New Preprint: Noah-Agent, a Multi-Expert AI Agent Framework for Fortran Climate Models
Published:
Parameterizing a large Fortran climate model by hand is slow, error-prone, and hard to validate. Noah-Agent asks whether a team of specialized AI agents can do it instead.
Author: Koutian Wu; GitHub: ktwu01
I have released Noah-Agent (v0.1) as a preprint on Zenodo. It is in preparation, so this is an early version rather than a final paper.
The Problem
Large-scale Fortran climate models are powerful but hard to work with. Parameterizing them and validating their behavior usually takes careful manual effort across a large, old codebase. That work is slow and easy to get wrong.
The Approach
Noah-Agent is a multi-expert AI agent framework for automated parameterization and validation of large-scale Fortran climate models. Instead of one model doing everything, several specialized agents coordinate:
- Some agents focus on selecting and adjusting parameters.
- Others focus on validating model behavior against reference outputs.
The goal is to automate steps that are otherwise done by hand, while keeping a human in the loop for judgment.
How It Connects to My Other Work
Noah-Agent sits alongside two related efforts:
- Noah-MP land surface modeling, my main PhD research direction on physics-based land surface models.
- ESM-bench, a benchmark for whether AI agents actually understand Earth System Model physics and code. Noah-Agent is a system that such a benchmark is meant to test.
Status and Links
This is a preprint, version 0.1, in preparation. There are no claimed results or acceptances yet.
- Preprint (Zenodo): https://zenodo.org/records/17862049
- Author: Wu, K. (2025)