Modernizing legacy Earth system models for emerging computing architectures remains a significant challenge. While interactive AI coding assistants have demonstrated the ability to accelerate software development, recent advances in agentic AI offer the potential to automate more complex software engineering tasks, including code translation, testing, validation and workflow orchestration. Realizing this potential requires coordinated multi-agent systems and rigorous methods for evaluating the scientific fidelity and trustworthiness of AI-generated code.
The ATOM team is developing and evaluating an agentic framework to modernize and enhance components of the GEOS atmospheric model. Our work focuses on orchestrating specialized AI agents that translate legacy scientific code to modern, GPU-accelerated implementations, that verify both functional and scientific correctness, and that assess the explainability of AI-generated software and workflows. This effort aims to reduce technical risk associated with large-scale model refactoring while establishing a foundation for AI-enabled Earth system modeling workflows.
In this presentation we will describe the agentic architecture, discuss our evaluation approach, present initial results, and outline the next steps toward scalable AI-assisted modernization of NASA Earth system models.