ENSO events can be predicted up to nearly two years in advance with the use of an AI-enhanced forecasting model.

Every few years, changes in how the ocean and atmosphere interact along the West Coast–stretching from southern California to Peru and reaching across the Pacific Ocean roughly to Fiji and the Solomon Islands–determine climate variability around the world. To better predict the El Niño-Southern Oscillation (ENSO), an international research team applied artificial intelligence (AI) to develop a model capable of predicting up to 22 months of the phenomenon.

They posted their approach on May 17th at Ocean and Atmospheric Research.

“ENSO dominates Earth’s year-to-year fluctuations in climate and can often cause severe environmental and socio-economic impacts globally,” said first author Haoyu Wang, PhD student at the China Institute of Oceanography’s Core Laboratory of Ocean Circulation and Waves. “. “However, despite continued advances in El Niño theory and modeling, global heat signature variations preceding ENSO events have not been fully understood, particularly for ENSO long-term forecasts more than 12 months in advance.”

This outward expectation of a year is limited in part by what the researchers call the spring constancy barrier, a reference to the variability of spring as it moves from the cold of winter to the steam of summer. As temperatures change over both the sea surface and the atmosphere, the data becomes muddier and provides a more confusing indication of what to expect from an ENSO.

said corresponding author Xiaofeng Li, a professor at the Principal Laboratory of Ocean Waves Circulation. Niño 3.4 indicates the central South Pacific Ocean, midway between the outer limits of ENSO. It is one of the four El Nino indicators.

“In addition, we designed an interpretable method for monitoring the relationship between global sea surface temperature and ocean heat content with ENSO from an AI perspective.”

They called their approach the spatiotemporal information extraction and fusion model (STIEF) and trained it on historical observations of sea surface temperatures and simulations of ocean heat content data. According to Li, it has two main components: the ability to extract features of space and time and the ability to combine these features together.

A deep learning model extracts properties of temporal and spatial ocean data in parallel. Then you use what you learn from those separate data points to understand how they relate based only on the immediate past data points. This allows the model to avoid the pitfall of assuming that a future data point is the result of an incremental progression, which compensates for rapidly changing changes in stability barriers in the spring.

According to Wang, the team also designed the model to understand how to process different data points retrospectively to make predictions. The processing is usually too complex to extract specific data and track how the model uses it in its predictions. Dubbed the “black box” case, researchers can see input variables and output predictions, but the process remains a mystery.

“We designed an interpretable approach to solve the ‘black box’ problem in the AI-based ENSO prediction model,” Wang said. “This allows us to observe correlations between different variables from an AI perspective, providing new insights for our theoretical research on ocean forecast phenomena.”

The researchers said they plan to further refine their model and eventually apply it to all four Niño indices to explore ENSO diversity. The ultimate goal is to create an interpretable AI model applicable to predictions of various ocean phenomena.