Testing Driving Mechanisms of Megathrust Seismicity With Explainable Artificial Intelligence

The correlation between subduction zone features and megathrust seismicity provides relevant clues on what controls the generation, location and clustering of mega-earthquakes (magnitudes M_w ≥ 8.0). Thus far, weak correlations are found between subduction zone parameters and seismicity through bivariate statistical analyses. Here, we used Explainable Artificial Intelligence (XAI) to assess the relevance of geophysical properties and tectonic motions along major subduction zones, paired with novel proxies of slab stress from calculations of buoyancy-driven subduction. The features derived from these data sets, describing the physical state, kinematics, and dynamics, served as inputs to a Fully Connected Network (FCN) trained to classify segments according to the largest earthquake magnitude that ruptured it. The subsequent use of Layer-wise Relevance Propagation, an XAI technique, on a trained FCN provides an estimate of the relevance of the input, identifying the features most relevant to the classification. The XAI procedure confirmed the importance of subduction interface curvature, sediment thickness, long wavelength bathymetric roughness, and free-air gravity anomalies, as previously proposed. Interestingly, our procedure revealed the importance of slabs extending to the upper mantle as well as the trench-parallel slab stress, showing how three-dimensional subduction forces may control large earthquakes. This suggests the preferential occurrence of large earthquakes on megathrust segments around slab steps and edges, where the slab depth measured along trench varies abruptly. At these steps, the trench-parallel forcing is maximized by the excess load of neighboring deeper slabs.