Unprocessed dataset used in the study: Testing driving mechanisms of megathrust seismicity with Explainable Artificial Intelligence
Our understanding of what may control the occurrence of mega-earthquakes at subduction zones is driven by the correlation between these margins’ large-scale features and their seismicity. We pair datasets describing the observable geophysical features with proxies of novel slab stress calculations, embedding the deep driving forces of subduction, to train a fully connected network (FCN) to classify whether segments of the megathrust can be ruptured by a large earthquake. After obtaining models that generalize well to unseen data, the use of layerwise relevance propagation (LRP) enables the quantification of the importance of the features to the model’s outputs. That is, what parameters are most influencial in the determination of the eventual size of earthquakes. Our explainable artificial intelligence (XAI) procedure finds that previously proposed features are consistently relevant, such as the megathrust curvature, sediment thickness, ocean floor roughness and upper plate tectonics, and illustrate the relevance of trench-parallel slab stresses. This confirms several previously proposed parameters and corresponding mechanisms. Importantly, the results suggest that the excess weight of slabs can force neighbouring megathrusts, providing a novel, three-dimensional view of what may cause large earthquakes.