Functional green illusion in global semi-arid carbon uptake revealed by physics-informed AI

Rising atmospheric vapour pressure deficit (VPD) is reshaping dryland carbon cycling, yet the widely held assumption that satellite-observed greening directly translates to enhanced gross primary productivity (GPP) remains untested at a global scale for water-limited ecosystems. Existing studies have not quantified the global extent of this ‘functional green illusion’, nor resolved the challenge of ensuring ecophysiologically consistent projections under extreme drought conditions. Pixel-wise Theil–Sen and Mann–Kendall trend classification of NDVI and GPP (2000–2024) shows that 5.3% of global semi-arid grasslands exhibit significant structural greening without corresponding GPP gains, with Central Asia as the primary hotspot (9.9% regional incidence). Separately, a physics-informed multimodal transformer (PIMT) that embeds light-use efficiency constraints is developed to map semi-arid grassland carbon uptake and to interpret the physiological mechanism of this decoupling: VPD-driven stomatal closure acting on minutes-to-hours timescales beneath a persistent green canopy. Our findings reveal a critical bias in terrestrial carbon sink assessments that rely on satellite greenness as a productivity proxy, and demonstrate that process-guided AI is essential for credible terrestrial carbon monitoring under intensifying atmospheric drought.