Toward Determining the Controls on Subduction Zone Seismic Behaviour with Machine Learning

Variations in earthquake frequency and magnitude across global subduction zones are thought to be influenced by a combination of geological and geophysical factors, such as the age and dip angle of the subducting plate. Despite numerous previous qualitative studies on the correlation between seismic behaviour and subduction zone characteristics, the parameters and mechanisms governing seismicity at subduction zones remain elusive. Our limited historical record of earthquakes further complicates this understanding. Finding underlying general correlations and mechanisms that are valid across different subduction trenches is critical for assessing seismic behaviour and earthquake hazards along subduction plate boundaries which are poorly monitored or have been seismically quiet during the short instrumental record. This study aims to bridge the knowledge gaps highlighted above by applying specific unsupervised machine learning techniques to publicly available data on subduction zone parameters and earthquake catalogues. This approach is particularly adept at uncovering hidden correlations in complex, high-dimensional datasets, which might not be discernible through traditional analysis methods. We suggest that seismic behaviour may be describable as a non-linear combination of subduction margin parameters and present a quantitative tool for comparing seismic behaviours across different margins. This may help assess seismic hazards in regions with scant seismic records or that have been historically quiescent. By doing so, we hope to contribute significantly to the predictive modelling of earthquake occurrences and their potential impacts globally.