AI Weather Agent Zephyrus Turns Natural Language Questions Into Instant Forecast Analysis

UC San Diego team introduces the Zephyrus AI weather agent

The research team from the University of California, San Diego will present the first AI weather agent they developed, named Zephyrus, at the 14th International Conference on Learning Representations (ICLR), April 23–27 in Rio de Janeiro.

AI weather models are improving forecasts but remain difficult to interpret

Recently, models driven by AI and deep learning have considerably improved weather forecasting. But analyzing the resulting data remains difficult and time-consuming. A main issue is that these types of AI models cannot describe their findings in plain language. A secondary issue is that these models cannot reason about textual information, such as meteorological reports and weather bulletins. The UC San Diego research team aims to address both.

"Our goal is to increase access to critical data and predictions by lowering the barrier to entry to analyzing these data," said Duncan Watson-Parris, a study co-author and faculty member at the UC San Diego Scripps Institution of Oceanography. "We want to increase the speed with which we can reason about multimodal data and learn about the Earth by making it easier for students and young scientists to interact with different datasets."

System converts natural language questions into code-based weather analysis

The researchers also hope the findings will lead to AI agents that can bring similar advances to other disciplines, especially climate science. Meteorology was a perfect test case because it combines large, complex datasets that change over time and the need to reason about these data in plain language. "Weather prediction is a critical scientific challenge, with profound implications spanning agriculture, disaster preparedness, transportation, and energy management," the researchers write.

To bridge the gap between a code-driven AI model and a language-based AI agent, the researchers set up an environment that allows the agents to interact with weather models and data via code. The AI agent can handle language-based queries, translate them into code, and then convert the code-generated answers into plain language.

Zephyrus performs well on basic tasks but struggles with extreme weather analysis

Zephyrus performed well on simple tasks, such as finding locations with specific weather conditions and generating weather forecasts for specific locations at specific times. But it struggles to find locations with extreme weather and to generate reports. Researchers tested four frontier LLMs to power Zephyrus, and all performed with similar accuracy.

Future development will focus on larger datasets and climate-specific training

For the next iteration of the AI agent, researchers plan to use larger training datasets. Next steps also include fine-tuning open-source models for climate-focused tasks.

"Our vision is to democratize earth science. Zephyrus is a crucial step toward creating AI co-scientists that dramatically lower the barrier to entry, allowing students and researchers everywhere to access and reason about critical weather and climate data at unprecedented speeds," said Rose Yu, study co-author and a faculty member in the UC San Diego Department of Computer Science and Engineering.

Funding and publication details

This work was supported in part by the U.S. Army Research Office under Army-ECASE award W911NF-07-R-0003-03, the U.S. Department of Energy, Office of Science, IARPA HAYSTAC Program, NSF Grants #2205093, #2146343, and #2134274, CDC-RFA-FT-23-0069, as well as DARPA AIE FoundSci and DARPA YFA.

Zephyrus: An Agentic Framework for Weather Science, ICLR 2026

Marshall Fisher, Jas Thakker, Yiwei Chen, Zhirui Xia, Yasaman Jafari, Ruijia Niu, Manas Jain, Veeramakali Vignesh Manivannan, Zachary Novack, Luyu Han, Srikar Eranky, Salva Rühling Cachay, Taylor Berg-Kirkpatrick, and Rose Yu, Department of Computer Science and Engineering, UC San Diego Jacobs School of Engineering

Duncan Watson-Parris, UC San Diego Scripps Institution of Oceanography and Halicioglu Data Science Institute, within the School of Computing, Information and Data Sciences at UC San Diego

Sumanth Varambally and Yi-An Ma, Halicioglu Data Science Institute within the School of Computing, Information and Data Sciences at UC San Diego.

*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as definitive, used to guide development decisions, or treated as established information in the field of artificial intelligence research.