Observation-Constrained Agroecosystem Model Inversion Reveals Continental-Scale Variation of Winter Wheat Traits

Understanding how crop trait variability shapes genotype × environment × management (G × E × M) interactions remains a key uncertainty in predicting agricultural performance under a changing climate. Continental-scale crop models commonly rely on spatially uniform parameters, limiting their ability to represent adaptive variation in phenology, allocation, and yield formation. Here we integrate the mechanistic agroecosystem model Ecosys with a deep learning-enabled model-data fusion inversion to infer spatially explicit physiological controls (trait proxies) of U.S. winter wheat directly from observations. By constraining simulations with satellite-derived photosynthesis and county-level yield records from 2008 to 2022 across ~1000 winter wheat-producing counties, the inversion recovers coherent patterns of maturity group, reproductive capacity, harvest index, and root-shoot allocation. The optimized simulations reproduce observed carbon uptake and yield variability (gross primary productivity r = 0.76-0.88; phenology bias < 2 weeks; > 90% of yields within ±20% of reports) and reveal distinct physiological profiles that align with the geographic distributions of major winter wheat market classes. The inferred controls explain class- and region-specific climate sensitivities: warmer winters reduce vernalization success in late-maturing cultivars, while elevated vapor pressure deficit causes strong yield losses in rainfed Hard Red Winter wheat. The results demonstrate that observation-constrained trait inversion within model-data fusion framework reveals biologically meaningful crop-class variation, thereby providing a scalable, physiologically grounded framework for diagnosing adaptive diversity and climate vulnerability across agroecosystems.