Physics-guided deep learning for global sea surface temperature forecasting: Balancing accuracy and stability across timescales

Accurate sea surface temperature (SST) forecasting across multiple timescales remains challenging. Daily forecasting frequently relies on autoregressive models prone to instability and over-smoothing, whereas monthly forecasting suffers from sparse data and the complex dynamics of ocean systems. Existing deep learning methods struggle to address these diverse challenges simultaneously. We introduce SSTFormer, a novel physics-guided deep learning framework that achieves leading results, with root mean squared error of 0.17 ℃ for daily forecasts and 0.60 ℃ for monthly forecasts, yielding lower bias and improved spatial coherence. The model’s core innovation is its unified and flexible architecture. For multi-step daily forecasts (1-15 days), it deploys as a “two-phase sequential ensemble” that replaces conventional autoregression and uses ocean current to solve instability and mitigate error accumulation. For single-step monthly forecasts, it is used in a direct forecasting configuration, proving effective at handling “sparse data” and “complex ocean dynamics.” SSTFormer demonstrates how a single architecture, through flexible deployment, can address the unique challenges of multi-scale SST forecasting, highlighting its potential as a unified and robust framework.