Distillation and Interpretability of Ensemble Forecasts of ENSO Phase using Entropic Learning
A new framework compresses ensemble forecasts into single models that preserve accuracy while revealing physical precursors.
Researchers Michael Groom, Davide Bassetti, Illia Horenko, and Terence J. O'Kane introduced a distillation framework for entropy-optimal Sparse Probabilistic Approximation (eSPA) models. It compresses complex ensemble forecasts of El Niño-Southern Oscillation (ENSO) phase into single, interpretable models that maintain state-of-the-art skill for predictions up to 24 months ahead. The distilled models enable spatial importance mapping, revealing known physical precursors and tracing event evolution from extratropical signals to mature ENSO states.
Why It Matters
This makes complex climate forecasts more transparent and actionable, aiding agriculture, disaster preparedness, and economic planning.