Abstract: The pattern of vegetation biomass allocation is a key mechanism for understanding the carbon sequestration function of grassland ecosystems and their responses to global change. By integrating multi-scale observations and global datasets, this study applied extreme gradient boosting (XGBoost) and used SHapley Additive exPlanations (SHAP) to quantify variable importance and partial dependence across environmental gradients, revealing how climatic, edaphic, and biological factors synergistically regulate large-scale grassland biomass allocation, and produced a global dataset at 0.05° spatial resolution. The XGBoost model estimated mean values of 282.05 ± 98.57 g m⁻² for aboveground biomass (AGB), 999.71 ± 264.17 g m⁻² for belowground biomass (BGB), and 6.41 ± 2.25 for root-to-shoot ratio (RS) at global-scale, all of which exhibited distinct spatial distribution patterns. Under water-limited conditions, plants adopted a root plasticity prioritized strategy; rising temperatures promote AGB but suppress BGB, while increased precipitation mainly promoted AGB. Species with high specific leaf area (SLA) did not substantially enhance AGB, whereas those with low SLA invested more in root development to cope with environmental stress. In near-neutral soils, BGB increased with rising pH, and moderate levels of exchangeable magnesium were beneficial to AGB. Our findings demonstrate that grassland biomass allocation represents an adaptive outcome of resource trade-offs and physiological plasticity under multifactorial stress. These insights provide an important theoretical foundation for optimizing carbon sequestration functions and guiding sustainable management of grasslands in the context of global change.

Key words: Grassland, Biomass allocation, Abiotic and biotic controls, Machine learning, Structural equation modeling