Scholarlink

<p>Various physics-based dynamical and data-based statistical models have been developed for uses in predicting sea surface temperature (SST) evolution in relation to the El Niño-Southern Oscillation (ENSO) over the tropical Pacific. At present, clear limitations remain in their ENSO predictions, with predicted SST anomalies (SSTAs) being widely spread across diverse models and considerable intermodel uncertainty. Fortunately, deep learning (DL)-based modeling has recently made promising advances in ENSO prediction tasks; numerous neural networks (NNs) have been constructed for ENSO predictions. However, most NNs themselves are purely data-driven and lack constraints of the necessary physical processes in the coupled system; there are few studies in which DL models are directly integrated with physics-based dynamical models. Previously, such a new type of intermediate coupled models (ICMs) was developed by directly integrating U-Net-derived sea surface wind stress models with an intermediate ocean dynamical model (denoted as ICM-UNet), with demonstrated success in simulating ENSO evolutions in freely coupled runs. It is thus natural to take a step further for prediction applications. In this study, this new ICM-UNet is applied for retrospective ENSO predictions, the first time that such a fusion of DL atmospheric model and dynamical oceanic model with different architectures can be achieved to make ENSO predictions. The overall evaluations indicate that the ICM-UNet yields valid retrospective predictions during the period 1995–2023, confirming that the ICM-UNet is a credible ocean-atmosphere coupled model for ENSO predictions. In case studies during 2020–2023, the ICM-UNet predictions reveal that SSTAs over the equatorial Pacific evolved into a second-year cooling in late 2021 and a warming tendency in 2023, forming a three-year La Niña and an El Niño event thereafter, which is consistent with the reality. The ICM-UNet successful fusion, taking advantage of both the physical constraints due to dynamical oceanic models and nonlinear representations of wind responses due to DL capacity, further underscores the high adaptability of integrating data-driven NNs into the ocean-atmosphere coupled modeling for ENSO-related studies.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Retrospective ENSO predictions using an intermediate ocean-atmosphere coupled model by integrating deep-learning sea surface wind stress

  • 期刊论文

  • 发表日期：

    2025年11月21日

查看全文
  • Shuangying Du,
  • Rong-Hua Zhang,
  • Chuan Gao

摘要

Various physics-based dynamical and data-based statistical models have been developed for uses in predicting sea surface temperature (SST) evolution in relation to the El Niño-Southern Oscillation (ENSO) over the tropical Pacific. At present, clear limitations remain in their ENSO predictions, with predicted SST anomalies (SSTAs) being widely spread across diverse models and considerable intermodel uncertainty. Fortunately, deep learning (DL)-based modeling has recently made promising advances in ENSO prediction tasks; numerous neural networks (NNs) have been constructed for ENSO predictions. However, most NNs themselves are purely data-driven and lack constraints of the necessary physical processes in the coupled system; there are few studies in which DL models are directly integrated with physics-based dynamical models. Previously, such a new type of intermediate coupled models (ICMs) was developed by directly integrating U-Net-derived sea surface wind stress models with an intermediate ocean dynamical model (denoted as ICM-UNet), with demonstrated success in simulating ENSO evolutions in freely coupled runs. It is thus natural to take a step further for prediction applications. In this study, this new ICM-UNet is applied for retrospective ENSO predictions, the first time that such a fusion of DL atmospheric model and dynamical oceanic model with different architectures can be achieved to make ENSO predictions. The overall evaluations indicate that the ICM-UNet yields valid retrospective predictions during the period 1995–2023, confirming that the ICM-UNet is a credible ocean-atmosphere coupled model for ENSO predictions. In case studies during 2020–2023, the ICM-UNet predictions reveal that SSTAs over the equatorial Pacific evolved into a second-year cooling in late 2021 and a warming tendency in 2023, forming a three-year La Niña and an El Niño event thereafter, which is consistent with the reality. The ICM-UNet successful fusion, taking advantage of both the physical constraints due to dynamical oceanic models and nonlinear representations of wind responses due to DL capacity, further underscores the high adaptability of integrating data-driven NNs into the ocean-atmosphere coupled modeling for ENSO-related studies.