Agent-Based Hull Design Optimization Using Graph Neural Networks and Reinforcement Learning
Daniel Akinmulewo, Moritz Schäfer Schiffbau-Versuchsanstalt Potsdam GmbH, Potsdam
In light of the ongoing decarbonization of the maritime industry and the associated need to reduce greenhouse gas emissions, the demand for energy-efficient and hydrodynamically optimized ship hulls is steadily increasing. However, conventional design and optimization approaches often reach practical limits due to the high dimensionality of hull geometries and the vastness of the design space. As a result, optimization processes are frequently terminated once contractual requirements have been met, or they remain restricted to local improvements because of time and cost constraints. AI-driven design methodologies offer the potential to explore complex design spaces more efficiently, automatically generate diverse hull variants, and rapidly evaluate their performance.
Within the research project *ShipNET – AI Ship Modelling Approach for Multidimensional Design* (Funding Code: 49MF2500089, INNO-KOM), an AI-driven, object-based, and modular design framework for ship hulls is being developed. The objective is to integrate intelligent AI agents as adaptive co-pilots into the design process, supporting and accelerating decision-making, variant generation, and hull-form optimization.
As part of the development phase, two specialized AI agents are being implemented:
• Agent for generating parametric hull variations
• Agent for evaluating hull variants
Agent for Generating Parametric Hull Variations:
The primary objective of this agent is to reduce the number of design variants that must be investigated, thereby minimizing development time, computational effort, and associated costs. Its functionality is intended to emulate the decision-making process of an experienced hull designer. Human designers are able to discard unpromising design alternatives early based on prior experience, taking into account factors such as manufacturability, aesthetics, and the performance of similar geometries.
Due to the high dimensionality of geometric hull data and its limited direct correlation with hydrodynamic performance indicators, raw hull geometries are only partially suitable for conventional machine-learning approaches. To address this challenge, the generative agent first transforms the hull geometry into a compact latent representation using a variational autoencoder.
Within this latent space, essential geometric characteristics such as bow shape, waterline profile, parallel midbody, and stern geometry are extracted and made parametrically accessible. This enables targeted and flexible modifications of individual hull sections while preserving global shape consistency and local surface quality.
The modified geometry is subsequently reconstructed into a coherent three-dimensional hull form using a decoder. Building upon this representation, Reinforcement Learning (RL) is employed to autonomously explore the design space. The agent learns to perform both global and local shape modifications while maintaining constructively feasible and hydrodynamically valid hull forms.
Optimization is guided by a reward function that incorporates CFD results, surrogate-model predictions, and human feedback. The reward function accounts for performance metrics such as resistance and flow quality, as well as designer preferences. Conceptually, the approach mirrors the iterative workflow of experienced hull designers while automating the process through data-driven learning.
Agent for Evaluating Hull Variants:
The evaluation agent addresses the significant computational expense associated with conventional viscous CFD-based performance analyses. To achieve this, a surrogate model based on Graph Neural Networks (GNNs) is being developed. The model directly processes hull geometries and learns the relevant relationships between geometric features and hydrodynamic behavior.
Unlike conventional surrogate models that rely on a limited set of global shape parameters, the GNN-based agent utilizes the complete geometric representation as input. This allows complex local shape effects to be captured and enables robust predictions of hydrodynamic performance during the early design stages. These predictions can be used to rapidly assess generated design variants and, if desired, to identify promising candidates for subsequent high-fidelity viscous CFD simulations.
The presented agent-based approach combines generative geometric modeling, Reinforcement Learning, and graph-based performance prediction models into an integrated AI framework for multidimensional hull optimization. The methodology has the potential to significantly accelerate future ship design processes while enabling a more comprehensive exploration of innovative hull concepts.
