Terrestrial evapotranspiration (ET) is widely mapped as a flux, but the vegetation–atmosphere coupling regime that regulates it, commonly summarized by the decoupling coefficient (Ω), is rarely represented in gridded data products. Existing Ω estimates are largely restricted to flux-tower studies, leaving a gap in long-term, spatially continuous, observation-constrained Ω data for the contiguous United States (CONUS). Here, we present a gridded Ω dataset for CONUS from 1982 to 2020 at 0.1° spatial and 8-day temporal resolution. The dataset was generated by combining tower-derived Ω estimates from 74 AmeriFlux sites with gridded meteorological, vegetation, and surface predictors in a specific plant-functional-type (PFT) physics-informed machine learning (PIML) framework. Product reliability was assessed using site-level cross-validation, comparison with a direct machine learning benchmark, feature extrapolation diagnostics, Monte Carlo uncertainty propagation, and pixel-level quality-control flags. The PIML model reproduced tower-derived Ω at withheld sites with a pooled coefficient determination (R<sup>2</sup>) of 0.77 and root mean square error (RMSE) of 0.13, and showed clearer gains over direct ML under high extrapolation risk. The gridded product captured expected differences among PFTs, with higher Ω in croplands and lower Ω in open shrublands and savannas, while retaining substantial within-PFT spatial and seasonal variability. Each annual NetCDF file includes three main variables: Ω estimates (‘omega’), associated uncertainty (‘omega_unc’), and quality-control flags (‘QC_flag’). The dataset is available at <a href="https://doi.org/10.5281/zenodo.21670853" target="_blank" rel="noopener">https://doi.org/10.5281/zenodo.21670853</a> (Xiong and Feng, 2026).