Estimating geodetic locking depth and long-term slip rate along the central San Andreas Fault using Neural Networks

Abstract:

Geodetic observations, such as from the Global Navigation Satellite System (GNSS), record deformation of the earth’s crust due to strain accumulation on locked faults. Along a vertical strike-slip fault such as the San Andreas Fault, a geodetically determined long-term slip rate and locking depth may be directly proportional to the moment of a future earthquake. Locking depth is generally interpreted to correspond to the base of the seismogenic zone, but it can be difficult to estimate. Typically, the locking depth in geodesy-based fault models is prescribed, complicating its interpretation in terms of the physical behavior of the earth. As an alternative, we use a neural network to independently estimate long-term slip rate and locking depth along the San Andreas Fault, using geodetic data from GNSS. Because there is insufficient real-world data available for training, we train the neural network using synthetic data generated from a model of a locked strike-slip fault. Then we apply the trained neural network to 1-D profiles of geodetic observations along the San Andreas Fault. We estimate an increase of locking depth from ~12 km in the northern Cholame segment to ~24 km on the southern Carrizo segment and an increase in slip rate from ~32 mm/yr to ~39 mm/yr on the same segments. Our estimated slip rates are consistent with previous studies that prescribe locking depth, and an estimated increase in locking depth corresponds to the maximum depth of microseismicity on these segments of the San Andreas Fault. This work serves as a proof-of-concept for the feasibility of estimating fault parameters with a neural network and will help us better understand the physical basis of locking depth, illuminate the fundamental nature of the seismogenic zone, and better anticipate a future San Andreas earthquake.