Rapid probabilistic source characterisation in 3D earth models using learning algorithms

Abstract:

Characterising earthquake sources rapidly and robustly is an essential component of any earthquake early warning (EEW) procedure. Ideally, this characterisation should:
(i) be probabilistic -- enabling appreciation of the full range of mechanisms compatible with available data, and taking observational and theoretical uncertainties into account; and
(ii) operate in a physically-complete theoretical framework.
However, implementing either of these ideals increases computational costs significantly, making it unfeasible to satisfy both in the short timescales necessary for EEW applications.

The barrier here arises from the fact that conventional probabilistic inversion techniques involve running many thousands of forward simulations after data has been obtained---a procedure known as `posterior sampling'. Thus, for EEW, all computational costs must be incurred after the event time. Here, we demonstrate a new approach---based instead on `prior sampling'---which circumvents this problem and is feasible for EEW applications. All forward simulations are conducted in advance, and a learning algorithm is used to assimilate information about the relationship between model and data. Once observations from an earthquake become available, this information can be used to infer probability density functions (pdfs) for seismic source parameters, within milliseconds.

We demonstrate this procedure using data from the 2008 Mw5.4 Chino Hills earthquake. We compute Green’s functions for 150 randomly-chosen locations on the Whittier and Chino faults, using SPECFEM3D and a 3D model of the regional velocity structure. We then use these to train neural networks that map from seismic waveforms to pdfs on a point-source, moment-tensor representation of the event mechanism. We show that using local network data from the Chino Hills event, this system provides accurate information on magnitude, epicentral location and source half-duration using data available 6 seconds after the first station triggers; if longer windows are used, constraints on focal mechanism can also be obtained. We demonstrate that the use of 3D wave propagation allows results to be constrained better than is possible when only 1D earth models are used.