Scientists used large language model (LLM) agents to rewrite classic cloud‑physics equations, re‑implement recent machine‑learning closures, and propose new forms for warm‑rain collision–coalescence. The study shows that the agents can produce code that runs, differentiates, and conserves mass in a test box model. By comparing these agent‑generated closures with existing analytical schemes and neural‑network emulators, the researchers found that many retuned analytical models match or beat the learned closures while using far fewer floating‑point operations. The authors tested three workflows: an expert guiding the agent, a single‑shot call to an LLM, and an iterative call cycle with feedback. Retuning classical parameterizations reduced error to below 5 % of the high‑fidelity reference and lowered computational cost by up to 80 %. However, unattended performance varied with the specific LLM; only iterative refinement consistently improved structural validity, and accuracy gains were modest and model‑dependent. These results suggest that LLM agents can accelerate, but not yet replace, human expertise in developing Earth‑system model physics.

The paper’s Figure 1 illustrates the overall pipeline and the three agentic workflows, showing how the agents synthesize literature, code, and evaluate against a super‑droplet reference. Figure 2 compares mean rain‑conversion error and rollout error versus floating‑point operations per step, displaying that retuned classical schemes achieve comparable or lower errors than neural‐network closures at substantially lower computational cost.