Surface to Seafloor: A Generative AI Framework for Decoding the Ocean Interior State

Andre N. Souza \aff a, Simone Silvestri \aff a, Katherine Deck \aff b, Tobias Bischoff \aff c, Glenn R. Flierl \aff a, Raffaele Ferrari \aff a · 2025

Read Surface to Seafloor: A Generative AI Framework for Decoding the Ocean Interior State by Andre N. Souza \aff a, Simone Silvestri \aff a, Katherine Deck \aff b, Tobias Bischoff \aff c, Glenn R. Flierl \aff a, Raffaele Ferrari \aff a (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-15308.txt.

Introduction

Souza and five co-authors present a generative AI method to infer what lies below the ocean surface, in a paper from twenty twenty-five. Satellites give rich data on surface height, warmth, and salt, but direct measures at depth are rare. The deep state still drives how heat and mass move through the ocean.

Their model learns to produce full three-dimensional fields of flow and buoyancy that fit the surface data. It also gives a sense of how unsure each guess is. Tests on a simulated two-gyre ocean show the method recovers the main currents and swirls. Skill falls in a steady way as surface data get coarser or the target depth grows.

Catalog record

CREATORAndre N. Souza \aff a, Simone Silvestri \aff a, Katherine Deck \aff b, Tobias Bischoff \aff c, Glenn R. Flierl \aff a, Raffaele Ferrari \aff a
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE63 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-15308

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Surface to Seafloor: A Generative AI Framework for Decoding the Ocean Interior State Andre N. Souza \aff a, Simone Silvestri \aff a, Katherine Deck \aff b, Tobias Bischoff \aff c, Glenn R. Flierl \aff a, Raffaele Ferrari \aff a Abstract Understanding subsurface ocean dynamics is essential for quantifying oceanic heat and mass transport, but direct observations at depth remain sparse due to logistical and technological constraints. In contrast, satellite missions provide rich surface datasets—such as sea surface height, temperature, and salinity—that offer indirect but potentially powerful constraints on the ocean interior. Here, we present a probabilistic framework based on score-based diffusion models to reconstruct three-dimensional subsurface velocity and buoyancy fields, including the energetic ocean eddy field, from surface observations. Using a 15-level primitive equation simulation of an idealized double-gyre system, we evaluate the skill of the model in inferring the mean circulation and the mesoscale variability at depth under varying levels of surface information. We find that the generative model successfully recovers key dynamical structures and provides physically meaningful uncertainty estimates, with predictive skill diminishing systematically as the surface resolution decreases or the inference depth increases. These results demonstrate the potential of generative approaches for ocean state estimation and uncertainty quantification, particularly in regimes where traditional deterministic methods are underconstrained or ill-posed. 1 Introduction Inferring subsurface ocean dynamics from surface observations remains a fundamental challenge in physical oceanography. The ocean interior plays a crucial role in regulating Earth’s climate by transporting heat, carbon, and other tracers, yet direct measurements at depth are sparse due to logistical and technological limitations. In contrast, satellite missions have provided decades of continuous, high-resolution observations of sea surface height (SSH) and temperature and large-scale salinity observations, offering a vastly richer dataset than what is available below the surface ( Liu and Hu, 2024 ) . Theoretical and numerical studies suggest that surface anomalies—such as sea surface height—strongly correlate with subsurface structures at mesoscale and submesoscale scales, meaning surface data could provide meaningful constraints on the ocean interior ( Hurlburt, 1986 ; Cooper and Haines, 1996 ) . However, existing methods for reconstructing subsurface states from surface observations are often limited by strong dynamical assumptions or high computational costs. This study focuses on reconstructing interior velocity and temperature fields, with a particular emphasis on their correlations, which are essential for estimating mass and heat transport throughout the ocean. Mesoscale eddies, which account for 20–50% of total heat transport ( Su et al., 2018 ; Saenko et al., 2018 ) , are particularly difficult to infer from surface observations alone. A deterministic reconstruction, which attempts to match a single “true” state, may fail to capture these dynamically significant features. Instead, we take a probabilistic approach, where a generative model reconstructs a distribution of plausible subsurface states statistically consistent with available surface data. This methodology ensures that eddy statistics, such as their variance, spatial correlations, and contributions to heat transport, are preserved without erasing dynamically important details. Several physically motivated methods have been developed to infer subsurface ocean states from surface data, each with distinct advantages and limitations. The Surface Quasi-Geostrophic (SQG) framework has been used for subsurface reconstruction. The theory works well when the inherent assumptions in QG theory are satisfied and there exists information about the vertical stratification ( Held et al., 1995 ; Lapeyre and Klein, 2006 ; Wang et al., 2013 ; Liu et al., 2019 ) . Theoretical extensions, such as the SQG model with a multivariate Empirical Orthogonal Function have been shown to improve upon standard SQG by accurately reconstructing both surface- and subsurface-intensified eddies, particularly in regions with complex vertical stratification like the Kuroshio Extension and the Peru-Chile upwelling system ( Yan et al., 2020 ) . However, recent satellite missions such as the Surface Water and Ocean Topography (SWOT) Mission ( Fu et al., 2024 ) achieve higher resolution than traditional altimeters (down to 10 km) and resolve scales where the the QG approximations are no longer valid, necessitating additional theoretical development or alternative approaches. Theoretical limitations can be overcome in the presence of data-abundance. In particular, machine learning offers a data-driven alternative to augment or completely replace physically constrained methods for ocean state estimation. Its applications range from model emulation ( Dheeshjith et al., 2024 ) to data assimilation ( Champenois and Sapsis, 2024 ) , with increasing success in leveraging satellite observations. Linear models have demonstrated the statistical predictability of large-scale ocean currents, such as the Atlantic Meridional Overturning Circulation (AMOC), using satellite-derived data ( Cromwell et al., 2007 ; DelSole and Nedza, 2022 ) . More recently, feed-forward neural networks (FNNs) and convolutional neural networks (CNNs) have been applied to reconstruct meridional overturning circulations from ocean bottom pressure, zonal wind stress, and sea surface properties ( Solodoch et al., 2023a ; Meng et al., 2024 ) . Additionally, studies incorporating satellite and Argo float data have demonstrated skill in inferring velocity at depth ( Chapman and Charantonis, 2017 ) , while CNNs trained on quasi-geostrophic (QG) simulations have successfully mapped surface stream function anomalies to subsurface flow fields in idealized settings ( Bolton and Zanna, 2018 ) . Complementary efforts using QG theory ( George et al., 2021 ) and data assimilation techniques ( Martin et al., 2023 ) have refined our understanding of large-scale ocean dynamics. Meanwhile, operationally constrained models integrate observational data with numerical simulations, leading to more accurate reconstructions of subsurface ocean conditions ( Solodoch et al., 2023b ) . While these studies have demonstrated that satellite-derived surface information can be used to infer the ocean interior, most have relied on supervised learning, where a model is trained to minimize residuals against a single “true” state. However, a given surface ocean state does not deterministically correspond to a unique deep ocean state. A more physically sound approach requires probabilistic modeling, which allows for generating an ensemble of plausible subsurface states constrained by surface observations, while preserving the uncertainties and variability inherent in ocean dynamics. Probabilistic methods can be considered an extension of supervised learning methods as they can often reproduce the inference from such methods, as a subset of their capabilities. For example, minimizing the root-mean-square error of a probabilistic output results in the mean of the underlying distribution which is the target of supervised learning methods (barring discrepancies in training and architecture choices). Importantly, probabilistic modeling allows for uncertainty quantification (UQ) ( Forget et al., 2015 ; Souza et al., 2020 ) : not only can samples from the inferred distribution be explored, but these samples also provide estimates of plausible subsurface fields when surface information is insufficient to constrain the output fully. Generative modeling offers a flexible approach to sampling from such distributions, with score-based diffusion models emerging as a particularly effective tool for high-dimensional statistical inference. Generative-adversarial models (GANs) have also been applied to predict ocean temperature profiles from sea surface temperature ( Meng et al., 2021 ; Zhang et al., 2023 ) , but these works did not leverage the potential for uncertainty quantification. The study in Martin et al. (2025) leveraged score-based generative methods in the context of ocean surface data assimilation but again, they did not focus on uncertainty quantification; however, there have been recent efforts at examining

Read the complete text