Systematic LLM Translation of Legacy Scientific Code to Differentiable Frameworks: Application to a Land Surface Model

Researchers developed an LLM-based pipeline that automatically translates legacy Fortran scientific code into JAX, a differentiable programming framework. Applied to a 19,000-line land surface model, the approach achieved 24x speedup and 8x faster parameter optimization while enabling gradient-based analysis through automatic differentiation.

This work addresses a critical bottleneck in computational science: migrating decades-old scientific codebases to modern differentiable frameworks without manual rewriting. The five-phase pipeline uses static analysis, iterative repair loops, and numerical validation to systematically translate Fortran into JAX while preserving correctness. The achievement matters because Earth system models underpin climate science, and differentiable versions unlock powerful optimization and uncertainty quantification techniques previously unavailable.

The research emerges as the AI and scientific computing communities increasingly recognize that gradient-based methods accelerate parameter estimation and sensitivity analysis. Legacy Fortran dominates climate, weather, and geophysics modeling—translating these systems has been prohibitively expensive. The LLM-agentic approach reduces manual effort by automating dependency resolution, error correction, and validation, creating a replicable template for other domains relying on legacy code.

For the scientific computing industry, this represents a pathway to modernization without abandoning institutional knowledge embedded in existing models. The 24x speedup at N=2,048 ensemble size and 8x reduction in optimization steps demonstrate tangible productivity gains. The 8x parameter recovery speedup is particularly significant for inverse problems in geophysics and climate modeling, where computational cost historically constrains uncertainty quantification.

Looking forward, the framework's reusability could accelerate differentiable implementations across Earth system models, weather prediction systems, and other physics-based simulations. Success here may drive similar pipelines for other legacy scientific languages and establish LLM-guided code translation as a standard practice. Broader adoption could reshape how institutions manage technical debt in computational science.

• →LLM-based pipeline successfully translates 19,000-line Fortran climate model to JAX with numerical correctness verification at module level.
• →Differentiable version achieves 24x wall-clock speedup and 8x faster parameter optimization compared to legacy Fortran.
• →Five-phase approach (dependency analysis, iterative repair, validation, integration, gradient verification) creates reusable framework for other scientific codebases.
• →Enables gradient-based sensitivity analysis and full Jacobian computation in single backward pass, previously unavailable in legacy models.
• →Released as open-source infrastructure, reducing barriers for modernizing other Earth system model components.