Fingerprinting subduction margins using principal component analysis profiles: A data science approach to assessing earthquake hazard

Abstract Giant earthquakes (Mw ≥8.5) along subduction margins pose great hazards to coastal societies. While it is generally accepted that geological margin properties play a role, controls on giant earthquake occurrence remain undetermined. The long intermittence times of giant earthquakes and the comparatively short earthquake record obscure any correlations between margin properties and seismicity. This work presents a new approach to relating margin properties to seismicity. We apply principal component analysis (PCA) to four margin properties to “fingerprint” margins by assigning them a PCA profile, which we compare to giant earthquake occurrence. This approach reduces bias from the short earthquake record because seismicity is not used as a PCA input feature. Using kernel PCA, a nonlinear PCA variant, we find nonlinear patterns in margin properties and suggest that links between properties and seismicity are nonlinear, which helps explain why they have been hard to establish. PCA clusters identify “active and moderate” and “quiet and extreme” margins. We argue that margin segments with “quiet and extreme” PCA profiles but no giant earthquakes since 1900 are as hazardous as those that have ruptured in giant earthquakes recently.