Water plays a critical role in controlling magma density, viscosity, and petrologic evolution, but accurate quantification of water contents in magmas remains challenging. This challenge is further complicated by the high diffusivity of hydrogen, which can lead to underestimation of water contents in melt inclusions. In this study, we present a machine learning (ML)−based hygrometer calibrated on experimental plagioclase-liquid equilibrium data compiled from the literature. To overcome the limitations imposed by the small-scale experimental datasets (n = 330), the Markov chain Monte Carlo approach was employed to generate additional, geologically plausible data (n = ∼3000) by sampling within the analytical uncertainties of the initial dataset. The augmented dataset was used to calibrate a new plagioclase-liquid hygrometer using the extremely randomized trees algorithm. The new hygrometer can be applicable to a wide range of compositions (from basalts to rhyolites), pressures (30−1600 MPa), and temperatures (563−1220 °C). The new hygrometer was further validated on an independent dataset of hydrous melting experiments. When applied with experimental temperature and pressure, the hygrometer achieved a root mean square error (RMSE; 1σ) of 0.55 wt% and an R2 of 0.92. For comparison, we also tested its performance using temperatures and pressures calculated with the MagMaTaB thermobarometer, which yielded an increased uncertainty (R2 = 0.74, RMSE = 0.97 wt%). Finally, the new hygrometer was applied to three volcanic eruptions in arc and hotspot settings and yielded more reliable water contents with a small error compared to existing hygrometers. By successfully quantifying water contents in magmas using plagioclase and melt compositions, our ML-based hygrometer enables effective comparison of magmatic water contents across different tectonic settings.
LIN et al. (Tue,) studied this question.