Every few years, changes in the ocean-atmosphere interaction along the west coast, which stretches from Southern California to Peru and the Pacific nearly to Fiji and the Solomon Islands, determine climate change around the world. To more accurately predict this El Niño Southern Oscillation (ENSO), an international research team applied artificial intelligence (AI) to develop a model that can predict this phenomenon for up to 22 months.
They announced their approach on May 17th: Ocean, land and atmosphere research.
“ENSO dominates the Earth’s year-to-year climate variability and can often cause severe environmental and socio-economic impacts on a global scale,” said lead author, Ocean Circulation Wave Focus Laboratory, Institute of Oceanography, Chinese Academy of Sciences. PhD student Haoyu Wang said. “However, despite continued advances in ENSO theory and modeling, the variability of global heat signatures preceding ENSO events remains completely unexplained, especially for long-lead ENSO projections more than 12 months in advance. is not understood by
This outside year projection is limited in part by what the researchers call the “spring persistence barrier.” This refers to the spring seasonal variation as it transitions from the freezing season of winter to the steam of summer. Changes in temperature in both the sea surface and the atmosphere will muddy the data, further obscuring what you would expect from ENSO.
“In this paper, we used an AI method to predict the ENSO event, achieving a 22-month valid prediction period for Niño 3.4 and minimizing the impact from prediction barriers in the spring,” said corresponding author, Ocean Circulation. Xiao Feng Li, Professor of Key Research Institute, said. wave. Niño 3.4 points to the middle of the South Pacific, halfway between ENSO’s outer boundaries. One of four Niño exponents.
“Furthermore, we have designed an interpretable method to observe the relationship between global sea surface temperature and ocean heat content from an AI perspective using ENSO.”
They named their approach the spatio-temporal information extraction and fusion (STIEF) model, which they trained on historical observations of sea surface temperature and simulated ocean heat data. According to Li, this includes his two key components. It is the ability to extract features of space and time, and the ability to fuse those features.
Deep learning models extract temporal and spatial properties of ocean data in parallel. It then uses what it learns from those individual data points to understand how the data points are correlated based solely on the previous data point. This allows the model to avoid the pitfall of assuming future data points are the result of gradual read-ups and to compensate for rapidly changing changes in the spring persistence barrier.
According to Wang, the team also designed the model to retrospectively understand how different data points were processed to make predictions. They are usually too complex to extract specific data and track how the model used it for predictions. Called the ‘black box’ problem, this problem allows researchers to see input variables and output predictions, but the process remains a mystery.
“We designed an interpretable approach to solve the ‘black box’ problem in AI-based ENSO forecasting models,” said Wang. “This will allow us to observe correlations between different variables from an AI perspective, providing new insights into theoretical studies of ocean prediction phenomena.”
The researchers said they plan to continue improving the model and eventually apply it to all four Niño indices to investigate ENSO diversity. The ultimate goal is to establish an interpretable AI model that can be applied to predict various ocean phenomena.
For more information:
Haoyu Wang et al., Interpretable Deep Learning Models for ENSO Prediction, Ocean, land and atmosphere research (2023). DOI: 10.34133/olar.0012
Courtesy of Ocean Land Atmosphere Research (OLAR)
