Earthfarseer: Versatile Spatio-Temporal Dynamical Systems Modeling in One Model

Efficiently modeling spatio-temporal (ST) physical processes and observationspresents a challenging problem for the deep learning community. Many recentstudies have concentrated on meticulously reconciling various advantages,leading to designed models that are neither simple nor practical. To addressthis issue, this paper presents a systematic study on existing shortcomingsfaced by off-the-shelf models, including lack of local fidelity, poorprediction performance over long time-steps,low scalability, and inefficiency.To systematically address the aforementioned problems, we propose anEarthFarseer, a concise framework that combines parallel local convolutions andglobal Fourier-based transformer architectures, enabling dynamically capturethe local-global spatial interactions and dependencies. EarthFarseer alsoincorporates a multi-scale fully convolutional and Fourier architectures toefficiently and effectively capture the temporal evolution. Our proposaldemonstrates strong adaptability across various tasks and datasets, with fastconvergence and better local fidelity in long time-steps predictions. Extensiveexperiments and visualizations over eight human society physical and naturalphysical datasets demonstrates the state-of-the-art performance ofEarthFarseer. We release our code athttps://github.com/easylearningscores/EarthFarseer.