An RCUNet-based sea surface wind stress model with multi-day time sequence information incorporated and its applications to ENSO modeling

Abstract

In traditional ocean-atmosphere coupled modeling for El Niño-Southern Oscillation (ENSO) studies, statistical methods are typically used to represent the instantaneous linear relationship between monthly-averaged anomalies of sea surface temperature (SST) and wind stress (τ). Recently, deep learning (DL) techniques have presented promising prospects for ENSO modeling, and the integration of neural networks (NNs) with dynamical models is an active research area. This study incorporates the Residual convolution blocks and a Convolutional Block Attention Module (the Res-CBAM block) into the original UNet configuration to build a new RCUNet-based $\tau$ model, denoted as ${\tau }_{\text{RCUNet}}$, which uses SST anomalies (SSTAs) during multi-day time intervals (TIs) to derive daily $\tau$ responses. Sensitivity tests to TIs are performed to illustrate how daily $\tau$ responses are dependent on the way multi-day SST forcings are used; the comparisons with different TIs show that when taking TI=3 days, the ${\tau }_{\text{RCUNet}}$ model can more precisely represent the relationship between SSTAs and $\tau$ anomalies. Next, daily $\tau$ anomalies obtained from the ${\tau }_{\text{RCUNet}}$ model are used to force an intermediate ocean model (IOM) in the ocean-only experiments, displaying coherent phase transitions and spatiotemporal evolutions of oceanic and atmospheric anomalies during typical ENSO events, which highlights the advantages of using the DL-based atmospheric $\tau$ model with multi-day SST time sequence information incorporated for ocean modeling. Furthermore, a new intermediate coupled model (ICM) is formed, named the ICM-RCUNet, in which the original atmospheric component represented by singular value decomposition (SVD) analyses is replaced by the ${\tau }_{\text{RCUNet}}$ model that is used as an atmospheric component, and a daily coupling is conducted with multi-day SST forcings. The ICM-RCUNet simulations exhibit interannual oscillations of atmospheric and oceanic states in the tropical Pacific, demonstrating the applicability of integrating physics-based dynamical ocean models with atmospheric NNs in ENSO-related studies. Further implications for ocean and coupled modelings using NNs are discussed.