Uncovering the Physical Controls of Slow Slip Events Using Machine Learning

Date: 4/24/2019

Time: 06:00 PM

Room: Grand Ballroom

The discovery of slow-slip events (SSE) from GPS monitoring has drastically altered our understanding of how accumulated tectonic stress at subduction zones is released. Unlike regular “fast” earthquakes, stress during SSE is released over a period of days to months and is often associated with tectonic tremor. When the geodetic and seismic signals are correlated spatially and temporally, this is referred to as a single event of episodic tremor and slip (ETS). Slow slip events have not been observed along all subduction zones and are sometimes confined to specific segments along a given subduction zone. Several subduction zone characteristics (geometric, kinematic, and physical) have been proposed to explain the occurrence of SSE. Here we attempt to uncover the conditions that favor the nucleation of SSE using machine learning (ML) techniques. ML has the unique ability to uncover complex data patterns from known (i.e., training) data. that can be used to predict outcomes in an untested dataset. This has the dual benefit of determining which subduction zone characteristics (or discriminant features in ML volcabulary) are correlated with SSE as well as predicting the probability of their occurrence in poorly monitored areas. Our approach is based on extracting subduction zone characteristics from geophysical data that are mapped globally. We examine features that characterize the incoming, subducting plate at the trench, including sediment thickness, plate age, plate roughness, plate deformation, and relative plate velocity. Global subduction zones are split into 50 km-wide trench-parallel segments. Each segment is assigned a scalar value corresponding with each characteristic, and a label corresponding to the occurrence of SSE along that segment. We apply various ML algorithms (Guassian Naïve Bayes, Random Forest, Logistic Regression, Support Vector Machine, K-Nearest Neighbour, and Linear Discriminant Analysis) to test the validity of these plate characteristics as predictors in the frinctional behaviour of a given plate segment.

Presenting Author: Morgan McLellan

Authors