Path Signatures and Graph Neural Networks for Slow Earthquake Analysis: Better Together?

The prediction of earthquakes has long been considered to be cumbersome due to its random and stochastic nature (Vere-Jones, 2011). The identification of frequently occurring slow slip events (SSE), otherwise known as slow earthquakes (Obara, 2002), at subduction interfaces has led to the premise that GPS time series measuring land surface displacement can recover the true nonlinear or even chaotic signal underpinning the occurrence of earthquakes (Poulet et al., 2014). Since we currently have dense networks of such GPS sensors, the spatio-temporal analysis of the signals requires the development of space-sensitive reduction tools. In this work, we emphasize the slow earthquake sequences of the north island of New Zealand (Wallace, 2020). This area features a dense network of GPS stations with displacement measurements recorded continuously for over a decade (see Figure 1). The north island of New Zealand sits at the intersection of three slow earthquake sequences (see Figure 2 of Wallace (2020)) caused by the shallow subducting Hikurangi trench on the east, the deep signal under the Taupo volcanic area in the west, and the alpine fault system in the south. These three sequences have distinctly different nonlinear time-sequences (i.e.