Can LLM Agents Improve Imperfect Parameterizations in Earth System Models?

Earth system models (ESMs) rely on simplified models, or parameterizations, to represent unresolved processes such as convection, cloud microphysics, and turbulence. However, translating new physical process understanding into stable, validated parameterizations and integrating them into ESMs often takes years. Agentic AI could accelerate this process, while enabling more systematic exploration of design choices. Using warm-rain collision–coalescence as a test case, we evaluate agentic AI to synthesize scientific literature, reimplement published bulk parameterizations, and propose new functional forms, implemented as differentiable JAX code. We compare conventional parameterizations, data-driven emulators, and agent-generated parameterizations against high-fidelity reference superdroplet simulations, and systematically compare performance. With expert guidance or iterative evaluation feedback, agents produced executable, differentiable, and mass-conserving code, but unattended performance remained strongly scheme dependent. Feedback reduced rollout errors, yet no agent-generated closure outperformed the retuned physical schemes, which often matched learned closures at substantially lower computational cost.