A Physics Inferred Deep Learning Retrospective of the Cane-Zebiak-Dolan 1986 Experimental El Nino Forecasts

In an ambitious experiment to conduct a 2-year forecast of the 1982/83 large El-Nino event by avoiding the intensive computational demand of a fully time dependent atmospheric coupled ocean model, M. Cane, S. Zebiak, S. Dolan [1986] employed a linear steady state, shallow water atmospheric model based on a monthly anomaly of SST forced by convergence of winds coupled to a time dependent linear ocean basin mixed layer anomaly model of A. Busalacchi, O’Brien [1981]. This experiment led to the first consistent demonstration of forecast predictability over all El Nino’s during the period 1970-1986.

After an examination of recent annual forecast results of the deep learning ensemble based DUNE model by P. Shukla and M. Halem [2024] of global forecasts of 2-meter temperatures (T2m) over land and SST over oceans , it was observed that SST forecasts were improved over climatology, persistence. As a result, it was considered feasible to explore retrospective machine learning simulations of the C-Z-D paper for this session dedicated to Prof. Mark Cane’s 80^th B-day. The DUNE model was trained on ERA5 monthly global T2m, SST, Winds at 10meters and climatological short wave radiation at the ToA at spatial resolutions of 0.25⁰x0.25⁰. To overcome AI tendencies to oversmooth predictions, we incorporate from C-Z-D the physics idea of wind convergence forcing as an additional channel input in DUNE training and also include it as an rms supplement in the loss function over the Nino 4 region.

DUNE has a spatial resolution of 0.25 degree latitude-longitude for each grid cell and for this experiment the use of monthly sliding windows was extended to 2-year forecasts (i.e. a 1 month forecast using the prior 24 months are updated by a monthly prediction replacing the prior last month repeatedly for successive 24 month predictions). Results for global oceans for a one year forecast using sliding window were better than Persistence and Climatology as calculated by RMS and ACC error statistics. Since 2-year forecasts take only seconds using the DUNE model weighted inferrences retrained from a 2 year sliding monthly window, such forecast results will be presented for every year from 1981 to 2024 in terms of RMS, ACC and maps of NINO3 and NINO4 regions.

Cane, M.A., S. E. Zebiak, & S. C. Dolan, " Experimental forecasts of EI Nino” NATURE VOL, 321 JUNE 26, 1986

Shukla P., M. Halem, “DUNE: A Machine Learning Deep UNet++ based Ensemble Approach to Monthly, Seasonal and Annual Climate Forecasting” arXiv:2408.06262v1 12 Aug 2024

Busalacchi, A.J., J. J.O’Brien," Interannual variability of the equatorial Pacific in the 1960's”, J. geophys. Res. 86, 10901-10907 (1981).