Researchers at Michigan State University and The Ohio State University built Mímir, an irrigation agent that lets a language model propose daily watering while a soil-water simulator checks the likely consequences before a separate controller chooses an action. In a retrospective evaluation across 19 crop-years and four growing environments, Mímir used about 51% less water than replayed historical schedules and recorded a lower combined control cost. The study did not test the system on a live farm or measure crop yields.
Irrigation decisions have delayed effects: too little water can build soil stress over several days, while too much can run off or drain below the roots. Mímir turns soil, crop-stage and weather information into a structured state, then uses a DeepSeek-14B language model to suggest how much water to apply. A numerical soil-water model simulates the consequences and can send feedback through up to three proposal-and-revision rounds. A deterministic selector then compares a bounded set of feasible actions, and a separate execution gate retains final authority; the language model never sends a command directly to irrigation hardware.
The system also stores recurring, verified lessons as context for later proposals. That memory can influence what the model suggests, but it cannot change the model's weights, the soil-water equations, the scoring rules, the allowed action range or the execution gate. The design tests whether an agent can adapt its reasoning over time while keeping physical prediction and action limits outside the language model.
The evaluation replayed about 6,938 daily records from Michigan potato, Texas maize, and Colorado maize and sunflower environments. Under one shared evaluator, Mímir's aggregate control cost was 100.94 ± 1.80, with 2.58 ± 0.12 millimeters of irrigation; replaying the historical schedules scored 790.51 ± 17.76 and applied 5.32 ± 0.57 millimeters. The weighted cost combines irrigation, accumulated soil-water stress, runoff and deep percolation or leaching, so using less water alone does not count as better performance. Mímir also had the lowest cost among the paper's evaluated reference methods, but the authors do not claim it beats every irrigation controller or the best deterministic optimizer.
These are retrospective results produced by the study's physical model and evaluator, not measurements of real-world water savings, crop yield or a deployed controller. The paper also reports that larger language models did not produce monotonic gains, and it does not claim that stored experience improves performance every season. Whether the architecture transfers to live farms remains an open test.
Sources
- Mímir: Physics-Grounded LLM Agents for Long-Horizon Irrigation Control (abstract)
- Full paper, including evaluation protocol, results and limitations
- Image: Figure 1 by the paper's authors, licensed CC BY 4.0. It is a conceptual diagram, not a photograph of a field deployment.