While deep learning models have shown promising capabilities in ENSO prediction, their inherent black-box nature often leads to a lack of physical consistency and interpretability. Here, we introduce ENSO-PhyNet, a Transformer-based model for ENSO prediction, which incorporates heat budget dynamical processes through self-attention computations. The model predicts sea surface temperature (SST) in the equatorial Pacific and achieves skillful predictions of the Niño 3.4 index with a lead time of up to 22 months. The self-attention maps reveal how the model makes predictions by focusing on specific processes in certain regions. Case analyses of recent El Niño and La Niña events underscore the impact of thermocline feedback and zonal advection feedback on the warming of the 2015 event, as well as the crucial role of anomalous easterlies in the emergence of the second-year La Niña in 2021. These findings demonstrate the model’s interpretability and its ability to identify signals that are physically consistent with the development of ENSO events.

虽然深度学习模型在 ENSO 预测方面表现出了良好的能力，但其固有的黑箱性质往往导致缺乏物理一致性和可解释性。在这里，我们介绍 ENSO-PhyNet，这是一种基于 Transformer 的 ENSO 预测模型，它通过自注意力计算结合了热收支动态过程。该模型对赤道太平洋海表温度（SST）进行了预测，并实现了Niño 3.4指数的熟练预测，前置时间长达22个月。自注意力图揭示了模型如何通过关注某些区域的特定过程来进行预测。对近期厄尔尼诺和拉尼娜事件的案例分析强调了温跃层反馈和纬向平流反馈对2015年事件变暖的影响，以及异常东风在2021年第二年拉尼娜出现中的关键作用。这些发现证明了该模型的可解释性及其识别与 ENSO 事件发展物理上一致的信号的能力。