New Hybrid Forecasting Approach Significantly Improves El Niño Predictions

Thatched roof houses. Rural life, a remote paradise island under blue skies and white clouds and palms. — Village on South Tarawa atoll, Kiribati, Gilbert islands, Micronesia, Oceania.

A recent international study published in Nature Communications has unveiled a new forecasting method that significantly enhances the prediction of El Niño–Southern Oscillation (ENSO) events, especially during the boreal spring, historically marked by diminished forecasting reliability. The strategy combines traditional dynamical climate models with deep learning methods.

The research, co-authored by The Ohio State University's Department of Geography Max Thomas Professor of Climate Dynamics, and Byrd Center Principal Investigator Zhengyu Liu, addresses a long-standing challenge in climate science: the accurate prediction of ENSO, a phenomenon that drives major weather patterns across the globe. While both deep learning (DL), trained on vast climate simulation datasets, including those using the Convolutional Neural Networks (CNN) and 3D-Geoformer, and dynamical models of the North American Multi-Model Ensemble (NMME), based on the physics of the ocean-atmosphere system, have advanced ENSO forecasting independently, each has notable limitations.

Three charts above and three heat maps below. — **Fig. 1: State of ENSO prediction skill for DL models and dynamical models. a**, b The ACC of DL models (thick red and pink lines) and dynamical models (black and other colors lines) from 1982 to 2018 for year-round (a) and for the initial month in March (b). c–e The seasonality prediction skill with shading of ACC during 1982-2018 for 3D-Geoformer (c), CNN (d), and dynamic-model mean (e) forecasts, respectively. The shading around the lines denotes the 95% confidence interval, based on the bootstrap method (see “Methods”). Note that the intervals of DL models are relatively small, making them less visually apparent in the figures. (Full size image).

The research team developed two innovative strategies to integrate these approaches.

The first method, Strategy 1, averages forecasts from deep learning models and the multi-model mean of dynamical forecasts. This hybrid approach achieved superior accuracy compared to either method alone, particularly across difficult forecast periods like the boreal spring, when traditional predictability declines.

Two graph with different color lines followed by a heat map. — **Fig. 2: Strategy 1, dynamical-DL based on ensemble mean forecast. a** The ACC of 3D-Geoformer (red), CNN (pink), dynamical-model mean (black), and strategy 1 (blue) for year-round. b Same as (a), except for the initial month in March. c The seasonality of the prediction skill with shading of ACC of Strategy 1 during 1982–2018. The shading around the lines denotes the 95% confidence interval, based on the bootstrap method. Note that the intervals of DL models are relatively small, making them less visually apparent in the figures. (**Full size image**).

The second method, Strategy 2, uses deep learning predictions as a "first guess" to select more accurate initial conditions for individual dynamical models. This approach significantly improved long-range ENSO predictions, especially beyond eight months, for most models tested.

Nine graphs in 3 rows with different color lines. — **Fig. 3: Strategy 2, dynamical-DL based on “First-Guess”.** The colorful and black lines indicate the prediction skill of each dynamical model (a–h) and the dynamic-model mean (i) for the year-round ACC during 1982–2018. The solid lines indicate the prediction skill of the all-ensemble mean. The thin dashed-dot lines indicate the prediction skill based on Strategy 2, where the “Truth” value (i.e., the observation) is an indicator (sub-ensemble members are 50% of all-ensemble members). And the dashed (CNN) and dot (3D-Geoformer) lines indicate the prediction skill based on Strategy 2, where the results of DL models are used as indicators (sub-ensemble members are 50% of all-ensemble members). The shading around the lines denotes the 95% confidence interval, based on the bootstrap method. (**Full size image**).

Both strategies demonstrated improved skill in predicting major ENSO events, such as those in 1997–98 and 2015–16, although the deep learning models still underestimated extreme intensities.

While the study marks a significant step forward, the researchers note that operational forecasting using these techniques will require further development. Current deep learning models depend on gridded reanalysis datasets that often lag behind real-world conditions, limiting timely forecasting. Developing deep learning–based climate forecasting models that rely on scattered observational data rather than gridded reanalysis datasets could reduce dependence on update latency, thereby enhancing the timeliness and effectiveness of real-time forecasting and improving their ability to simulate extreme ENSO events.

Enhanced ENSO forecasting has wide-reaching societal and economic benefits. By integrating artificial intelligence into climate science and enhancing early-warning systems for weather and climate extremes, from droughts and floods to disrupted agricultural cycles, models strengthen the boundaries of seasonal forecasting and can improve global preparedness for climate-driven events.

Learn more about "Combined dynamical-deep learning ENSO forecasts."

News Filters:

Research

Climate Change

Publications