The 2025 hurricane season has marked the beginning of a “new era” of meteorology as official agencies now use artificial intelligence (AI) models to pinpoint storm formation, track and intensity, providing the insurance industry with more comprehensive and accurate near-term weather forecasts, according to Gallagher Re.
While experts caution that one season does not guarantee long-term success, the technology has shown immense promise, suggesting AI will only become more essential to global weather forecasting, the reinsurer’s 2025 Natural Catastrophe and Climate Report stated.
The US National Hurricane Centre (NHC) is a notable user of data-driven AI models, which have augmented confidence in longer-range Atlantic hurricane track forecasts by multiple days.
However, the firm noted, these models were less skilled at predicting other storm factors, like intensity and precipitation.
Google’s DeepMind tropical cyclone (TC) model, the most visible AI model used in 2025, debuted in June via the Google Weather Lab, producing ensemble forecasts up to 15 days out.
During the 2025 season, the DeepMind TC model consistently outperformed conventional numerical weather predictions (NWP) models, like the US GFS, on Atlantic track forecasts.
Its performance was mixed in other basins, especially the Western Pacific, where its track skill lagged behind physics-based systems like the European ECMWF. Intensity forecasts, particularly regarding rapid intensification (RI) cycles, also showed less skill, analysts noted.
“While one season is not enough to make definitive declarations of success or failure for any model, there was enough promise in AI performance to recognize that such technology will only become more ingrained in weather forecasting,” Gallagher Re analysts stated.
Adding: “The primary advantages of AI models are their quick run times, minimally required computing power, and stability across successive runs. However, the quality of AI output is entirely dependent on the depth and quality of the historical and reanalysis data on which they are trained.”
Models require regular retraining and calibration to maintain essential baseline accuracy, particularly in the face of moderate to rapid atmospheric or oceanic shifts, including those caused by a changing climate.
Relying on even seemingly robust datasets carries risk, as many are still limited, especially when applied across an entire basin. This limitation, according to the report, can result in gaps in output, notably an underrepresentation of rare (tail) events.
Despite the attention AI prediction models are garnering, “scientists and other public and private sector users will continue to heavily rely on conventional NWP and tropical cyclone models,” Gallagher Re analysts stated.
The integration of AI models is expected to further educate catastrophe modelers and risk managers, enhancing their ability to grasp the range of potential outcomes for an event and ultimately improving the timeliness and accuracy of real-time communication.
Analysts concluded: “While it is not expected that AI models will replace physics‑based models completely, we continue to see record investment into this area.
“This includes the Artificial Intelligence / Integrated Forecasting System (AIFS) run by the ECMWF and NOAA’s Project EAGLE. Private companies such as Microsoft, IBM, and NVIDIA have also invested heavily in the AI weather forecasting space.”
Adding: “Beyond forecasting, the rapid growth of AI and its data centre footprint raises broader energy/electricity concerns. How the world invests in green energy to meet this demand will be vital in determining whether greenhouse gas emissions can be reduced.”
