Global land carbon sink halved in 2024, AI model suggests (phys.org)

Peking University researchers using a process-based AI model called Carbon Mind report that the global land carbon sink in 2024 dropped to less than half the decade-long average, driven by an abrupt, extreme jump in global temperature. Published in Science Bulletin, the study maps the loss at high spatial resolution and finds the biggest proportional declines across the tropics — notably in semi-arid grasslands and savannas rather than in rainforests — implicating heat- and drought-induced declines in vegetation productivity as the principal drivers. Technically, Carbon Mind blends mechanistic understanding of terrestrial carbon processes with AI learning to produce interpretable, traceable estimates and to overcome the typical one-year lag of conventional assessments, enabling near–real-time detection and diagnosis of carbon-cycle responses. The result signals that tropical land systems may be more vulnerable than assumed, with immediate implications for atmospheric CO₂ growth and climate projections. Integrated with atmospheric inversions and ground observations, this AI-enabled approach can improve stress-testing of climate pathways, guide adaptive land-management, and support more responsive policy interventions as climate extremes become more common.