The demand for reliable and replicable short-term probabilistic earthquake forecasts is becoming increasingly compelling, as we continue to witness seismic sequences with occasionally multiple disturbing or damaging earthquakes. Purely statistical models of earthquake clustering adequately capture the patterns of triggered seismicity and currently represent the standard approach for different operational earthquake forecasting systems. On the other hand, developing and testing physics-based forecast models let us validate the most popular physical hypotheses for earthquake triggering and clustering. These models couple complex stress interactions between faults with laboratory-derived frictional laws providing a framework for earthquake forecasting in the context of continuum mechanics. However, while featuring the unique characteristic of integrating many products of observational seismology, they are extremely data-intensive especially in near real-time settings where their applicability is still contentious. Over the last decade, the scientific advancements in seismology have provided higher resolution seismic catalogues as well as improved fault characterisations; this presents us with great opportunities to (1) evaluate their usefulness in improving the short-term performance of models of both forecast categories, (2) explore which specific modelling choices driven by real-time data quality and availability boost our forecasting skills and by how much, and (3) assess what are the data products required for such model improvements to be operationally delivered. To answer the above points, this thesis presents three forecasting experiments offering a novel experimental strategy, where the absolute and relative performance of statistical and physics-based models is formally quantified under different forecasting modes and modelling choices, in both cases of tectonic and induced seismicity. Looking ahead to the future improvements in near real-time input data quality promised by the most recent progresses in artificial intelligence techniques, the results of these experiments suggest what are the pathways that should be undertaken for future model developments.
|Date of Award||21 Jan 2021|
- The University of Bristol
|Supervisor||Max Werner (Supervisor) & Nicholas A Teanby (Supervisor)|