1 of 9

Southern Ocean Dynamics Under Climate Change: New Knowledge Through Physics-Guided Machine Learning

NeurIPS 2023 Workshop: Tackling Climate Change with Machine Learning

1Dept. of Computer Science/Mathematics, Harvey Mudd College, Claremont, USA

2Computational Climate & Ocean Group, Dept. of Computer Science, University of California, Davis, USA

3NOAA/OAR Geophysical Fluid Dynamics Laboratory, Princeton, USA

4ECMWF, Bonn, Germany

5IPSL/IRD, Paris, France

William Yik1,2,3, Maike Sonnewald2,3, Marianna Clare4, Redouane Lguensat5

2 of 9

The ocean and climate change

  • The ocean, covering over 70% of the globe, has absorbed more than 90% of recent warming.
  • Models predict changes in complex ocean systems.
  • Example: shifts in location/strength of the Antarctic Circumpolar Current (ACC)
  • However, the physical drivers behind these changes are not well understood.

Image: NOAA/Atlantic Oceanographic & Meteorological Laboratory

Böning, C. W., Dispert, A., Visbeck, M., Rintoul, S. R., & Schwarzkopf, F. U. (2008). The response of the Antarctic Circumpolar Current to recent climate change. Nature Geoscience, 1(12), 864-869.

3 of 9

Global climate modeling

  • Coupled Model Intercomparison Project Phase 6 (CMIP6)
    • Standardized experimental design and distribution protocol
    • Massive amounts of data (23.4 PBs shared, still sparse)
  • Hard to disseminate
    • Understanding how the underlying physics of the ocean is changing is difficult

Image: Lawrence Livermore National Laboratory

O'Neill, B. C., Tebaldi, C., Van Vuuren, D. P., Eyring, V., Friedlingstein, P., Hurtt, G., ... & Sanderson, B. M. (2016). The scenario model intercomparison project (ScenarioMIP) for CMIP6. Geoscientific Model Development, 9(9), 3461-3482.

4 of 9

Tracking global Heating with Ocean Regimes (THOR)

5 of 9

Step 1: Mesoscale unsupervised clustering

  • Native Emergent Manifold Interrogation (NEMI) utilizes Uniform Manifold Approximation and Projection (UMAP) paired with agglomerative clustering

​

​

​

​

​

​

  • Partitions the ocean grid cells into clusters (dynamical regimes) based on their physics

McInnes, L., Healy, J., & Melville, J. (2018). Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426.

Sonnewald, M. (In review). A hierarchical ensemble manifold methodology for new knowledge on spatial data: An application to ocean physics. JAMES.

6 of 9

Step 2: Learning regimes from readily available fields

  • Inputs
    • Sea surface height (ZOS) + x/y gradients
    • Depth (column height) + x/y gradients
    • Wind stress curl (∇⨉𝜏)
    • Depth summed monthly mass transport (umo_2d + vmo_2d)
    • Coriolis parameter (f)
  • Labels: 6 dynamical regimes identified by NEMI
  • Ensemble of 50 feedforward MLPs for uncertainty quantification

7 of 9

Step 3: Predicting regimes under climate change

THOR reveals a shift in physics where the Antarctic Circumpolar Current (ACC) meets the Pacific Antarctic Ridge (PAR).

8 of 9

Key contributions

  • Mesoscale inference of subsurface dynamical regimes
  • THOR guides further exploration where the Antarctic Circumpolar Current (ACC) meets the Pacific-Antarctic Ridge (PAR)
  • THOR reveals a shift in dynamics
  • Due to increased wind stress, the ACC moves away from the rough surface of the PAR into a flatter region where it flows more freely

9 of 9

Future directions and conclusion

  • Comparing CMIP models could give insight into how different representations of ocean physics affect predictions
  • Predicting dynamical regimes with spatially aware neural networks, without trading off with explainability
  • Questions? Contact wyik@hmc.edu

Read our paper!