Can LLM Agents Improve Imperfect Earth System Model Parameterizations?
Organizations: Department of Earth and Environmental Engineering Learning the Earth with Artificial Intelligence and Physics Center Columbia University New York, NY 10027
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.
Figures & tables
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
| Research question | Answer | |
|---|---|---|
| (1) Can LLM agents translate published schemes into differentiable code? | Yes with light expert guidance; unattended success depends on the scheme and the model. | |
| (2) Does automated feedback improve agent-generated closures? | Structural validity improves reliably; accuracy gains are modest and depend on the model. Twelve-round searches improved performance for the two models examined. | |
| (3) Do agent-generated or learned closures outperform retuned physical schemes? | Two agent-generated closures do in rain-mass error, both from Claude Opus 5.5; learned closures only for the timing of rain onset. |
| Quantity | Symbol | Values |
|---|---|---|
| Initial liquid water content | ||
| Initial mean droplet radius | ||
| Gamma shape parameter | (unitless) |
| Reference | Moment order | Processes |
|---|---|---|
| Kessler (1969) | 1 | AU, AC |
| Berry and Reinhardt (1974a) | 2 | AU, AC |
| Khairoutdinov and Kogan (2000) | 2 | AU, AC |
| Seifert and Beheng (2001) | 2 | AU, AC, SC r , SC c |
| Scheme | Parameter | Published | Fitted | Ratio |
|---|---|---|---|---|
| Kessler | ||||
| Berry & Reinhardt | ||||
| Train loss | Held-out test loss | |||
| Scheme | published | fitted | published | fitted |
| Kessler | 3.127 | 0.800 | 3.988 | 0.776 |
| Berry & Reinhardt | 1.737 | 0.733 | 2.301 | 0.739 |
| Khairoutdinov & Kogan | 4.556 | 0.645 | 6.021 | 0.632 |
| Seifert & Beheng | 1.096 | 0.283 | 1.464 | 0.272 |
| Component | Sessions | Prompts | Substantive | Corrective | Active (h) | Commits |
|---|---|---|---|---|---|---|
| Translate 4 literature schemes | 1 | 5 | 4 | 2 | 0.5 | 3 |
| Seifert & Rasp (2020) MLPs | 3 | 25 | 19 | 4 | 4.0 | 2 |
| SuperdropNet (2025) | 3 | 36 | 24 | 8 | 7.1 | 3 |
| Physics-constrained Neural ODE | 5 | 131 | 101 | 14 | 22.0 | 19 |
| Model | Translation | Single-shot | Iterative | Iterative | Total |
|---|---|---|---|---|---|
| (40 calls) | (10 calls) | (10 4 rounds) | (5 12 rounds) | ||
| Gemini 2.5 Pro | 5.62 † | 1.77 † | 4.15 † | – | 11.54 |
| Gemini 3.1 Pro (preview) | 5.48 ‡ | 3.65 | 13.23 | – | 22.37 |
| Gemini 4 Argon | 1.95 § | 0.28 § | 2.22 § | 1.75 § | 6.21 |
| Claude Haiku 5.5 | 0.66 | 0.08 | 0.15 | – | 0.89 |
| Claude Sonnet 5.5 | 3.72 | 3.48 | 9.15 | – | 16.34 |
| Process | Seifert & Beheng (2001) | Claude Opus 5.5 closure | Relation |
| Autoconversion, moment dependence | Same form | ||
| Autoconversion, rain-fraction correction | , | , | Different |
| Accretion, kernel | , | , | Same form |
| Accretion, correction | (rain fraction) | (cloud-drop mass) | Different |
| Cloud number, autoconversion + self-collection | , , | Different split | |
| Rain self-collection | , | Same form, large drops limited |
AI Agents for Scientific Discovery
AI agents for scientific discovery use language models, tools, and experiments to generate, test, and refine scientific knowledge.
See all 323 papers