Presentation Information
[hyhteQZr-03]Reconstructing the ocean state from observations through Physics-Informed Neural Networks
*Bruno Buongiorno Nardelli1 (1. Consiglio Nazionale delle Ricerche - Istituto di Scienze Marine (Italy))
No observing system can independently resolve the full ocean state across the spatial and temporal scales relevant to ocean dynamics. At the same time, numerical models remain affected by uncertain initial conditions, unresolved processes, parameterizations, and numerical errors. This contribution examines how physics-informed artificial intelligence can provide observation-based dynamically consistent reconstructions suitable for integration into future operational Digital Twins of the Ocean.
We present a series of applications combining satellite and in situ observations, numerical simulations, and neural-network architectures. These range from the reconstruction of three-dimensional structure of the ocean by combining surface observations and hydrographic profiles, to the super-resolution of sea-surface height and temperature for improved surface-current estimation and physics-informed learning of lateral diffusion, to AI emulators of the diabatic quasi-geostrophic Omega equation solver.
Physical knowledge is incorporated through observational and architectural choices, as well as loss-function constraints enforcing governing equations, thermodynamic consistency, and static stability. Results show that these approaches can improve the representation of dynamically relevant structures while maintaining closer consistency with independent observations than conventional statistical reconstructions. In fact, neural networks should not be regarded merely as data-analogue generators: they learn nonlinear relations among features and can be constrained by established physical principles as well as be used to improve physical parameterizations. Physics-informed learning can therefore complement mechanistic modelling and data assimilation, supporting faster and higher-resolution ocean-state estimation. By connecting heterogeneous observations, physical models, and scalable AI components, these methods offer practical building blocks for operational, interoperable, and observation-driven ocean digital twins.
We present a series of applications combining satellite and in situ observations, numerical simulations, and neural-network architectures. These range from the reconstruction of three-dimensional structure of the ocean by combining surface observations and hydrographic profiles, to the super-resolution of sea-surface height and temperature for improved surface-current estimation and physics-informed learning of lateral diffusion, to AI emulators of the diabatic quasi-geostrophic Omega equation solver.
Physical knowledge is incorporated through observational and architectural choices, as well as loss-function constraints enforcing governing equations, thermodynamic consistency, and static stability. Results show that these approaches can improve the representation of dynamically relevant structures while maintaining closer consistency with independent observations than conventional statistical reconstructions. In fact, neural networks should not be regarded merely as data-analogue generators: they learn nonlinear relations among features and can be constrained by established physical principles as well as be used to improve physical parameterizations. Physics-informed learning can therefore complement mechanistic modelling and data assimilation, supporting faster and higher-resolution ocean-state estimation. By connecting heterogeneous observations, physical models, and scalable AI components, these methods offer practical building blocks for operational, interoperable, and observation-driven ocean digital twins.
Comment
To browse or post comments, you must log in.Log in