Multi-Fidelity Design Framework Integrating Compositional Kernels to Facilitate Early-Stage Design Exploration of Complex Systems

Early-stage design of complex engineering systems is critical since most of the major design decisions are being made at this stage while engineers have minimal knowledge about the design. The importance of early design stages has been recognized in various engineering fields, including ship [1] and aircraft design [2]. More specifically, design decisions that dictate the vessel’s overall configuration reduce design freedom and commit a significant portion of the overall cost. Therefore, engineers must conduct comprehensive design space exploration to identify design trends and trade-offs, leading to informed design decisions.

Throughout the design process, engineers employ a range of analysis methods with varying fidelities to evaluate different designs. At opposite ends of the spectrum, high-fidelity (HF) methods, often based on first principles, provide highly accurate results but entail significant computational expenses, while low-fidelity (LF) methods yield less accurate results while requiring lower computational costs. Traditionally, engineers have used LF methods and tools to explore the broad multidimensional design space of complex systems [2]. These methods include simplified physical models, empirical and semi-empirical methods [3], and trends based on historical data [4]. Based on engineering experience, these methods are suitable for conventional designs for which we have a lot of gained knowledge.

However, nowadays, there is a drive toward innovative designs driven by the need to address challenges such as the integration of autonomous systems [5], enhancing safety [6], and promoting sustainability [7]. This necessitates the design of systems with capabilities that surpass our current state-of-the-art designs. Some examples of novel hull designs in naval architecture are the tumblehome wave-piercing hull [8] employed in the Zumwalt-class destroyer, the aluminum trimaran design [9] used for the USS Independence frigate, the AXE bow [10] used in various yacht designs, and the asymmetric hull configuration found in the Baltika icebreaker [11]. In aerospace, the flying-V [12] represents an innovative aircraft design that integrates the fuselage into the wing structure, aiming to enhance its performance.

Incorporating new features into systems introduces added uncertainty, primarily because engineers lack prior experience with these unique systems [13]. This introduced uncertainty needs to be quantified and taken into account in order to support decision-making [14]. Thus, in order to mitigate this heightened uncertainty, it becomes important to introduce HF analysis at an earlier stage in the design process. This is essential because relying solely on LF models and tools to examine the design space may result in designs that are not feasible or suboptimal when the underlying physics is not appropriately captured [15]. Additionally, in the case of ship design, engineers rely on established rules and guidelines prescribed by classification societies to guide their design practices. However, when dealing with novel ships featuring unconventional shapes and sizes, blindly following the class society formulations proves insufficient [16], as these guidelines are primarily based on conventional ship knowledge. A similar challenge is present across different design domains. Currently, advanced HF models are able to capture the underlying physics that dictate a design’s performance, yet they require a significant amount of computational resources, human input, and extended lead times [17,18]. However, it is unrealistic to evaluate a sufficient number of designs using expensive HF methods in order to construct an approximation of the design space. Therefore, a significant challenge in the optimization of vessel design is the efficient increase of modeling accuracy required to capture the crucial physics that restricts or allows for a specific concept [19].

To address these issues, a promising approach is to construct multifidelity (MF) models, which are defined as models combining various fidelities [20]. According to Ref. [19], models are considered multifidelity when there is synergistic use of different mathematical descriptions (i.e., different physics typically represented by different governing equations, boundary conditions, or parametric attributions) in the analysis or design procedure. The goal of MF models is to achieve computational efficiency through the use of LF models while maintaining precision through the use of the HF model [21]. MF models are state-of-the-art for solving complex problems. For a comprehensive understanding of the various MF techniques, an extensive review of the various MF techniques can be found in Refs. [20,21]. MF models have demonstrated notable achievements across various engineering domains such as engineering design and analysis, and applied mathematics. They have been used to tackle diverse problems, encompassing solving partial differential equations [22], aiding in complex analyses like estimating the wave-induced vertical bending moment on ships [23,24], and supporting design applications [25].

Several state-of-the-art methods for design applications have been extended to an MF scheme. Examples of such methods include Monte Carlo (MC), Gaussian processes (GPs), and neural networks (NNs). For example, Ng and Willcox [26] proposed an MF estimator based on the control variate MC method. Diverse methods exist in the literature for constructing MF NNs architectures. For instance, Meng and Karniadakis introduced a composite NN leveraging MF data. In their approach, the first NN approximates the LF data, while the second and third NNs model the correlation between the low- and high-fidelity data [27]. Examples of different approaches for building MF-NN architectures can be found in Refs. [28,29]. According to the findings reported in Guo et al. [29], the co-Kriging model demonstrated superior performance compared to the MF-NN architectures in the case of the Forrester function. Conversely, in the case of more complex functions such as the Jump Forrester function, the opposite trend was observed [29]. To estimate the uncertainty associated with the prediction, Meng et al. [30] introduced a MF Bayesian NN framework. Within this framework, uncertainty was estimated by predicting the uncertainty of learning the correlation between the low- and high-fidelity data. GPs have found extensive application in design problems. Their primary advantages lie in their ability to provide uncertainty quantification pertaining to predictions, which can be effectively used in the context of design optimization. Another advantage is their effectiveness in addressing problems characterized by small data regimes.

Despite the ongoing developments in MF models, there are still significant areas in design applications that remain unexplored. For example, for multidisciplinary design problems, Mainini et al. [15] argue that there is no mathematical framework that is capable of determining (1) which design disciplines, (2) the degree of coupling for analysis tools, (3) the level of accuracy necessary to capture the crucial physics of a specific design, (4) where the data is best collected, and (5) how to make optimal design decisions with limited computational resources. Furthermore, Peherstorfer et al. [21] emphasize that for design frameworks, it is crucial to construct frameworks that do not solely focus on models but include additional information sources, so that decision-makers can effectively utilize a wide range of available information.

In design applications, a primary challenge lies in the necessity for a larger HF dataset to attain precise predictions, particularly for complex and higher-dimensional problems. Expanding the HF dataset poses difficulties since each data point is a product of computationally expensive analyses or, in some cases, physical experiments. Consequently, the acquisition of a substantial HF dataset is constrained by the limited computational budget available. To address this challenge, the present work proposes the integration of compositional kernels in a design framework based on the autoregressive scheme of MF GPs. The main idea is that a more accurate approximation of the design space can be attained by utilizing knowledge about the underlying structure of the design space revealed by the compositional kernels. The goal is to build a framework where fewer HF evaluations are necessary to create an accurate MF model, resulting in a reduction in computational cost. The proposed method has been applied to five benchmark problems, as well as to a cantilever beam design problem as a first step toward its application to early-stage design of complex engineering systems such as naval vessels.

Section 2 provides an overview of relevant work on the exploration of design space structure and uncertainty quantification for design applications. Section 3 provides the technical background of the proposed method. Section 4 examines the results from the conducted case studies. Finally, Sec. 5 outlines the conclusions derived from the research and offers recommendations for future research.

The article addresses early-stage design exploration of complex systems by introducing an MF-GP-based method. In this section, the fundamental concepts of the topic are discussed individually, and relevant studies are presented. These concepts include exploration of design space’s structure, uncertainty quantification in design exploration, and GPs.

Early-stage design of complex systems deals with multidimensional design spaces, and this forms a significant challenge. In practical applications, the design space often involves a large number of design variables. Therefore, the scalability of modeling methods becomes a crucial consideration. For instance, the design variables can range from a few dozen, as seen in the hydrostructural optimization of hydrofoils with a 17-dimensional design space [31], to thousands of variables in the case of an aerostructural optimization benchmark problem for commercial transport aircraft [32]. Another important challenge to address during this design stage is the quantification of uncertainty (UQ). As highlighted by Ref. [14], “UQ creates value to the extent that it holds the possibility of changing a decision that would otherwise be made differently.”

The literature has examined numerous approaches aimed at facilitating design exploration. For example, Di Fiore et al. [33] proposed incorporating both information extracted from data and domain knowledge to facilitate the conceptual design of re-entry vehicles. Furthermore, Singh and Willcox [34] developed a framework grounded in Bayesian statistics and decision theory. This framework integrates information from different stages of a product’s lifecycle to enhance decision-making in the design process. The method proposed in the present article is founded on the premise that the design space under investigation possesses a specific structure, which can be uncovered and leveraged to enhance the efficiency and effectiveness of design exploration.

The idea of uncovering the patterns within the design space has been studied by Melati et al. [35]. The researchers proposed a machine learning-based approach rooted in pattern recognition to effectively map and characterize the multidimensional design space of nanophotonic components. Through pattern recognition techniques, the authors successfully unveiled relationships among an initial sparse set of optimized designs, thereby reducing the number of characterized variables.

In the context of GPs, the covariance matrix via the kernel function can be used to define patterns in the design space. The efficacy of employing an appropriate kernel function for Bayesian optimization was demonstrated by Moss et al. [36]. In their work, the authors introduced a Bayesian optimization method for raw strings that seamlessly incorporates a string kernel, showcasing the power and effectiveness of this approach. In addition, the study conducted by Palar et al. [37] explores the potential of composite kernel learning and model selection in enhancing the accuracy of bi-fidelity GPs.

While the main focus of the current work aligns with the research paper of Palar et al. [37], the approach of defining the kernel functions differs. The approach of Palar et al. [37] is to build the compositional kernels as a weighted sum of basis kernels, and the weights are treated as hyperparameters. In contrast, the present study proposes an optimization routine where the kernel functions of the different fidelities are sequentially built. Thus, the kernel function for the $i$th fidelity is built based on the kernel function of the lower fidelity model $i−1$⁠. The proposed method can be extended to an $s$th fidelity problem setup.

UQ is a widely discussed topic within the realm of design applications [38]. Defining uncertainty poses challenges since it relates to a lack of knowledge. Uncertainty is commonly categorized into two types: aleatory and epistemic [39]. Aleatory uncertainty pertains to natural randomness that is inherent in the system such as the outcomes of tossing dice and drawing cards, and thus, it cannot be reduced or eliminated. On the other hand, epistemic uncertainty arises from a lack of knowledge or information. This type of uncertainty can be reduced as one gains a better understanding of the problem through further investigation and learning. However, there are arguments asserting that uncertainty can only be epistemic. As described in Ref. [40], uncertainty is an inherent aspect of the human brain and is not an inherent property of the external world. Probability, serving as a measure of uncertainty, reflects an individual’s state of mind rather than an absolute state of affairs. In this article, we follow the common categorization to aleatory and epistemic uncertainty.

A typical example of aleatory uncertainty in this context of hydrodynamic analysis is the probabilistic formulation of ocean waves. Due to the inherent randomness and variability of wave conditions, a probabilistic approach is essential to adequately address the associated uncertainties in the hydrodynamic analysis. On the other hand, Mavris et al. [2] discuss several examples of epistemic uncertainties that are pertinent to early-stage design, including the treatment of assumptions, ambiguous requirements, the fidelity of different codes, economic uncertainties, and technological risks. Furthermore, according to Collete [41], the key epistemic uncertainties for ship structures involve operational profiles and behavior, model uncertainty, and the influence of the human engineer. Hence, within the domain of early-stage design for complex engineering systems, engineers confront both aleatory and epistemic uncertainties. In certain instances, as highlighted by Ref. [41] in the context of ship structures, numerous studies tend to simplify or overlook epistemic uncertainties.

UQ can be used to improve early-stage design exploration of novel and complex systems. First, UQ can be used to make more informed and optimal design decisions [38,42]. Second, when dealing with innovative concepts, there is an inherent introduction of additional uncertainty into the design exploration problem. The additional uncertainty arises due to the inherent lack of knowledge concerning the performance of such systems. A real-world example in industries like automotive and aerospace involves the use of prototyping to acquire additional insights into the performance of new engineering systems. It is noteworthy to emphasize that, particularly in the domain of ship design, the large physical dimensions, the complexity, and the fact that ship are not being built in large series render the construction of full-scale prototypes unfeasible [1]. Therefore, it becomes important to consider and account for this uncertainty in order to effectively navigate the design space and make reliable design decisions.

A viable mathematical framework for addressing epistemic uncertainty lies within the family of Bayesian methods. Bayesian methods possess an inherent capacity to quantify uncertainty arising from our limited knowledge. This is because Bayesian statistics quantify the degree of belief in the truth of a particular proposition, rather than the frequency of event occurrence. For a more comprehensive discussion on the distinction between frequentist and Bayesian statistics, readers are referred to Ref. [39].

The present article focuses on GPs, a subset of Bayesian methods. In general, a GP is a collection of random variables such that any subset of these variables is jointly Gaussian [43]. The MF schemes of GPs incorporate data obtained from various fidelities. One of the schemes is the linear autoregressive scheme (AR1) proposed by Kennedy and O’Hagan [44]. In addition to this, two nonlinear schemes have been introduced: the nonlinear autoregressive Gaussian process (NARGP) proposed by Perdikaris et al. [45] and the deep Gaussian processes (deep GPs) proposed by Damianou and Lawrence [46]. GPs and MF-GPs have demonstrated their effectiveness in addressing design optimization problems across different engineering fields such as aircraft design [47], ship design [48], and materials design [49]. Their popularity widely comes from the fact that they are well suited for small data regimes [50], which aligns them to the problem of early-stage exploration of novels systems where there is inherently limited data available. In contrast to AR1, which assumes a linear dependency between the fidelities, NARGP, and deep GPs are capable of capturing more complex nonlinear dependencies among the fidelities. A general trend, based on aerospace-related engineering problems as shown in [51], is that linear schemes exhibit superior performance in scenarios with limited data compared to nonlinear schemes, which require a larger amount of data for effective training. The present research targets engineering problems in the small data regime; thus, the AR1 scheme was selected as the basis for this work.

While design optimization has seen successful applications, ongoing research efforts are necessary to fully harness the potential of GPs for early-stage exploration purposes. This study investigates the concept of exploring the structure of the design space with the aim of leveraging this information to enhance prediction accuracy and reduce uncertainty. Such an approach holds potential for design applications in domains such as ship and aircraft design, where HF analyses, including computational fluid dynamics computations and experiments, can be significantly (computationally) expensive. Consequently, reducing the number of HF simulations would yield benefits for tackling such problems. The article builds upon the work of Charisi et al. [52] by expanding the model beyond the bi-fidelity case. This is achieved by developing compositional kernels for each fidelity level sequentially, starting from lower to higher fidelity. Furthermore, extensive analysis of the model’s performance is conducted across multiple case studies involving analytical benchmark problems with varying attributes, enhancing our understanding of its applicability to diverse design problems.

This section contains the technical details of the proposed method in Sec. 3.1, which is composed of two primary components: (1) the MF-GPs and (2) the compositional kernels. The mathematical formulation of GPs and MF-GPs is provided in Sec. 3.2, while the optimization process for the compositional kernels is described in Sec. 3.3.

The authors propose integrating compositional kernels to the linear autoregressive scheme (AR1) to facilitate design exploration. The integration of the compositional kernels aims to capture the shape of the underlying HF design space with the goal of making improved predictions with less HF analysis data.

The core concept revolves around seeking the optimal compositional kernel, comprising kernels that effectively capture distinct characteristics of the design space, such as linear or periodic patterns. Constructing these compositional kernels involves solving a discrete optimization problem. The optimization of the compositional kernel is guided by the analysis data utilized as the training set for the MF-GP.

Let us assume that the design problem involves models with fidelities ranging from $1$ to $s$⁠, where fidelity $1$ represents the lowest model and fidelity $s$ represents the highest fidelity model. The first step is to build a compositional kernel for the data of the lowest fidelity (fidelity $1$⁠) based on a single fidelity GP model (technical details in Sec. 3.2). For each fidelity $i$ ranging from $2$ to $s$⁠, a compositional kernel is built based on the bi-fidelity GP model (technical details in Sec. 3.2) using fidelity $i$ data and fidelity $i+1$ data. The process is summarized in Algorithm 1 and Fig. 1.

The kernel, a measure of similarity between data points [43], incorporates the prior beliefs and knowledge about the function $f$⁠. Kernel validity demands symmetry and positive semi-definiteness. Previous studies have produced basis functions for constructing valid covariance matrices, such as the periodic kernel for modeling repeating functions [55].

Duvenaud et al. [56] introduced compositional kernels, which are formed by combining a limited number of basis kernels through addition or multiplication. The idea of the method was to decompose the function to be learned into interpretable components.

This section presents the case studies and the results. The case studies considered encompass a range of analytical benchmark problems proposed by Ref. [15], as well as an engineering problem involving a cantilever beam. The case studies involving the analytical functions are the following:

• —

  the Forrester function (used for conceptualization)

• —

  the Jump Forrester function

• —

  the ND Rosenbrock function

• —

  the Heterogeneous function

• —

  the 2D shifted-rotated Rastrigin function

This case discusses and demonstrates the benefit of compositional kernels in early design exploration. For this analysis, 5 HF and 25 LF analysis data were used. Figure 2 shows the LF approximation represented by the blue line, the HF approximation depicted in orange, and the prediction illustrated by the black dashed line. As shown in Fig. 2(a), the prediction using the AR1 scheme with the squared exponential kernel and the given observational data does not effectively model the HF function. On the other hand, as shown in Fig. 2(b), the prediction using the AR1 scheme and the given observational data give an accurate prediction of the HF function, which represents the design space. In this instance, the kernel for the LF data was represented as a product of a linear kernel and a white noise kernel, while the squared exponential kernel was employed for the HF data. This specific case visually demonstrates that the additional information provided by the compositional kernel about the structure of the HF function can improve the performance of the framework. Thus, it is possible to make more accurate predictions with less HF data, thereby reduce the required computational cost. It is clear that the proposed approach incurs additional computational cost for developing the compositional kernels; however, the computational time and costs of running the MF analysis are less than obtaining additional HF computational or physical experimental data.

In this particular bi-fidelity case study, we used a total of 25 LF points while varying the number of HF points in the range of 5–15. The Latin hypercube sampling method was employed to generate 20 different datasets to determine the statistics of the error measures. The results are presented in Tables 1 and 2.

The results demonstrate that both, the proposed and reference models, exhibit similar performance in the cases of five and eight HF points. However, it outperforms the reference model when using 10 and 15 HF points. The improvement ranges from 1 to 22% depending on the number of HF points, but it is negative (–3%) in the case of eight HF points. This suggests that while the proposed model shows the potential for significant advancements in scenarios with limited data, where both the SF model and the reference model struggle to accurately represent the underlying design space, the amount of HF data needs to be adequate to properly capture the function’s structure. The most substantial improvement is observed when using 15 HF points. A representative case is illustrated in Fig. 3. The design space includes 10 HF and 25 LF data points. The SF model in Fig. 3(a) fails to capture the function. The reference model (Fig. 3(b)) is better but not entirely accurate. In contrast, the proposed model (Fig. 3(c)) demonstrates superior accuracy in predicting the function. In this case, the kernel function for the LF data was a multiplication of a linear and a Brownian kernel, while for the HF data, the Matérn 5/2 kernel was chosen. Based on these findings, it can be concluded that neither model adequately captures the discontinuity; however, the proposed model exhibits enhanced capability in capturing the function within the remaining domain.

In this specific bi-fidelity case study, we systematically increased the volume of HF and LF data in alignment with the number of dimensions. The quantity of HF data ranged from 45 to 125, while LF data varied from 140 to 300 data points. Importantly, this approach has no impact on the relative performance of the models, as all three models were trained with identical data volumes. To assess the statistics of error measures, LHS was employed to generate 20 distinct datasets. The resulting outcomes are presented in Tables 3 and 4. In the case of all three models, the results reveal a decline in their performance with the increasing dimensionality. This outcome aligns with our expectations, as higher dimensionality brings about greater complexity, necessitating a larger volume of training data to achieve the same level of accuracy. Despite augmenting the training data as the problem scaled, the extent of this increase did not compensate for the heightened complexity. The proposed model outperforms the reference model in all the examined cases and showed a significant improvement. It is noteworthy that in the context of 20 dimensions, the reference model proves ineffective in predicting the function (⁠$R2=−0.1913$⁠), while the proposed model maintains a satisfactory level of accuracy (⁠$R2=0.6381$⁠).

One of the well-known challenges to address when using Gaussian processes is that the computational complexity of training a GP model is known to be $O(N3)$⁠, where $N$ represents the number of data points [62]. This cubic complexity poses challenges when dealing with large datasets or high-dimensional design spaces. This significantly impacts the procedure of constructing compositional kernels, making it computationally expensive for high-dimensional input spaces. Figure 4 displays the escalating computational costs plotted against the dimensions of the function. To provide an indication of this increase, the average time of building the compositional kernels for the 4D Rosenbrock function is 4728 s, which was 13 times lower than the time required for the 20D Rosenbrock function, which was 64,557 s.

Notably, the MF approach demonstrates a significant advantage over the SF approach, particularly across the range of tested HF points. Moreover, the proposed model exhibits improved prediction accuracy across all the tested DoEs. The improvement in the predictions of the proposed model compared to the reference model ranges from $26%$ to $32%$⁠. Insights into the performance of the models can be gained from Fig. 5, which specifically focuses on the case where a dataset of $5$ HF points and $25$ LF points is shown. In Figs. 5(a) and 5(b), it is evident that both the SF model and the reference model struggle to accurately predict the shape of the function. The proposed model employed a kernel function comprising the multiplication of the linear and Brownian kernels for the LF data, while a squared exponential kernel was used for the HF data. Notably, the proposed model achieves a more precise representation of the function throughout the entire domain. However, one drawback of the method is that the uncertainty bounds are reduced even in the area close to $x=0$⁠, where the model fails to capture the structure of the function.

The outcomes are displayed in Tables 7 and 8. It is evident that the GP HF model falls short in capturing the underlying function. In contrast, both the proposed and the reference model demonstrated substantial enhancements. The reference model shows improvements ranging from 21% to 34%, while the proposed model delivers enhancements between 61% and 75%. In terms of statistical error analysis, we generated RMSE plots for the three models over 20 iterations to assess error convergence. As illustrated in Figs. 7(a)–7(c), it is clear that the error values reach a plateau around iteration 16. The case study was extended to a three-fidelity scenario, and the outcomes are presented in Tables 9 and 10. These results exhibit analogous trends to the bifidelity case, with the distinction that both the proposed and reference models show greater improvements in predictions, ranging from 32% to 63% and from 72% to 79%, respectively. Emphasizing the impact of incorporating additional models, it is worth noting that this results in escalating computational costs. Figure 8 illustrates the computational cost for both the bifidelity and trifidelity scenarios.

The proposed framework was tested on a structural design problem involving a cantilever beam. This particular problem was chosen because it serves as a simplified representation of real-world, complex engineering problem, such as estimating lifetime loads on intricate structures like aircraft or ships. The formulation of the problem was taken from Ref. [51] and modified.

The cantilever beam is shown in Fig. 9(a). The square-section beam is fixed to the wall on one end, while a concentrated load is applied to the opposite end. In addition, there is a hole on the side that is anchored to the wall. The aim is the calculation of the developed von Mises (VM) stress. The problem is set up as a bi-fidelity problem, where the LF method is the analytical estimation of the maximal von Mises (VM) stress and the HF method is the numerical estimation of the maximal VM stress. Furthermore, the problem was modeled as a two-dimensional (2D) problem, with the independent variables being the beam’s length (⁠$L$⁠), and diameter (⁠$d$⁠). The problem domain was defined within the ranges of $Lε[2.0,3.0]$ m, and $dε[0.25,0.4]$ m. The applied force was established as a constant value of $950$ kN.

The outcomes are presented in Tables 11 and 12. These results indicate that the reference model yields results that are on par with the SF model. This can primarily be attributed to the considerable disparity between the LF fidelity model and the HF model, which is attributed to the presence of the hole. In contrast, the predictions of the proposed model are closer to the HF design space. Based on Table 12, the improvement compared to the SF model ranged from 15% to 24%. An example of the problem can be visualized in Fig. 10.

In summary, the findings demonstrated that the incorporation of compositional kernels significantly enhanced the predictive capabilities of the AR1. Various analytical benchmark problems were simulated to thoroughly test the proposed model.

The 1D Jump Forrester function, representing a discontinuous space, was treated as a bi-fidelity problem. The number of HF points ranged from 5 to 15, while LF points remained constant at 25. The proposed model yielded an improvement in predictions, reaching up to 22% in the case of 15 HF points. Similarly, the 1D heterogeneous function case study followed a modeling approach akin to the 1D Jump Forrester scenario. The proposed model exhibited improvement, with predictions reaching up to 32% in the case of 15 HF points. The 2D shifted-rotated Rastrigin function, employed to assess multimodal behavior, was modeled both as a bi-fidelity and tri-fidelity problem. In the bi-fidelity scenario, the HF points ranged from 5 to 30, while LF points remained constant at 100. An improvement of 62% was observed with 30 HF points. In the tri-fidelity case, HF points ranged from 5 to 30, medium-fidelity points were held constant at 50, and LF points remained constant at 100. In this case, the improvement was measured 49% for the majority of the cases. For the cantilever beam problem, the reference model produced results comparable to the SF GP. However, the proposed model demonstrated improved results, showing enhancements ranging from 15% to 22%. Overall, these results hold promise for the application of the model in addressing complex design problems within multidimensional spaces.

The objective of the proposed method is to facilitate early-stage design exploration of complex vessels. While the presented case studies involved a lower complexity level, it is crucial to critically reflect on the scalability of the method. The authors believe that the method has undergone extensive testing in various analytical case studies, which simulate challenges sharing similarities with real design problems, such as the presence of discontinuities and complex patterns. Furthermore, the chosen case studies demonstrate good alignment with benchmark problems that hold wide acceptance within the research design community. An important consideration when applying the suggested method to high-dimensional realistic design problems is the increased computational costs associated with the development of the compositional kernels. The associated computational cost depends on the dimensionality of the problem, the size of the training set, and the number of analysis methods used in the MF model. The latter is not of interest because inherently these design problems deal with small data regimes. The impact of problem dimensionality was explored in the ND Rosenbrock case study. Additionally, the effect of integrating additional fidelity models was examined in the case of the Rastrigin function. Overall, the integration of compositional kernels introduces a trade-off between the computational benefits arising from reduced training dataset sizes and the supplementary computational expenses stemming from kernel optimization. The determining factor in this trade-off is contingent upon the nature of the design problem. Particularly in scenarios characterized by design exploration tasks featuring KPIs that are expensive to evaluate, we assert that the integration of compositional kernels presents a promising avenue.

The expansion of the method to tackle high-dimensional problems will inevitably result in increased computational expenses, presenting a notable challenge. However, it is important to note that HF analysis techniques in ship design problems such as computational fluid dynamics (CFD) analysis and model tests can be significantly (computationally) expensive. Therefore, the authors believe that the application of the proposed method and the subsequent reduction in the required number of HF simulations will yield computational benefits for design space exploration problems.

Several suggestions for future research efforts are worth considering. First, the method should be tested to real engineering engineering problems to ensure its capability to effectively capture the complex patterns inherent in real multidimensional design spaces. In the context of real-world applications, another significant consideration involves evaluating the enhancements in accuracy and the reduced uncertainty in predicting the design space. In addition, it is important to assess whether the computational benefits stemming from the reduced necessity for HF simulations are offset by the computational costs associated with optimizing the kernel function.

This publication is part of the project “Multi-fidelity Probabilistic Design Framework for Complex Marine Structures” (project number TWM.BL.019.007) of the research program “Topsector Water & Maritime: The Blue Route,” which is (partly) financed by the Dutch Research Council (NWO). The authors thank DAMEN, the Netherlands Defence Materiel Organisation (DMO), and MARIN for their contribution to the research.

There are no conflicts of interest.

The datasets generated and supporting the findings of this article are obtainable from the corresponding author upon reasonable request.