Global sensitivity analysis of models for volcanic ash forecasting

Abstract

Volcanic ash is a widespread and destructive volcanic hazard. Timely and accurate forecasts for ash deposition and dispersal help mitigate the risks of volcanic hazards to society. Producing these forecasts requires numerous simulations with varying input parameters to encapsulate uncertainty and accurately capture the actual event to deliver a reliable forecast. However, exploring all possible combinations of input parameters is computationally infeasible in the lead up to an eruption. This research explores the input space of two volcanic ash transport and dispersion models, Tephra2, which is based on a simplified analytical solution, and Fall3D, which is a computational model based on more general assumptions, in the context of forecasting an unknown future eruption. We use the exemplar of Taranaki Mounga (Mount Taranaki), Aotearoa New Zealand, which has an estimated 30% to 50% chance of an explosive eruption in the next 50 years. We statistically determine how much each input parameter contributes to model output variance through a global sensitivity analysis via Sobol’ indices and the extended Fourier Amplitude Sensitivity Test (eFAST). Our findings show that grain size distribution, diffusion, plume shape, and plume duration (Fall3D only) have a substantial first-order impact on model output variance. In contrast, mass, particle density, and plume height have minimal impact in the first-order but become influential when considering parameter-parameter inter-relationships (total-order). The results not only enhance our understanding of model sensitivities but also point to improved efficiency in forecasting efforts.