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

Summary In 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 method while taking into account simplified source process information. We focus on using information can be expected available shortly after impacting earthquake, e.g. moment-tensor simple finite-fault parameters. utilize physics simulations (PBSs) which include effects orientation or finite faults, like rupture directivity. order keep computational effort within feasible bounds apply approach global scale, restrict ourselves setup (standard 1D layered earth model 2 Hz sampling frequency) for either moment tensor kinematic fault model. From simulated records then extract parameters interest arbitrary locations area impact display spatial patterns motion. Although are kept simple, results from this parameter (e.g. peak-ground-displacement) in good agreement with observations two well-studied earthquakes partially more accurate than traditional, empirical approaches deviation <0.3 log10 units). However, waveform calculation subsequent extraction is computationally expensive. For significant speedup rapid assessment, directly train neural network models sets their corresponding distribution. show trained networks able reproduce related effects, directivity focal mechanism patterns, any case. Given set parameters, obtain prediction errors smaller 0.05 units (ca. 11 per cent) magnitude dependent increase speed 1,000 times compared initial modeling. The proposed procedure enables thus immediately compute probabilistic maps uncertainties estimates, by distributions ensemble solutions.