Relating Megathrust Seismogenic Behavior and Subduction Parameters via Machine Learning at Global Scale

The ability of megathrust fault segments to generate devastating interplate earthquakes (and triggered tsunamis) has been long recognized as partially controlled by one or more factors related to the plate tectonics configuration of subduction zones. However, there is still debate regarding the actual contribution of each factor and possible combinations of them that could favor the occurrence of large earthquakes. We investigated (Crisosto and Tassara, GRL2024) the relationship between the seismogenic behavior of megathrusts segments at a global scale and various subduction parameters (subducting plate age and roughness, slab dip, convergence speed and azimuth, distances to closest ridge and plate boundary). For each of 157 trench-perpendicular transects covering most of the subduction zones worldwide we estimate one value of the afford mentioned parameters and one b-value of the frequency magnitude relationship (Gutenberg and Richter, 1946) that parameterizes the relative amount of large to small earthquakes. For this we use the ISC global seismicity catalogue between 1900 and 2022 considering events located less than 10 km around the SLAB2.0 model (Hayes et al., 2018) and computed the b-value for each transect implementing the b-positive estimator (van der Elst, 2021), which helps avoiding contamination of the estimates by transient changes during aftershock sequences. With this dataset we performed a parametric approach by implementing three decision tree‐based Machine Learning (ML) algorithms to predict the b‐value as a non‐linear combination of subduction variables. Using the Shapley Additive exPlanation (SHAP) values to interpret the ML results, we observe that plate age and subduction dip are the most influential variables, as also noticed by previous authors (e.g. Nishikawa and Ide, 2014). However, our results contradict these previous views because we observe that older, not younger slabs, that are associated to shallow‐dipping plates correlates with low b‐values, pointing to higher megathrust stress (using the b-value as a stressmeter, as proposed by Schoerlemer et al., 2005). This pattern is attributed to the higher rigidity of older plates, increasing flexural strength that opposes to bending, generating a shallow penetration angle, increasing the frictional interplate area and therefore augmenting the likelihood of larger earthquakes. These findings shed light on the complex dynamics of seismic activity on a global scale and provide valuable information for understanding the megathrust earthquake behavior and its hazard assessment worldwide