1. Introduction
Apart from the ship’s hull, a series of other free-form functional surfaces based on foil profiles, such as hydrofoils, stabilizers, rudders, and propeller blades, plays a significant role in naval architecture and marine engineering. Therefore, being in a position to involve their design and shape optimization in the early ship design stages will have an inarguably positive impact on ships’ and marine structures’ performance characteristics, energy efficiency, and environmental footprints. Functional surfaces in ship design possess complex geometries, and therefore, advanced modeling and analysis techniques are needed to appropriately represent, evaluate, and modify their shapes. In this work, we focused on hydrofoil-based surfaces, such as the ones mentioned above, and we aimed to employ a high-level parametric model for their modeling and representation. High-level parametric modeling permits the accurate description of complex and intricate shapes encountered in diverse fields such as naval architecture, ocean engineering, turbo-machinery, and automotive design [
1,
2]. Parametrization and parametric modeling for foil shapes (airfoils and hydrofoils) have been extensively studied in the pertinent literature. Masters et al. [
3,
4] presented a comprehensive review and comparison of the different parametrizations employed in the representation of foil profile shapes over the last few years. At the same time, several computational methods exist in the literature for the evaluation of an airfoil’s or hydrofoil’s performance. For example, panel methods have been widely used in the aerodynamic and hydrodynamic analysis of airfoils and hydrofoils. These numerical techniques permit the development of efficient software tools for computing flow characteristics around complex geometries and facilitating the design and optimization processes. The pioneering work of Hess and Smith [
5] in the 1960s laid the foundation for panel methods in airfoil analysis. Their approach, known as the Hess–Smith panel method, employed constant-strength doublets as the fundamental unknowns on the panels. Subsequent studies by Drela and Giles [
6] introduced the vortex panel method, incorporating vortex singularities on the panels to capture flow features such as leading-edge separation and trailing-edge vortices. Several enhancements have been proposed over the years to improve the accuracy and efficiency of such methods. These include modifications to account for viscous effects, unsteady flow conditions, and dynamic stall phenomena. The development of higher-order panel methods, such as the higher-order vortex panel method and the panel-vortex method, has allowed for more-accurate predictions of flow separation and wake dynamics. Additionally, higher-order boundary element methods (BEMs), based on B-splines or NURBS representations for the geometry and/or the solution, have been presented for the analysis of flow around marine propellers; see, e.g., Lee and Kerwin [
7], Kim et al. [
8], Gao and Zou [
9], along with high-order advanced BEMs for flows around foils and wings based on isogeometric analysis; see, for example, [
10,
11]. Panel methods have been extensively validated against experimental data and more-advanced computational fluid dynamics (CFD) techniques. Numerous studies have demonstrated the capability of panel methods to accurately predict airfoil and hydrofoil characteristics, including lift, drag, and flow separation, and they have been applied in a wide range of engineering fields, including aircraft design, wind turbine analysis, ship hydrodynamics, and underwater vehicles; see [
12,
13] for an overview of classical and modern methods. While panel methods offer significant advantages, they have also certain limitations [
14,
15]. They assume potential flow and neglect viscous effects, making them less-suitable for situations involving high Reynolds numbers or complex flow phenomena. However, for the initial design stages and types of applications we are targeting in this work, they are certainly adequate. Nevertheless, we should note that this is a rather restrictive limitation in our study, which restricts its usage solely to the early design stages. A more-general and -appropriate approach would be to include the results of advanced CFD simulations, which would expand the applicability of our proposed method, as well as allow the ML-enabled techniques to perform to their full potential by obviating the errors and uncertainties inherent to engineering-based CFD simulations and results; see, for example, the relevant discussions in [
16,
17].
Since both foil parametric modeling and computational analysis of foil performance are topics that have been extensively studied and matured over the last decades, we need to go beyond them to reach our overarching goal of including hydrofoil design and optimization in the early stages of ship design. To this end, we propose the employment of supervised machine learning techniques for the generation of artificial neural network (ANN) models that would be able to address in real-time and with adequate accuracy two design problems: (a) the prediction of the aerodynamic or hydrodynamic performance of a given foil design and (b) the identification and construction of the foil design that exhibits a desired, user-defined performance profile. The ultimate goal here was to cost-effectively explore the design space and easily identify appropriate designs in the conceptual design phase. At the same, we would like to be in a position to perform preliminary design optimization and identify plausible and near-optimal solutions before reaching the detailed analysis design stage. In this work, we assumed that a sufficiently accurate foil parametric model, with a relatively small number of parameters, is appropriate for the representation and construction of foil profiles, while the lift and drag coefficients, over a range of angles of attack, constitute the performance profiles, which are utilized in the early design stages. Therefore, ML model design and training were performed on the basis of the corresponding geometric and performance features with the resulting ANNs being subsequently utilized in the development of a “Foil Design Assistant” application, which allows designers to easily and effectively explore a vast design space, as is presented and discussed in
Section 3. Finally, the same models were used to conduct performance-based design optimization, and as demonstrated in the same section, they can be easily employed in the early design stages with results that are adequate and comparable to classical performance-based optimization procedures.
Before proceeding with a detailed description of the proposed approach, we will briefly present here the current status regarding the employment of ML models and techniques in foil design and performance optimization and identify the gaps in the pertinent literature that we aspire to cover. Machine learning techniques have gained significant attention in recent years for their potential to enhance the design and performance prediction of airfoils and hydrofoils. ML models, such as neural networks and support vector machines, coupled with global optimization algorithms, have been successfully employed for airfoil and hydrofoil design optimization. For example, Hui et al. [
18] presented a Convolutional Neural Network (CNN) to predict the pressure distribution of airfoils for subsonic speeds. In their work, they employed the signed distance function to represent foil profiles, and the developed CNN was trained on these pictorial representations. In addition, surrogate modeling, deep learning techniques, and other approaches have been also used; see [
19,
20] for some recent relevant works and [
21] for a state-of-the-art review and current trends in aerodynamic shape optimization and ML. Additionally and although the pertinent literature is less extensive when considering ML implementations for the inverse problem, there has been a relatively small number of research works aiming to predict the shape of foils by the given performance data. For example, Kharal and Saleem [
22] presented a method to predict the airfoil shape when its pressure distribution is given. In their work, they employed a parametric model based on PARSEC and Bézier-PARSEC parametrizations (see [
23,
24]) to represent shapes with a reduced set of parameters. This selection bears some similarity to our proposed approach, although we avoided the use of pressure distributions as the input data since they can be hardly acquired or described at the early design stages. Finally, in the context of high-dimensional simulation-driven functional-surface-optimization problems, designers can greatly benefit from the application of design-space-dimensionality-reduction methods. In such approaches, one assesses the variability presented within the given design space and subsequently reduces its dimensionality by the employment of pertinent methods. For example, Diez et al. [
25] introduced a method based on the Karhunen–Loéve expansion to assess shape modification variability and establish a latent space with a reduced dimension for the shape modification vector of ship hulls; see also [
26], which built upon this approach for a wider range of functional surface designs. Generative Adversarial Networks (GANs) have been also used in the pertinent literature to model the forms of airfoils and hydrofoils. For example, Chen et al. in [
27] demonstrated the effectiveness of their approach in enhancing the optimization results when compared to classical foil parameterization techniques, and Du et al. in [
28] combined B-spline shape representations with GANs to construct a generative tool for foil design. However, the lack of prior knowledge regarding the proper number of modes in dimension reduction strategies may lead to inappropriate results when only a low number of modes is incorporated, while, on the other hand, using too many modes undermines the benefits gained by dimensionality reduction. Li et al. [
29] developed a set of marginal functions to represent the link between dominant foil modes and higher-order modes derived from the University of Illinois Urbana-Champaign (UIUC) airfoils database [
30]. These functions effectively eliminated unwanted airfoils from the design space. However, defining acceptable constraint boundaries ahead of time is frequently difficult, and there was no guarantee that the margin-based validity functions did not also eliminate perfectly valid foil designs by accident. Therefore, such methods are not without drawbacks since they may fail to fully maintain shape complexity and geometric structure, resulting in reduced-dimension subspaces that lack the ability to generate sufficiently diverse and/or valid shapes during shape optimization. In summary, ML-based methods can explore vast design spaces, considering multiple parameters and constraints, to identify optimal configurations that maximize performance metrics such as the lift-to-drag ratio, along with other performance criteria and design constraints. Several studies, as the ones mentioned above, have demonstrated improved designs compared to traditional methods, enabling the discovery of novel and efficient geometries. By training on large datasets of experimental or computational data, ML algorithms can extract and learn complex relationships between design parameters and performance metrics. These models can then provide quick estimations of lift, drag, pressure distribution, or flow separation behavior, eliminating the need for costly and time-consuming wind tunnel experiments or numerical simulations. Studies have explored the integration of ML methods with traditional approaches, such as panel methods or CFD, to combine the strengths of both techniques. ML models can be even trained on CFD data to accelerate simulations or replace computationally expensive parts of the analysis, while still capturing the essential physics.
This paper is divided in two main sections:
Section 2, where the employed tools and methods are discussed, and
Section 3, containing the achieved results and comparisons. Specifically, we begin by describing the parametric model used for the representation of foil designs (see
Section 2.1), followed by the presentation of the foil performance evaluation procedure in
Section 2.2. We then discuss the geometric moments (see
Section 2.3) that will be used to augment the training datasets, and we proceed with their construction in
Section 2.4.
Section 2 concludes with the description of the ML models’ training in
Section 2.5. The constructed models’ comparison is presented and discussed in
Section 3.1, followed by the presentation of the developed “Foil Design Assistant” application in
Section 3.2.
Section 3 concludes with a comparison of performance-based design optimization conducted with and without regression models in
Section 3.3. The paper concludes with a summary of the achieved results and a brief discussion of future research directions in
Section 4.
2. Materials and Methods
Although several ML-enabled approaches have been widely used in the pertinent literature to predict foil performance and employ these predictions in design optimization, some characteristics of the works presented in
Section 1 require extensively large datasets or equally extensive precomputed results from advanced computational approaches, which adversely affect their application in the early design stages. The four major hindrances we would like to address in this work relate to:
The lengthy design vectors or matrices (high-dimensional initial design spaces) that are commonly found in dimensionality reduction approaches and ML models based on pattern matching and foil shape discretizations;
The existence of a non-negligible percentage of invalid designs in latent spaces produced by dimensionality-reduction techniques;
The requirements of huge training labeled datasets;
The low-level performance characteristics, such as pressure distributions, which are not convenient inputs during the early design stages.
These were addressed by employing a low-dimensional, robust, and accurate foil parametric model for the foil geometry description coupled with simple performance indicators and the lift and drag profiles over a range of angles of attack, which permit a user-friendly and intuitive definition of geometry or performance requirements by an engineer even at the early design stages. An obvious drawback in this setup is the lack of meta-information commonly used in advanced physics-informed neural networks (PINNs), which embed foil physics and are able to more-accurately predict foil performance and, consequently, enhance design optimization; see, for example, some recent relevant PINNs used in foil performance prediction in [
31,
32]. To address this last issue, we opted to include the integral characteristics (geometric moments) of foil profiles, which are commonly correlated with the performance criteria and can, therefore, provide a physics-informed flavor without requiring expensive calculations, as was recently demonstrated by Shah et al. in [
26].
The employed regression models require a rich space of foil profile designs, , with the corresponding label data, i.e., performance profiles, residing in a second set, . The former set in this work corresponds to the foil profile geometric descriptions, and it comprises the design vectors, , which correspond to the parameters of the selected parametric model. This initial design vector was augmented, as we will describe in the sequel, by an additional vector of foil geometric moments. With regard to the latter set, contains the matching foil performance profiles, which are mainly encoded with a 18-component vector. The first nine components are the values of the lift coefficient, , for nine angles of attack, whereas the remaining nine components correspond to the drag coefficient, , for the same angles of attack. Once again, moments were also used to augment this second vector. For all models considered, angles between 0 and 8 with a 1 step were used for the and calculations. Note that these two sets were used as the input or output training sets depending on whether we were modeling the forward or inverse design problem.
2.1. Foil Parametric Model
One of the main objectives in this work was to employ a foil parametric model that can robustly and accurately describe a large family of designs with a small set of parameters. This was performed to avoid well-known problems with high-dimensional geometric descriptions and the effect of foil design parametrizations in design space representation and ML models’ training; see [
18]. A hydrofoil parametric model, firstly introduced in [
10] and later expanded in [
33] to accurately represent a wider range of foil designs, constituted the basis of the parametric model used in this work. As mentioned in the introductory section, a wide variety of methods have been proposed in the pertinent literature for the parametrization of airfoils and hydrofoils, and the works of Masters et al. [
3,
4] presented a comprehensive review and extensive comparison for a big number of them. As reported in [
33], the 17-parameter parametric model we adopted here permits the accurate representation, within the Kulfan tolerance [
34,
35], of 100% of the NACA dataset in [
4] and more than 80% of the foil designs in the UIUC database [
30]. As a side note, we should mention here that the calculations were conducted with the modified Kulfan tolerance proposed in [
4], which involves an inflation of the computed error and, therefore, is stricter than the original Kulfan tolerance. The typical wind tunnel tolerance, also known as the Kulfan tolerance, corresponds to the maximum allowable geometric deviation between two foil designs, which does not alter their performance evaluation in wind tunnel experiments. Specifically, the initial Kulfan tolerance is defined as
where
c denotes the foil’s chord length and
x the longitudinal position of the measured geometric deviation. This definition was modified in [
4] to
resulting in a single requirement, i.e.,
. Furthermore, the employed parametric model was specifically designed to accommodate the specific needs of design optimization, i.e., relatively small number of parameters, encoded in the so-called design or
shape vector, along with a high accuracy in the representation of the profile shapes, which coincides with the requirements for the ML model training we posed in this work. Finally, the parametric model’s robustness stems from the fact that all control point movement constraints are embedded in the construction of the parametric model with the parameters being nondimensionalized in
and all combinations of the parameter values in the unit interval corresponding to valid profile designs with no self-intersections, unwanted oscillations, or any other invalid geometric feature.
The employed parametric model generates foil instances with a shape vector of 17 parameters, with each instance being represented by a cubic NURBS curve (see [
36]), defined as follows:
where
is the set of rational B-spline bases of order 4 (degree 3) defined over the knot sequence,
and possessing non-negative weights
,
, while
are the corresponding control points determined by the parameter vector
. The parameters’ definition is graphically presented in
Figure 1 with their descriptions and types of effect on the foil shape, i.e., global or local, being summarized in
Table 1. Finally, the shape modification effect of each parameter on the foil profile is individually depicted for each parameter in
Figure 2. The solid thick line in
Figure 2 corresponds to the foil instance with
, whereas the red-colored foil instances correspond to varying parameter values,
, for each of the first fifteen parameters. The interested reader may refer to [
33] for more details on the construction and performance of the adopted parametric model.
The incorporation of the parametric model in the development and training of the ANNs used in this work has additional benefits in the behavior of these ML models when predicting the foil shape or its performance. Specifically, as we will demonstrate in
Section 3, the parametric model limits the user to producing foil designs that are both valid and within the boundaries of the employed design space, which safeguards us against highly erroneous predictions, which may occur from ML methods [
37,
38], as well as a design exploration that goes far beyond the training set boundaries [
39]. Moreover, it offers an easy way of checking the validity of shapes returned by the inverse design model, since its output must be a valid parametric model shape vector. Another important positive impact coming from the employment of the parametric model is the indirect dimensionality reduction, since we can describe complex shapes with a small set of parameters, and even more importantly, we can do that in a way that does not generate invalid designs in the reduced-dimension design space. This is a clear advantage of our approach, since common dimensionality-reduction techniques cannot eliminate the existence of invalid designs in the latent spaces; see the relevant discussions in [
26,
40].
2.2. Foil Performance Analysis
The lift and drag coefficients, for each design instance, were estimated with the aid of the well-known XFOIL (
https://web.mit.edu/drela/Public/web/xfoil/ accessed on 10 May 2023) computational package, which is a popular tool used in assessing the performance of subsonic airfoils and hydrofoils. XFOIL employs a fast panel method with a fully coupled viscous–inviscid interaction method used in the ISES code developed by Drela and Giles; see [
6,
41]. XFOIL requires a polygonal approximation of the foil profile, and thus, for each foil design in our dataset, we generated, through the employed parametric model, 200 points that closely approximated the foil’s shape. Specifically, in most cases, 160 points, with a curvature-based distribution that minimizes the distance between the actual curve and the polygonal approximation, sufficed to produce convergent and accurate performance results. However, we decided to increase that number to 200 to avoid inaccuracies and potential problems with more-complex foil profile geometries. A
matlab (
https://www.mathworks.com/products/matlab.html accessed on 10 May 2023) script was generated to read the design vectors for all foil designs in the dataset, save a polygon approximation, and finally, automatically call XFOIL to compute the lift and drag coefficients for all angles of attack in
with a Reynolds number equal to
. We should also note here that we increased the viscous solution iteration limit (
ITER) in XFOIL so that up to 50 Newton iterations were executed when assessing convergence. This was a significant step, which helped us avoid missing results and non-converging calculations with some more-intricate foil designs. A sample XFOIL output for the calculations described here is depicted in
Figure 3. As a final note, we should also add here that a small number of performance calculations (less than 3% of the foil designs in the database) failed for one or two angles of attack. In such cases, we performed a cubic spline interpolation with the remaining
and
values and estimated the missing value(s) with the resulting spline curve.
2.3. Geometric Moments
The geometry and performance information captured by the parameters of the parametric model and the lift/drag coefficients may not be sufficient to fully capture the relation between design and performance, which can result in inadequate regression or generally inaccurate prediction models. In the area of performance-based design optimization, it is not uncommon to embed, in the design vector, information about the quantities of interest and/or incorporate knowledge about the underlying physics in the ML model, as, for example, in the case of physics-informed neural networks; see [
31,
42,
43]. However, these approaches tend to be computationally intensive and require detailed analysis, which imposes an additional computational burden, which is especially unwelcome in the early design stages. Motivated by the work of Shah et al. [
26], we attempted to enrich the design and performance description with the integral properties of the foil shape geometry and performance profiles, respectively. Specifically, for the case of the foil’s shape, the geometric information, encoded in the parameters’ vector, was coupled with foil shape moments, leading to an enhanced design vector, which, as will be demonstrated, can offer a more-accurate description of the shape’s inherent characteristics and its relation to the performance metrics. Thus, the incorporation of geometric moments acts as an indirect way of incorporating quantities that are correlated to performance criteria in a cost-effective manner. In other words, while most performance criteria (e.g., lift or drag) require generally costly simulations, these quantities are generally dependent on geometric moments, which can be easily computed with a significantly lower computational cost.
Geometric moments for performance profiles do not necessarily incorporate any inherent information about the foil design, but they can still help the model better predict relations between performance characteristics and the corresponding foil geometry. Therefore, the geometric moments of the lift and drag coefficient profiles were also considered for the inverse design problem.
Assuming now that
corresponds to the 2D domain bounded by the foil profile curve
, we may compute its
-,
, order moments by:
where
when
and 0 otherwise. Obviously, the 2D domain area corresponds to the 0
th-order moment, i.e.,
, and its centroid can be calculated as
. A proper combination of geometry and its moments allows the construction of a vector that captures the shape’s inherent characteristics and can serve as its unique identifier, the so-called signature vector in [
44].
For the foil profile case, we may compute the Geometric Moments (GMs) with the two-dimensional version of the divergence theorem, i.e., Green’s theorem. Specifically, using Green’s theorem and the representation of the foil profile in (
2), we may equivalently write:
and so on.
For the lift and drag coefficient profiles, we constructed a spline interpolating the computed values, and we can then proceed with the calculation of the corresponding moments under the interpolant
, where
x corresponds to the angle of attack and
y to the coefficient values. Note that we used the signed area here, since we may have foil shapes with negative lift coefficients, and therefore, using the signed area would be a more-meaningful performance feature to differentiate between positive- and negative-lift-producing designs. Hence, for the performance profiles, we may calculate the moments as follows:
and so on.
The performance criteria considered in this work, i.e., the lift and drag coefficients, are invariant to translation and uniform scaling, whereas the moments described above change value with Translation (T) and Scaling (S). Therefore, in the general case, we would need to employ TS-invariant Geometric Moments (TS-iGM), as described in Xu and Li in [
45] where the authors derived Translation, Scale, and Rotation invariant Geometric Moments (TSR-iGM). However, in our dataset, we only considered foil designs with chords aligned with the longitudinal axis (x-axis), fixed placement, and normalization to [0,1], which permitted us to use the moments derived above without introducing any biases; therefore, we may skip the use of moment invariants.
2.4. Training Datasets
As previously mentioned, the basis for the construction of the foil design space, and subsequently the resulting training datasets, was the publicly available UIUC database of foil designs; see [
30]. This database comprises approximately 1500 foil profiles encoded via point sets of varying lengths, i.e., the number of points on the profile. The first step in the construction of the design dataset is a preprocessing step and pertains to ordering points to one common orientation and normalization of the foil shapes so that all of them lie on the x-axis and have a chord length equal to one unit. Subsequently, we determined the parameter values of the foil parametric model, which reconstruct each foil profile within the modified Kulfan tolerance; see Equation (
1). This was done by solving the following minimization problem:
where
is the number of points,
, defining the
-th foil in the UIUC database, and
are the ordinates of the first and last foil point (Recall that a preprocessing step was performed to ensure that all foil points were stored in a counterclockwise orientation with the first point being the trailing edge point of the upper foil side), respectively.
We solved the minimization problems appearing in Equation (
6) using a sequential quadratic programming technique implemented in
matlab with an initial design vector resulting from a rough estimation of the design parameters based on the input profile points. This resulted in approximately 1300 design vectors that approximated the corresponding number of foils within the modified Kulfan tolerance. The identified parameter vectors were saved to form the initial design dataset
. If we now performed a Principal Component Analysis (PCA) [
46] of
and use the first two principal components to visualize the designs in
, we obtained a picture with isolated designs or clusters of designs with relatively big gaps among them. Although this did not directly correspond to the distance between the high-dimensional design vectors, it provided us with an indicative picture of the distribution of the foil designs in the design space. Additionally, given the number of parameters and performance targets we planned to work with, the number of design vectors was not sufficiently large to ensure a good fit by standard regression or neural network models. To address these issues, we generated 10 design variations in the neighborhood of each design in
by varying the identified parameter values
around their initial value. In this way, we obtained a significantly larger dataset with fewer gaps between designs, as can been seen by the blue dots in
Figure 4. This figure was produced by performing PCA of the augmented set and visualizing both the initial and additional designs (variations) in the same picture with the help of the first two principal components. As one may easily observe in
Figure 4, the red circles, corresponding to
, are isolated or form clusters of designs with relatively big distances between them, whereas the included variations, blue dots, fill in most gaps and construct a significantly richer dataset that we denote as
. The next and final step in the process of design dataset construction was performed with the help of Equation (
4) and pertained to the calculation of moments,
to
, for each design in
. These moments were subsequently added to form the final version of
.
The initial performance dataset,
, was generated by assessing the performance of each design in
with XFOIL, as described in
Section 2.2, and storing the resulting
and
values in
. As in the case of the design dataset, the performance dataset was augmented with the moments for the performance curves,
and
, using Equation (
5), producing the final performance dataset
. We should note here that a small number of designs failed to produce complete performance profiles (cases with more than a couple of
and/or
values missing), and they were excluded from both
and
. Therefore, the final datasets, which were used in the regression and training of the neural network models, comprise
12,615 designs with their design vectors and moments in
and the corresponding number of performance and performance moment values in
. In summary, the matrix structures corresponding to
and
are depicted in
Table 2. We should also note here that we posed certain limitations to the type and number of performance inputs. If we were to use different values or even a different number of angles of attack for each foil design, we would probably need to revert to a Recurrent Neural Network (RNN), which would be able to handle variable-length sequences of inputs. This would become even more important if we were to consider more nuanced performance profiles, including pressure distributions or other low-level performance metrics.
2.5. Regression Models
For both cases, i.e., forward and inverse design problems, we considered 3 different regression models, which were compared with a range of different training datasets:
A Multivariate Linear Regression (MLR) [
47,
48], which constitutes our basis (reference) model;
Two families of feedforward Artificial Neural Networks (ANNs) trained with (a) Levenberg–Marquardt (LM) back-propagation, and (b) LM with Bayesian Regularization (BR); see
Table 3 for more details.
As a reference model, for both problems, we employed the multivariate normal regression of either the 18-dimensional performance responses (
) on the design matrix corresponding to the 17 design parameters (
), or vice versa for the inverse design problem. The model is formulated as
where
n is the number of samples, 12,615 in our case,
is the 18- or 17-dimensional vector of responses,
is the design vector of predictor variables, and
corresponds to the resulting regression coefficients with
denoting the vector of error terms. Note that we added a column of ones in matrix
, formed by the 12,615
vectors, to include constant terms (intercepts) in the regression. The root-mean-squared error for the prediction of the performance values with MLR is reported in
Table 4, whereas the corresponding MLR error for the inverse problem is found in
Table 5.
The remaining models were two-layer feedforward neural networks with an input layer, a hidden layer, and an output layer. For the hidden layers, hyperbolic tangent sigmoid activation functions were used, whereas linear transfer functions were employed in the output layer, as we performed regression in all cases. We should also note here that positive linear (poslin, also known as ReLU—Rectified Linear Unit in the pertinent literature) and logistic sigmoid transfer functions were tested in the hidden layers with inferior results in all tested cases. For the majority of the trained models, we observed that 20 neurons in the hidden layers produced good results, but for some of the employed training datasets, this number was raised to 25 or 30, as can be seen in
Table 3. Specifically, we performed model training tests starting with 10 neurons, in the hidden layers, and went up to 70 neurons with a step of 5 neurons. For all cases, the results showed an improvement from 10 to 20 or 30 neurons, followed by a decline in performance after 30 or 35 neurons. As mentioned above, two training approaches were used, LM back-propagation and LM with Bayesian regularization, with “LM” included in the model name in the former case and “BR” for the latter one; see
Table 3. Additionally, the scaled conjugate gradient algorithm was used in the training, but we did not include it in our comparisons, as it was outperformed across all tests by the other two training methods.
LM is a standard approach used in training ANNs, and the following steps are applied in its implementation:
Network weights’ and biases’ initialization.
Input data are fed into the network, and the activation function values for each neuron in the hidden and output layers are calculated.
The error between the network’s predicted output and the actual target output is computed using a suitable error metric, the mean-squared error (MSE) in our case.
The error is then propagated backwards to adjust the weights and biases. For this, the Levenberg–Marquardt algorithm is used to optimize the network parameters. The LM algorithm combines the benefits of fast convergence from the Gauss–Newton algorithm and the robustness of steepest descent. It dynamically adjusts a parameter known as the damping factor during training, controlling the step sizes taken in the weight space.
The weights and biases of the network are updated using LM.
Steps 2 to 5 are repeated till convergence has been achieved (or when a maximum number of iterations has been reached).
Bayesian regularization updates the weight and bias values based on the same LM optimization algorithm. It then minimizes a linear combination of squared errors and weights and determines the correct combination so as to produce a network that generalizes well, i.e., does not over-fit the given training data. This was performed by extracting a validation set from the training data and monitoring the model’s performance to prevent over-fitting. Bayes’ theorem was used to combine the prior distributions with the likelihood to obtain the posterior distributions. The linear combination of the squared errors and weights was modified accordingly to achieve good generalization qualities. The interested reader may refer to [
49,
50] for a more-detailed description of the Bayesian regularization process. We should also note here that an additional test set, separate from the validation set mentioned above and corresponding to 15% of the datasets, was set aside, for all models, and used at the end to confirm the generalization qualities of the trained models. With regard to the training set employed in LM training, a 5-fold cross-validation was utilized to avoid over-fitting. In all cases presented in this work, coefficients of determination,
, exceeding 90% for both the training and test sets were achieved. Note that the value reported here for the training set corresponds to the average value from the 5-fold cross-validation.
Before we proceed with the presentation of the achieved results, we need to also briefly discuss the use of different datasets in the training of the models included in
Table 3. Specifically, we used a wide range of combinations, starting from the simplest,
and
, dataset pair, to the full datasets,
and
, described in
Table 2. This was performed to assess the effect of full or partial moments’ inclusion in either the design or performance datasets along with the effect of transferring components from the input to the output sets and vice versa. The area of the foil profile is one such component example, as it can be considered either as the input feature or the output target.
Table 3 presents a summary of the trained models along with the composition of the input and output sets that were used for training.
4. Conclusions and Future Work
In this work, we demonstrated a successful application of Artificial Neural Network models in the forward and inverse foil design problems for early foil design stages. This was accomplished by integrating a foil parametric model for an intuitive, accurate, and compact representation of foil designs and an appropriately prepared rich dataset of realistic foil designs. The incorporation of the parametric model in the development of the model served also as a safeguard against invalid input, highly inaccurate predictions, and design exploration that went far beyond the training dataset. Specifically, the parametric model was constructed in way that guaranteed the construction of valid foil shapes for the complete range of its parameter values with the employed instances distributed in the design space, as indicated in
Figure 4. At the same time, the output of the models used in the inverse design problem needed to produce valid shape vectors, which permitted us to easily identify erroneous predictions. In addition, geometric moments were incorporated in both the input and output datasets and were shown to significantly improve the accuracy of the ANNs in the prediction of the foil’s performance. Similarly, the inclusion of performance moments was shown to enhance the predictions for the inverse design problem, although their effect was less pronounced. The developed ANNs were compared against a reference linear regression model to benchmark their performance, and a wide range of variations with regard to their training procedure and the composition of the training datasets was examined to identify the best-performing models. Finally, the best-performing models were used to demonstrate the proposed approach in a real-time, intuitive “foil design assistant” software prototype, with a performance-based shape optimization example. The results obtained from the ML-enabled shape optimization were comparable to the traditional foil shape optimization approaches with the additional benefit of requiring only a fraction of the time needed for traditional approaches.
Although we demonstrated good prediction accuracy with the developed procedures, there were several limitations that we would need to address in the future. One of these limitations pertained to the usage of fixed performance metrics, i.e., the lift and drag coefficients for a fixed number of angles of attack, which limited the applicability of our approach to the early design stages. We could potentially employ recurrent neural networks and/or 1D convolutional neural networks to include variable-sized performance input and lower-level performance metrics, such as pressure distributions, which would expand the applicability of our approach beyond the early design stages. In a similar manner, the assessment of foil performance using panel methods was a restrictive limitation in our work, as it was noted in the introductory section. As we have now established the beneficial impact of the geometric moments’ inclusion, as well as the dimensionality reduction and robustness, which was attributed to the employed parametric model, we may move forward and include high-fidelity CFD simulations for the foil performance assessment, which would obviously expand the applicability of the proposed design tools. Moreover, we plan to conduct a study in the near future regarding the use of the developed Foil Design Assistant to assess its impact on the design effectiveness of students and professionals assigned foil design assessment and optimization tasks. Finally, we plan to extend this approach for the 3D case with wind turbine and propeller blades being the next possible objects of interest. Apart from the obvious tasks of extending the type of ML models and constructing appropriate datasets, the inclusion of geometric moment invariants should be also considered to alleviate the need for fixed placements and design normalization, which might not be possible for general functional surfaces.