Empowering agricultural decision making with CLM: Mechanism guided crop large model

Model-based reinforcement learning (MBRL) provides an attractive paradigm for optimizing irrigation and fertilization in precision agriculture. However, its reliability depends significantly on crop simulation models, which often struggle to predict crop status under out-of-distribution (OOD) conditions. This issue is usually attributed to poor OOD generalization, which is largely caused by insufficient representation of crop growth mechanisms in learned simulation models. To solve this issue, we propose CLM, a mechanism-guided Crop Large Model constructed through an autoregressive generative pretraining framework for action-conditioned crop dynamics forecasting. By distilling crop growth mechanisms from different mechanistic crop simulators, CLM captures a wider range of mechanism-driven crop dynamics and improves the generalization of crop growth forecasting under OOD conditions. Leveraging the generalization of CLM, we further integrate CLM into an MBRL framework as a pretrained crop-specific prediction model for OOD environments, generating more reliable multi-step imagined trajectories to support the optimization of fertilization and irrigation policies. Experiments conducted with the WOFOST simulator show that CLM achieves a modest but consistent average improvement of 5.0% in MAE for OOD crop growth forecasting over the evaluated baselines under severe distribution shifts. Further simulated fertilization-irrigation decision-making experiments demonstrate an improvement of more than 27% in net income and a reduction of nearly 17% in fertilization costs, with no statistically detectable reduction in simulated yield across the five paired evaluations. These simulated results suggest that our mechanism-guided CLM improves the accuracy of crop growth forecasting under OOD conditions and offers a more generalized basis for agricultural policy optimization, while real-world field validation remains an important direction for future research.