Introduction
The shipping industry is undergoing a fundamental transformation towards net-zero carbon emissions following the International Maritime Organization (IMO) 2050 Net-Zero Strategy.
Among the emerging low-carbon technologies, ammonia-hydrogen hybrid propulsion systems are widely recognised as a promising solution for long-distance maritime transport due to their high energy density, carbon-free operation and flexible power management. However, these systems exhibit highly nonlinear dynamics, multiple interacting subsystems, and stringent operational safety requirements. Their performance is further complicated by stochastic marine environments, including varying sea states, changing propulsion demands and equipment degradation. Developing intelligent control strategies that are both safe and data-efficient therefore remains a major scientific and engineering challenge.
Recent advances in Artificial Intelligence, particularly Deep Reinforcement Learning (DRL), have demonstrated considerable potential for autonomous decision-making in complex engineering systems. Unlike conventional optimisation approaches, DRL can learn adaptive control policies directly from operational data without requiring an accurate mathematical model. Nevertheless, existing reinforcement learning algorithms suffer from three important limitations that hinder their deployment in safety-critical marine applications.
First, current methods typically require millions of training interactions before achieving satisfactory performance, making them prohibitively expensive for real-world marine systems. Second, most AI algorithms provide limited guarantees that operational safety constraints will always be respected during learning and deployment. Finally, purely data-driven models often struggle to generalise to operating conditions that differ from those encountered during training, such as extreme weather or component degradation.
Research Aim and Questions
This PhD project aims to address these challenges by developing Physics-Aware Generative World Models for intelligent energy management of ammonia-hydrogen marine propulsion systems.
Instead of relying solely on data-driven learning, the proposed framework will integrate established physical principles, including thermodynamics, electrochemistry and energy conservation, into modern generative AI models. By embedding physical knowledge directly within the learning architecture, the project seeks to substantially improve learning efficiency while maintaining physically consistent predictions.
The project will investigate three fundamental research questions:
How can physical laws be incorporated into generative world models to improve sample efficiency and prediction accuracy?
How can mathematical safety guarantees be integrated into AI decision-making to ensure safe operation under complex marine conditions?
How can causal representation learning improve robustness and enable generalisation to previously unseen operating environments?
Methodology
Methodologically, the project will combine several rapidly developing research areas, including physics-informed machine learning, generative world models, reinforcement learning, control theory and causal artificial intelligence. The candidate will develop new machine learning algorithms and validate them using realistic simulations of ammonia-hydrogen marine energy systems. Opportunities may also exist to collaborate with industrial partners working on future zero-emission shipping technologies.
Expected Outcomes
The project is expected to contribute both fundamental AI methodologies and practical engineering solutions. Expected outcomes include new algorithms for trustworthy AI-based control, open-source simulation tools, benchmark datasets, and publications in leading journals spanning artificial intelligence, marine engineering and sustainable energy. The research will ultimately support the development of intelligent propulsion systems capable of operating safely and efficiently under highly uncertain real-world conditions.
References:
Hafner, D., Pasukonis, J., Ba, J., & Lillicrap, T. (2025). Mastering diverse control tasks through world models. Nature, 1-7.
Hao, Z., Liu, S., Zhang, Y., Ying, C., Feng, Y., Su, H., & Zhu, J. (2022). Physics-informed machine learning: A survey on problems, methods and applications. arXiv preprint arXiv:2211.08064.
García, J., & Fernández, F. (2015). A comprehensive survey on safe reinforcement learning. Journal of Machine Learning Research, 16(1), 1437-1480.
MEPC, R. (2023). 2023 IMO strategy on reduction of GHG emissions from ships.
Dong, D. T., Schönborn, A., Christodoulou, A., Ölcer, A. I., & González-Celis, J. (2024). Life cycle assessment of ammonia/hydrogen-driven marine propulsion. Proceedings of the Institution of Mechanical Engineers, Part M: Journal of Engineering for the Maritime Environment, 238(3), 531-542.