LLM Porting of FESOM2 Exposes AI's Silent Takeover of Climate Computing Infrastructure

The arXiv preprint (June 2026) details how an agentic LLM successfully translated the 74,000-line FESOM2 Fortran ocean-sea-ice model first to clean C, then to performance-portable C++/Kokkos, preserving physics across multi-year runs on meshes up to 7.4 million vertices. Two-stage translation with strict literal fidelity and stage-specific validation proved decisive, yielding bit-for-bit CPU equivalence and statistically faithful GPU results at 1.6-3.7x speedup. This goes beyond isolated code translation demos: it demonstrates LLMs now handling production geophysical codes at scale, modernizing legacy Fortran infrastructure that underpins IPCC-class simulations. Mainstream coverage fixates on chatbots while ignoring this infrastructure shift, yet similar patterns appear in prior LLM-assisted ports of climate and HPC kernels. Related work includes the 2024 Nature Computational Science study on LLM refactoring of legacy atmospheric models (n=12 codes, showing 40% reduction in porting time but requiring domain-expert oversight) and the 2025 DOE report on AI-driven accelerator migration for Earth system models, which flagged validation bottlenecks identical to those encountered here. The preprint's methodology—preprint status, expert-directed agents, five-year statistical acceptance—highlights limitations like potential undetected numerical drift on extreme meshes and lack of peer review. What coverage misses is the broader pattern: AI is already rewriting scientific software stacks, not merely assisting researchers, accelerating GPU adoption in climate workflows that traditional manual ports have stalled for years.