Generative data assimilation highlights fronts as key regulators of ocean energy cascade

Scott A. Martin; Georgy E. Manucharyan; Patrice Klein

Source-linked AI summary

Generative data assimilation highlights fronts as key regulators of ocean energy cascade

Scott A. Martin, Georgy E. Manucharyan, Patrice Klein

arXiv:2608.14955v1Aug 15, 2026physics.ao-phcs.AI

TL;DR

How submesoscale motions shape mesoscale eddy energetics remains uncertain because observations cannot broadly resolve these currents. The paper combines satellite observations with generative data assimilation to reconstruct kilometer-scale surface states, resolving small eddies and fronts that conventional products miss.

Problem
How submesoscale motions shape mesoscale eddy kinetic energy and its variability remains insufficiently quantified because observations lack broad, submesoscale-resolving coverage.
Method
The study combines sparse satellite observations with a generative deep-learning prior learned from high-resolution simulations to reconstruct gap-free, kilometer-scale surface ocean states.
Results
GenLLC reconstructs 5–10 km eddies and fronts with submesoscale-resolving fidelity, including strongly ageostrophic dynamics absent from geostrophic and existing data-driven products.
Takeaways & Limitations
The framework provides an observation-constrained basis for characterizing submesoscale dynamics and evaluating their contribution to mesoscale eddy energetics.

Abstract

from arXiv · show

Mesoscale eddies are fundamental to the ocean circulation, yet the extent to which submesoscale motions, a few kilometers across, influence mesoscale eddy energetics through a kinetic energy cascade remains uncertain. High-resolution simulations predict that submesoscale fronts are key regulators of the cascade, transferring energy both downscale towards dissipation and upscale to sustain and shape the seasonality of mesoscale eddies. Testing these predictions has remained difficult because existing observations and state estimates cannot resolve submesoscale currents over sufficiently broad domains. Here we map the ocean's submesoscale energy cascade by combining multi-source satellite observations with a generative deep learning framework, reconstructing gap-free, kilometer-scale surface currents with physically plausible dynamics learned from simulations. Applying this to the eddy-rich Agulhas Current system, we find that submesoscales energize the mesoscale through an upscale energy cascade above 10 km, contributing to the seasonality of mesoscale eddies. Below 10 km, convergence at submesoscale fronts drives a downscale cascade towards dissipation. Both upscale and downscale pathways concentrate within fronts, where cross-scale transfer is up to an order of magnitude more efficient. Despite their limited extent, fronts account for a substantial fraction of the domain-integrated cascade, establishing them as key regulators of the cascade and targets for next-generation eddy parameterizations.

Introduction

Submescale fronts may regulate mesoscale eddy energetics through seasonal upscale and downscale kinetic-energy cascades, but observations and existing state estimates have lacked broad, kilometer-scale current coverage. The study addresses this gap with generative data assimilation that combines satellite observations and simulation-learned physical priors to reconstruct the Agulhas system and map its cascade.

Motivation: Submesoscale fronts are proposed as intermediate engines that transfer kinetic energy upscale to sustain mesoscale eddies and contribute to their seasonal cycle.Wintertime mixed-layer deepening energizes submesoscales, with mesoscale eddy energy often lagging the submescale signal.
Limitations: Evidence for submesoscale control has mainly come from simulations whose energetics may be distorted by resolution limits, parameterized physics, and neglected air–sea coupling.Regional observations constrain the dynamics but do not provide broad, synoptic coverage of submesoscale currents.
Approach: GenLLC combines sparse satellite observations with a high-resolution simulation prior to generate gap-free, kilometer-scale surface ocean states constrained toward the real ocean.Generative data assimilation produces physically plausible small-scale realizations rather than a single artificially smoothed state.
Reconstruction capability: GenLLC resolves 5–10 km eddies and fronts with vorticity reaching O(1) Rossby number, structures absent from geostrophic altimetry currents and unresolved by NeurOST.It reconstructs submesoscale vorticity without assimilating direct current observations by exploiting relationships with SSH and SST.
Cascade findings: In the Ring Path, the upscale cascade peaks at −3.5 µWm−3 in winter and spring, 40% stronger than in summer and autumn, while downscale transfer extends below O(10) km.The wintertime cascade coincides with mesoscale KE increasing from 10 Jm−3 in early winter to 15 Jm−3 by summer at scales above 80 km.

Frontal regions as key regulators of the bidirectional cascade

Frontal regions around mesoscale eddies organize both limbs of the bidirectional kinetic-energy cascade. Convergence drives downscale transfer at 5 km, while high strain drives localized upscale transfer at 20 km.

Frontal regions as key regulators of the bidirectional cascade: Submescale fronts at mesoscale-eddy peripheries are sites of bidirectional cross-scale kinetic-energy transfer.Frontogenesis drives the downscale cascade, while elevated strain drives the upscale cascade at larger scales.
Frontal regions as key regulators of the bidirectional cascade: At 5 km, episodic downscale kinetic-energy fluxes align with fronts around mesoscale eddies.The conditional mean flux varies inversely with divergence: convergent regions transfer energy downscale, whereas divergent regions transfer energy upscale.
Frontal regions as key regulators of the bidirectional cascade: At 20 km, high-strain frontal regions surrounding mesoscale eddies organize the upscale kinetic-energy flux.The upscale flux is strongest in high-strain regions and sharply localized.
Frontal regions as key regulators of the bidirectional cascade: 25% of the gross upscale flux comes from regions with σ > 0.4f occupying only 7.5% of the area.This represents a three-fold gain in efficiency over the domain average.

Discussion

The results provide synoptic observational evidence that submesoscale frontal dynamics transfer kinetic energy both upscale to sustain mesoscale eddies and downscale toward dissipation. GenLLC also enables space-based characterization of frontal dynamics, while reliance on simulation priors and contaminated observations limits interpretation.

Key findings: Submesoscale frontal dynamics transfer significant kinetic energy upscale to sustain mesoscale eddies, while frontogenetic convergence routes energy downscale toward dissipation.The seasonally intensified upscale cascade extends from the mesoscale down to O(10) km.
Limitations: GenLLC’s simulation-trained diffusion prior may preserve structural biases in symmetric instability, eddy–wave interactions, and summertime mixed-layer instability.The estimate relies on the prior where observations are absent, despite high-resolution observations constraining the output.
Limitations: 70%: the satellite-observation cascade is up to 70% stronger than in LLC4320 in the Agulhas Ring Path.The comparison covers a single seasonal cycle, so interannual variability may contribute to the difference.
Future directions: GenLLC’s divergence fields provide access to submesoscale vertical velocities and enable future monitoring of frontal heat fluxes from space.Vertical velocities at submesoscale fronts drive heat fluxes significant for the broader climate system.

Methods … Submesoscale-permitting ocean model simulation training data

The study combines score-based data assimilation with a video diffusion prior trained on high-resolution LLC4320 simulations to reconstruct physically plausible, submesoscale-resolving surface ocean states from sparse observations. The method assimilates observations through reverse-time sampling while learning sequential surface dynamics from five-step simulation sequences.

Score-based data assimilation: Score-based data assimilation uses an unconditional diffusion prior learned from simulations to infer complete surface dynamical states from sparse satellite observations.The inferred state includes SSH, SST, SSS, and zonal and meridional surface currents.
Score-based data assimilation: The diffusion process transforms plausible simulated ocean states into Gaussian noise, then reverses this transformation using the learned score function.Sampling proceeds along a monotonically increasing time coordinate, τ, from τ = 0 to τ = T.
Score-based data assimilation: Posterior sampling augments the unconditional score with an observation-likelihood approximation evaluated using the score network and observation operator.The posterior score is ∇x log p(x(τ)|y), and the denoised state and heuristic variance guide the likelihood approximation.
Score-based data assimilation: Reverse-time integration from random noise jointly follows the learned prior and observations, producing physically plausible states consistent with the data and enabling ensemble estimates.Different random-noise initial conditions generate ensemble state estimates.
Learning surface ocean dynamics through video diffusion: Video diffusion addresses limited single-time observation coverage by assimilating multiple consecutive time steps, exploiting SWOT’s precessing orbit and changing cloud coverage.The SWOT swath is 120 km wide, restricting single-time domains for characterizing the submesoscale cascade.
Learning surface ocean dynamics through video diffusion: The model generates five consecutive time snapshots, implicitly learning surface-ocean dynamics including eddy interactions, merging, splitting, and strain-driven deformation.These phenomena are represented in the high-resolution simulation training data.
Submesoscale-permitting ocean model simulation training data: Training uses surface data from NASA’s global 1/48° LLC4320 MITgcm simulation, which provides physically plausible dynamics across a full annual cycle.LLC4320 is described as the highest-resolution global ocean simulation with realistic forcing run long enough to cover a full annual cycle.
Submesoscale-permitting ocean model simulation training data: The network generates SSH, SST, SSS, and both surface-current components on 256 by 256-pixel patches, using 1.5-2 km grid spacing and five 12-hour averages.SST and SSS patch means are removed, while SSH is detrended with a time-varying linear plane; internal tides are filtered from SSH and currents above 1.1f before averaging.

Observation operator for submesoscale surface ocean state estimation · Geostrophy and characterizing dynamics through surface current gradients

The observation operator assimilates multi-modal satellite information while relying on large-scale geostrophic currents, enabling GenLLC to estimate submesoscale surface states. Surface dynamics are characterized through velocity-gradient components, with ageostrophic motions becoming important as geostrophic balance breaks down at submesoscales.

Observation operator for submesoscale surface ocean state estimation: The operator combines SWOT, nadir-altimeter, and NeurOST SSH observations with satellite SST and SSS information.SSH uses submesoscale-resolving wide-swath SWOT, along-track nadir altimetry, and coarse gridded NeurOST estimates.
Observation operator for submesoscale surface ocean state estimation: SEVIRI SST is coarse-grained to ∼0.05° before pixels occluded by clouds are masked.SSS assimilation is limited to coarse objective-analysis estimates because high-resolution satellite SSS observations are unavailable.
Observation operator for submesoscale surface ocean state estimation: Surface-current assimilation uses only large-scale NeurOST geostrophic currents because satellites do not observe small-scale surface currents.The operator assumes large-scale dynamics are mostly geostrophic.
Geostrophy and characterizing dynamics through surface current gradients: Surface ocean dynamics are characterized using relative vorticity, strain rate, and divergence from the velocity-gradient tensor.Together, these components characterize tracer evolution and deformation in geophysical flows.
Geostrophy and characterizing dynamics through surface current gradients: Relative vorticity characterizes local flow rotation, while strain rate characterizes fluid-element stretching and is associated with frontogenesis and cross-scale KE exchange.The velocity components are zonal and meridional currents, and the coordinates are zonal and meridional directions.
Geostrophy and characterizing dynamics through surface current gradients: Divergence measures velocity-field convergence or divergence and connects closely to local vertical velocity.δ < 0 indicates convergence and downwelling, whereas δ > 0 indicates divergence and upwelling.
Geostrophy and characterizing dynamics through surface current gradients: Geostrophic balance describes Coriolis and pressure-gradient force balance and is widely used to infer surface currents from SSH.The geostrophic components are ug and vg, with η denoting SSH, g gravity, and f the Coriolis frequency.
Geostrophy and characterizing dynamics through surface current gradients: GenLLC predicts total velocities, including ageostrophic contributions that SSH alone cannot retrieve, because geostrophic balance increasingly breaks down at submesoscales.Ageostrophic contributions become significant when Ro = ζ/f exceeds ∼0.1.

Coarse-graining computation of cross-scale kinetic energy fluxes

The section defines cross-scale kinetic-energy fluxes as coarse-grained transfers caused by nonlinear eddy interactions and describes their computation with FlowSieve on a 1 km Cartesian grid. Positive flux denotes downscale transfer, whereas negative flux denotes upscale transfer.

Kinetic-energy definition: The reference density used for surface-current kinetic energy is 1025 kg m−3.This density is denoted by ρ0.
Flux definition: Cross-scale kinetic-energy fluxes, Π_l, quantify nonlinear kinetic-energy transfer from scales larger than l to smaller scales.The flux is determined by the interaction between the large-scale strain tensor and subfilter-scale stress.
Flux interpretation: Positive Π_l indicates downscale transfer, while negative Π_l indicates upscale transfer across scale l.Downscale transfer moves energy from larger to smaller scales; upscale transfer reverses that direction.
Numerical implementation: FlowSieve computes the strength and direction of the kinetic-energy cascade through the cross-scale energy flux, Π_l.The code is parallelized for coarse-graining and cross-scale flux calculations.
Numerical implementation: All data are re-gridded to a regular local Cartesian grid with 1 km resolution using an ENU projection before coarse-graining.High-frequency interpolation artifacts introduced for coarse-gridded products are removed before cross-scale calculations.

Quantifying the contribution of frontal regions to domain-integrated KE cascade

The analysis quantifies how frontal dynamics regulate cross-scale kinetic-energy fluxes and how concentrated those transfers are within limited regions of the domain. It conditions cascade fluxes on strain and divergence, then compares cumulative transfer with the corresponding occupied area.

Frontal-dynamics dependence: Conditional means of Π_l in strain-rate and divergence bins assess how frontal dynamics co-vary with cross-scale kinetic-energy fluxes.The analysis aggregates across all reconstructions and evaluates Π_l against strain σ and divergence δ.
Cascade-direction analysis: Downscale and upscale cascades are analyzed separately by sorting grid points according to strain rate or divergence and accumulating Π_l.Downscale points have Π_l > 0, whereas upscale points have Π_l < 0.
Domain-integrated contribution: Cumulative cascade transfer is compared with the fraction of domain area below the same conditioning-variable threshold.The area fraction is computed from grid points with χ < χ∗, where N denotes the number of grid points.
Domain-integrated contribution: When cumulative transfer exceeds cumulative area, the cascade is disproportionately concentrated in a small portion of the domain.This concentration can be expressed as increased cross-scale transfer efficiency relative to the domain average.

Experiments

Experiments assess GenLLC using simulated observations with known ground truth and apply it to real observations in two distinct Agulhas-system regions. Real-world reconstructions are evaluated in well-observed windows with withheld observations enabling independent validation.

Simulated-observation experiment: An observing system simulation experiment evaluates GenLLC against simulated satellite observations sampled from a simulation with known full ground truth.Synthetic observations include altimetry, cloud-masked and coarse-grained SST, SSH, SSS, and surface geostrophic currents.
Study regions: The experiments focus on the Agulhas Return Current and Agulhas Ring Path, selected for distinct background oceanographic conditions and differing energetic environments.The Return Current is strongly energetic and bisected by the meandering Agulhas Return Current jet.
Observation coverage: The study selects space-time windows with good SSH and SST coverage, requiring at least 15% SWOT-observed pixels and 30% SEVIRI-observed pixels.The selected sequences correspond roughly to at least three SWOT passes through each domain during the 2.5-day reconstructed sequence.
Real-world experiment: GenLLC is applied to real-world state estimation in the same well-observed windows using 2024 SWOT, nadir altimetry, and SEVIRI SST observations.Real observations replace the simulated satellite observations and coarse-gridded products used in the simulation experiment.
Independent validation: Independent validation withholds SWOT observations at the second and fourth reconstructed time steps and applies additional SEVIRI cloud masks.The withheld observations provide independent high-resolution SSH and SST data for validating the reconstructed fields.

Extended Data

The Extended Data illustrates GenLLC reconstructions against real and simulated observations, ground truth, operational products, and independent drifter data. Additional diagnostics evaluate state-estimation accuracy, flow structures, kinetic-energy cascades, observation constraints, and relationships between cascade fluxes and strain or divergence.

Reconstruction examples: GenLLC reconstructions are compared with operational gridded satellite products and sparse high-resolution satellite observations across SSH, SST, SSS, vorticity, and regional vorticity structure.The SWOT pass shown was not assimilated into the state estimation.
Reconstruction examples: A simulated-observation experiment compares the same SSH, SST, SSS, vorticity, and regional vorticity fields with LLC4320 ground truth and the GenLLC state estimate.
Evaluation: State-estimation evaluation aggregates power spectra and point-wise accuracy across 22 observation windows with 20-member ensembles in the Return Current.The spectra cover SSH, SST, SSS, and kinetic energy, comparing LLC4320, GenLLC ensemble members and mean, and coarse gridded satellite products.
Evaluation: GenLLC surface-current velocities are evaluated against independent Global Drifter Program observations and compared with NeurOST using velocity distributions and predicted-versus-observed scatter plots.
Cascade diagnostics: Extended Data compares vorticity, strain, divergence, and kinetic-energy cascades between GenLLC predictions and LLC4320 ground truth across Agulhas regions and tests high-resolution observation constraints.The observation-constraint analysis contrasts winter/spring cross-scale fluxes with and without SWOT SSH and high-resolution SST assimilation, using 95% bootstrap confidence intervals.
Cascade diagnostics: The relationship between cross-scale kinetic-energy fluxes and flow structure is examined through conditional means and area contributions involving strain and divergence.The analyses use the 20 km flux Π5km against strain σ and the 20 km flux Π20km against divergence δ.

Supplementary Information · 1 Additional evaluations of GenLLC surface state estimation

Supplementary evaluations assess GenLLC state-estimation accuracy in observing-system simulation and real-world settings, including comparisons with satellite, SWOT, and drifter observations. Additional tests find relatively minor degradation from super-inertial SSH contamination while noting unresolved filtering needs and model-specific caveats.

1.1 Extended discussion of state estimation validation results from main text: CRPS evaluates probabilistic state-estimation accuracy and calibration by comparing the predicted ensemble distribution with the observed value.Its decomposition combines ensemble-mean absolute error with a spread-related term that rewards uncertainty and penalizes overconfidence.
1.1.1 OSSE: 22 well-observed Return Current windows are reconstructed with 20-member ensembles to evaluate GenLLC beyond a single snapshot.The reconstructions are compared with simulated coarse gridded satellite products in an OSSE context.
1.1.1 OSSE: Coarse satellite products strongly underestimate small-scale variance across all evaluated variables.This comparison motivates assessing GenLLC spectra over multiple well-observed windows rather than relying on one snapshot.
1.1.2 Real World: Independent withheld SWOT observations reveal submesoscale eddy-like structures coinciding with aggregated GenLLC features, although direct vorticity validation remains difficult.The real-world validation compares satellite-observable variables against withheld observations and uses geostrophic balance for the SWOT comparison.
1.1.2 Real World: GenLLC produces stronger small-scale variability than gridded products for SSH, SST, SSS, and surface currents relative to NeurOST geostrophic currents.For SST, both CRPS and mean absolute error are significantly lower than the gridded satellite product against withheld SEVIRI observations.
1.1.2 Real World: Drifter comparisons in the Ring Path provide an additional velocity check but are complicated because drifters measure total surface velocity, including processes beyond GenLLC’s targets.The Ring Path is used because it is more densely sampled than the Return Current, where energetic background currents rapidly advect drifters away.
1.2 Testing the effect of super-inertial SSH observation contamination on GenLLC state estimates: Super-inertial SSH contamination causes relatively minor degradation in GenLLC state estimates, increasing confidence in observational constraints on the submesoscale cascade.Filtering these signals from SWOT remains an open need, and the real-world effect could be smaller because LLC4320 has overly energetic internal tides.

2 Additional Methods

The method combines a diffusion-model architecture and a mixture-of-experts strategy with score-based data assimilation, using sparse multi-source satellite observations and scale-specific constraints. Sampling stabilization and carefully processed SST, SSH, and coarse satellite products support physically plausible high-resolution state reconstruction.

Model architecture: GenLLC uses a UNet backbone augmented with bottleneck self-attention, sinusoidal noise-level conditioning, and layer normalization for training stability.The conditioning embeds scalar noise levels into multi-frequency image tensors, while layer normalization replaces batch normalization when noise levels vary substantially.
Model architecture: Separate diffusion models handle low and high noise levels because one backbone performed unsatisfactorily across mesoscale-to-submesoscale dynamics.The mixture-of-experts approach lets each model focus on a distinct scale range.
Inference: Score-based data assimilation guides diffusion-prior generation with sparse observations after translating the EDM prior into the VPSDE formulation required by SDA.The translation uses a formula derived by Manshausen et al.; SDA inference settings are provided in Supplementary Table 2.
Inference: Adding Langevin corrector steps only below t = 0.5 stabilized sampling, whereas omitting them caused generated states to lose submesoscale variance.The low-noise corrector schedule retained the benefits of correction while avoiding the reported smoothing of submesoscale variance.
Observations: The assimilation uses sparse high-resolution SST from SEVIRI and SSH from nadir altimeters and SWOT, alongside coarse gridded satellite products as large-scale constraints and comparison state estimates.SEVIRI SST uses the OSI SAF Level 3 collated product at 0.05° grid and hourly resolution; SWOT uses the AVISO Level 3 2.5 km Expert product.
Observations: NeurOST surface geostrophic currents are coarse-grained with a 100 km Gaussian filter before assimilation, while the product’s effective resolution is 90-120 km in wavelength.The coarse-graining restricts assimilated geostrophic currents to larger scales where geostrophy and NeurOST estimates are considered most reliable.

3 Validity of estimating the submesoscale kinetic energy cascade in small spatio-temporal windows

The study evaluates cascade strength and seasonality within well-observed, approximately 400-km patches because the diffusion prior is restricted to 256x256 LLC4320 grid points. Agreement between small-patch and full-domain calculations supports robust cascade estimates below 100 km despite edge-effect exclusions and larger-scale uncertainty.

Domain and observational constraints: The diffusion prior generates restricted 256x256 LLC4320-grid-point domains, approximately 400 km wide, requiring good SWOT SSH and SEVIRI SST coverage.Cascade strength and seasonality are diagnosed using a coarse-graining framework within well-observed patches.
Domain and observational constraints: Excluding a border region as wide as the coarse-graining filter prevents edge effects from contaminating domain-averaged cross-scale flux estimates.Larger-scale uncertainty increases because of this border exclusion.
Validation of cascade estimates: Close correspondence between small-patch and full-domain computations supports robust kinetic-energy-cascade estimates below 100 km, without sampling artifacts.The comparison increases confidence that the main-text results are not influenced by sampling artifacts.