A Machine Learning Approach for Volcanic Eruption Mass Estimation

Abstract Estimation of total volcanic erupted mass—the primary metric of eruption magnitude—is typically performed post‐eruption relying on dense monitoring of ground‐based seismology, gravity and deformation instrumentation, and therefore exists for only ∼100 of ∼1280 volcanoes worldwide. Here we present the first global‐scale assessment of potential erupted mass using a machine learning framework that integrates geophysical and eruption‐history data. The model is trained on a heavy‐tailed eruption‐history data set, yet the Gradient Boosting Regression Tree (GBRT) model using quantile log‐transformation with cost‐sensitive learning (Huber) successfully recovers both low‐mass and high‐magnitude events with acceptable precision with an uncertainty of 0.32 Gt. The model applies to predict erupted mass for 135 globally distributed volcanoes active between 1982 and 2024, assuming medium‐to long‐duration eruptions. Three primary metrics, together with out‐of‐bag (OOB) error analysis, were used to comprehensively evaluate the performance of the predictive model. Permutation importance analysis, which directly evaluates feature impact on model performance, identifies eruption start/duration, last volcanic activity, Moho depth, gravity, and elevation as the dominant predictors. Our results indicate that large inferred eruptions may occur at long‐dormant volcanoes, while substantial erupted volumes can also be associated with persistently active systems. We present the first global hazard map highlighting regions with potential for massive eruptions.