Generative Algorithms in Early Ship Design: An Exploration of Hull Subdivision Generation

182 Generative Algorithms in Early Ship Design: An Exploration of Hull Subdivision Generation Diego De León, NHL Stenden, Leeuwarden/Netherlands, [email protected] Herbert Koelman, NHL Stenden, Leeuwarden/Netherlands, [email protected] Abstract This paper explores the potential for a data-driven tool to aid in the early ship design process, through the generation of subdivisions for the general layout via a proof-of-concept prototype which leverages a GAN to create plausible layout alternatives. The software implementation integrates a BSP tree structure for parametrisation, and a CAD geometry implementation. To work within the intrinsic limitations of generative algorithms, the decision-making is made by a naval architect, targeting facilitating the evaluation of multiple concepts and broadening the design possibilities. The paper describes the functioning of the proof-of-concept prototype, considerations on its creation and applicability. 1. Introduction The current state of ship design is caught in between the rapid development of new computational technologies, and the challenges of a fundamental change in the industry propelled by environmental and legislative pushes towards decarbonisation and sustainability. The diversity of solutions needed for sustainable propulsion and the implementation of energy saving technologies means that the new design processes should allow for a faster evaluation of multiple solutions, which is a change in paradigm from previous methodologies that had an immutable constant in their source of energy. This multiple solution paradigm leads to the question of how to use these new computational technologies to enhance the ship design process. Within this project the proposed answer is the fast ideation at the beginning of the ship design process using generative algorithms, for the generation of multiple initial ship layouts as a base for naval architects and engineers to evaluate and work on, accelerating the initial process and allowing for the consideration of more possibilities. 2. Ship subdivision and layout rationale The general arrangement of a ship plays a critical role in determining its functional and operational performance. In the early design stages, layout decisions establish the foundation for how spaces interact, how systems are integrated, and how future technologies, such as alternative propulsion can be accommodated. Despite this central role, the general layout remains one of the least digitally supported areas in the ship design process. This gap is especially evident as the maritime industry shifts towards greener propulsion systems and more modular, adaptive vessels. Alternative propulsion solutions, often come with unique spatial and engineering requirements. Traditional design processes, relying heavily on expert intuition and manual iteration, struggle to efficiently explore the new design spaces that these technologies introduce. A tool that can rapidly generate and evaluate a wide variety of layout configurations becomes increasingly valuable in this context. 2.1. Energy transition challenges The maritime industry has seen a series of changes in its long history as new technologies become available and offer more practical means of moving a ship. Multiple transitions mean multiple study cases that show how the industry and technological landscape has taken every change, Herdzik (2023) but certain parallels can be observed. 183 Every transition has been initiated by a change in technology, the availability of a new solution that supersedes the previously dominant technology due to practical, economical or technological reasons, creating a solution-driven change DNV (2019). These historical precedents of changes in the industry differ from the one presented by the current decarbonisation challenge, where the urge to replace the dependency on fossil fuels lies not in the technical limitations of fuel oil itself, but in the external impacts it creates. This problem-driven change presents a higher degree of uncertainty, as the proposed solutions are very application dependent, creating a new design paradigm where one of the constants assumed during the conception of the vessel becomes a variable to be evaluated. 3. Technological considerations and algorithm rationale The integration of digital tools such as Computer-Aided Design (CAD), Computer-Aided Engineering (CAE), and simulation platforms has long been a cornerstone of modern ship design and engineering. These technologies enable detailed modelling, performance analysis, and iterative refinement across multiple domains: from hydrodynamics and structural integrity to machinery layout and stability, Roh (2018). However, despite their established role, these tools are typically deterministic in nature: they operate under defined rules with limited capacity to adapt or generalize beyond their programmed scope. This deterministic character, while essential for validation and certification, can impose practical limitations when exploring vast and complex design spaces or responding to emerging performance criteria. The process is often constrained by the need for explicit specification, sequential workflows, and expert interpretation, leaving little room for data-driven intuition, fuzzy logic, or emergent design discovery, Gaspar (2018). Optimisation frameworks have addressed some of these challenges by introducing iterative refinement and goal-oriented processes, yet they still rely on predefined parameters and objective functions. In contrast, data-driven approaches, particularly those enabled by machine learning, offer a complementary perspective. These methods shift the focus from calculating exact outputs to learning patterns, generalising behaviour, and uncovering structure from data. Where deterministic tools aim for precision and reproducibility, data-driven systems can support design exploration and design variation, enabling new forms of support in early-stage design and conceptual phase. Crucially, these approaches do not replace traditional engineering tools but rather extend their capabilities, bridging the gap between physical modelling and computational intuition. 3.1 Data as a resource A critical issue is that most machine learning models, particularly supervised learning approaches, require labeled data to function effectively, Huang (2024). This means that raw data must often be accompanied by annotations that indicate what it. Without such contextual labeling, data lacks the structure necessary for models to learn meaningful associations or make accurate predictions, Markova (2022). The process of labeling data is often time-consuming and resource-intensive, particularly in domains that require domain expertise as marine engineering. As a result, the availability of labeled data can become a bottleneck in the development and deployment of machine learning systems. In engineering and design contexts, including the maritime domain, data for machine learning applications originates from a range of sources. The two most common sources in practice are operational data and historical design data. Operational data refers to information generated during the actual use of a system or product, such as sensor readings, performance logs, maintenance records, and environmental conditions. In maritime applications, this could include engine performance metrics, fuel consumption rates, route tracking, or structural responses under various sea states deGeus-Moussault (2024). This data provides valuable 184 insights into how designs perform in real-world conditions, supporting tasks like predictive maintenance, performance optimization, and adaptive control. Existing design data, derived from past projects and legacy systems, whether stored as CAD models, simulation results, or design tables, represent a repository of engineering knowledge and design intent. Machine learning models can use this information as a basis for pattern recognition, benchmarking, or generating new concepts inspired by proven solutions. This form of data reuse supports the notion of "learning from experience," enabling algorithms to build upon decades of accumulated engineering practice. As an alternative to data from existing ships, synthetic data is artificially generated rather than collected from real-world observations. In engineering and design domains, this data is often produced using simulation environments or mathematical models that replicate the behaviour of physical systems under controlled conditions. Synthetic data serves as a valuable complement to real-world datasets, particularly in scenarios where empirical data is scarce, incomplete, sensitive, or expensive to obtain. However, synthetic data also comes with limitations, Picard (2023). A key concern is fidelity, whether the synthetic data accurately reflects the complexity and variability of real-world phenomena. If the data does not capture important nuances, models trained on it may fail to generalize or may overfit to artificial patterns. Another issue is the potential for bias introduced by the assumptions or simplifications embedded in the generation process. 4. Prototype development The first prototype developed in this project demonstrates a full pipeline for procedural ship layout generation based on a Binary Space Partitioning (BSP) tree structure. At its core, this prototype illustrates how early-stage layout decisions can be algorithmically generated, geometrically modeled, and evaluated, all within an open, modular framework designed for future extensibility. 4.1. CAD geometry background and BSP To support the development of a generative layout prototype, the choice of a suitable computer-aided design (CAD) backend is a critical decision. In this context, OpenCASCADE offers a powerful toolkit for 3D modeling and computational geometry. This solution is also particularly well-suited for implementing a Binary Space Partitioning (BSP) tree approach, which is central to this project’s layout generation logic. BSP trees provide a structured way to recursively subdivide a design space into functional compartments, making them ideal for representing compartmentalized layouts such as ship interiors, De Koning et all (2011). OpenCASCADE’s geometric and topological modeling capabilities allow for precise and pragmatic creation, manipulation, and visualization of the partitions defined by the BSP structure. This compatibility enables seamless integration between abstract spatial logic and concrete geometric representation. Each node in a BSP tree can be directly mapped to a volumetric shape or compartment in OpenCASCADE, ensuring that layout generation remains both computationally efficient and geometrically meaningful. This synergy allows the prototype to transition smoothly between datadriven logic and engineering with a valid geometry, an essential feature for early design workflows that combine algorithmic generation with CAD integration. 4.2. Generative Adversarial Network To address the challenge of generating plausible and diverse general layouts for ships, this project implements a Generative Adversarial Network (GAN) architecture. GANs are a class of machine learning models particularly well-suited for generative tasks, with some existing applications within ship design, Khan (2023). 185 At the core of a GAN are two neural networks with opposing goals: the generator and the discriminator. The generator attempts to create candidate layouts that mimic real designs, starting from random input (noise) or guided design parameters. The discriminator, in contrast, evaluates the validity of the generated layouts by either comparing them to a dataset of real examples or using other types of filters. The two networks engage in a zero-sum game: as the generator improves its ability to fool the discriminator, the discriminator simultaneously becomes better at detecting synthetic designs. This adversarial training dynamic drives both networks to improve continuously, resulting in progressively higher-quality generated outputs, Cresswell (2018). A key advantage of this approach lies in its modularity and adaptability. The generator and discriminator can be developed, trained, and adjusted independently. This flexibility allows to explore a variety of architectures, learning strategies, and input features without rebuilding the entire system. The discriminator is not limited to being a single neural network. The modularity of the model allows the exploration of the use of non-neural, computational discriminators. Existing engineering tools can act as evaluators that provide a judgment of generated layouts. This approach aligns well with the hybrid nature of early-stage design, where domain knowledge and engineering logic still play a crucial role. By allowing the discriminator to include engineering mathematical models, simulation tools or rulebased validators, the system gains a significant advantage in producing realistic, and also functional and constraint-compliant designs. This type of evaluation also facilitates the user to enter specific types of constraints to obtain the desired output. Another important benefit is the versatility of the model framework. Both networks can be replaced or enhanced with other machine learning methods, such as autoencoders, reinforcement learning agents, or decision trees, depending on the goals of a specific design task. The GAN framework becomes a flexible experimental sandbox in which new layout strategies can be tested and improved iteratively. The use of a GAN allows this prototype to go beyond rule-based generation by learning design patterns directly from data. This makes it possible to support early-stage designers not just with static templates, but with dynamically generated layouts that respond to learned design preferences and can evolve through training. Combined with a human-in-the-loop workflow, this approach opens the door to a powerful new class of design tools that support both automation and expert oversight. 5. Implementation and workflow A design input is expected from the naval architect utilising the tool and from the engineering files, at this stage the .iges file. For this proof of concept prototype the input expected includes the ship dimensions, basic operational requirements such as expected minimal range and desired cargo volume, and basic expected design decisions to test, such as the fuel type to evaluate. These required designs will be subject to change depending on the implementation of data exchange with other software tools within the platform and future capabilities of the requirement and engineering check modules of the discriminator. The process begins with the random generation of a BSP tree, which defines a hierarchical spatial subdivision of a cubic design volume. This volume represents the internal space of the ship, abstracted to allow flexible partitioning without yet being constrained by the hull shape. The generator is currently driven by random number generators but it is designed to be replaced by a neural network in future iterations, enabling better data-driven partitioning strategies. Once the BSP structure is generated, it is mapped into a set of 3D compartments using OpenCASCADE’s solid modeling tools. The resulting subdivided solid serves as the first stage of the layout representation. To enforce geometric realism, a Boolean subtraction operation is performed using a given hull form. This operation trims the subdivisions to fit within the available internal volume of the vessel, ensuring that generated layouts remain within feasible spatial bounds. 186 Fig.1: D^3: SEA general functioning principle Following this, each compartment is evaluated in terms of position, size, and volume, providing a basis for early-stage layout analysis. These metrics serve as feedback to the generator, forming a simplified form of discriminator logic. The discriminator, in this initial prototype, applies a set of rule-based and numerical filters, such as minimum compartment volume or geometric alignment, which help identify and discard invalid or poorly formed layouts. This first step sets the stage for future improvements where discriminators may incorporate historical layout data, engineering rules, or even existing digital tools, including those being developed in the EU Horizon SEUS platform. All layout data, including the BSP tree and resulting geometry, are stored in a .json format, offering human readability, logging, and integration with other systems such as SARC PIAS subdivision tools. For CAD interoperability, the OpenCASCADE architecture the prototype is built upon also supports export in .STEP and .IGES formats, ensuring compatibility with standard engineering workflows and tools. This flexibility positions the prototype as a contributor to the broader SEUS digital platform, enabling downstream integration with other tools for evaluation, simulation, or visualization. Where each subdivision is described by the plane to which it is parallel and the fraction of the volume at which it happens. The base subdivision continues through a "branch" divided by right and left paths of the BSP tree until it finishes on a "leaf", where extra information can be added, in this case volume and the intended use of the subdivision, for engineering use. 6. Results and analysis The first experiments with the prototype were carried out in three stages. In the first stage, BSP trees were generated without applying any constraints, producing purely random subdivisions within the design volume. These results illustrated the baseline behaviour of the generator, but also confirmed that completely unconstrained subdivisions have limited value even for early-stage ideation. In the second stage, a constraint was introduced to ensure a prescribed volumetric balance between fuel and cargo spaces. In the third stage, a minimum deck height constraint was added alongside the volume requirement; this configuration was tested but not evaluated in depth within the current work. 187 Fig.2: Simple volume distribution filter based subdivisions The layouts generated at this stage remain preliminary and abstract. While they comply with the imposed geometric filters, they do not yet exhibit the coherence or functionality expected of practical ship layouts. In their current form, the outputs are not directly applicable to engineering use. This distinction underlines the difference between algorithmically generated results and layouts designed by human labour. A naval architect would not only respect geometric and volumetric constraints, but would also incorporate a wide range of implicit, “common sense” considerations. For example, simple concepts such as a basic level of symmetry, which is a common feature in ship design due to both aesthetic and practical considerations, are not accounted for in this version of the algorithm. These layers of design logic are difficult to capture through purely numerical filtering and highlight the gap between algorithmic feasibility and professional design practice. The experiments also reveal the inherent complexity of internal subdivision in ship design. Effective compartmentalisation is a multi-variable problem, where while decisions can be reduced to a set of numerical rules as by engineering practice, a complete systematic analysis would require a very high number of considerations. Although hard constraints provide a necessary foundation, the level of complexity may benefit from more abstract reasoning. For this reason, the use of existing subdivision data as reference patterns emerges as an important complementary strategy. By learning from established examples, the algorithm may approximate more complex design rules that cannot easily be formalised, exploiting the known advantages of data-driven approaches. Another observation is that the prototype is not intended as an optimisation framework. Traditional optimisation methods target performance measures and operate within fixed sets of constraints and objectives. In contrast, the present approach seeks to support faster ideation at the earliest design stages. Its value lies in the capacity to generate multiple plausible alternatives quickly, enabling designers to evaluate a broader range of possibilities than would be practical through manual iteration alone. In this sense, the prototype is positioned as an augmentation of human design capabilities, not a replacement for them. While the results confirm the feasibility of BSP-based generative subdivision and demonstrate the potential of the approach as a tool for rapid concept generation, they also highlight the necessity of additional layers of evaluation logic, incorporation of reference data, and sustained human oversight. The tool’s role is not to deliver final or optimised layouts, but to accelerate the ideation process and augment the designer’s capacity to explore multiple pathways in the early phases of ship design. 188 7. Conclusions and future work The prototype demonstrates the feasibility of using BSP trees for algorithmic generation of ship subdivisions and validates the approach as a foundation for further development. Its primary contribution is not the delivery of optimised or directly applicable layouts, but the demonstration of a generative framework capable of producing rapid concept alternatives. By accelerating early-stage ideation, such tools can augment the work of naval architects, providing a broader range of candidate layouts to consider at the outset of the design process. The results also confirm that human supervision remains indispensable. While the generated layouts are geometrically consistent, they lack the context, experience and other implicit “common sense” knowledge that human designers naturally apply. This reinforces the need for a human-in-the-loop workflow in which the algorithm acts as a generator and the designer as evaluator and decision-maker. Future development will focus on three complementary directions. First, the integration of external software solutions, already part of the broader project platform, offers a path to rapidly extending the set of evaluative filters available. By drawing on established tools rather than developing every component in isolation, it becomes possible to considerably raise the quality and realism of the generated results. Increasing the number and sophistication of filters is expected to directly translate into more coherent and practically relevant layouts, within reason and the capabilities of the tools to explore. Second, the inclusion of training data from existing subdivisions will enable the system to move beyond purely numerical constraints. Learning from historical or reference layouts allows the generator to approximate the complexity of design logics in a similar way that human experience serves as a shortcut to the complex engineering mathematics in the design process. However, one of the main limitations in this direction is the availability of subdivision data in formats that are suitable for direct training. Overcoming this barrier will be essential for fully leveraging the potential of data-driven approaches, and potential approaches exploit the use of synthetic data and machine-learning-based data tagging and extraction tools. The prototype establishes a solid foundation for BSP-based generative subdivision and demonstrates its potential as a tool for accelerating ideation in early-stage ship design. Its further development will depend on enriching the discriminator through integration with external software and embedding knowledge derived from existing subdivisions, ultimately enabling a hybrid approach where computational generation and human expertise work in tandem. Acknowledgements With strong gratitude to the EU Horizon SEUS project and its partners: NTNU, NHL Stenden, UTU, SARC B.V., CADMATIC, Gondan Shipyards & Ulstein Shipyards. To the European Union as the main sponsor of the project. References CRESWELL, A.; WHITE, T.; DUMOULIN, V.; ARULKUMARAN, K.; SENGUPTA, B.; BHARATH, A.A. (2018), Generative adversarial networks: An overview, IEEE Signal Processing Magazine 35/1, pp.53–65 DE GEUS-MOUSSAULT, S.; SEUBERS, H.; LINSKENS, H.; CORADDU, A.; PRUYN, J. (2024), Operational Data for Sea Margin Calculations in Early Ship Design, 15th IMDC Conf., Amsterdam DE KONING, D.; KOELMAN, H.J.; HOPMAN, H. (2011), A novel ship subdivision method and its application in constraint management of ship layout design, 10th COMPIT Conf., Berlin, pp.292–304 189 DNV (2019), Comparison of alternative marine fuels, https://sea-lng.org/wp-content/uploads/2020/04/ Alternative-Marine-Fuels-Study_final_report_25.09.19.pdf GASPAR, H.M. (2018), Data-driven ship design, 9th COMPIT Conf., Pavone HERDZIK, J. (2023), Marine Fuel From The Past to The Future, Scientific Journals of the Maritime University of Szczecin, https://repository.am.szczecin.pl/handle/123456789/2780 HUANG, Q.; ZHAO, T. (2024), Data Collection and Labeling Techniques for Machine Learning, https://arxiv.org/abs/2407.12793 KHAN, S.; GOUCHER-LAMBERT, K.; KOSTAS, K.; KAKLIS, P. (2023), Shiphullgan: A generic parametric modeller for ship hull design using deep convolutional generative model, Computer Methods Appl. Mech. Eng. 411 MARKOVA, V. (2022), Impact of Data Preprocessing on Machine Learning Performance, Academia.edu, https://www.academia.edu/67070662/Impact_of_Data_Preprocessing_on_Machine_ Learning_Performance PICARD, C.; SCHIFFMANN, J.; AHMED, F. (2023), DATED: Guidelines for Creating Synthetic Datasets for Engineering Design Applications, arXiv preprint arXiv:2305.09018 ROH, M.-I.; LEE, K.-Y. (2018), Computational Ship Design, Springer VEELO, B.N. (2004), Variations of Shape in Industrial Geometric Models, PhD Thesis, NTNU, Trondheim