Can LLM Agents Improve Imperfect Earth System Model Parameterizations?

Kara D. LambarXiv 2026 · arXiv

TL;DR

This paper explores whether LLM agents can automate the slow, manual process of developing "parameterizations"—the mathematical shortcuts Earth System Models (ESMs) use to represent complex sub-grid processes like cloud formation. While agents excel at the structural "plumbing" of code (making it executable and differentiable), they struggle to outperform simple physical models that have been properly retuned, though iterative feedback allows the best agents to occasionally discover novel, highly accurate closures.

What Problem Does This Paper Address?

Earth System Models are our primary tools for predicting climate change, but they have a resolution problem. They can’t explicitly simulate every cloud or ocean eddy; instead, they use parameterizations to approximate these "unresolved" processes.

The bottleneck is that translating new physical insights into stable, validated code typically takes years of manual effort by experts. Furthermore, the field is currently split between traditional physical schemes (interpretable but potentially oversimplified) and modern machine learning emulators (accurate but computationally expensive and "black-box"). The authors ask: Can AI agents bridge this gap by automatically synthesizing literature, writing high-performance code, and iterating on new model designs?

Methods

The researchers focused on a specific test case: warm-rain collision-coalescence (how small cloud droplets merge into raindrops). They evaluated three agentic paradigms:

  1. Human-in-the-loop (HITL): An expert guides an agent (Claude Opus 4.8) to translate literature into code or build new neural architectures.
  2. Single-shot API: Ten different LLMs were asked to translate a scientific paper (PDF) into functional JAX code in one go.
  3. Iterative Feedback: Agents received automated feedback on their code's performance against reference data and were allowed to refine their "invented" closures over multiple rounds.

All code was generated in JAX, a framework that allows for automatic differentiation—a key requirement for modern "differentiable physics" models that can be calibrated using gradient-based optimization.

Overview of the analysis and methodology, showing the transition from explicit droplet distributions to bulk moments and the agentic experiment workflow.

Datasets

The study used the McSnow dataset, which provides high-fidelity "ground truth" simulations.

  • Reference: McSnow uses a "superdroplet" approach, explicitly tracking the stochastic collisions of millions of droplets. It is physically accurate but far too slow for climate models.
  • Scale: 819 initial conditions (719 for training/validation, 100 for testing), varying liquid water content, droplet radius, and distribution shape.

Key Insights and Ideas

  • The "Retuning" Reality Check: A major "aha" moment in the paper is that many fancy machine learning emulators only seem better than old-school physics because the physics models haven't been updated for decades. When the authors "retuned" classical schemes (like Kessler or Seifert & Beheng) to the McSnow data, the performance gap between physics and ML narrowed significantly.
  • Structural vs. Predictive Success: Agents are surprisingly good at the "hard" parts of software engineering—ensuring code is mass-conserving, differentiable, and stable for long simulations. However, they are less successful at "inventing" new physics that beats existing human-derived theories.
  • Feedback is the Secret Sauce: Unattended agents often fail on their first try. However, providing iterative feedback (e.g., "your model has a high rain-mass error") allowed median performance to improve for 8 out of 10 models tested.
  • Efficiency Matters: Traditional physical schemes are 1,000x to 10,000x cheaper in terms of floating-point operations (FLOPs) than neural network emulators. The paper shows that agent-refined physical schemes can achieve near-ML accuracy at a fraction of the cost.

Key Results

  • Translation Speed: In the HITL setting, an agent translated four complex historical schemes into JAX in less than an hour of active expert time.
  • Model Rankings: Claude Opus 5.5 was the top performer in unattended tasks. In the iterative refinement experiments, 2 out of 100 agent-generated closures actually achieved lower rain-mass error than the best retuned physical baseline.
  • Accuracy vs. Cost: As shown in the figure below, while data-driven models (like SuperdropNet) are highly accurate, agent-generated closures (like "opus 01") achieved comparable accuracy to retuned physical schemes while remaining computationally efficient.
Figure 2: Parameterization performance. (a) Mean error in rain-conversion timescales. Published and retuned coefficients are shown for the four literature schemes. (b) Rain-mass rollout error versus computational complexity, measured in floating-point operations per model step. Here, Superdropnet and Process-rate MLP refer to our reimplemented versions.

Limitations and Future Work

  • Box Model Simplicity: This study used a "box model" (a single point in space). In a real climate model, these schemes must interact with complex 3D dynamics and thermodynamics, which can amplify small errors.
  • Hallucination Risks: Agents can produce complex code that looks correct but contains subtle physical "hallucinations" or errors that are difficult for humans to spot without exhaustive testing.
  • Future Directions: The authors suggest moving toward "self-improving models" where agents continuously incorporate new satellite observations or high-fidelity data to refine ESM components autonomously. They emphasize the need for "safeguards," including automated conservation and stability tests, before AI-generated code enters production climate models.

Summary updated

Abstract

Earth system models (ESMs) rely on parameterizations to represent unresolved processes such as convection, cloud microphysics, and turbulence. However, translating advances in physical understanding into stable, validated parameterizations and integrating them into ESMs often takes years. Agentic AI could accelerate this process and enable more systematic exploration of design choices. Using warm-rain collision-coalescence as a test case, we evaluate agents' ability to synthesize scientific literature, reimplement published bulk parameterizations, and propose new functional forms as differentiable JAX code. We compare conventional parameterizations, data-driven emulators, and agent-generated closures against high-fidelity superdroplet simulations, repeating unattended experiments across ten language models. With expert guidance or iterative feedback, agents produced executable, differentiable, mass-conserving code, although unattended performance depended strongly on the scheme and language model. Iterative feedback lowered median rollout error for eight of ten models. Retuned physical schemes matched or outperformed many learned parameterizations at substantially lower computational cost, but two of 100 unattended agent-generated closures achieved slightly lower rain-mass error than the most accurate retuned scheme on held-out simulations.
BibTeX
@misc{lamb2026can,
  author = {Kara D. Lamb},
  title = {{Can LLM Agents Improve Imperfect Earth System Model Parameterizations?}},
  year = {2026},
  eprint = {2611.03543},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.03543},
}
Figure