Redefining Engineering Design: Generative AI in Ship Building
The maritime industry (Mar-I) is currently facing unprecedented pressure from environmental regulatory bodies to radically redesign its maritime assets for decarbonisation. This requires the adoption of intelligent approaches to design and the use of efficient simulation tools to revolutionize and reconfigure existing maritime systems. The push for new, more sustainable systems is propelled by significant regulatory changes, such as the IMO 2020, which mandates a reduction in emissions, and the advent of disruptive technologies within the frameworks of Industry 4.0, and 5.0.
Currently, for design tasks in the Mar-I, designers and engineers use extensively off-the-shelf parametric modellers and computational tools. These tools are characterised by conservatism, for they are built to generate shapes lying in the neighbourhood of a successful baseline/parent shape. Next, these modellers are coupled with optimisers for improving the baseline shape against performance criteria (e.g., ship wave resistance, sea-keeping, structural strength, etc.), which involve time-consuming simulations, e.g., computational fluid dynamics (CFD). At the end of the process, the new design is likely a local optimum whose shape is a minor variation of the existing one. Conclusively, the coexistence of conservative parametric modellers with high-cost simulations and a large number of design parameters needed for shape optimisation of complex shapes leads to non-efficient simulation-driven design pipeline that suffers from the curse of high-dimensionality and a limited capability to explore design spaces efficiently for delivering variant, innovative, user-centred and truly optimal designs.
Role of AI in design
The expansion of AI (Artificial Intelligence) have brought digitalisation and smartness to the core of above described contemporary design paradigm of Mar-I. Recently, machine learning, particularly in the form of scientific machine learning (SciML) has become increasingly prevalent in engineering design. This, on several occasions, has lightened the computational load from traditional solvers by building efficient low or high-fidelity surrogate models that predict performance almost instantly, thus accelerating the simulation-driven design process. Although the efforts of integrating SciML in Mar-I are increasing, the pace is extensively slow compared to other engineering fields.
Until recently, the ability of SciML models to generate innovative solutions was limited. This limitation arose from the fact that they were only built to predict the performance criteria of designs coming from very narrow design spaces. Generative models such as generative adversarial networks (GANs), variational auto-encoders (VAEs), diffusion models, and transformers have changed this by providing rich design spaces that allow for the creation of innovative shapes in addition to performance prediction.
In engineering design tasks, generative models are gaining attention for creating vast generative design spaces (GDSs). These models learn a set of latent features from the given training dataset of existing designs, which are used as design parameters to form GDSs. GDSs are not only low dimensional to expedite shape optimisation but, if properly trained, can also produce novel and valid design alternatives beyond the spectrum of the training dataset. Additionally, efforts are underway to enhance the quality of GDSs to make them physics-informed and user-centred. Physics-informed GDSs can leverage physical laws to ensure that generated designs satisfy certain performance criteria, while user-centred GDSs can incorporate user preferences and constraints to generate designs that are more aligned with the user's needs.
In short, these models have facilitated designers and engineers for a rich exploration of physics-informed design spaces to generate optimum alternatives beyond their imagination or conservatism while embedding high-level design goals.
Present design approach
There have been substantial efforts in computer-aided ship design for building robust parametric tools, but they can only handle a specific hull type. Despite their efficiency in creating valid and smooth ship-hull geometries, they cannot be readily used to generate instances of ship types that deviate significantly from their target ship types.
The current approach to ship design, primarily during the preliminary stages, relies heavily on using existing databases and designs to meet new requirements. Naval architects and designers typically begin by identifying potential designs from these databases. They then construct a parametric model based on a suitable ship-hull surface representation. This process often results in a limited design space, allowing only minor variations from a baseline design. Designers draw inspiration from existing designs, using their features and components to create a handful of potential alternatives.
However, embedding these features into a new design is complex and constructing a new parametric description for a unique shape is both expertise-driven and time-consuming. While this method has been effective for well-established ship types, it falls short in situations that demand more innovative design ideas. These situations might arise from uncommon requirements that necessitate exploring a broader design space, or from the need to revolutionize and redesign existing ship types. Such needs could stem from major regulatory changes, like the IMO 2020 mandate for reduced emissions, or the advent of disruptive technologies in the context of Industry 4.0. Examples include the adoption of non-fossil fuels like ammonia and hydrogen, and the design and operation of autonomous vessels.
Adopting a strategy that embraces more radical design ideas would not only benefit novel design tasks, such as special purpose vessels, but also provide a competitive edge for traditional players in the industry.
领英推荐
Generative AI to aid design
To tackle the aforementioned design approaches, the University of Strathclyde and the Massachusetts Institute of Technology (MIT) have each developed advanced tools based on deep generative models: ShipHullGAN and ShipGen, respectively.
ShipHullGAN
ShipHullGAN represents an innovative tool in the field of ship design, utilizing deep convolutional generative adversarial networks (GANs) for the creation and representation of various ship hull types. This tool aims to overcome the limitations of traditional parametric ship design methods that are typically restricted to specific ship types.
The model has been trained on a substantial dataset comprising 52,591 physically validated designs, encompassing a diverse array of ship types such as container ships, tankers, bulk carriers, tugboats, and crew supply vessels. To prepare this data for training, a novel shape extraction and representation strategy was employed. This process transformed all training designs into a uniform geometric representation with the same resolution. This step is crucial because GANs typically require input vectors of fixed dimensions.
ShipHullGAN incorporates a space-filling layer immediately following the generator component. This layer ensures that the generated designs encompass all design classes. During the training phase, designs are presented as shape-signature tensors (SSTs). These SSTs effectively utilize geometric moments to provide a compact geometric representation, facilitating the integration of physics-informed elements into ship design in a cost-effective manner.
The technology has undergone extensive comparative studies and optimization cases. These evaluations have demonstrated ShipHullGAN's capability to generate a wide range of designs, both traditional and novel, with geometrically valid and practically feasible shapes. This results in augmented features and versatile design spaces that significantly advance the field of ship design.
ShipHullGAN in action:
ShipGen
ShipGen is an gerneative tool designed for ship design, using a guided denoising diffusion probabilistic model. This deep generative model efficiently generates high-performing and feasible parametric ship hull designs, outputting design parameters in a tabular format. It focuses on optimizing seven performance metrics, resulting in hull designs with reduced drag and increased cargo capacity.
A key aspect of ShipGen is the use of classifier and performance guidance during the sampling process, which helps in producing hulls with better performance. The tool was trained using the ShipD dataset, a publicly available collection of parametric ship hulls.
It's important to note that the hulls generated by ShipGen are primarily for conceptual analysis and may not closely resemble traditional hull forms. They are designed with features that combine to enhance performance, reflecting a thoughtful approach in the field of ship design.