CAM-NET: An AI foundation model for whole atmosphere predictions

Numerical weather prediction (NWP) models typically produce deterministic forecasts that depend on initial conditions, without explicitly accounting for the uncertainties inherent in atmospheric evolution. This deterministic nature limits their ability to represent the probabilistic distributions of future states. Furthermore, atmospheric dynamics are governed by highly nonlinear, high-dimensional partial differential equations, whose accurate solutions require immense computational resources and precise initialization. These challenges constrain the accuracy of long-range forecasts and limit the feasibility of real-time applications that demand both speed and reliability.

To address these limitations, we present Compressible Atmospheric Model-Network (CAM-NET), a novel AI foundation model designed to predict both neutral atmospheric and tracer variables from the Earth’s surface to the ionosphere with high fidelity and efficiency. CAM-NET leverages the Spherical Fourier Neural Operator (SFNO) to model global-scale atmospheric dynamics while preserving the Earth's spherical geometry. Trained on over a decade of data from the Whole Atmosphere Community Climate Model with thermosphere and ionosphere extension (WACCM-X), CAM-NET matches WACCM-X in predictive accuracy while achieving a 3–4 order-of-magnitude speedup. Spectral analysis reveals CAM-NET's limitations in resolving small-scale structures; we discuss potential solutions to improve spectral fidelity and enhance fine-scale atmospheric representation.