Abstract
The Digital Twins of the Ocean (DTOs) represent an emerging framework for monitoring, simulating, and predicting ocean dynamics, supporting a range of applications relevant to understanding and responding to the global climate system. By integrating large-scale, multi-sourced datasets with advanced numerical models, DTOs provide a powerful tool for climate science. This review examines the role of machine learning (ML) in advancing DTOs applications, addressing the limitations of traditional methodologies under current conditions of increasing data availability from satellites, in situ sensors, and high-resolution numerical models. We highlight how ML serves as a versatile tool for enhancing DTOs capabilities, including real-time forecasting, correcting model biases, and filling data gaps where conventional approaches fall short. Furthermore, we review surrogate models that aim to complement or replace traditional physical models, offering increasing accuracy and the appeal of much faster inference for forecasts, and the insertion of hybrid models, which couple physics-based simulations with ML algorithms and are proving to be continuously improving in accuracy for complex oceanographic tasks as bigger datasets become available and methodologies evolve. This paper provides a comprehensive review of ML applications within DTOs, focusing on key areas such as water quality and marine biodiversity, ports, marine pollution, fisheries, and renewable energy. The review concludes with a discussion of future research directions and the potential of ML to foster more robust and practical DTOs, ultimately supporting informed decision-making for sustainable ocean management.