Accelerating low-frequency ground motion simulation for finite fault sources using neural networks

Abstract

SUMMARYIn the context of early emergency response to moderate and large earthquake shaking, we present a simulation based low-frequency ground motion estimation workflow that expedites an existing simulation method while taking into account simplified source process information. We focus on using source information that can be expected to be available shortly after an impacting earthquake, for example moment-tensor and simple finite-fault parameters. We utilize physics-based simulations which can include effects based on source orientation or finite faults, like rupture directivity. In order to keep the computational effort within feasible bounds and to apply the approach on global scale, we restrict ourselves to a low-frequency setup (standard 1-D layered earth model and 2 Hz sampling frequency) for either a moment tensor or a simple kinematic finite fault model. From the simulated records we then extract ground motion parameters of interest for arbitrary locations within the area of expected impact and display the expected spatial patterns of ground motion. Although simulations are kept simple, the results from this low-frequency ground motion parameter simulation (e.g. for peak-ground displacement) are in good agreement with observations from two well-studied earthquakes and partially more accurate than traditional, more empirical approaches (standard deviation <0.3 log10 units). However, waveform calculation and subsequent ground motion parameter extraction is computationally expensive. For a significant computational speedup in the context of rapid ground motion assessment, we directly train neural network (NN) models from large sets of source model information and their corresponding spatial ground motion distribution. We show that the trained NNs are able to reproduce the earthquake source related effects, like directivity and focal mechanism patterns, of the ground motion in any case. Given a set of source parameters, we obtain prediction errors smaller than 0.05 log10 units (ca. 11 per cent) and a magnitude dependent increase in computational speed of more than 1000 times compared to the initial waveform modelling. The proposed procedure enables thus to immediately compute probabilistic ground motion maps related to uncertainties in source parameters estimates, for example by sampling distributions based on parameter uncertainties or directly from an existing ensemble of focal parameter solutions.