Machine learning in ocean data assimilation: Advances, gaps and the road to operations

This review examines recent advances in the application of machine learning to ocean data assimilation, covering contributions published between 2020 and 2025. We identify emerging trends, recurring limitations, and critical open questions, structuring the discussion around four scientific challenges: observation integration, boundary treatment, fine-scale process representation, and physical consistency. While convolutional neural networks remain widely used, particularly in bias correction and super-resolution tasks, recent studies increasingly employ multilayer perceptrons, long short-term memories, transformers and neural operators for error estimation, sequential bias correction, and latent-space assimilation. Despite this architectural diversity, most contributions remain confined to idealized configurations or offline modules, with limited evidence of generalization and integration into operational pipelines. We conclude that hybrid systems combining embedded physical knowledge with systematic validation across different oceanic regimes will be essential to unlock the full potential of machine learning-enhanced ocean data assimilation. • Review of ML-ODA for irregular data, boundaries, fine scales and physics embedding. • Mapping ML architectures by spatial scale and model complexity. • Introduction of a maturity ladder for operational readiness of ML–ODA. • Cross-cutting bottlenecks: benchmarks, uncertainty, computational cost. • Strategic development roadmap: physics-aware, multi-scale and explainable ML.