Joint volcanic source term estimation and SO2 dispersion forecasting by combining model emulators, observations, and empirical relationships within a hierarchical Bayesian model.

We present a novel source term estimation and SO2 ground concentration forecasting system for the Reykjanes peninsula, developed as a Digital Twin Component for the DT-GEO project. The joint distribution over the source term and ground concentrations is estimated by drawing samples from the posterior specified by a Bayesian hierarchical model, using Hamiltonian Monte Carlo. The Bayesian model consists of a physics based forward model that gives ground concentrations as a function of the source term, and an empirical model that describes the influence of the atmosphere on the source term. The forward model has to be (1) fast and (2) differentiable for the sampling procedure, so we replace the forward model (Fall3D) with an emulator. We can use simple linear emulators by taking advantage of the linearity of ground concentration with flux for basic scenarios, while neural network based emulators are being developed for more complicated model physics. Observations of SO2 flux, plume height, and ground concentration constrain the hierarchical model parameters that determine the forecast ground concentrations. Application to the recent basaltic fissure eruptions on Reykjanes, Iceland, will be presented.