In a significant leap for computational meteorology, a research team at the University of California San Diego has developed Zephyrus, the first “agentic” AI framework designed to bridge the gap between complex atmospheric data and natural language. Unlike traditional forecasting models that require expert-level programming to navigate, Zephyrus allows users to query weather datasets using plain English. The system translates these queries into executable code, processes vast amounts of spatiotemporal data from models like GraphCast and FourCastNet, and returns findings in accessible prose. Presented at the 14th International Conference on Learning Representations (ICLR) in Rio de Janeiro, the project aims to democratize access to Earth science data, enabling students and non-specialists to reason about weather patterns and climate simulations at unprecedented speeds.
The Challenge of Modern Meteorological Data
As of 2026, the landscape of weather forecasting has been radically transformed by artificial intelligence. Leading models such as Google DeepMind’s GraphCast and Huawei’s Pangu-Weather now routinely outperform traditional numerical weather prediction (NWP) systems, offering 10-day forecasts in under a minute—a task that previously required hours of supercomputer runtime. However, a persistent bottleneck remained: these “black box” models output massive, high-dimensional arrays of data that are difficult for humans to interpret without specialized coding skills in languages like Python or Julia.
The UC San Diego research team, led by computer scientist Rose Yu and atmospheric physicist Duncan Watson-Parris, identified a dual-pronged problem. First, existing AI weather models cannot explain their internal logic or results in natural language. Second, they lack the ability to integrate and reason about textual information, such as historical weather bulletins or real-time meteorology reports.
“Our goal is to increase access to critical data and predictions by lowering the barrier to entry to analyzing these data,” said Watson-Parris, a faculty member at the UC San Diego Scripps Institution of Oceanography. Standing in a lab filled with monitors displaying swirling aerosol patterns, his demeanor was one of focused urgency. “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.”
Architecture of the Zephyrus Framework
To solve the communication gap, the researchers created an environment known as ZephyrusWorld. This “agentic” setup acts as a middleman between a Large Language Model (LLM) and the raw meteorological data. When a user asks a question—for example, “Where in the North Atlantic will wind speeds exceed 50 knots over the next 48 hours?”—Zephyrus does not simply guess. It follows a structured operational loop:
- Natural Language Understanding: The agent parses the English query to identify specific parameters (location, variable, timeframe).
- Code Generation: Using an interface to the WeatherBench 2 dataset, the agent writes Python code to filter and analyze the relevant atmospheric variables.
- Execution and Observation: The code runs in the ZephyrusWorld environment, interacting with forecasting models like Stormer or physics-based climate simulators.
- Refinement: If the code returns an error or if the data appears scientifically implausible, the agent—specifically the “Zephyrus-Reflective” variant—iterates on its own logic to correct the mistake.
During testing, the researchers evaluated the agent using ZephyrusBench, a new benchmark comprising over 2,200 question-answer pairs spanning 49 distinct meteorological tasks. The tasks ranged from simple data lookups to complex “counterfactual reasoning,” where the AI must predict how a storm’s path might change if sea surface temperatures were one degree higher.
Performance and Limitations in Extreme Scenarios
The results presented at ICLR 2026 indicate that Zephyrus is remarkably adept at foundational tasks. It achieved high accuracy in identifying regional weather conditions and generating localized forecasts. However, the study also highlighted the current ceiling of agentic AI in the field. Zephyrus continues to struggle with “extreme event detection”—the rare, high-impact anomalies like 1-in-100-year floods or rapid hurricane intensification—where data is sparse and the physics are most volatile.
“Weather prediction is a critical scientific challenge with profound implications spanning agriculture, disaster preparedness, and energy management,” the researchers noted in their conference paper.
The team tested four different frontier LLMs to power the Zephyrus brain, finding that while all performed with similar general accuracy, they all shared a tendency to struggle with generating long-form, multi-page technical reports. This suggests that while the AI can “find” the data, the synthesis of that data into expert-level scientific narrative remains a work in progress.
Democratizing the Future of Earth Science
The implications of Zephyrus extend far beyond the walls of the University of California. By creating a system that understands both the rigors of atmospheric physics and the nuances of human language, the researchers hope to empower a new generation of “AI co-scientists.”
“Our vision is to democratize Earth science,” said Rose Yu, an associate professor in the UC San Diego Department of Computer Science and Engineering and a 2025 Samsung AI Researcher of the Year. Her tone during the project’s unveiling was one of collaborative optimism. “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.”
Looking ahead, the team plans to scale the project by fine-tuning open-source models specifically on climate-focused datasets. This would move the technology away from expensive, proprietary LLMs and toward a community-driven tool that can be used by researchers in the Global South and other regions disproportionately affected by climate change but often lacking in supercomputing resources.
The research was supported by a broad coalition of federal agencies, including the U.S. Army Research Office, the U.S. Department of Energy, and DARPA, signaling the high strategic value placed on the intersection of generative AI and climate resilience.
