LLM-Assisted Workflow Orchestration for Decadal Prediction Analysis

Large language models (LLMs) offer new opportunities to make climate data analysis and prediction workflows more accessible by enabling interactive, natural language–driven interactions. Recent studies have shown that LLM-based assistants can support exploratory analysis and improve reproducibility, but operational climate prediction—particularly on seasonal to decadal time scales—often involves more complex workflows. These include standardized evaluation procedures, model–observation comparisons, calibration steps, and custom post-processing, which typically require deeper technical expertise and familiarity with specialized tools and high-performance computing (HPC) environments.

In this work, we present an LLM-assisted interface designed to support decadal climate prediction analysis by orchestrating existing evaluation and post-processing tools through natural language prompts. The system allows users to initiate multi-step workflows on HPC systems, automatically generating configuration files, handling lead-time–dependent data selection, comparing predictions against observational references, and applying calibration methods. By integrating retrieval-augmented generation (RAG), the LLM is also informed by the underlying analysis code bases, enabling scientists to flexibly define, adapt and extend workflows by composing existing functions and generating lightweight custom routines.

Our results demonstrate how LLM-driven orchestration can act as a co-pilot for complex climate prediction workflows, lowering technical barriers while preserving scientific rigor. This approach supports faster iteration, greater transparency, and improved accessibility for researchers working across seasonal to decadal prediction challenges. We discuss opportunities, implications and challenges for future climate services that arise with this new way of creating and managing complex climate-scentific workflows. Likewise, we argue that natural language interfaces have the potential to reshape how scientists interact with prediction data, models, and computational infrastructure—aligning closely with the goals of current climate prediction research and applications.