Constrained Carbon Partitioning: A Self-Trained Physics-Informed Machine Learning Model Refines GPP Estimates From Eddy Covariance Measurements.

Gross primary productivity (GPP) is the largest term in the global carbon budget but cannot be directly observed. We present a knowledge-guided machine learning (KGML) framework that partitions eddy covariance-measured net ecosystem exchange (NEE) into gross primary production (GPP) and ecosystem respiration (RECO) with partitioned water vapor fluxes and CO2 flux source areas from 36 U.S. National Ecological Observatory Network (NEON) towers. The KGML is guided by hard physical constraints that enforce mass balance and 'soft' theoretical expectations including optimal stomatal response to vapor pressure deficit (VPD) and links between GPP and transpiration (T) through stomatal function. The model achieves strong physical consistency (NEE R2 = 0.99) while capturing expected ecophysiological relationships including GPP-T coupling (R2 = 0.58) and stomatal responses to light and VPD. Compared to conventional partitioning methods, KGML infers lower GPP and RECO estimates on average, with the largest negative biases occurring at low light levels (0-200 μmol photons m-2 s-1). These differences likely reflect a combination of mechanisms including light-induced respiration suppression consistent with the Kok effect, stomatal-transpiration coupling constraints, and dynamic allocation between respiration components. The flux differences vary across plant functional types (PFTs), where forested ecosystems (deciduous broadleaf, evergreen needleleaf, and mixed), savannas and grasslands show the largest negative annual GPP deviations (-10% to -18% versus nighttime partitioning), while croplands and open shrublands show moderate negative deviations (-5% to -10%). The lower GPP estimates by PFT are closer to those inferred by Keenan et al. that explicitly considers limitations on RECO from the Kok effect. We discuss implications for our understanding of ecosystem and global carbon cycle processes, as well as ways to further benefit from the full information content of eddy covariance observations by combining physics with knowledge of biological processes.