SWIFT: Spatiotemporal Weather Integration and Forecasting Transformer

SWIFT (Spatiotemporal Weather Integration and Forecasting Transformer) is an experimental weather forecasting architecture that investigates whether atmospheric states can be compressed into structured spatial latent representations and forecast efficiently within that compressed space. A convolutional encoder learns and compresses local spatial relationships, while a temporal transformer forecasts the evolution of the latent atmospheric state through time. The initial Stage 1 experiment uses Bureau of Meteorology ACCESS-SY data as a coherent atmospheric source state to evaluate whether this approach can preserve forecast-relevant information while reducing the computational burden of high-resolution autoregressive forecasting. Later stages are intended to investigate replacing the ACCESS-derived state with representations constructed directly from heterogeneous atmospheric observations.