# Long-Term Prediction of Mesoscale Sea Surface Temperature and Latent Heat Flux Coupling Using the iTransformer Model

## TL;DR

A new model called iTransformer improves predictions of ocean and air interactions in key current regions, offering insights into climate changes.

## Contribution

The iTransformer model introduces a novel approach for predicting mesoscale air–sea coupling using machine learning.

## Key findings

- The iTransformer model achieved 3.0% relative improvement in predictions by incorporating sea surface temperature data.
- The model successfully reproduced linear trends and mean values of coupling coefficients.
- Predictions using iTransformer were comparable to those from traditional climate models.

## Abstract

Mesoscale air–sea interaction, which is active in Western Boundary Currents (WBCs), has a non-negligible effect on mid-latitude climate variability. The analysis and prediction of the mesoscale air–sea interaction rely on high-resolution observation datasets and mesoscale-resolving climate models, which often require long processing times to estimate future changes and have several limitations. Therefore, in this study, we used a newly developed iTransformer model, which integrates mesoscale sea surface temperature anomaly (SSTa) and latent heat flux anomaly (LHFa) coupling coefficient data to predict future changes in SSTa–LHFa coupling. First, we individually trained the model using data corresponding to 1–15 past winters from ERA5 dataset. Thereafter, we used the trained model to predict SSTa–LHFa coupling coefficient for the next 10 winters. Compared with the predictions using only the coupling coefficient, the prediction yields 3.0% relative improvements when SST data were incorporated. The iTransformer model also showed the ability to reproduce the linear trend and mean value of mesoscale SSTa–LHFa coupling coefficients. Furthermore, we chose the optimal input length for each WBC and used the model to predict changes in mesoscale SSTa–LHFa coupling in the future. The results thus obtained were comparable to those obtained using mesoscale-resolving climate models, indicating that the iTransformer model showed satisfactory prediction performance. Therefore, it provides a novel pathway for exploring mesoscale air–sea interaction variations and predicting future climate change.