A Data-Driven Tropical Cyclone Model Based on Lagrangian Limited-Area Transformers: Investigating Physical Fidelity through Real-Case Simulations and Idealized Experiments

Data-driven weather prediction (DWP) offers advantages in computational efficiency and demonstrates superior synoptic-scale performance compared with numerical weather prediction. However, global DWP models typically operate at coarser resolutions and remain limited in their ability to accurately predict tropical cyclone (TC) wind structure and rainfall. Therefore, the Lagrangian Limited-Area Transformer for typhoon (LLAT.ty) is developed for TC-following DWP; it is specifically designed to learn TC evolution with 3-hourly intervals and is trained using TC-centered ERA5 reanalysis. Notably, LLAT.ty explicitly predicts the TC center location for the next time step. Therefore, LLAT.ty can generate three-dimensional forecasts on continuous time steps, provided that boundary conditions from the global DWP are used via one-way boundary updating or two-way domain interactions. LLAT.ty also provides additional predicted variables essential for TC forecasting, including rainfall and vertical motion. This study subsequently uses LLAT.ty as a scientific testbed to examine whether a purely data-driven prediction system reproduces the established dynamical behavior of TCs. Simulations of all 2024 typhoons show that LLAT.ty improves TC track predictions by approximately 30% relative to the baseline FourCastNETv2 global DWP. Moreover, idealized experiments demonstrate that LLAT.ty produces realistic TC responses to variations in sea surface temperature, latitude-dependent Earth’s rotation, and vertical wind shear governing the convective asymmetry. Also, LLAT.ty successfully simulates the evolution of beta gyres, which exert a physically reasonable influence on TC motion. Overall, this study suggests that DWP models can better represent TC-specific dynamics by adjusting the learning framework to the scales and weather systems of interest.