Understanding where landslides are likely to occur is an important step in reducing disaster losses, especially where unstable terrain overlaps with residential development. A study published in Scientific Reports on 4 October 2026 applies machine learning to landslide susceptibility mapping in eThekwini, South Africa. By combining environmental data with an assessment of housing exposure, it provides spatial evidence that could inform local disaster risk planning.
Tropical cyclones, sometimes referred to as hurricanes or typhoons, are powerful rotating storms that form over warm ocean waters and can cause widespread devastation when they make landfall. In the past 50 years, they have been responsible for more than 700,000 deaths and US$1.4 trillion in economic losses globally4. Accurate and timely forecasts are crucial for issuing effective warnings, enabling evacuations and preparing emergency responses. Three principal factors influence the downstream effects of a tropical cyclone: its track (where a cyclone forms and where it goes), its intensity (its maximum wind speed) and its structure (its size and shape). For decades, public forecasting agencies such as the US National Hurricane Center (NHC) have provided forecasts of these parameters and associated tropical cyclone hazards several days into the future, to forewarn about incoming impacts. Forecasting these complex systems is notoriously difficult. The track of a cyclone is governed mainly by large-scale atmospheric steering currents and cyclone–environment interactions, and its intensity is thought to also be strongly dictated by fine-scale thermodynamic processes within and around its core5,6. This has led to a long-standing scale dilemma in operational forecasting, because numerical weather prediction (NWP) models must simulate the atmosphere at high spatial resolutions to capture intensity changes, but this is too expensive to do over a very large domain7,8. Global NWP models, such as the European Centre for Medium-Range Weather Forecasts (ECMWF)’s ensemble system (ENS)9, excel at capturing this large-scale atmospheric flow and thus generally produce the most skilful track forecasts, but lack the resolution to accurately forecast intensity. Conversely, specialized and deterministic high-resolution regional models, such as the US National Oceanic and Atmospheric Administration (NOAA)’s Hurricane Analysis and Forecast System (HAFS)10, better capture intensity, but at the expense of track accuracy. In recent years, AI has emerged as a powerful new paradigm for weather prediction, and models have achieved superior forecast accuracies and speeds, compared with NWP systems, for core atmospheric state variables11,12,13,14. These models show state-of-the-art track prediction for tropical cyclones14, but inherit the low-intensity bias of 0.25° global analysis owing to low resolution (around 28 km). Poor intensity skill has been noted as a major limitation of AI weather models15. Several works have combined AI with NWP16,17, and show improved track forecasts compared with NWP models and improved intensities compared with AI models, whereas other approaches have explored end-to-end approaches based on observational data18,19,20. However, these approaches have not achieved state-of-the-art results in operational tropical cyclone forecasting.
Here we introduce WeatherNext Cyclones (WN-C), an AI system that excels at both track and intensity forecasting. WN-C is a probabilistic system that generates an ensemble of 50 (or more) forecasts of possible scenarios to quantify uncertainty. It is jointly trained on a combination of two data sources (a process we refer to as co-training): decades of global atmospheric analysis data from the ECMWF1, capturing global weather dynamics, and cyclone data from the International Best Track Archive for Climate Stewardship (IBTrACS)2,3, a specialized, expert-curated database of nearly 5,000 observed tropical cyclones, which provides detailed information on their track, intensity and wind radii.
Sources: Nature
Comment list ( 0 )
