Dr. Benavente-Bravo, Roberto

Estimating locking degree at faults is important for determining the spatial distribution of slip deficit at seismic gaps. Inverse methods of varying complexity are commonly used to estimate fault locking. Here we present an innovative approach to infer the degree of locking from surface GNSS velocities by means of supervised learning (SL) algorithms. We implemented six different SL regression methods and apply them in the Central Chile subduction. These methods were first trained on synthetic distributions of locking and then used to infer the locking from GNSS observations. We tested the performance of each algorithm and compared our results with a least squares inversion method. Our best results were obtained using the Ridge regression, which gives a root mean square error (RMSE) of 1.94 mm/yr compared to GNSS observations. The ML-based locking degree distribution is consistent with results from the EPIC Tikhonov regularized least squares inversion and previously published locking maps. Our study demonstrates the effectiveness of machine learning methods in estimating fault locking and slip, and provides flexible options for incorporating prior information to avoid slip instabilities based on the characteristics of the training set. Exploring uncertainties in the physical model during training could improve the robustness of locking estimates in future research efforts.

We propose a method for defining the optimal locations of a network of tsunameters in view of near real-time tsunami forecasting using sea surface data assimilation in the near and middle fields, just outside of the source region. The method requires first the application of the empirical orthogonal function analysis to identify the potential initial locations, followed by an optimization heuristic that minimizes a cost-benefit function to narrow down the number of stations. We apply the method to a synthetic case of the 2015 Mw8.4 Illapel Chile earthquake and show that it is possible to obtain an accurate tsunami forecast for wave heights at near coastal points, not too close to the source, from assimilating data from three tsunameters during 14 min, but with a minimum average time lag of nearly 5 min between simulated and forecasted waveforms. Additional tests show that the time lag is reduced for tsunami sources that are located just outside of the area covered by the tsunameter network. The latter suggests that sea surface data assimilation from a sparse network of stations could be a strong complement for the fastest tsunami early warning systems based on pre-modelled seismic scenarios.