Yue Sun: Scaling AI for Earth Science: Deployment, Extension and Evaluation of Zephyrus on Gadi
Recent advances in tool-using language agents have created new opportunities for automating research workflows. Because many Earth system modelling (ESM) tasks already rely on structured datasets, scripts, and simulation pipelines, the community is well positioned to explore agentic approaches for integrating dataset analysis, forecasting, and modelling tasks.
Recently released at ICLR 2026, Zephyrus is the first agentic research framework for weather and climate science. It enables language models to orchestrate Earth science tools through iterative tool use and reasoning. The current system integrates five core capabilities: dataset indexing, geolocation, climate simulation, weather forecasting, and climatology analysis.
We deployed Zephyrus on Gadi and evaluated it on multi-step ERA5 tasks using both open-source and proprietary language models. The agent can autonomously chain tools, for example, retrieving ERA5 data, locating a point of interest, running a forecast, and comparing outputs, without manual scripting or workflow engineering.
Results show that Zephyrus agents substantially outperform text-only baselines, improving performance accuracy metrics by up to 37% on representative ERA5 workflows. The framework dynamically composes task-specific workflows, reducing tool-integration overhead and accelerating iterative experimentation in computational Earth science
This talk will interest practitioners exploring AI-assisted workflow automation for reducing repetitive data-processing and tool-integration tasks while enabling more time for scientific analysis and hypothesis development.
We will demo Zephyrus live, share early lessons from running agentic research workflows on HPC systems, and discuss the path toward a community-curated Earth science tool library.
Keywords: AI for Science, Agentic Workflows, Tool-Using Language Models, High Performance Computing, Autonomous Research Systems