Machine Learning and Modeling for Ship Design

Machine Learning (ML) is a sub-field of Artificial Intelligence (AI), devoted to understanding and building methods that leverage data to improve performance on sets of tasks. Over the last decade, as a result of installing geospatial data systems, measuring and monitoring onboard ships, and proliferation of simulation and optimization algorithms, Big Data has become an established technology in shipping, providing a steadily expanding data flow to industry and research. As a result, the literature distribution of ML applications in shipping has undergone an exponential growth since 2005, reaching thousands of citations per year. One of the first attempts to review the relevant literature in this exponentially growing field was conducted by Huang et al. [1]. The authors highlighted the potential of ML to enhance shipping through various applications while underscoring the need to understand its current limitations and to ensure its reliability by integration with physical methods.

The aim of this Special Issue (SI) was to profile the current status of research versus the next major aim of ML-based research in shipping, namely the need for a deeper embedding in AI of ML technologies. Hence, although the SI was focused on ship design, contributions addressing the aspect of the ship’s operational lifecycle were welcome. In this context, we invited contributions in the following areas:

• ML for design and analysis, including supervised/unsupervised techniques and physics-informed models;
• Dimensionality reduction in homogeneous or heterogeneous design spaces, latent spaces ship design, and sensitivity analysis in optimization;
• ML for operational modeling, control and autonomous systems, among others.

The call for papers for this Special Issue was announced in January 2023 and closed after approximately one year. A large number of submissions were received during that time, with 17 articles (16 research articles and 1 review article: contribution 4) undergoing a rigorous review process and being selected for inclusion in this SI of the Journal of Marine Science and Engineering (JMSE). To the best of the Guest Editors’ knowledge, this is the first SI that has successfully collected relevant contributions to the area of ML technologies for ship design and operations, providing the readers of JMSE with versatile USPs (unique selling points). Following this Special Issue, a number of relevant reviews appeared in the pertinent literature which identified similar gaps, trends, and future research directions, as we will discuss later; see [2,3] and contribution 1.

The published articles fall into five main categories as briefly described below. Some articles address multiple topics, which is why a couple of articles appear in more than one of the main categories.

1.

Ship powering and performance prediction with a focus on predicting fuel consumption, engine power, resistance and propulsion factors using data-driven techniques and a wide range of ML models; see contributions 1,8,10,15.

Specifically, contribution 1 developed both ML (black-box) and mathematical (white-box) models to predict ship fuel consumption. Simultaneously, the Kwon approach for data cleaning was applied to improve prediction accuracy and minimization of white-box-produced deviations. Authors in contribution 8 compared a physics-based white-box model with a deep neural network (DNN—black-box) to predict ship engine power. The DNN, trained on extensive operational data, showed favorable agreement with real-world data. Power prediction constituted the main topic in contribution 10 as well. Specifically, multi-layer and convolutional NNs were employed to assess hull form geometries with respect to their performance, including residual-resistance coefficient, wake fraction, thrust deduction fraction, as well open-water characteristics for propellers. The results obtained were in good alignment with experimental towing-tank data, which were also used, in part, for the training process. Finally, the authors of contribution 15 addressed catamaran resistance prediction via ML models, which were subsequently utilized in hull form design optimization against resistance and structural weight while enhancing battery utilization in the electric propulsion system.

2.

Hull form and hydrofoil design and optimization using parametric models, dimensionality reduction techniques, and AI-driven generative or exploratory methods; see contributions 5,6,7,11,17.

In contribution 5, the authors used NNs to address both the forward (predicting performance from shape) and inverse (finding shape from performance) design problems for hydrofoils. The datasets were constructed with the aid of appropriate parametric models while the shape and performance encoding was augmented with geometric moments. A real-time software tool was developed to assist designers in preliminary and hydrofoil shape optimization. A similar study was conducted in contribution 7, but this focused on overcoming challenges from high-dimensional design spaces in hydrofoil optimization. A method was proposed to generate low-dimensional spaces that retain key geometrical and physical information, and it tested the shape optimization of hydrofoil profiles against maximizing lift. Design space reduction was also addressed in contribution 6 with the design space correction applied to hull forms. The proposed method employed Rough Set theory [4] to sequentially reduce the design space. Applied to KRISO container ship (https://www.nmri.go.jp/study/research_organization/fluid_performance/cfd/cfdws05/gothenburg2000/KCS/container.html (accessed on 18 November 2025)) model, the method efficiently minimized wave resistance and proved its feasibility. Finally, the studies conducted in contribution 11 and contribution 17 involved generative models applied to ship hull shapes and hydrofoil profiles, respectively. A diffusion-based generative model was proposed in contribution 11 to automatically construct novel ship hull designs that meet multiple objectives like low wave drag and increased displacement volume. Finally, authors in contribution 17 systematically compared generative against non-generative models for hydrofoil designs. They claimed that a physics-augmented non-generative model can be more cost-effective and produce more valid designs than even advanced generative models.

3.

Data processing and model enhancement for a variety of maritime applications with a focus on addressing complexity, uncertainty, and data quality issues, leveraging limited data through synthetic generation or transfer learning, and selecting optimal methods for specific classification or prediction tasks; see contributions 4,9,12,16.

The literature review conducted in contribution 4 targeted the complexity and uncertainty found in modeling turbulent phenomena while identifying limitations and suggesting data-driven future research directions. In contribution 9, the authors examined transfer learning in the context of deep neural networks to predict wake flow and resistance for containerships with flow control fins. Their method accurately predicts performance for different ship sizes using limited CFD data, thus reducing the need for new costly CFD simulations. Data quality issues are addressed in contribution 12, where several Machine Learning classification techniques were used to classify the ship’s main engine model, along with different imputation methods for handling missing values and dimensionality reduction methods. Methods using dimensionality reduction like PCA, UMAP and t-SNE with an ExtraTreeClassifier delivered the best results. Finally, overcoming the issue of limited data with synthetic data generation is the main theme in contribution 16, where the authors showcase their approach to improving ML-based determination of a containership’s main particulars at the early design stage.

4.

Fluid dynamics and wake analysis with a focus on understanding fluid phenomena like propeller wakes and turbulence using advanced data analysis and clustering techniques; see contributions 3,4,9.

Propeller wake analysis is performed in contribution 3 with a combination of particle image velocimetry and clustering techniques. This study reveals the predictable and chaotic characteristics of wake flow across different scales, providing a deeper understanding of vortex dynamics. Wake flow, viscous resistance, and propeller axial velocity were central to the study in contribution 9, where deep neural networks and transfer learning were employed in assessing the flow characteristics of container ships with control fins. Finally, turbulent flows are systematically studied under the prism of traditional and data-driven approaches in contribution 4.

5.

Navigation, control, and stability with a focus on autonomous vessel navigation, route optimization, and assessing ship stability; see contributions 2,13,14.

This final category touches more on the operational aspects of shipping with studies addressing autonomous navigation systems, ship routing, and stability assessment of small ships. A simulator is developed in contribution 2 to test an autonomous sailboat’s navigation system through waypoints. The simulator, which includes the actual pilot hardware, was validated against real sea trials and was shown to outperform human crews, while also identifying areas for algorithmic improvement. Routing is also the main topic in contribution 13. However, this time, a method for generating optimal ship routes by mining historical trajectory data and applying the A* search algorithm is suggested. The proposed approach produces shorter routes than existing methods, promising cost and emission savings. Finally, stability assessment is discussed in contribution 14, which presents a deep learning model that predicts hydrostatic curves from hull form features. Their method provides critical data for calculating capsizing risk in small ships which are generally exempt from adhering to stability regulations.

We consider that the SI was very well received based on the relevant bibliometric information that has been collected at the time of writing; see

Table 1. Bibliometric information for articles appearing in this SI (October 2025).

Article	Citations (MDPI/Google Scholar/Scopus)	Views
Contribution 1	26/27/26	10,214
Contribution 2	-/-/-	2246
Contribution 3	6/7/5	2482
Contribution 4	30/61/29	10,291
Contribution 5	6/8/5	2668
Contribution 6	1/1/1	1630
Contribution 7	4/7/4	2191
Contribution 8	6/7/6	1760
Contribution 9	4/4/4	2333
Contribution 10	6/8/6	2515
Contribution 11	20/34/20	4207
Contribution 12	1/1/1	3851
Contribution 13	11/14/10	2546
Contribution 14	2/2/2	2176
Contribution 15	8/16/8	2336
Contribution 16	7/10/7	1927
Contribution 17	3/6/2	2596
Total	141/213/136	57,969

below. However, at the same time, some gaps were identified, and future plausible methodological extensions and research avenues can be drawn from this comprehensive collection of research articles, as will be discussed in a following edition.

The maritime industry is undergoing a digital transformation, driven by the need for enhanced efficiency, reduced emissions, and increased autonomy. While traditional physics-based methods remain foundational, the proliferation of data and advances in AI and ML bring unprecedented opportunities. It is our view that this SI has managed to capture some of these tendencies in pertinent research, while at the same time highlighting the state of the art in hybrid methodologies, integrating physical principles with data-driven models, to solve complex challenges in ship design, performance prediction, and operational optimization.

As previously mentioned, we consider that this SI helps identify research gaps, plausible future research directions as well as potential extensions of the suggested methodologies. To this end, the following research avenues appear particularly promising for future research work in similar contexts:

• From hybrid to “grey-box” modeling: Development of more advanced frameworks that seamlessly integrate physical laws (white-box) with data-driven patterns (black-box). This is a gap that was identified from the very first relevant review performed in 2022; see [1]. Several authors consider Physics-Informed Neural Networks and other hybrid models to be particularly apt for addressing this issue, although other approaches might also need to be considered.
• Generalizability and transfer learning across ship types and scales: Research could focus on creating foundational models for naval architecture that can be fine-tuned for different ship types (tankers, LNG carriers, offshore vessels) or scales (from small boats to mega-ships), reducing the dependency on vast, ship-specific datasets.
• Generative AI for multi-objective, life-cycle design: Research should move beyond optimizing for a single or limited performance metrics towards generative AI models that can balance a wider set of objectives across a vessel’s life cycle. This includes not only hydrodynamic performance but also manufacturing constraints, structural integrity, operational efficiency in varying conditions, and end-of-life recyclability.
• Real-time digital twins for autonomous and decision-support systems: Research can focus on integrating the various predictive models (for power, performance, and stability) into real-time digital twins of vessels. These twins would use live sensor data to continuously update their state, enabling predictive maintenance, real-time voyage optimization that considers weather and hull fouling, and enhanced decision-support for both autonomous systems and human operators.
• Data quality, and standardization for maritime AI: The fundamental data challenges highlighted in several articles should be addressed. This includes developing robust methods for handling noisy data, creating high-fidelity synthetic data generators for rare events or novel designs, and proposing industry-wide data standards to facilitate collaboration and improve model reliability. This is one of the issues identified in all relevant reviews [1,2,3], which call for the development of public datasets to accelerate research and allow fair comparisons of different methods. Such attempts have recently started to appear in the literature; see [5] for an example.
• Knowledge transfer from-to Applied and Basic Research in AI/ML: Editors and publishers should encourage the publication of papers that study and reveal low performance, failures, or gaps encountered when applying SOTA AI/ML methods in complex engineering problems, as these may be related to the problem of shape optimization and are subject to geometric, operational, and environmental constraints. Such a policy could facilitate a productive interaction between basic and applied research communities towards developing and testing novel AI/ML methodologies.