AI helps climate scientists answer urgent climate questions

Researchers are using large language models and other AI methods for coding, data analysis and literature synthesis in climate work; teams report faster reviews and new tools such as Google's Groundsource for flash-flood prediction, while noting AI does not replace physics-based models and cannot see beyond its training data.

Researchers are integrating large language models and other AI techniques into climate science workflows. Scientists use these tools for coding, visualization, data processing and literature synthesis. Teams report faster reviews and new localized tools for risk estimation. Experts say AI methods are likely to complement, not replace, traditional physics-based climate models. Key developments: - Climate scientist Zeke Hausfather and others have used LLMs for brainstorming, coding and creating visualizations alongside conventional work. - A team led by Google DeepMind used the Gemini model to help synthesize 79 papers on the Atlantic Meridional Overturning Circulation; the AI contributed about 42% of the final text and the team revised the work 104 times over 46 person-hours, roughly ten times faster than usual. - Google Research released Groundsource, a flash-flood prediction tool developed using Gemini to identify millions of news reports; the tool is publicly available, yields valid results about 82% of the time in the team’s tests, and has not yet been peer-reviewed. - Researchers are exploring hybrid approaches that combine AI with physics-based models to downscale global simulations to local risk estimates and to improve representations of clouds and other small-scale processes. - Studies show AI can outperform some traditional methods in specific regional problems and weather phenomena, but scientists emphasize the need to understand how AI reaches its results because AI cannot see beyond its training data. Summary: AI-based methods are accelerating parts of climate research and enabling more localized and efficient analyses. Researchers emphasize that these tools complement physics-based models and require careful oversight. Ongoing benchmarking, peer review and efforts to understand model behavior remain priorities.