The storm cell is now being asked to announce its own life cycle

StormScope (arXiv, January 2026, NVIDIA) is a family of transformer-based generative diffusion models trained on geostationary satellite imagery and ground radar over the continental United States. It forecasts at 10 minute and 6 km resolution, which the abstract calls competitive with state-of-the-art mesoscale NWP models for lead times up to 6 hours, with ensembles for uncertainty. It says it leads on deterministic and probabilistic metrics from 1 to 6 hours against extrapolation and operational mesoscale NWP.

OneForecast (ICML 2025) is a global-regional nested weather framework built on graph neural networks. It uses a multi-scale graph, densifies the target region to capture local high-frequency features, adds adaptive message passing for extreme events, and a neural nested grid to limit boundary information loss.

The IET paper models thunderstorm gale forecasting as a competitive multi-agent problem. A dynamic spatio-temporal graph model, DSTGFP, builds an adaptive adjacency matrix among station agents for wind speed forecasts, a second network handles radar echo extrapolation, and Kriging with fuzzy logic produces a gale potential index. The overall framework reports a CSI of 0.5641, a POD of 0.6471 and a FAR of 0.1852.