AI-Based Reconstruction of Eddy-Resolved Subsurface Currents in the Indo-Pacific Convergence Zone
ID：814
Oral Presentation
2027-01-13 09:00 （China Standard Time）
Session：Session 32 - AI for Ocean Dynamics: advancements, discoveries and challenges
Abstract
Non-equilibrium processes are prevalent in the ocean and play a dominant role in governing energy exchange, vertical transport, and turbulent mixing. However, robust characterization of these processes remains challenging due to sparse observations and uncertainties in numerical models, particularly in dynamically active regions. This study develops a deep learning–based subsurface current reconstruction (DLSCR) model for the Indo-Pacific Convergence Zone (IPCZ), a dynamically complex region with multiscale circulation. The model reconstructs eddy-resolving (1/12°) three-dimensional subsurface ocean currents down to 643 m from sea surface information—including sea surface currents, height, temperature, salinity, and wind stresses. DLSCR demonstrates robust reconstruction skill throughout the upper 400 m of the ocean, with low root mean square errors and high spatiotemporal correlations. This capability is further illustrated by the model’s representation of key upper-ocean dynamical features in the IPCZ, including the spatiotemporal evolution of subsurface mesoscale eddies and the three-dimensional structure of the subtropical gyre. Beyond reconstruction skill, a perturbation-based interpretability analysis reveals surface currents (U/V) and SSH as the two primary contributors. Surface currents (U/V) affect subsurface current reconstruction across the full depth range, with the strongest sensitivity occurring in the mixed layer (0–50 m). In contrast, SSH exerts its primary influence in the thermocline (50–150 m). These results demonstrate that DLSCR captures physically meaningful surface–subsurface relationships in the IPCZ, providing an interpretable, data-driven framework for reconstructing eddying ocean interior.
Speaker
Author
Important Dates
-
Jan 30
2026
Session / Event proposals open
-
Mar 31
2026
Session / Event proposals close
（Extended to April 7） -
Apr 30
2026
Session acceptance notifications
-
Apr 30
2026
Abstract submissions & registration open
-
Jun 30
2026
Abstract submissions close
（Extended to July 21） -
Aug 31
2026
Abstract acceptance notifications
-
Sep 15
2026
Scientific program released
-
Oct 15
2026
Early bird registration closes
-
Jan 12–15
2027
XMAS 2027