LAFT — LLM-Assisted Fortran Translation

Code and data for the LAFT paper 1: translating the Kessler warm-rain microphysics scheme from Fortran to JAX with four large language models, and validating every translation against the Fortran reference with a model-agnostic pipeline.

The repository has two parts:

Folder	What it is	Start with
LAFT/	The framework: Fortran parsing, bridge generation, multi-pass prompt authoring, lint, semantic audit, runtime validation, driver comparison, and GPU profiling. Project-agnostic; every project-specific value comes from a config/project.toml.	LAFT/docs/README.md for the overview, LAFT/ORCHESTRATOR.md to run it
kessler/	The paper-1 project: the Kessler Fortran reference and data, the pipeline snapshot, the bridge tests, the four LLM translations with their full validation records, and the cross-model comparison outputs.	kessler/README.md

The four translations

Archive	Model	Fortran comparison
kessler/translations/claude-sonnet46-jd/	Claude Sonnet 4.6	all pass, machine precision
kessler/translations/gemini31pro-jd/	Gemini 3.1 Pro	all pass, machine precision
kessler/translations/gpt54thinking-jd/	GPT-5.4 Thinking	all pass, machine precision
kessler/translations/qwen25-32b-jd/	Qwen2.5-32B (Coder-Instruct, via vLLM)	all pass, machine precision

Each archive holds the JAX code, the prompts, and every gate's output (lint, semantic audit, runtime validation, bridge test suite, driver run, and the comparison against the Fortran outputs). See kessler/README.md for how the archives were produced and how to read their provenance notes.

Setup

Two conda environments, built from the pinned recipes in LAFT/envs/:

conda env create -f LAFT/envs/ts.yml            # tree-sitter: Stage 0 parsing only
conda env create -f LAFT/envs/jax-validate.yml  # JAX 0.6.2: everything else

jax-validate installs the CUDA 12 build of JAX for NVIDIA GPUs. For a CPU-only machine change jax[cuda12]==0.6.2 to jax==0.6.2 in the recipe; the Fortran comparisons still pass, only the GPU timings cannot be reproduced. Details, including the optional vLLM environment for locally hosted models, are in LAFT/docs/README.md §Environments.

Running the pipeline

The pipeline is driven by an AI coding agent, not by a human at a shell. ORCHESTRATOR.md and the per-stage workflow documents are written as instructions to an agent such as Claude Code, the Gemini CLI, or OpenAI Codex: the agent reads the orchestrator, runs the phase scripts and submits the PBS jobs in order, authors or repairs the JAX translation from the generated prompts, reads each gate's output, and stops for human approval between stages. A person supervises and approves; the agent executes.

cd kessler          # the project root: config/project.toml lives here
claude              # or your agent of choice, started in this directory
# then ask it: "Follow ORCHESTRATOR.md and run the pipeline for this project."

The individual phase scripts are ordinary Python and can also be run by hand in the environments above (Stage 0 in ts, everything after in jax-validate), but the workflow documents assume an agent is the reader.

The batch jobs in LAFT/pbsJobs/ are written for NCAR's Derecho and Casper systems (PBS Pro, NVIDIA A100). LAFT/pbsJobs/README.md lists what to change on another cluster.

Reproducing the paper's numbers

• Accuracy (kessler/_compare_results/outputs/reports/performance_comparison_2026-04-08.md, plots/numerical_accuracy.png): rerun any archive's driver and comparison through the pipeline; the mean absolute error in theta is about 7e-16 K for all four models.
• The Qwen translation itself (kessler/translations/qwen25-32b-jd/): LAFT/pbsJobs/pbs_qwen25_translate.sh re-runs Qwen2.5-Coder-32B-Instruct through vLLM with the paper's settings on the same prompts (see LAFT/workflow_translator/qwen/qwen.md §Paper-1 job). Sampling at temperature 0.1 is not bit-reproducible, so expect the configuration, not the archived bytes. The Claude, GPT, and Gemini translations were authored in-context by their coding agents and have no batch job.
• GPU scalability and profiling (kessler/_compare_results/outputs/plots/, reports/*_nsys_report_*.md): produced on A100 GPUs; the reference data and the JAX code are in the repository, the timings depend on the hardware. The paper's timing data behind the scalability figure (serial Fortran on one CPU core and the four translations on one GPU) is archived under kessler/_compare_results/outputs/scalability/. The evaluation scripts that produced the cross-model figures and profiling reports are not part of LAFT and are not included; the serial Fortran benchmark source is in kessler/data/exp2_jd/src/.
• The OpenACC GPU reference (kessler/data/exp2_jd/acc_src/): the manually accelerated Fortran used in the paper's scalability comparison, with its driver, Makefile, and profiling runner, so its timings can be reproduced and compared with any JAX translation. See its README.md.

Citation

If you use this code or data, please cite the paper and the repository. The machine-readable metadata is in CITATION.cff (GitHub's "Cite this repository" button reads it).

Linck, I., Dennis, J., Schreck, J., Stengel, K., & Gagne, D. J. (2026). Evaluating LLMs for Translating Legacy Fortran Atmospheric Physics to GPU-Accelerated JAX. Preprint submitted to arXiv; manuscript submitted to the Journal of Advances in Modeling Earth Systems (JAMES). Code and data: https://doi.org/10.5281/zenodo.22696825

@article{linck2026laft,
  title   = {Evaluating {LLMs} for Translating Legacy {Fortran} Atmospheric
             Physics to {GPU}-Accelerated {JAX}},
  author  = {Linck, Iris and Dennis, John and Schreck, John and
             Stengel, Karen and Gagne, David J.},
  year    = {2026},
  note    = {Preprint submitted to arXiv; manuscript submitted to JAMES},
  url     = {https://github.com/NCAR/LAFT},
  doi     = {10.5281/zenodo.22696825}
}

The Zenodo DOI above is the concept DOI and always resolves to the latest release. The arXiv identifier and journal DOI will be added here and in CITATION.cff once assigned.

License

Apache License 2.0, see LICENSE.