Machine Learning-Primed Inference for Multiple-Event Location and Phase Association in Global Seismic Monitoring
Name
Albarghouty-lubnakb-phd-eaps-2026-thesis.pdf
Size
10.22 MB
Format
Adobe PDF
Checksum (MD5)
7d246f58b0bd4ff9899c2304c06a09af
Author(s)
AlBarghouty, Lubna Khalid
Advisor(s)
Ravela, Sai
Date Issued
May 2026
Publisher
Massachusetts Institute of Technology
Abstract
Global seismic monitoring requires detecting and characterizing seismic events from large volumes of arrival observations recorded across a station network. The central difficulty is that event location and phase association are fundamentally coupled: determining which arrivals belong to an event requires knowledge of its location, while estimating its location requires knowledge of which arrivals originate from it. In the presence of multiple events, sparse observations, false picks, and missed detections, this coupling leads to a highly combinatorial inference problem that remains a major obstacle to scalable monitoring.
This dissertation develops a machine-learning-guided, physics-based computational framework that addresses three fundamental challenges faced by seismic monitoring systems: the combinatorial complexity of arrival association, the representation of highly non-Gaussian origin uncertainty, and the enforcement of exclusive event–arrival correspondences.
First, to manage the combinatorial structure of the association problem, candidate event inference is formulated as a latent-variable problem in which arrival assignments are treated probabilistically within an expectation–maximization framework. Second, because association ambiguity yields complex, often multimodal uncertainties in origin, the dissertation develops a particle-based ensemble formulation to capture candidate event distributions. Third, to enforce the physical constraint that each arrival originates from at most one event, the dissertation develops an exclusivity-promoting extension of expectation–maximization that produces physically admissible one-to-one assignments. Together, these developments yield candidate event locations, estimates of the number of events, and a distributional representation of the resulting event bulletin.
The dissertation also addresses the computational burden of applying such inference globally. Fully physics-based inversion remains essential because it can discover events not represented in training data, but applying it over the entire globe is intractable. To resolve this tension, the dissertation develops a machine-learning hotspot detector that acts as a high-recall computational prefilter. Using an online analog-based learning strategy and an Ensemble Conditional Gaussian Process (Ens-CGP), the detector maps windowed global arrival patterns to coarse spatial hotspot fields that identify candidate regions of seismic activity. These regions restrict downstream inference to a small subset of the global search domain. Experiments show that most events are recovered while activating only a small fraction of the global grid, indicating that arrival geometry alone contains substantial information about the spatial distribution of seismic sources.
The resulting architecture combines machine learning, probabilistic inference, exclusivity enforcement, and non-parametric uncertainty quantification towards a unified monitoring framework. More broadly, the dissertation argues that scalable seismic monitoring is best approached not through purely association-first pipelines or purely black-box learning systems, but through a layered architecture in which machine learning guides where detailed physical inference should occur. This framework provides a principled path toward more tractable, interpretable, and uncertainty-aware global seismic monitoring.
MIT Department
Massachusetts Institute of Technology. Department of Earth, Atmospheric, and Planetary Sciences
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
