This paper proposes TelePiT, a novel deep learning architecture, to address the challenges of intraseasonal to seasonal (S2S) forecasting, which involves forecasting climate conditions weeks to months in advance. TelePiT accurately encodes global atmospheric variables into spherical geometry via spherical harmonic function embeddings, explicitly captures atmospheric physical processes across a variety of learnable frequency bands via multiscale physics-informed neural ODEs, and explicitly models teleconnection patterns via a teleconnection-aware transformer to model critical global climate interactions. Experimental results show that TelePiT outperforms state-of-the-art data-driven baseline and operational numerical weather forecasting systems across all forecast horizons.
