NEWS & EVENTS

Understanding where magma is stored and how it moves before eruption is important for reconstructing volcanic plumbing systems and interpreting eruptive behaviour. Temperature and pressure are key constraints on these processes and are commonly estimated using mineral-based geothermobarometers. Clinopyroxene (cpx), a common mineral in many igneous rocks, has a wide stability field and a large experimental calibration dataset, making it an important phase for developing thermobarometers. Recent machine-learning studies have produced a number of cpx-based thermobarometer models calibrated using different datasets and algorithms, creating a practical challenge in identifying models suitable for a given magmatic system. Model performance has commonly been assessed using in-distribution data (i.e., data drawn from the same underlying distributions as the calibration dataset), which may underestimate uncertainties when the models are applied to out-of-distribution (OOD) samples. We propose a machine-learning-based framework to rank clinopyroxene-based thermobarometers for given natural samples. The ranking system comprises two parts: (1) applying an OOD detection scheme to exclude unsuitable models and (2) applying deviation functions obtained for individual thermobarometers to predict P–T deviations and identify models giving the lowest deviation. The ranking system was tested using a dataset compiled from published experiments with no overlap with the calibration datasets of the models considered in this study. The P–T estimates of the selected models show lower root-mean-square errors than any single model, with improvements of ~0.4–0.6 kbar and ~10–14 °C. In the Merapi case study, the selected results suggest a broader and more vertically extensive storage system in 2010 than in 2006. Together with previous petrological and geophysical evidence, these results are consistent with deeper, hotter, and more volatile-rich recharge prior to the more explosive 2010 eruption.

For additional information, please contact Miss Xiaoyu LIU, liuxy23@connect.hku.hk.