Efficient physics-interpretable deep learning for skillful tropical stratosphere dynamics modeling

Tropical stratospheric dynamics, dominated by the quasi-biennial oscillation (QBO), exert profound influences on global atmospheric circulation, climate variability, and the occurrence of extreme events. Despite decades of development, current dynamical models still struggle with interannual prediction and climate simulation of the QBO, exhibiting pronounced deficiencies in representing its amplitude and downward propagation in the lower stratosphere. Here, we propose SETA-DL, a parameter-efficient deep learning algorithm designed to model the temporal evolution of tropical stratospheric dynamics, which captures QBO disruption events with enhanced precision relative to state-of-the-art operational systems. Interpretability analyses reveal that SETA-DL encodes tropical wind variability into two anti-phased patterns, consistent with the modulation associated with equatorial Kelvin waves and mixed Rossby-gravity wave activity. Guided by these learned dynamical features, SETA-DL reconstructs historical simulations and future climate projections with improved physical consistency, reducing inter-model spread while recovering previously underrepresented dynamical wave modes. Our artificial intelligence method generates projections spanning 85 years within 3 node-seconds, highlighting the potential of data-driven approaches to enhance the fidelity of stratospheric simulations at substantially lower computational cost and enable more reliable assessments of future QBO behavior under global warming.