NeuralShipper: Generative AI for the Next Generation of Ship Design and Manufacturing
Interview with Shahroz Khan, CEO – Compute Maritime
This interview with Shahroz Khan, CEO of Compute Maritime explores the company’s upcoming presentation at CDFAM Barcelona and the development of NeuralShipper, a generative AI system for ship design. The discussion covers AI-native geometric modeling, CAD-ready outputs, and how foundation models are being applied to early-stage maritime engineering workflows.
Can you introduce Compute Maritime and explain what you’ll be presenting at CDFAM in relation to NeuralShipper and its role in generative AI for ship design?
Compute Maritime is a UK-based deep tech company building the next generation of AI-native design tools for the maritime industry. Our flagship platform, NeuralShipper, is the world’s first generative AI co-pilot specifically developed for ship design, optimisation, and simulation-ready engineering workflows.
At CDFAM, we will be presenting how NeuralShipper introduces a new computational design paradigm for maritime: instead of relying on slow, manual parametric modelling, engineers can generate thousands of high-performance vessel concepts within minutes, directly from a small set of specifications.
What makes this particularly relevant to the CDFAM community is that NeuralShipper is built around a Large Geometric Foundation Model that produces CAD-native outputs, primarily as valid NURBS surfaces, enabling immediate downstream simulation and manufacturability. It is a practical example of how generative AI can transform one of the most conservative engineering sectors while addressing sustainability at scale.
You describe NeuralShipper as a generative co-pilot for maritime systems. What types of inputs does it require, and how are user-defined constraints and performance criteria handled in the design process?
NeuralShipper is designed for early-stage concept exploration, so it requires only minimal inputs to begin generating meaningful design candidates.
Typically, users provide high-level specifications such as vessel type, principal dimensions, displacement targets, speed requirements, operational constraints, and sometimes mission-specific objectives like fuel efficiency or seakeeping priorities.
From there, the platform allows designers to define custom constraints and performance criteria, which are integrated directly into the generation and optimisation loop. This means NeuralShipper does not simply generate random shapes, it generates candidates that satisfy engineering feasibility and are aligned with the designer’s intent.
The result is a human-in-the-loop workflow where AI accelerates exploration, but the naval architect remains in control of design direction and decision-making.
What does the data infrastructure behind your Large Geometric Foundation Model look like, and how do you ensure the consistency and applicability of training data across diverse vessel types?
The foundation of NeuralShipper is a large-scale geometric dataset spanning over 100,000 vessel designs across nearly all major ship categories and hull typologies.
A key challenge in maritime design is that ship geometry is highly heterogeneous: a yacht, a workboat, and a naval vessel follow very different design rules and constraints.
To address this, our data infrastructure focuses on consistent geometric representation, topology normalisation, and physics-aware filtering. We ensure the model learns from designs that are not only diverse, but also structurally valid and representative of real engineering practice.
This is what allows NeuralShipper to generalise beyond narrow parametric families and operate as a true foundation model for maritime geometry.
How do you manage the generation of CAD-ready outputs, particularly with regard to maintaining NURBS surface quality and ensuring models are valid for simulation and manufacturing?
One of NeuralShipper’s most important breakthroughs is that it is the first generative AI system capable of directly producing CAD-grade geometry, primarily in the form of smooth NURBS surfaces.
Surface quality is a fundamental limitation in most existing 3D generative models. Low-level representations often introduce artefacts or discontinuities that make the output unsuitable for CFD, optimisation, or production workflows.
We address this by making geometric validity and smoothness central objectives of the model itself. NeuralShipper is not just generating shapes, it is generating engineering-ready surfaces that can immediately be used in simulation pipelines and manufacturing-oriented design environments.
This CAD-native capability is critical because in hydrodynamics, even minor surface imperfections can significantly distort performance predictions.
In what ways do you see AI acting as a collaborative partner in early-stage maritime design, especially when information is incomplete or requirements are evolving?
Early-stage ship design is inherently uncertain. Requirements evolve, constraints shift, and designers often have incomplete information, especially when working with new propulsion systems or alternative fuels.
NeuralShipper is built to operate precisely in that space.
Rather than replacing the engineer, AI acts as a collaborative design partner: rapidly proposing feasible alternatives, expanding the design space, and enabling real-time exploration that would be impossible through manual workflows.
This allows naval architects to spend less time building baseline geometry and more time evaluating trade-offs, testing innovative concepts, and making high-impact decisions earlier.
The goal is not automation for its own sake, but acceleration of creativity and engineering intelligence.
What do you hope to share with and learn from the CDFAM community through your participation this year?
CDFAM brings together the leading community working on computational design, geometric deep learning, simulation-informed generation, and advanced manufacturing.
What we hope to share is that maritime transport, despite being responsible for moving over 90% of global trade, has been largely absent from mainstream generative design innovation.
NeuralShipper demonstrates how CAD-native foundation models can unlock sustainable, manufacturable, and high-performance design workflows in one of the world’s most critical industries.
At the same time, we are excited to learn from the CDFAM community, especially around new developments in 3D generative modelling, physics-guided AI, and scalable design-to-manufacturing pipelines.
We see this as a unique opportunity for cross-domain collaboration and for bringing maritime into the broader computational design conversation.
“We’re moving from generative AI that creates shapes, to generative AI that creates engineering-ready geometry, directly compatible with simulation, optimisation, and manufacturing.”
Register to attend CDFAM Computational Design Symposium in Barcelona to connect with Shahroz Khan and other leaders applying AI and machine learning in engineering and architecture. Join the discussion on practical computational design methods, simulation-driven workflows, and the future of AI-augmented design tools.
