Invited Talk
Tianxing: a Linear Complexity Transformer Model with Explicit Attention Decay for Global Weather Forecasting
Shijin YUAN, Tongji University

Biography

Dr. Shijin Yuan is a Professor at the School of Computer Science and Technology, Tongji University. She has been engaged in interdisciplinary research spanning artificial intelligence, atmospheric science, and oceanography, focusing on large weather models, AI-based parameterization schemes, AI data assimilation, AI post-processing for numerical models, and intelligent forecasting and predictability studies of high-impact weather and climate events. Under her leadership, the AI + Atmospheric & Oceanic Sciences Research Team at Tongji University has grown into a highly impactful group, with a series of pioneering and influential papers published in top international journals such as npj CAS and JAMES.

Professor Yuan’s AI-driven predictions have been utilized by IRI and SIPN for ENSO and Arctic sea ice forecasting, providing AI-powered scientific support for national and regional economic and societal development.

She will introduce the TianXing model, a transformer-based data-driven model designed with physical augmentation for skillful and efficient global weather forecasting. TianXing applies a linear complexity mechanism, with an explicit attention decay mechanism in the linear attention derived from physical insights, ensuring proportional scalability with input data size while significantly diminishing GPU resource demands. Meanwhile, a stacked autoregressive forecast strategy is employed to enhance its performance in medium-range forecasting. TianXing’s performance in latitude-weighted RMSE and ACC, particularly in Z500 and T850, surpasses several previous data-driven models and operational models such as NCEP GFS and ECMWF IFS. TianXing has also demonstrated capabilities in predicting extreme weather events, such as typhoons.