First International Symposium on Fishing Vessel Energy Efciency E-Fishing, Vigo, Spain, May 2010

Automated Marine Propeller Design Combining Hydrodynamics Models and Neural Networks

D. Calcagni1 , F. Salvatore1 , G. Bernardini2 , M. Miozzi1

1 Italian Ship Model Basin (INSEAN), Rome, Italy 2 University Roma Tre, Rome, Italy AbstractAn automated computational methodology for the preliminary design of marine propellers is presented. A Neural Network architecture is developed to describe the performance of propeller models from a virtuallygenerated systematic series. Propeller performance predictions are based on a propeller hydrodynamics model based on a inviscid-ow model. Next, propeller geometry optimization is investigated through a Genetic Algorithm formulation. The structure of the resulting propeller design environment is illustrated and Neural Network, optimization models are validated through comparisons with existing experimental data and results from alternative approaches. Numerical applications to a notional design excercise addressing a propeller retrot study for an aged shing vessel are also presented. Marine Propulsion; Ducted Propellers; Neural Networks; Numerical Optimization (key words) I. INTRODUCTION The quest for reducing operating costs and the increasing awareness of environmental issues related to fuel consumption and NOx emissions raise pressing demands to improve the energetic efciency of vessels. The propulsion system is among the most important factors contributing to determine vessel performance, and hence consistent efciency improvements are expected from a successful design of propulsion components. Dealing with large transport ships, enhanced design techniques taking advantage of state-of-the-art experimental and computational techniques can be applied in view of the large investments shipowners can afford for that kind of vessels. This is not the case with smaller vessels like shing boats where budget available for design is typically very limited and standard, sometimes obsolete, analysis and prediction approaches are used. A viable strategy to achieve efciency improvements for shing boats is to develop fast and automated modelling techniques to rapidly explore propulsion retrotting solutions as well as to design new vessels compliant to more and more stringent regulations. In the present paper, a propeller design environment combining multi-disciplinary models is described. Specically, three different tools are integrated: (i) propeller hydrodynamics modelling through robust techniques, (ii) virtuallygenerated systematic propeller series processing, and (iii) automated optimal geometry renement. Proposed modelling and analysis tools are described in the following sections and applications to sample design problems are presented. Computational hydrodynamics modelling is based on an inviscid ow approach developed over the last decade and extensively validated for marine propeller ow studies. This approach provides fast and robust performance predictions once propeller geometry and operating conditions are dened. Applications of this methodology to both open and ducted propellers are presented. Next, a procedure to identify general relationships between propeller performances and geometry/operating conditions is developed by using an Articial Neural Network (ANN) model. Combining computational hydrodynamics and ANN models, a virtual systematic propeller series is created and new design congurations are found through fast automated procedures. The capability of the proposed ANN model to describe a whole propeller series is demonstrated by considering the Wageningen B-Series for open propellers as a reference case. Finally, an optimal propeller design excercise is proposed, in which the computational hydrodynamics/ANN model is interfaced to a numerical optimization code based on a Genetic Algorithm methodology. The capabilities of the resulting design environment are described and evaluated through a notional design excercise addressing a propeller retrot study for an aged shing vessel.

II. COMPUTATIONAL MODELS Core of the present automated design technique is the computational hydrodynamics model used to predict propeller performance under given operating conditions. In view of the application to shing vessels, emphasis is given here to numerical models capable to describe ducted propellers. With respect to conventional open screw propellers, ducted propellers have the advantage of increasing thrust and efciency at low speed when an accelerating duct is used, whereas the risk of blade cavitation is reduced by mounting decelerating ducts. The hydrodynamic interaction between propeller and duct is very complex and is characterized by viscosity driven phenomena such as blade tips immersed into the inner duct boundary layer, ow separation on outer duct surface and thick viscous wakes downstream blunt duct trailing edges. Nevertheless, an inviscid-ow approach is capable of capturing global interactional phenomena related mainly to thrust-induced vorticity and three-dimensional unsteady ow effects. Current approaches for the computational analysis of ducted propellers include both inviscid (potential) and viscous ow solvers. Inviscid methods are widely used because of their robustness and limited computational effort required. Examples of these type of models are described in [1], [2], and in recent work as [3]. Major drawbacks of inviscid-ow approaches are that ow details where viscosity effects are dominant are neglected. This may traduce into unreliable predictions of propeller performance of screw operating in off-design conditions. The exact description of viscous phenomena characterizing duct/blade interaction requires viscous ow models based on the solution of Navier-Stokes equations. An example of RANS ow eld calculations on a complete ducted propulsor is reported in [4]. Of course, sophisticated ow modelling and highly-accurate results are obtained at computational costs that are typically unaffordable for design procedures where several candidate congurations are evaluated and compared to determine the best one. A. Propeller Hydrodynamics Model In view of the analysis above, a computational approach suitable for design scopes and based on inviscid-ow modelling is proposed here and referred to as BEM hydrodynamics model. Limitations typical for these types of methods are minimized by using a general formulation in which blade trailing vorticity dynamics, boundary layer ow and cavitation are described. Specically, an approach for open screw propellers developed and validated over the last decade (see [5] and [6]) has been extended to analyse ducted propulsors [7]. Description of blade/duct interaction is enhanced through trailing vortices alignment [8] and gap-ow modelling proposed in [9].

Here, the computational model is briey reviewed, whereas details can be found in referenced papers above cited. Assuming the onset ow is inviscid and incompressible, propulsor-induced perturbation velocity v may be expressed in terms of a scalar potential as v = and is governed by the Laplace equation 2 = 0. A classical mathematical model (e.g., [11]) yields that the perturbation velocity can be determined through a boundary integral equation where is express as a superposition of singularities distributed over the uid domain boundary represented by the propulsor solid surface and a surface describing the vortical wake generated by propeller blades. Impermeability conditions on solid boudaries and vorticity convection along the wake surface are imposed to dene suitable boundary conditions. A key feature of the methodology is that unknowns are distributed only over these boundary surfaces and hence the additional burden of unknowns distributed throughout the uid region typical of viscous-ow solvers is prevented.

Figure 1: Computational grid for ducted propeller ow analysis through the BEM hydrodynamics model.

By discretization, the problem is recast as the solution of a linear system of equations. Once the perturbation potential is determined, the total velocity eld follows as q = vI + , where vI has different expressions to describe the incoming ow to each propulsor component (screw and duct). Velocity and pressure elds are related through Bernoullis theorem p 1 2 p0 1 + q2 + + gz0 = vI + , t 2 2 (1)

where p is the pressure, q = q , vI = vI , and gz0 is the hydrostatic head.

As a result of the hydrodynamic analysis, open propeller thrust TP and torque QP are calculated by integrating pressure p and viscous friction over the propeller surface SP : TP = QP = p nx dS +
SP SP

1. information incoming to a neuron from all neurons of the preceeding layer is summed and biased in order to allow/disable activation of the neuron itself; 2. a proper transfer function modies the income signal, generating the output signal of the neuron; 3. the output signal is sent, through synapses, to all the neurons in the following layer. The output Yi(l) of a single neuron may be expressed as Yi(l) = F ai(l) = F
Nneur j=1

tx dS

(2)

p (r n)x dS +
SP SP

(r t)x dS

where n and t are respectively unit normal and tangent to SP , and subscript x denotes vector projection along the propeller axis direction. In (2) above, the pressure is evaluated via Bernoullis equation (1), whereas the friction coefcient is estimated by coupling BEM and boundary-layer models or simply from semi-empirical formulas for a at plate in turbulent ow, [10]. Replacing propeller surface with duct surface, similar expressions determine duct thrust TD . Then, total ducted propulsor thrust TT is (using ITTC symbols) TT = TP + TD (3)
4 whereas the thrust coefcient is KT T = TT / 1 n2 DP , with 2 n and D respectively propeller rotational speed (rps) and diameter. Similarly, ducted propeller torque coefcient is 5 KQP = QP / 1 n2 DP . Figure 1 shows a BEM model com2 putational grid describing a ducted propeller and vortical wakes emanated by blades and by the duct trailing edge.

j 0 wi(l) X j + wi(l) X 0 )

0 l Llayer ; Yi(0) = X i ; X 0 = 1 where ai(l) represents the activation function, X j the input j data to neurons weighted by quantity wi(l) , related to the connection between neurons. Quantity F is a monotone, continous and differentiable transfer function. A sigmoid function is used here, see right Fig. 2. In order to reproduce the input/output structure characterizing a given system, ANNs require suitable training. The problem of training a neural network consists in determining the optimal set of weights w that allows to describe a response surface over the entire domain of the independent variables with a minimum error; they are updated through a recursive procedure that ends when neural numerical and desired outputs differ less than a prescribed threshold. Specically, a Levenberg-Marquardt algorithm, based on a gradient search direction calculated via a back-propagation of the error, is here used [12]. Once training is completed, the ANN is ready to simulate and predict the behaviour of any conguration described as arbitrary combination of input variables. The only requirement is that queried input variables do not take values that are signicantly outside the range of variation of input variables used in the training phase. A major advantage of using trained ANNs is that computational time required to predict system output is negligible. Thus, this technique is appealing when combined with design and optimization procedures where recursive calculations are necessary. C. Numerical Optimization Model In the design process of a marine propeller, one of the main goals is the denition of the optimal combination of the design parameters, dening screw geometry and operating conditions, in accordance to required performances. To have a fast and automated design tool, an optimization procedure based on trained ANNs to describe the physical behaviour of the system is here proposed.

B. Neural Network Model BEM hydrodynamics modelling can be used to evaluate the performance of a given propeller once geometry and operating conditions are assigned. Dealing with design, it is also important to develop procedures to nd general relationships between propeller performances and geometry/operating conditions. To achieve this, BEM hydrodynamics model is combined with an Articial Neural Network (ANN) model. Neural Networks denote a general approach to describe a complex system through the relationships among Nin input and Nout output variables without any explicit mathematical input/output model, but simply through a learning process. Developed over the last ve decades, today ANNs are widely used in many disciplines, and applied with different aims such that of function approximation, classication and data processing [20]. ANN architectures try to mimic biological systems: multiple layers of neurons are linked each other through synapses that transfer information from the input layer to the output layer. A feed-forward architecture implies that information ux is mono-directional. Each neuron receives and processes information from all neurons of the preceding layer through the following steps (see Fig. 2):

Figure 2: Neural network architecture: multi-layer conguration and single neuron structure.

Numerical optimization applied to propeller design typically implies that propeller performance gures (objective functions) have to be maximized (or minimized) through variation of a given set of parameters dening propeller geometry and operating conditions (design parameters). Physical or practical limitations to values that variables assume during the optimization procedure are called constraints. In the following sections, examples of design parameters, objective functions and constraints are given. In the present work, a numerical optimization procedure based on genetic Algorithms (GAs) is considered. Basic features of the methodology are given here, whereas details can be found e.g. in [13], [14] and [15]. The underlying idea of GAs is to mimic natural selection in the process of nding the optimal solution. Candidate solutions during the optimization procedures are called individuals and the whole set of individuals is a population. Here, since binary-based GAs are used, each individual is associated to a string of binary digits (genes) ordered in a given sequence (chromosome). At each step of the process, solutions corresponding to all individuals are evaluated in terms of the objective functions. A penalty-function technique [16] is used to include the constraints in the optimization process. A selection of the best individuals, based on a tness measure calculated from the objective function and constraints, is done and chromosome strings are saved to build a new generation. New offsprings starting from the actual population are generated through chromosome string manipulations (crossover). A mutation operator is also applied to avoid premature convergence to local optima, instead of global ones. The amount of chromosone variations during the evolutionary process is controlled in order to reduce the impact of random mutations as

the solution converges to an optimum. Furthermore, an elite selection strategy is used to prevent new generations producing worse results than elder ones. The optimization procedure is iterated untill the chromosomes similarity (bit-string afnity) reaches a given value [13]. D. Parametric Search An alternative approach to nd the optimum inside a population is represented by parametric search. Specically, the optimum is searched by investigating the region of design variables space where all constraints are satised [17]. In this case the search procedure is simply based on parametric variation of all design variables in order to span all the domain of interest. The optimum is found by comparing objective function values in all tested design variables space points. Because of the dummy search technique, such a procedure typically requires to test a larger number of conditions than those investigated by numerical optimization algorithms and hence it tends to be computationally inefcient, expecially when the number of design variables growths. In the present work parametric search is used to compare results from optimization via GAs. III. VALIDATION OF NEURAL NETWORK MODEL A necessary step before application of ANNs to optimal design consists in assessing the capability of the proposed Neural Network Model to simulate performances of a general class of propellers [18]. As an example of training dataset, the Wageningen B-Series open propellers [19] are considered here. Specically, available experimental data for the Wageningen B-Series allow to build an extensive dataset of 230

propeller models obtained from different combinations of geometry parameters as number of blades Z, pitch/diameter ratio P/D, expanded area ratio EAR. Each model is tested at advance ratio Ja = VA /nDp from zero (bollard pull) to zero thrust. Input/output relationship are sketched in Fig. 3. A total of 4000 combinations of geometry and operational conditions is then considered, for which measured propeller KT , KQ and efciency = (Ja /2)KT /KQ are known.

[19] in particular, the gure shows the inuence of two input variables, J and P/D, for given value of other ones, Z = 4 and Ae /A0 = 0.85. ANN prediction error is below 4% over a wide range of input variables P/D, J. Only exception is 10% error peaks in the boundary region with lowest P/D and highest J values (thrust close to zero). Results by the two ANN models are in good agreement. Similar results are obtained for other output variables KQ and . The quantitative comparison of original vs. ANN-predicted data is conrmed by Figure 5, where actual KT results are shown for three different values of P/D. B-Series experimental data [19] and ANN results are within plotting accuracy. In particular, results by the present ANN model and those from commercial softwareNeuroSolutionsT M software [21] are in excellent agreement.

Figure 3: Neural Network training based on Wageningen BSeries open propeller dataset: input/output relationships. Two cases are considered: (a) training dataset based on experimental data from the original Wageningen series [19]; (b) training dataset for the same propeller series, with experimental data replaced by computational results from BEM hydrodynamic model. In order to perform Neural Network training rst and validation then, the whole dataset is splitted into two subsets. The subset used in the training phase (training dataset) is obtained from the whole dataset by considering only dataset entries at even values of Ja , whereas dataset entries at odd Ja values form the subset used for ANN results validation. Results by the ANN architecture developed in the present work are compared to those obtained using software NeurosolutionsT M [21]. In order to make the two ANN models comparable, input data arrangement and neurons transfer function used are the same, and both models use a Levenberg-Marquardt procedure for the gradient search in the back propagation step [20]. Moreover, a check on the correctness of the hidden layer compositions in terms of number of neurons has been made, by adopting a genetic algorithm scheme in the determination of the optimal layers dimension [15]. This form of optimization requires that the network be trained several times in order to nd the combination of inputs that produces the lowest error. The comparison between original Wageningen B-Series open propeller data and those obtained from Neural Network test simulations is described in Figs. 4, 5. Figures 4 depict error maps of calculated propeller KT versus original data

Figure 5: ANN prediction of Wageningen B-Series propeller KT : Z = 4, AE /A0 = 0.85, P/D = [0.6, 1.0, 1.4]. Original data [19] compared to present ANN model and NeuroSolutionsT M software [21].

IV. AUTOMATED DESIGN: CASE STUDY The integrated BEM hydrodynamics-Neural Networks model is coupled to numerical optimization via Genetic Algorithms and applied to a propeller design exercise. Akin to classical propeller series like the Wageningen BSeries, a basic assumption here is that a propeller series may be built by considering only four design parameters: the propeller diameter DP the number of blades, Z the expanded area ratio, EAR pitch to diameter ratio, P/D

Figure 4: Prediction by Neural Networks of Wageningen B-Series propeller thrust coefcient KT : Z = 4, AE /A0 = 0.85. Error map between calculated and original data [19]. Left: present ANN model; right: NeuroSolutionsT M software [21].

Other blade geometry details including radial distributions of chord, pitch, skew, sectional prole thickness and camber are implicitely dened through combinations of the four design variables above. The design objective is to maximize hydrodynamic efciency while keeping xed delivered torque Qd (and power Pd , with n also kept constant). Then, design constraints are imposed as follows: 1. torque absorbed by propeller at given rps matches delivered torque within a 2.5% allowed deviation: |(Q Qd )/Qd | 0.025; 2. propeller thrust at given ship speed matches hull resistance within a 2.5% deviation: |(T R)/R| 0.025; 3. propeller diameter is limited to t into an existing afterbody layout: DP Dmax ; 4. cavitation condition based on Kellers Formula: EAR EARmin , with [19]: EARmin = (1.3 + 0.3Z)T 2 +k (p0 pv )DP (4)

The reference conguration chosen here reects the propeller of a shing vessel of about 1400 T gross weigth, 70 m length, 2000 kW/375 RPM engine power at MCR. Considering freerunning conditions, design parameters for this reference conguration are: reference ship speed, VS = 7.31 m/s delivered power, Pd = 1714 KW engine rotational speed, n = 4.17 rps maximum diameter allowed, DP = 2.7 m Combining model test and sea-trials data from the project above, hull resistance curve R = R(VS ), wake fraction (1 wt ) and thrust deduction (1 t) factors are known for this reference conguration. The optimum search is based on numerical optimization via genetic algorithms and results are compared to those obtained by the parametric search approach. In particular, the parametric search procedure can be summarized as sketched in Fig. 6: (i) determine the condition for DP satisfying the KQ -constraint; (ii) determine the condition for VS satisfying the KT -constraint; (iii) check other constraints affecting EAR in combination to DP , VS . Considering alternative optimum search techniques (parametric search Vs. numerical optimization) and NN training dataset (experimental Vs. BEM hydrodynamics), four different combinations are obtained and addressed here below. With respect to design variables, a set of optimal conditions, for which the choice of the best is made only by comparison of the chosen objective function, is shown in Figs. 8 to 10.

where k = 0.2 is used for single-screw vessels. Constraint no. 2 implies that vessel speed cannot be lower than reference speed VS . Then, hydrodynamic efciency optimization means that thrust (and vessel speed) is increased. A real-life case study is derived from existing vessel data analysed during the EU-FP6 research project SUPERPROP, [7]. Aim of the project (2005-2008) was to analyse tecno-economical aspects related to retrotting propellers of aged boats like trawlers and tugs.

Figure 6: Parametric search algorithm to determine the optimal propeller operating in the hull wake.

diameter is in the nearby of its maximum allowable value, whereas the expanded area ratio is very close to its minimum required value. Discrepancies between results obtained using parametric search and GA optimization are mainly due to different strategies used to update input variable values in both approaches. In particular, EAR is updated with step (EAR) = 0.01 whereas a smaller step is used in the GA procedure. As a consequence, parametric search and GA solutions are shifted whereas predicted slopes are very similar. Next, the effect of using numerical or experimental data to train the NN model is analysed. This kind of comparison is very important because it reveals the error that occurs when a numerical model is used instead of experimental data. Discrepancies are now much more evident, and vary between 6% for quantitity P/D to only 2% in case of predicted optimum efciency. The above differences yield that for the present case small performance differences are obtained between a four-bladed propeller and a ve-bladed propeller. The selection between these two candidate solutions is then postponed to results of careful analysis addressing cavitation risks, acoustic and vibratory response of the propeller. V. EXTENSION TO DUCTED PROPELLERS The propeller design exercise described above refers to open screw congurations. In view of application of the proposed methodology to working boats like shing and supply vessels, the extension to include ducted propeller modelling capabilities is necessary. To this aim, the integrated BEM hydrodynamics/NeuralNetworks/numerical optimization tool described above is easily generalized once open propeller hydrodynamics by BEM is extended to ducted propellers. Work underway in this area is described in the present section. In order to assess the proposed methodology, the present numerical model has been used to predict global performances of the original propeller mounted on the shing vessel taken as the reference case. Its model propeller is denoted E1622 in the INSEAN nomenclature and is characterized by a pitch to diameter ratio, P/D, equal to 1.0 and an expanded area ratio Ae /A0 0.85. Experimental data are available for the propeller working in ducted as well as in unducted conguration. In Fig.11 predicted thrust and torque are compared to experimental data for the INSEAN E1622 isolated propeller case. A good agreement is shown for all values of the advance coefcient. Propeller thrust coefcient KT is slightly underpredicted at high values of J. Next, in Fig.12 the ducted propeller conguration is considered and results obtained by using different trailing wake models are compared. Specifically, results are labelled respectively as P W and F W , where P W stands for numerical calculations considering a prescribed helicoidal wake shape, whereas the notation F W

Figure 7: Parametric search algorithm to determine the optimal propeller operating in the hull wake: efciency versus the pitch to diameter and the expanded area ratio.

All gures show a sequence of points representing local minima when Z is xed and input variables P/D and EAR are changed. For the case Z=5, Fig. 7 shows the subset of solutions satisfying all constraints. Among all points, one is the optimum in that efciency is maximised. Typically, the minimum lies on the boundaries of the domain; in particular, results in Figg. 9 and 10 show the propeller

Figure 8: Automated design procedure. Comparison between parametric search and GA optmization. NNs trained with numerical and experimental data. Top: efciency vs. blade number. Bottom: pitch ratio vs. blade number.

Figure 9: Automated design procedure. Comparison between parametric search and GA optmization. NNs trained with numerical and experimental data. Top: diameter vs. blade number. Bottom: ship speed vs. blade number.

refers to results by using a ow-aligned trailing wake ?? determined from open propeller simulations at equivalent loading conditions. At high values of J (above 0.5) the total thrust coefcient is overpredicted by using any wake shape, and this is mainly due to an overestimation of the duct contribution. At low values of J, the propeller contribution to thrust is overpredicted, whereas the duct contribution is underpredicted. As far as duct contribution to thrust is concerned, the use of a ow-aligned wake improves numerical predictions, whereas propeller thrust is generally not inuenced (with slight overprediction of experimental data at low values of J) if a free wake model is included. The above mentioned tests reveal that the introduction of a ow-aligned wake is a crucial issue in order to accurately

predict ducted propellers performance. The inclusion of a wake-alignement technique for the complete propeller/duct conguration is then deemed necessary and represents the objective of ongoing work. VI. CONCLUDING REMARKS An automated computational methodology for the preliminary design of marine propellers has been described. The methodology combines inviscid-ow hydrodynamics modelling, Neural Networks and numerical optimization via Genetic Algorithm into a fast and robust tool that can be used at reduced computational costs. In the paper, the computational methodology has been outlined and ongoing development and validation work has been presented.

Figure 11: INSEAN E1622 unducted propeller open water performance. Comparison between BEM results and experimental data.

erful tool to predict input/output relationships to describe a virtual systematic propeller series built on userdened characteristics. Applications of the Neural Network/BEM hydrodynamics model to simulate experimental results for the Wageningen B-series demonstrate virtually-generated propeller series can be easily built. The proposed automated technique is able to determine the optimal conguration given operating conditions and design constraints. Future work will focus on improving BEM hydrodynamics modelling of ducted propeller congurations and creating virtual ducted propeller systematic series by using the proposed Neural Network strategy.

Figure 10: Automated design procedure. Comparison between parametric search and GA optmization. NNs trained with numerical and experimental data. Top: EAR vs. blade number. Bottom: deviation from EARmin .

In particular, simple numerical applications have been presented to assess the capability of the whole methodology to correctly predict optimal congurations for given operating conditions and other design constraints. Although the notional nature of the design excercise addressed, model capabilities are demonstrated and the following conclusions can be drawn. The computational hydrodynamics model based on BEM provides reliable predictions of propeller performances over a wide range of operating conditions. Ducted propeller modelling reveals critical aspects at low advance ratio that require further investigations. The proposed Neural Network model provides a pow-

ACKNOWLEDGEMENT The present work has been partly supported in the frame of Project SUPERPROP, Superior Life-time Operation Economy of Ship Propellers, under EU-FP6 grant TST4-CT2005-516219, and by Italian Ministry for Transports in the frame of project Programma Ricerche INSEAN.

[5] Pereira, F., Salvatore, F., Di Felice, F. (2004). Measurement and Modelling of Propeller Cavitation in Uniform Inow. Journal of Fluids Engineering, Vol. 126, 2004, pp. 671-679. [6] Salvatore, F., Testa, C., Ianniello, S., Pereira, F. (2006). Theoretical Modelling of Unsteady Cavitation and Induced Noise, Proceedings of CAV 2006 Symposium, Wageningen (The Netherlands) September 2006. [7] Salvatore, F., Calcagni, D., Greco, L. (2006). Ducted propeller performance analysis using a boundary element model, INSEAN Technical Report / 2006-083, 2006. [8] Greco, L., Salvatore, F., Di Felice, F. (2004). Validation of a Quasipotential Flow Model for the Analysis of Marine Propellers Wake. Twenty-fth ONR Symposium on Naval Hydrodynamics, St. Johns, Newfoundland (Canada), August, 2004. [9] Hughes, M. J. (1993) Analysis of Multi-Component Ducted Propellers in Unsteady Flow, PhD Thesis, M.I.T., May 1993. [10] Harvald, S. A. (1992). Resistance and Propulsion of Ships. Wiley Interscience Pubilication, New York, USA. [11] Morino, L. (1993), Boundary Integral Equations in Aerodynamics, Applied Mechanics Reviews, Vol. 46, No. 8, pp. 445466. [12] Principe, J.C., Euliano, N.R., Lefebvre, C. (2000) Neural and Adaptive Systems: Fundamentals Through Simulations, ISBN: 0471351679, John Wiley and Sons Inc., New York, USA. [13] Raymer, P.D. (2002) Enhancing Aircraft Conceptual Design Using Multidisciplinary Optimization, PhD Thesis, ISBN: 9172832592 Royal Institute of Technology, Stockholm (Sweden), May 2002.

Figure 12: INSEAN E1622 ducted propeller open water performance. Comparison between BEM results and experimental data. Top: screw and duct thrust. Bottom: total thrust and torque.

[14] Holland, J.H. (1975) Adaptation in Nature and Articial Systems, Univers. of Michigan Press, Ann Arbor. [15] Goldberg, D.E. (1989), Genetic Algorithms in Search, Optimization and Machine Learning, Addison Wesley. [16] Hatfka, R.T., G rdal, Z. (1992) Element of Structural Optiu mization, Kluwer Academy Publishers, Dordrecht.

References
[1] Kerwin, J.E., Kinnas, S.A., Lee, J.-T., Shih, W.-Z. (1987) A surface panel method for the hydrodynamic analysis of ducted propellers, Transaction SNAME, vol. 95, 1987. [2] Hughes, M.J. (1997). Implementation of a Special Procedure for Modeling the Tip Clearance Flow in a Panel Method for Ducted Propellers, Propellers/Shafting 97 Symposium, Virginia Beach (USA). [3] Baltazar, J., Falc o de Campos, J.A.C. (2009). On the a Modelling of the Flow in Ducted Propellers with a Panel Method, First International Symposium on Marine Propulsors, SMP09, Trondheim (Norway), June 2009. [4] S nchez-Caja, A., Pylkk nen J.V., Sipil , T.P. (2008) Simua a a lation of the Incompressible Viscous Flow around Ducted Propellers with Rudders Using a RANSE Solver, 27th Symposium on Naval Hydrodynamics, Seoul (Korea), October 2008.

[17] Myers, R.H., Montgomery, D.C. (1995) Response Surface Methodology, ISBN: 0471581003, John Wiley and Sons Inc., New York, USA. [18] Roddy, R.F., Hess, D.E., Faller, W. (2006) Neural Network Predictions of the 4-Quadrant Wageningen Propeller Series, David Taylor Technical Report NSWCCD-50-TR-2006/004, April 2006. [19] Kuiper, G. (1992), The Wageningen Propeller Series, Marin Pubblication No. 92-001. [20] Haykin, S. (1999) Neural Networks: A comprehensive foundation, Second edition, Prentice-Hall. [21] NeuroSolutions: Getting Started Manual, Version 5, NeuroDimension, Inc., Gainesville, FL, USA.