Introduction:

Global meteorological agencies provide seasonal forecasts based on indicators such as El Niño–
Southern Oscillation and Indian Ocean variability. The NOAA operates forecasts up to 9 months,
while the ECMWF offers extended predictions up to 13 months. However, these physics-based
forecasting systems suffer from the chaotic nature of the atmosphere, leading to a rapid decline
in forecast accuracy as the lead times increase. This study proposes a novel seasonal forecasting
framework utilizing neural ordinary differential equations (Neural ODEs). Through comparative
experiments with the North American Multi-Model Ensemble model, we demonstrate that Neural
ODEs significantly reduce computational costs and mitigate cumulative errors, thereby enhancing
long-term forecasting capability. Furthermore, we employ multifractal detrended fluctuation analysis to quantify the memory characteristics, revealing the intrinsic limits of predictability in the climate system.

pdf : Lee et al.(2026)