Machine learning clinopyroxene hygrometry reveals slab-driven mantle hydration in the genesis of the Emeishan large igneous province, southwestern China

Water plays a critical role in the formation of large igneous provinces (LIPs) and deep-Earth material cycling. However, directly quantifying water content in ancient mantle sources remains a challenge. Here we developed a novel clinopyroxene-based hygrometer using a comprehensive dataset of 316 experimentally crystallized clinopyroxene samples with known equilibrium melt water contents. Based on this dataset, we trained and tested multiple machine learning models to predict water content from the major element composition of clinopyroxene. The S-MKL algorithm, a combination of a stacking-based representation learning framework (S-RL) and multiple kernel learning (MKL), achieved superior predictive accuracy and was selected as the optimal model. We applied this new hygrometer to clinopyroxene phenocrysts from the Permian Emeishan LIP in SW China. The results reveal a hydrous mantle source for the Emeishan LIP, with water contents comparable to those of backarc basin basalt. A pronounced decrease in the gradient of water content was identified from southwest to northeast, reflecting mantle source heterogeneity linked to Paleo-Tethyan oceanic slab subduction and stagnation in the mantle transition zone. Fractional crystallization modeling confirms that this gradient represents source-level variations rather than shallow processes. We conclude that fluids released from the stagnant slab metasomatized the mantle, enhancing its fertility and contributing to LIP generation. This study quantitatively establishes a link between slab-driven mantle hydration and LIP emplacement, highlighting the importance of plume-slab interactions and deep-Earth water recycling.