Abstract
Short-term earthquake forecasting provides crucial hazard information during aftershock sequences. State-of-the-art models, such as the Epidemic-Type Aftershock Sequence (ETAS)model, are formulated as point processes - statistical models that represent earthquakes as
points in time and space. Advances in earthquake detection now allow for the recording of much
smaller magnitude events, leading to new earthquake catalogs containing a ten-fold increase in
the number of recorded earthquakes. This volume of new data presents both opportunities and
challenges for existing forecasting models, as well as calling for the exploration of innovative
modeling approaches. This thesis explores the utility of Neural Point Processes (NPPs), a
machine learning variant of point processes, to enhance earthquake forecasting capabilities.
I begin by extending an existing temporal NPP to the magnitude domain, adapting it to
forecast earthquakes above a target magnitude threshold while depending on smaller magnitude
events. I apply this model to a catalog of the Central Apennines earthquake sequence in Italy,
demonstrating significant information gain over ETAS at the low magnitude thresholds of this
enhanced catalog.
Next, I apply Simulation Based Inference (SBI) to Bayesian parameter estimation for the
ETAS model, improving the scalability of inference from O(n
2
) to O(n log n) with the number
of earthquakes. By specifying a model through simulation rather than the likelihood, SBI
broadens the scope of available models to encompass greater complexity. This would enable
inference of earthquake models with intractable likelihoods, including data incompleteness and
physics based simulators.
Finally, I develop EarthquakeNPP, a benchmarking platform for evaluating NPPs against
state-of-the-art models from the seismology community. This platform provides benchmark
datasets from California, along with a widely accepted implementation of the ETAS model,
making these resources accessible to the machine learning community. The platform highlights
the potential of NPPs and outlines a road-map for future implementations to provide more
impact in earthquake forecasting.
By bridging the gap between statistical machine learning and seismology, this thesis provides
a foundation for future interdisciplinary research, offering valuable insights for researchers in
both fields.
| Date of Award | 4 Feb 2025 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Max Werner (Supervisor) & Daniel John Lawson (Supervisor) |