1539. AI-Driven Weather Forecasting: Greater Accuracy and Emerging Challenges

Advances in artificial intelligence (AI) are rapidly transforming weather forecasting. In a recent episode of Science Sessions, the podcast of the Proceedings of the National Academy of Sciences (PNAS), atmospheric scientists, oceanographers, and environmental researchers discussed the latest developments in AI-enabled forecasting, from short-term weather prediction to long-term climate projections. By learning complex patterns from vast amounts of observational data, AI is emerging as a powerful complement to traditional physics-based forecasting models. At the same time, researchers emphasize that the performance of AI models remains highly dependent on the quality of observations and historical data used for training.

Recent studies suggest that improvements in forecasting accuracy can generate substantial societal benefits. Research led by Columbia University found that day-ahead weather forecasts improved by approximately 34% between 2005 and 2023, contributing to reductions in heat-related mortality. Expert assessments indicate that forecast skill is likely to continue improving throughout this century, enhancing societies’ ability to cope with increasing heat risks under climate change.

The application of AI extends beyond short-term weather forecasting. Researchers at Google have developed generative AI approaches that convert coarse-resolution outputs from global climate models into high-resolution regional climate projections. Their approach achieves accuracy comparable to traditional physics-based methods while reducing computational costs by about 90%, potentially enabling more detailed assessments of risks such as wildfires, floods, and other climate-related hazards.

AI is also showing promise in forecasting extreme weather events. A research team from the Chinese Academy of Sciences developed a machine-learning model capable of predicting rapid intensification of tropical cyclones, achieving approximately 92% detection accuracy while keeping false alarm rates below 10%. Similarly, researchers at the Hong Kong University of Science and Technology reported improvements in thunderstorm nowcasting, using a deep diffusion model to predict storm development and dissipation up to four hours in advance.

Despite these advances, important limitations remain. Researchers at the University of Chicago examined the ability of AI models to predict so-called “gray swan” events—physically plausible but previously unseen extreme events. Their findings suggest that when similar examples are absent from training datasets, AI models tend to underestimate the intensity of severe tropical cyclones. While AI excels at recognizing known patterns, its ability to extrapolate beyond past experience remains limited.

Taken together, these studies highlight both the opportunities and challenges associated with AI-enabled forecasting. AI has the potential to significantly improve weather and climate prediction, but its success depends on continued investment in high-quality observations, robust data infrastructure, human expertise, and physics-based understanding. As climate change increases the frequency and severity of heatwaves and extreme weather events, AI is poised to become an increasingly important component of forecasting systems, provided that its limitations are properly understood and managed.

Reference:

 PNAS Science Sessions Podcast, Using AI to Predict the Weather (2026)

 https://www.pnas.org/about/science-sessions-podcast

Contributor: Miyuki IIYAMA, Strategic Coordination Office