Systematic LLM Translation of Legacy Scientific Code to Differentiable Frameworks: Application to a Land Surface Model
Researchers have developed a novel five-phase pipeline utilizing LLM-based agents to translate legacy Fortran code into the JAX framework. This system automates the process of code migration, including dependency analysis, autonomous error correction, and numerical parity enforcement. The pipeline was successfully applied to CLM-ml-v2, a 19,000-line land surface model, resulting in a differentiable version that significantly speeds up computation and parameter recovery. AI
IMPACT Enables rapid differentiation of complex scientific models, accelerating research and parameter estimation.