples, we checked the hull forms in the dataset to tance, which does not involve image processing, we choose the
ensure there are no unreasonable hull forms. After calculating the FCNN model as the DL model used in this work . Besides, unlike
performance of hull samples, we checked the results and the wave classification tasks, it is a better choice to keep the information
pattern of all hull samples and eliminated all incorrect hull sam- channels of all neurons connected to uncover the actual mapping

23
Y. Ao, Y. Li, J. Gong et al. Journal of Ocean Engineering and Science 8 (2023) 15–32

Fig. 6. The schematic diagram of FCNN topology structure and single neuron processing.

Fig. 7. Simplified schematic of the FCNN model.

relationship of regression tasks. Therefore, we choose the FCNN setting the activation function is to add nonlinear properties to the
model as the DL model used in this paper. DL model. To make the DL model easy to train without the vanish-
ing gradient problem, the neurons in the input layer and hidden
4.2. FCNN model construction layer adopt the ReLU activation function [30], which is defined as
the positive part of its argument,
The DL model adopted in this paper is the FCNN model. A sim- a = f (x ) := x+ = max(0, x ). (20)
plified schematic illustration of FCNN model is shown in Fig. 7. The
input layer of the DL model contains 501 neurons, including length Compared with the sigmoid function, ReLU is not easy to cap-
change rate, width change rate, height change rate, draught change ture the performance of vanishing gradient, in which even when
rate, Froude number change rate, and the movement amount of all the activation function input reaches the boundary, the gradient of
control points. It is worth noting that the order of all input neurons the neuron will not be too small. We adopted a linear activation
must be the same, and when we want to verify and use the DL function for neurons located in the output layer to ensure predic-
model, we must also ensure that the order of input neurons does tion accuracy, which is given as follow,
not change. The output layer of the DL model contains one neuron,
a = f (x ) = x. (21)
which exports the total resistance of the current hull model.
Before tuning the FCNN model, we need to set several hyper- The loss function is an indicator for evaluating the completion
parameters to make the FCNN model ready to be tuned, includ- degree of DL model training, where the lower the value of the loss
ing activation function, loss function, and the method to judge the function, the higher the training completion of the DL model. The
prediction accuracy. The activation function is a mathematical con- loss function plays the role of supervisor and guides the update
verter that can transform the information of the current neuron of weights and biases. This paper adopted the Mean Absolute Per-
into the information flowing into the next neuron. The purpose of centage Error (MAPE) [31], which is given as follow to direct the

24
Y. Ao, Y. Li, J. Gong et al. Journal of Ocean Engineering and Science 8 (2023) 15–32

DL model to update to reduce the gap between the predicted re- used in this work, we need to determine hyperparameters includ-
sult and the actual result, ing hidden layer structure, regularization parameters, dropout rate,
learning rate, batch size, and training epochs.
1  yˆi − yi
N
Loss = | | (22) For a supervised learning problem, it usually needs to divide
N yi the dataset into two parts: training dataset and testing dataset.
i=1
In this research, we take 75% of samples in the dataset as the
where: training dataset while using the remaining 25% of samples in the
dataset as the testing dataset. In the actual computation, the to-
• N - total number of samples in dataset,
tal 31,058 hull structure samples were divided randomly into the
• yˆi - ith predicted value,
ratios 23,293:7765.
• yi - ith actual value.
The hidden layer structure determines the ultimate prediction
The loss function value is an evaluation index that is easy for capability of the DL model. Considering that we have selected
computers to understand since the only point that the machine FCNN as the DL model, we only need to determine the number of
needs to pay attention to reduce the value of the loss function as hidden layers and neurons in each hidden layer. We adopted L1, L2
much as possible. However, when comparing the prediction accu- regularization methods and the dropout technique to avoid over-
racy between DL models, the value of the loss function becomes fitting, in which we add a correction term to the loss function as
incomparable considering the different regularization parameters shown in Eq. (25),
and different hidden structures. In this paper, we choose the mean  
ˆ = Loss + λ1
Loss |β j | + λ2 β 2j (25)
and standard deviation of the absolute percentage difference of
j j
samples in the testing dataset as the indicators to determine the
prediction accuracy between different DL models to avoid the in- where:
fluence caused by the different loss functions, which is shown as
• ˆ - revised loss function,
Loss
follows,
n v • λ1 - L1 regularization rate,
i=1 pi • λ2 - L2 regularization rate,
Emean = (23)
nv • β j - jth parameter in the DL model.

 There are other techniques that one may use to avoid overfit-
nv
i=1 ( pi − μ )2 ting, such as inactivating neurons randomly.
and Estd = (24) The learning rate controls the learning speed of the DL model,
nv
where the lower the learning rate, the slower the learning speed of
where: the DL model. We can regard the training process of the DL model
as to find extreme values in the n-dimensional parameter space (n
• nv - total number of samples in the testing dataset,
is the number of the DL model parameters), and the learning rate
• pi - ith absolute percentage error,
determines the step size of the finding process. Considering the
• μ - population means. DL model is far away from the extreme point at the beginning of
Note that Emean represents the average of the absolute percent- the training process, then an appropriately large learning rate can
age differences of all samples in the testing dataset, representing help shorten the epoch number required to approach the extreme
the average prediction accuracy of the DL model. The smaller Emean value. Once the DL model is close to the extreme value, a lower
of a DL model, the higher the prediction accuracy the DL model is. learning rate is assigned to the DL model, which ensures that the
Estd is the standard deviation of the absolute percentage difference DL model can continuously get close to the extreme value instead
of all samples in the testing dataset, representing the degree of of oscillating near the extreme value. This process is described by
normalization of a DL model when predicting different hull forms. the following equation,
The smaller Estd of the DL model, the higher the degree of normal- ini_lr
ization of a DL model. lri = (26)
1 + i · dr
The construction, training, and testing of the DL models were
all developed based on the Keras framework [32] with the Tensor-
Flow backend [33]. In the developed python code of the computer ini_lr
and dr = (27)
program, the (keras.models.Model) was adopted to construct the te
FCNN model, with (keras.layers.core.Dense) to add the hidden layer. where:
The regularization parameters were added into the DL model by
adopting (keras.regular izer s.l1_l2) and (keras.layers.core.Dropout).
• i - the current epoch number,
The optimizer was established based on (keras.optimizers.Adam),
• lri - the actual learning rate of ith epoch,
with the access to adjust the initial learning rate and decay rate.
• ini_lr - the initial learning rate,
All the tuning, training, and testing processes were conducted
• dr - the decay rate,
on a personal laptop computer with the following specifications:
• te - the total number of training epoch.
Alienware Area-51 m, CPU: Intel Core I9-9900 K, 3.60 GHz, RAM: Here, we adopt a variable learning rate strategy that has the
32.0 GB, and GPU: Nvidia Geforce RTX 2070. advantage to achieve large learning rate at the early stage while
attain a small learning rate at the late stage, as shown in Eq. (26),
4.3. Hyperparameters tuning where the current learning rate is related to the initial learning
rate, and the learning decay rate is determined by the epoch num-
The tuning process is one of the most time-consuming stages ber. As shown in Eq. (27) the learning decay rate will decrease with
in developing a deep learning method because almost all the hy- the increase of the epoch number.
perparameters are independent, and various hyperparameter com- Note that the number of epochs is a hyperparameter that de-
binations need to be tested to obtain the best or optimum com- fines the number of times that a learning algorithm will work
bination. In the hyperparameters tuning process for the DL model through the entire training dataset. One epoch means that each

25
Y. Ao, Y. Li, J. Gong et al. Journal of Ocean Engineering and Science 8 (2023) 15–32

Table 7
Default hyperparameters at the begin-
ning of the tuning process.

Hyperparameters Default value

L1 regularition rate 0
L2 regularition rate 0
Dropout rate 0
Batch size 4096

sample in the training dataset has had an opportunity to update


the internal model parameters.
An epoch has one or more batches, and the batch size is the
number of samples fed in the DL model in each epoch. The pur-
pose of the application of batch size is to reduce the memory re-
quired during training. Batch size is the amount of training data
that the computer needs to process simultaneously, making the
lower the batch size, the less memory required during the train-
ing process. Nevertheless, too small a batch size may result in
inaccurate gradient values, leading to possible oscillations during Fig. 8. Loss vs. learning rate curve.
training. Training epochs represent the total epoch number of the
dataset through the DL model during the training process. In this Table 8
research, the batch size was fixed to 4,096, considering the sample Parts of tuning result of hidden structure.
number in the training dataset. The training epochs are always the
Hidden sturcture Emean (% ) Estd (% )
larger, the better, and we should set the training epochs as high
as possible and decide whether to stop based on the loss function (64) 24.898 24.131
(256, 4) 15.386 12.712
value.
(32, 32) 13.473 12.071
In the hyperparameter tuning process, the most straightforward (128, 4, 2) 9.882 8.357
method is to try all hyperparameter combinations to select the
best one. However, if we try all the possible combinations in actual
operation, then we need to try tens of thousands of combinations, Table 9
Parts of tuning result of regularization parameters.
which is unacceptable in workload. Thus, we used some techniques
to reduce the number of attempts required when doing the hy- L1 rate L2 rate dropout rate Emean (% ) Estd (% )
perparameters tuning process. In the beginning, we first keep all case 1 0 0 0 26.167 194.706
hyperparameters unchanged except the hidden layer structure and case 2 0.004 0.001 0.001 8.314 6.985
change the hidden layer structure of the DL model to obtain the case 3 0 0.128 0 6.512 5.752
case 4 0.016 0 0.002 5.389 5.182
most suitable hidden structure for predicting the total resistance
of hull models, where the default hyperparameters are shown in
Table 7.
We first determined the range of hidden layers number and the the DL model has a tiny change, expressed the increase of learn-
range of neurons number after comparing the capability and train- ing rate here can not improve too much training speed. By the
ing time. We first set the number of hidden layers between 1 and time when the learning rate reaches 10−1 , the loss value of the
3. We also stipulate that the neurons in each layer are between 2 DL model begins to increase, meaning the learning rate is far too
and 501 and must be a power of two. Simultaneously, considering large. Based on the curve, we choose 5E − 4 as the learning rate
that the number of input parameters is greater than the number and choose 20 0 0 as the training epoch number used for hyperpa-
of output parameters, we stipulate that the number of neurons in rameters tuning, which gives the DL model a fast learning speed.
the back layer will not larger than the number of neurons in the Depending on the hidden layer structure, every step costs 9 and 10
front layer. We can obtain all the hidden layer structure possibili- us. We calculated 164 possible hidden layer structures, and Table 8
ties through the above regulations and obtain the optimal hidden shows parts of tuning results with different hidden structures.
layer structure by training all DL models and comparing the indi- Fig. 9 (a)–(d) show the training history of the four DL models
cators. mentioned in Table 8. As can be seen from these four images, we
In this research, we adopted a rate finding method developed can find that the testing loss begins to increase, especially at epoch
by Leslie Smith in [34] to automatically find the range of optimal 2,0 0 0, compared with the decrease of the testing loss, meaning the
learning rates, where the whole process of the rate finding method DL model has overfitted, which can be avoided by tuning suitable
was explained in detail in Smith [34]. Based on the rate finding regularization parameters.
method, we create a loss vs. learning rate curve to show how the We choose the hidden layer structure with relatively low Emean
learning rate affects the value of the loss function, shown in Fig. 8. and Emean , no obvious overfitting, and low training loss as the final
As shown in Fig. 8, we can visualize how the learning rate af- choice, which is (256, 4, 2 ). Then we used a hidden layer struc-
fects the training speed of the DL model, where the DL model loss ture (256, 4, 2 ) to do further tuning to determine the regulariza-
value stays stable unless the learning rate exponentially increased tion parameters, which can help us avoid overfitting and further
from 10−10 to 10−8 , which means when the learning rate is too train the DL model. Before tuning the regularization parameters,
low cause the DL model can not learn anything. When the learn- we set the range of all regularization parameters between 0 and
ing rate arrived at 10−8 , the learning rate is just large enough that 0.128. We calculated 729 possible regularization parameter combi-
our model can start learning, and when the learning rate arrived at nations, and Table 9 shows parts of tuning results with different
10−3 , the DL model has a considerable learning speed. When the regularization parameters. Fig. 10(a)–(d) shows the training history
learning rate locates between 10−3 and 10−2 , the learning speed of of the four DL models mentioned in Table 9.

26
Y. Ao, Y. Li, J. Gong et al. Journal of Ocean Engineering and Science 8 (2023) 15–32

Fig. 9. (a) Training history for (64), (b) Training history for (256, 4), (c) Training history for (32, 32), and (d) Training history for (128, 4, 2).

Table 10 ducing the learning rate when needed. The entire training history
Hyperparameters of the DL model at
of the DL model is shown in Fig. 11. Considering the training pro-
the beginning of the training process.
cess lasted for 20,0 0 0 epoch, we divide it into two parts for bet-
Hyperparameters Value ter observation, showing the application of stopping and resuming
Hidden sturcture (256, 4, 2 ) technique.
L1 regularition rate 0.032 Fig. 12 (a) shows the training history of the DL model from
L2 regularition rate 0 epoch 0 to epoch 30 0 0, where we set the initial learning rate as
Dropout rate 0
10−3 and decrease the learning rate from 10−3 to 10−4 at epoch
Batch size 4096
20 0 0. As shown in Fig. 12(a), both the training loss and the testing
loss drop very quickly at the beginning, where the loss value de-
As shown in Fig. 10(a)–(d), the regularization parameters play creases about 98% in the first 100 epoch compare with the whole
an important role in preventing overfitting. By comparing the indi- training history, which means a relatively large learning rate can
cators Emean and Emean and ensure that overfitting does not occur, accelerate the initial training process. During epoch 100 to epoch
we finally set the L1 regularization parameter as 0.032, set the L2 20 0 0, both the training loss and the testing loss appear stochas-
regularization parameter as 0, and set the dropout rate as 0. tic shock, but the overall trend is still declining, showing the DL
Finally, after tuning hidden structure and regularization param- model can still learn based on the current learning rate. The learn-
eters, all the hyperparameters of the final DL model are deter- ing rate was decreased to 10−4 at epoch 20 0 0, followed by a sig-
mined, which is shown in Table 10. nificant drop-down for both the training loss and the testing loss
after epoch 20 0 0.
4.4. Training and testing Fig. 12 (b) shows the training history of the DL model from
epoch 30 0 0 to epoch 20,0 0 0, where we decrease the learning rate
In this research, during the training process of the DL model, from 10−4 to 10−5 at epoch 10,0 0 0. As shown in Fig. 12(b), the
we used the technique of stopping and resuming the training pro- learning rate continued to decrease to 10−5 at epoch 10,0 0 0, fol-
cess so that we can monitor the loss and reduce the loss by re- lowed by a relatively significant drop-down for both the training

27
Y. Ao, Y. Li, J. Gong et al. Journal of Ocean Engineering and Science 8 (2023) 15–32

Fig. 10. (a) Training history for case 1, (b) Training history for case 2, (c) Training history for case 3, and (d) Training history for case 4.

Table 11
Training results of the final DL model.

Loss Emean Estd

Training 3.8844 – –
Testing 5.1630 3.9754% 4.7562%

We choose to stop the training process at epoch 19,900, where


the testing loss reached the minimum in the entire training his-
tory, and we also store the DL model located at epoch 19.900 for
further use. As shown in Table 11, the final training and testing loss
values were close to each other. The Emean shows that the average
difference between the predicted and actual total resistance in the
testing dataset is 3.9754%. The Estd shows that the standard devia-
tion between the predicted and actual total resistance in the test-
ing dataset is 4.7562%. The results show that the DL model defined
and trained in this work performs well in predicting the testing
Fig. 11. The entire training history of the DL model. data, and no severe overfitting problem exists. The final trained DL
model can be used to predict the total resistance of the hull form,
which can evaluate the total resistance of the hull form modified
loss and the testing loss after epoch 16,0 0 0. Although it is not ap- from the KCS to assist the hull design stage.
parent, it can be noticed that the testing loss has risen slightly
between epoch 10,0 0 0 to epoch 20,0 0 0, which means the appear- 5. Model verification
ance of overfitting. The DL model has been continued trained till
it reached epoch 20,0 0 0, where there has no pronounced decrease The weights and biases of the deep learning model can be
for both the training loss and the testing loss. saved to be further used to predict the total resistance of the

28
Y. Ao, Y. Li, J. Gong et al. Journal of Ocean Engineering and Science 8 (2023) 15–32

Fig. 12. (a) The training history from epoch 1 to epoch 3,0 0 0, and (b) The training history from epoch 30 0 0 to epoch 20,0 0 0.

Table 12
The details of the testing cases in step 1.

Testing case Length rate Breadth rate Height rate Draft rate Froude number rate

Case 1 (0.7, 1.3) 1 1 0 1


Case 2 1 (0.7, 1.3) 1 0 1
Case 3 1 1 (0.7, 1.3) 0 1
Case 4 1 1 1 (−0.45, 0.05) 1
Case 5 1 1 1 0 (0.6, 1.4)

hull structure form generated by the geometric modification of the Fig. 13 (a) shows the curves for total resistance predicted by
KCS. The most convenient of the DL model is that once we im- the potential flow method and the DL model when the length
port all geometry modification parameters to the DL model, then change rate in the scope range (0.7, 1.3) but the scope range of
the DL model can immediately evaluate the performance of the the dataset is (0.8, 1.2). For the entire range, the maximum er-
hull form, which can save the modeling and calculating workload, ror is 27.85%, located at 0.7, with the average error of all sam-
and designers do not need to consider any hydrodynamic the- ple points in the entire range equals 6.90%. Moreover, for the do-
ory. In real situations of engineering design of the hull form, the main, the maximum error is 7.90%, located at 0.85, with the av-
hull form after geometry modification will be randomly located erage error of all sample points in the domain is 4.08% which
in the design space, and it is most likely on the interval or gap can meet the engineering requirements of the hull form design
space among training samples. However, the total resistances of stage.
hull forms located in the interval space are not contained in the Fig. 13 (b) shows the curves for total resistance predicted by the
training dataset, and they are very different from training samples. potential flow method and the DL model when the breadth change
Predicting a case not located in the range of training samples is al- rate in the scope range (0.7, 1.3) but the scope range of the dataset
ways a challenge, especially for the DL model in this paper, which is (0.8, 1.2). For the entire range, the maximum error is 13.93%,
needs to deal with the design space with 501 dimensions. To test located at 1.3, with the average error of all sample points in the
the generality of the trained DL model, we performed several val- entire range equals 5.44%. Moreover, for the domain, the maximum
idation cases by comparing the result evaluated through the po- error is 10.85%, located at 1.2, with the average error of all sample
tential flow method and the DL model, in which the hull form af- points in the domain is 4.92%.
ter geometry modification is located in the interval space between Fig. 14 (a) shows the curves for total resistance predicted by the
training samples. potential flow method and the DL model when the height change
Considering that the geometry modification parameters contain rate in the scope range (0.7, 1.3) but the scope range of the dataset
the principal dimension modification parameters and the move- is (0.8, 1.2). For the entire range, the maximum error is 15.64%,
ment amount of control points, we verify the prediction accuracy located at 1.3, with the average error of all sample points in the
of the DL model in three steps in different ways. In the first step, entire range equals 4.92%. Moreover, for the domain, the maximum
we verify the prediction accuracy of the DL model by fixing the error is 6.95%, located at 1.2, with the average error of all sample
movement amount of control points and when only the principal points in the domain is 4.25%.
dimension modification parameters can be changed. To test the Fig. 14 (b) shows the curves for total resistance predicted by the
generality of the DL model, in step 1, we performed all testing potential flow method and the DL model when the draft change
cases when all the movement amount of control points equal to rate in the scope range (−0.45, 0.05) but the scope range of the
0.05, ensuring the test cases do not have a chance to appear in dataset is (−0.4, 0). For the entire range, the maximum error is
the dataset. The details of the testing cases in step 1 are shown in 9.62%, located at −0.28, with the average error of all sample points
Table 12. To be statistically representative, we also check the case in the entire range equals 4.22%. Moreover, for the domain, the
in which the principal dimension is out of the domain by extend- maximum error is also 9.62%, located at −0.28, with the average
ing the range scope of principal dimensions. error of all sample points in the domain is 4.63%.

29
Y. Ao, Y. Li, J. Gong et al. Journal of Ocean Engineering and Science 8 (2023) 15–32

Fig. 13. (a) Curve of total resistance vs. length change rate, and (b) Curve of total resistance vs. breadth change rate.

Fig. 14. (a) Curve of total resistance vs. height change rate, and (b) Curve of total resistance vs. draft change rate.

points in the entire range equals 8.35%. Moreover, for the domain,
the maximum error is 19.12%, located at 1.2, with the average error
of all sample points in the domain is 5.67%.
In step 2, we performed all testing cases for different move-
ment amounts of control points to test the generality of the DL
model, ensuring the test cases do not have a chance to appear
in the dataset. Considering that there are 406 parameters related
to control points among the geometric modification parameters,
it will be complicated to choose the movement amount of con-
trol points one by one. Hence, to test the generality of the DL
model, we generated testing cases using linear functions to gen-
erate the movement amount of control points, avoiding the testing
cases have a chance to appear in the training dataset. Assuming
the movement amount of the first control point parameter is m1 ,
and the movement amount of the last control point parameter is
m496 , then the movement amount of the ith control point param-
Fig. 15. Curve of total resistance vs. Froude number change rate. eter mi is given in the following equation,

mi = (m496
495
−m1 )
( i − 1 ) + m1 . (28)

Fig. 15 shows the curves for total resistance predicted by the Fig. 16 shows the curves for total resistance predicted by the
potential flow method and the DL model when the Froude num- potential flow method and the DL model when m1 was in the
ber change rate in the scope range (0.6, 1.4) but the scope range scope range (−0.1, 0.1) and the m406 was fixed to 0.1. The de-
of the dataset is (0.7, 1.3). For the entire range, the maximum er- tails of Case 6 in step 2 are shown in Table 13. Since the mini-
ror is 33.87%, located at 0.6, with the average error of all sample mum scope range of the movement amount of control points is

30
Y. Ao, Y. Li, J. Gong et al. Journal of Ocean Engineering and Science 8 (2023) 15–32

Table 13
The details of the testing cases in step 2.

Testing case Length Breadth Height Draft Froude m1 m406

Case 6 1 1 1 0 1 (−0.1, 0.1) 0.1


Case 7 1 1 1 0 (0.7, 1.3) 0.1 −0.07
Case 9 1 1 1 0 (0.7, 1.3) −0.07 0.1

Table 14
Predicted errors in method verification.

Testing case Max (entire) Ave (entire) Max (domain) Ave (domain)

Case 1 27.85% 6.90% 7.90% 4.08%


Case 2 13.93% 5.44% 10.85% 4.92%
Case 3 15.64% 4.92% 6.95% 4.25%
Case 4 9.62% 4.22% 9.62% 4.63%
Case 5 33.87% 8.35% 19.12% 5.67%
Case 6 12.50% 3.09% 12.50% 3.09%
Case 7 14.37% 4.68% 14.37% 4.68%
Case 8 15.70% 4.48% 15.70% 4.48%

Fig. 17 (b) shows the curves for total resistance predicted by


the potential flow method and the DL model when m1 was fixed
to −0.07 and the m406 was fixed to 0.1, with the Froude num-
ber change rate in the scope range (0.7, 1.3). For the domain
Fig. 16. Curve of total resistance of step 2.
range, the maximum error is 15.70%, located at 0.7, with the
average error of all sample points in the domain range equals
4.48%.
(−0.1, 0.1), all samples tested in step 2 are in the design space. The predicted errors of all testing cases in method verification
For the entire range, the maximum error is 12.50%, located at are shown in Table 14. As shown in Table 14, the DL model can
−0.1, with the average error of all sample points is 3.09% which predict the total resistance of the hull form model generated by
can meet the engineering requirements of the hull form design the geometry modification of the KCS with high prediction accu-
stage. racy. It is worth noting that all cases tested in this chapter have
In step 3, we verify the prediction accuracy of the DL model never appeared in the dataset, meaning that the DL model we con-
by simultaneity change both the principal dimension parameters structed and trained here can predict a sample it is never seen,
and the movement amount of control points to ensure the con- which gives designers an excellent method to assist the hull de-
sistency for different kinds of geometry modification parameters. sign stage. We can find that the maximum error and average error
The details of Case 7 and Case 8 in step 3 are shown in Table 13. in the entire range are almost always larger than the maximum
Fig. 17(a) shows the curves for total resistance predicted by the error and average error in the domain range through the above
potential flow method and the DL model when m1 was fixed test cases. Meanwhile, the maximum error in the domain range al-
to 0.1 and the m406 was fixed to −0.07, with the Froude num- ways happened in the boundary of the domain range, which shows
ber change rate in the scope range (0.7, 1.3). For the domain that the prediction accuracy is related to the sample number in the
range, the maximum error is 14.37%, located at 0.7, with the training dataset and the distribution density of the samples in the
average error of all sample points in the domain range equals design space, and the higher sample density, the higher prediction
4.68%. accuracy.

Fig. 17. (a) Curve of total resistance of Case 6, and (b) Curve of total resistance of Case 7.

31
Y. Ao, Y. Li, J. Gong et al. Journal of Ocean Engineering and Science 8 (2023) 15–32

6. Conclusions References

In this work, we have developed a deep learning neural net- [1] D. Watson, Practical Ship Design, vol. 1, Elsevier, 1998.
[2] H. Schneekluth, V. Bertram, Ship Design for Efficiency and Economy, vol. 201,
work model to predict the total resistance of ship hull form gen- Butterworth-Heinemann Oxford, 1998.
erated based on the geometry modification of the KRISO Container [3] S. Harries, Tradition and Future of Ship Design in Berlin, Technical University
Ship (KCS), which can assist the hull form design at the prelimi- of Berlin, 2008.
[4] M. Roh, K. Lee, Comput. Ind. 58 (6) (2007) 539–557.
nary ship design stage. [5] A. Papanikolaou, Ship Design: Methodologies of Preliminary Design, Springer,
Expressly, we set the KRISO Container Ship (KCS) as the initial 2014.
hull model and simplified the geometry modification process into [6] F. Mistree, W. Smith, B. Bras, J. Allen, D. Muster, Trans., Soc. Naval Arch. Mar.
Eng. 98 (1990) 565–597.
a set of geometry modification parameters by combining the lin-
[7] J.D. Anderson, J. Wendt, Computational Fluid Dynamics, vol. 206, Springer,
ear transformation method and Free-Form Deformation (FFD) tech- 1995.
nique. By randomly generating hull form samples and adopting a [8] Y. LeCun, Y. Bengio, G. Hinton, Nature 521 (7553) (2015) 436–444.
[9] E. Alpaydin, Introduction to Machine Learning, MIT Press, 2020.
potential flow method to calculate the total resistance of all sam-
[10] Y. Chen, Z. Lin, X. Zhao, G. Wang, Y. Gu, IEEE J. Sel. Top. Appl. Earth Obs. Re-
ples, we established a dataset used to train and test the DL model. mote Sens. 7 (6) (2014) 2094–2107.
The generality test shows that the trained DL model can predict [11] A. Kendall, R. Cipolla, Geometric loss functions for camera pose regression with
the total resistance of hull forms accurately and quickly in the de- deep learning, in: Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, 2017, pp. 5974–5983.
sign stage, which is a remarkable success for applying deep learn- [12] J. Friedman, T. Hastie, R. Tibshirani, The Elements of Statistical Learning, vol. 1,
ing techniques in the hull form design process. Based on this study, Springer Series in Statistics New York, 2001.
we have concluded that by using deep learning techniques, one can [13] C.M. Bishop, Pattern Recognition and Machine Learning, Springer, 2006.
[14] A. Milaković, F. Li, M. Marouf, S. Ehlers, Ships Offshore Struct. 15 (9) (2020)
predict the performances of the hull form with high accuracy to 974–980.
meet the engineering requirements of the initial design. [15] J. Wackers, M. Visonneau, A. Serani, R. Pellegrini, R. Broglia, M. Diez, 33rd Sym-
The DL-based hull form performance prediction model devel- posium on Naval Hydrodynamics, 2020.
[16] E. Gunpinar, U.C. Coskun, M. Ozsipahi, S. Gunpinar, Computer-Aided Des. 111
oped in this work adopts the potential flow method generated data (2019) 65–79.
sample as the virtual training data to train a DL model and then [17] A.K. Shaeffer, W. Wilson, C. Yang, SNAME Maritime Convention, OnePetro,
uses it to predict actual hull performance. Considering that the 2020.
[18] T. Cepowski, Ocean Eng. 195 (2020) 106657.
initial design stage belongs to the preliminary design, which has
[19] K. Grabowska, P. Szczuko, Ship resistance prediction with artificial neural net-
a lower accuracy requirement than those in other design phases, works, in: 2015 Signal Processing: Algorithms, Architectures, Arrangements,
the potential flow method adopted in this work has sufficient ac- and Applications (SPA), IEEE, 2015, pp. 168–173.
[20] P. Prebeg, V. Zanic, B. Vazic, Ocean Eng. 84 (2014) 259–272.
curacy to meet the design requirements and has a faster speed of
[21] E.F. Campana, D. Peri, Y. Tahara, M. Kandasamy, F. Stern, Trans. - Soc. of Naval
constructing the sample dataset. Furthermore, we think that the Arch.Mar. Eng. 117 (2009) 30.
successful application of deep learning technology in this work of- [22] T.W. Sederberg, S.R. Parry, Free-form deformation of solid geometric models,
fers an efficient method to assist the ship design process, with in: Proceedings of the 13th Annual Conference on Computer Graphics and In-
teractive Techniques, 1986, pp. 151–160.
currently commonly used design methods incurring large time re- [23] M. Tezzele, N. Demo, A. Mola, G. Rozza, PyGeM: Python geometrical morphing,
quirements due to the massive calculation workload of ship design. Software Impacts 7 (2021) 10 0 047.
In this work, we only take the total resistance as the only pre- [24] J. Holtrop, G. Mennen, Int. Shipbuild. Prog. 29 (335) (1982) 166–170.
[25] C. Dawson, A practical computer method for solving ship-wave problems, in:
dictive target. The dataset only contains the hull form generated Proceedings of Second International Conference on Numerical Ship Hydrody-
based on the geometry modification of the KCS, which means the namics, 1977, pp. 30–38.
final DL model can only predict the total resistance of specific hull [26] Y. Li, J. Gong, Q. Ma, S. Yan, Eng. Anal. Bound. Elem. 95 (2018) 266–285.
[27] X. Chen, R. Zhu, C. Ma, J. Fan, Ocean Eng. 114 (2016) 142–153.
forms. We can further expand the application range of the deep [28] John. Burkardt, in: The truncated normal distribution, Department of Scientific
learning model by adding more hull models to the dataset, and Computing Website, Florida State University, 2014, pp. 1–35.
we can also increase the prediction accuracy of the deep learning [29] I. Goodfellow, Y. Bengio, A. Courville, Deep learning, MIT Press, 2016.
[30] P. Ramachandran, B. Zoph, Q.V. Le, Searching for activation functions, arXiv:
model by increasing the sample density. In work in progress, we
1710.05941(2017).
may extend the prediction method to other performance of hull [31] A. De Myttenaere, B. Golden, B. Le Grand, F. Rossi, Neurocomputing 192 (2016)
forms and further increase the design space by increasing the va- 38–48.
[32] F. Chollet, in: Keras: The python deep learning library, Astrophysics Source
riety and number of dataset samples.
Code Library, 2018, pp. ascl–1806.
[33] M. Abadi, P. Barham, J. Chen, Z. Chen, A. Davis, J. Dean, M. Devin, S. Ghemawat,
Declaration of Competing Interest G. Irving, M. Isard, Tensorflow: a system for large-scale machine learning, in:
12th {USENIX} Symposium on Operating Systems Design and Implementation
({OSDI} 16), 2016, pp. 265–283.
The authors declare that they have no known competing finan- [34] L. Smith, Cyclical learning rates for training neural networks, in: 2017 IEEE
cial interestsor personal relationships that could have appeared to Winter Conference on Applications of Computer Vision (WACV), IEEE, 2017,
influence the work reported in this paper. pp. 464–472.

Acknowledgments

The work was performed in the University of California, Berke-


ley. YA is supported by a fellowship from China Scholar Council
(No. 201806680134), and this support is greatly appreciated.

32

You might also like