A Physics-consistent Foundation Model for Learning Earth Surface Dynamics

Accurate prediction of climate-driven land–surface responses is crucial for effective natural resource management, hazard mitigation, and adaptation to growing societal pressures. Existing environmental models, including both process-based approaches and task-specific machine learning methods, often exhibit limited spatial transferability due to sparse observations, structural rigidity, or sensitivity to non-stationary climate conditions. Recently, foundation models have demonstrated emergent capabilities that surpass those of task-specific systems, offering a unified paradigm adaptable to diverse Earth surface processes. However, most existing Earth foundation models (e.g., TerraMind, Prithvi, DOFA, Pangu, and Aurora) primarily scale model size without adequately addressing computational efficiency or embedding the intrinsic physical laws within the large data.

We introduce EarthDynamics, a physics-consistent foundation model for learning Earth surface dynamics that integrates physical priors with computational efficiency. EarthDynamics comprises three interrelated components. First, multi-modal encoding schemes are developed to jointly represent dynamic meteorological forcings, such as precipitation and temperature, and static geophysical attributes, including watershed properties and terrain characteristics. Second, a physics-consistent Transformer architecture is designed to explicitly embed physical constraints, including conservation laws and first-order derivatives, within the pretraining framework, thereby enhancing generalization, improving computational efficiency, and reducing dependence on large training datasets. Third, task-specific head networks enable multi-scale and multi-task inference of key environmental variables, including water levels, streamflow, and landslide occurrence.

Through the integration of these components, EarthDynamics provides a unified and extensible framework for process-informed forecasting across Earth surface systems. The model demonstrates robust performance across a wide range of dynamic tasks, including spatiotemporal simulations of geodynamic processes (e.g., shallow water equations and the Navier–Stokes equations), as well as real-world applications such as flood dynamics, landslide dynamics, rainfall–runoff process, and soil moisture forecasting. EarthDynamics consistently outperforms state-of-the-art supervised learning approaches and fine-tuned vision-based foundation models. EarthDynamics has the potential to serve as foundational infrastructure for water resource management, flood risk assessment, and environmental protection, enabling reliable and scalable predictions under climate change from regional to global scales.