COGENT: Continuous Graph Emulators with Neural Ordinary Differential Equations for Long-Term Physical Forecasting

A new model named COGENT leverages continuous graph emulators and Neural Ordinary Differential Equations (NODEs) to enhance long-term physical forecasting on irregular geospatial meshes. This novel framework processes a finite historical record of system states, along with associated forcing fields and external influences, through a specialized graph-based history encoder. This mechanism generates node-specific context vectors, designed to effectively capture both local spatial interactions and complex temporal dependencies within the data. By encoding a comprehensive history of system states and external forcings, COGENT aims to provide a more nuanced understanding of system dynamics and their evolution. The integration of continuous dynamics via NODEs allows for modeling continuously evolving physical systems, mitigating the inherent discretization errors found in traditional discrete-time approaches. This design offers enhanced stability and accuracy over extended prediction horizons ¹. Such a robust framework holds significant potential for improving predictive fidelity across diverse applications involving dynamic and complex physical environments, from detailed climate projections to sophisticated infrastructure planning, where long-term accuracy is paramount for strategic decision-making.