13 June Group Meeting

[andrewfagerheim]

OTG Weekly Meeting: 13 June 2025

Attendees: Dhruv Balwada, David A. Lee, Jack Pope, Yan Feier, Prani Nalluri, Andrew Marshall Fagerheim, Manuel Othon Gutierrez Villanueva

Resources:

• paper: https://github.com/ocean-transport/groupmeeting_notes/blob/main/papers/Martin%20et%20al%202024.pdf
• slides: https://github.com/ocean-transport/groupmeeting_notes/blob/main/slides/Martin%20et%20al%202024.pptx

Discussion:

• Paper Deep learning improves representation of eddy dynamics, Martin et al 2024
• Mesoscale eddies (L ~ 100km) are in geostrophic balance: balance between Coriolis force due to rotation and pressure gradients; characterized by small Rossby numbers (Ro = U/fL << 1)
• Energy transfer occurs in both directions due to eddies: forward cascade transfers energy from larger to smaller, inverse cascade transfers energy from smaller to larger
• Altimeters are instruments that measure SSH from satellites (must correct for dynamics, tides, seafloor topography, etc)
  • Method of going from individual satellite tracks to a gridded product is “objective mapping” (Bretherton et al 1976), the product for SSH is called DUACS: Data Unification and Altimeter Combination System (Le Traon et al 1998)
• Limitations of DUACS (ie the standard product): temporal decorrelation scales are prescribed (and not published?), small eddies (L < ~100m) are unresolved, nonlinear interactions are underestimated
• Neural Ocean Surface Topography (NeurOST SSH-SST) is this paper’s new product to go from sparse altimeter observations to a gridded SSH product using SST (which is available at higher resolution)
  • Train on all but one altimeter, then use this withheld data for testing
  • Unique contribution of this paper: models are never examined at all, instead they’ve withheld observational data for evaluation
  • Train one model globally, but encoders operate on smaller subdomains (~1000km x 1000km)
  • Model output is based in subdomains, smoothing is applied to create final outputs of global SSH maps
• CNNs: () 30 frames of data go into the model and 30 frames are then predicted from this, in other words use +-15 days worth of data to predict each timestep; method is “video prediction” but they only take the middle for the final output
  • Spatial encoder: takes a large amount of data and “compresses” it into a smaller size
  • Spatiotemporal translator: not only concerned about spatial information but also temporal change
  • Spatial decoder: takes compressed that were predicted by the model and turns it back into a fully gridded product
• Question we keep coming back to: does final smoothing methods make physically unrealistic predictions at the boundaries between subregions? Is this part of the reason geostrophic velocity has much higher RMSE than SSH does?
• Question: Can a model trained globally learn the dynamics of many different regions? If it’s trying to optimize both western boundary currents and eastern coastal upwelling, can it learn both well?
• Also compare NeurOST to surface drifters: hard for me to know how to interpret this because they don’t compare DUACS to surface drifters
• NeurOST has much larger KE transfers than DUACS, inverse cascade is larger in the spring and large enough to explain the winter-to-spring change in KE
• Forward cascade in the N Atlantic?

[manuelogtzv]

Also, I wonder how the NeurOST SSH and geostrophic velocities compare to in situ observations of currents from shipboard ADCPs. Lenn et al. (2018) conducted comparisons between DUACS SSH and geostrophic velocities, as well as underway total velocity measurements, in the Drake Passage, and a good correlation was observed between the two datasets (see Fig. 3 and 4 of Lenn's paper). I wonder if the correlation improves when using NeurOST fields.