Towards an improved model-data fusion framework to track and forecast the volcanic system state at arc volcanoes in [near] real-time

Abstract:

Once volcanic unrest is determined from surface observations, it is important to understand what initiated the unrest and what conditions can lead to reservoir failure in order to deliver an accurate and timely forecast. This requires tracking the state of the volcanic system in real-time while relating the system failure parameters to surface observations using realistic models. For example, at arc volcanoes where complex transcrustal magmatic processes occur, these parameters include: the magma influx rate, magma and continuum rheology, the geometry and location of storage zones, and volatile exsolution.

However, combining observations and models efficiently to determine as much as possible the true state of the system remains a challenge. On one hand, observations are discrete in time and space whereas the true state of the system is continuous. This difference often results in gaps in information. Observations are also always accompanied by perturbations that are related to natural error sources such as atmospheric perturbations as well as measurement errors related to the instrumentation, data acquisition and processing. On the other hand, models provide the physical basis on how the true system evolves and gives rise to the observations. Although we seek to develop realistic and tractable models, it is impossible to completely represent the true state of the system, therefore our models will always incorporate inaccuracies. Here, I will discuss some of the state-of-the-art model-data fusion strategies used in volcano geodesy and provide relevant examples of their application (e.g. rapid response during the 2020 Taal eruption). I will talk about volcanic data assimilation—the future of real-time eruption forecasting—and the need to develop a fast framework that will incorporate realistic and unifying dynamic models, multi-parametric observations, and their respective errors. Finally, I will demonstrate how a simple and first-order volcanic data assimilation framework can be used and possibly be integrated into operational practice and decision event trees.