Advancing global sea ice prediction capabilities using a fully-coupled climate model with integrated machine learning

TL;DR

A hybrid modeling framework that embeds machine learning inference into the Geophysical Fluid Dynamics Laboratory Seamless System for Prediction and Earth System Research (SPEAR) climate model for online sea ice bias correction during a set of global fully coupled 1-year retrospective forecasts demonstrates that ML can demonstrably improve numerical sea ice prediction capabilities.

Abstract

We showcase a hybrid modeling framework which embeds machine learning (ML) inference into the GFDL SPEAR climate model, for online sea ice bias correction during a set of global fully-coupled 1-year retrospective forecasts. We compare two hybrid versions of SPEAR to understand the importance of exposing ML models to coupled ice-atmosphere-ocean feedbacks before implementation into fully-coupled simulations: Hybrid_CPL (with feedbacks) and Hybrid_IO (without feedbacks). Relative to SPEAR, Hybrid_CPL systematically reduces seasonal forecast errors in the Arctic and significantly reduces Antarctic errors for target months May-December, with >2x error reduction in 4-6-month lead forecasts of Antarctic winter sea ice extent. Meanwhile, Hybrid_IO suffers from out-of-sample behavior which can trigger a chain of Southern Ocean feedbacks, leading to ice-free Antarctic summers. Our results demonstrate that ML can significantly improve numerical sea ice prediction capabilities and that exposing ML models to coupled ice-atmosphere-ocean processes is essential for generalization in fully-coupled simulations.

Figures (11)

• Figure 1: Sea ice concentration (SIC) prediction error, 2018--2024. (A) SPEAR pan-Arctic RMSE. (B) RMSE difference between HybridIO and SPEAR. (C) Same as (B) but for HybridCPL. (D--F) Same as for (A--C) but for pan-Antarctic. Stippling in (C,F) show where HybridCPL outperforms HybridIO at the 95% confidence level. Errors are relative to NSIDC NASA Team observations DiGirolamo2022.
• Figure 2: March-initialized reforecast bias, 2018--2024. (A) Mean pan-Arctic sea ice extent. (B--D) Sea ice concentration (SIC) bias across entire 1-year reforecasts for SPEAR, HybridIO and HybridCPL, respectively. (E--H) Same as (A--D) but for Antarctic. Biases are relative to NSIDC NASA Team observations DiGirolamo2022. Regions in (B) are 1. GIN Sea, 2. Barents Sea, 3. Kara Sea, 4. Laptev Sea, 5. East Siberian Sea, 6. Chukchi Sea, 7. Beaufort Sea, 8. Central Arctic, 9. Baffin Bay and Labrador Sea, 10. Hudson Bay, 11. Bering Sea, 12. Sea of Okhotsk. Regions in (F) are 1. Weddell Sea, 2. Indian Ocean, 3. Pacific Ocean, 4. Ross Sea, 5. Amundsen and Bellingshausen Sea.
• Figure 3: Ice-ocean (IO) and coupled (CPL) sea ice concentration (SIC) increments and ML inputs. (A) Mean March--July SIC increments from data assimilation between 1982--2017, from the reanalysis-forced IO simulation. (B) Same as (A) but for the CPL simulation with atmospheric nudging. (C,D) Mean 2018--2024 March--July machine learning increments from March-initialized reforecasts with HybridIO and HybridCPL, respectively. (E,F) Same as (C,D) but for normalized sea-surface salinity. (G--J) Same as (A--D) but for Antarctic June--August. (K,L) Same as (I,J) but for normalized sea ice thickness.
• Figure 4: March-initialized HybridIO--SPEAR anomalies, 2018--2024. (A) Mean Weddell Sea anomalies in sea ice concentration (SIC), sea ice thickness (SIT), net shortwave radiation (SWn), net longwave radiation (LWn), turbulent heat flux (THF), and mixed-layer depth (MLD). THF sign convention is positive up, while LW and SW are positive down. (B--D). Average HybridIO anomalies in MLD, SST, and THF across the 1-year reforecasts. Contour shows region of anomalies in (A).
• Figure 5: February-mean Antarctic sea ice and ocean biases. (A,B) 36-year (1982--2017) sea-surface temperature (SST) and sea ice concentration (SIC) biases from the coupled DA experiment, respectively. SST bias relative to Optimum Interpolation SST data Banzon2016 and SIC bias relative to NSIDC NASA Team observations DiGirolamo2022. (C,D) Same as (A,B) but for November-initialized SPEAR reforecasts of February, 2018--2024. (E,F) February sea ice edge locations from November initialized reforecasts without and with SST nudging, respectively.