资源环境科技发展态势分析平台

We use a machine learning approach to build a ground motion model (GMM) from a synthetic database of ground motions extracted from the Southern California CyberShake study. An artificial neural network is used to find the optimal weights that best fit the target data (without overfitting), with input parameters chosen to match that of state-of-the-art GMMs. We validate our synthetic-based GMM with empirically based GMMs derived from the globally based Next Generation Attenuation West2 data set, finding near-zero median residuals and similar amplitude and trends (with period) of total variability. Additionally, we find that the artificial neural network GMM has similar bias and variability to empirical GMMs from records of the recent Mw7.1 Ridgecrest event, which neither GMM has included in its formulation. As simulations continue to better model broadband ground motions, machine learning provides a way to utilize the vast amount of synthetically generated data and guide future parameterization of GMMs.

Plain Language Summary A limited number of recorded earthquakes used to build models describe ground motion amplitude with variations of source mechanism, magnitude, and distance from the fault (termed ground motion models). These ground motion models help describe seismic hazard and ultimately influence the structural design of buildings, pipelines, and other forms of infrastructure in earthquake-prone regions. Instead of waiting for more earthquakes to occur, we can synthetically generate both past and hypothetical earthquakes to supplement the recorded database of ground motions. Here, we use a machine learning approach to model synthetically generated earthquake ground motion amplitudes across a wide range of frequencies. The ground motion database is compiled from a suite of forward simulations of possible fault ruptures focused on Southern California. We evaluate our model on observational data from around the world that have the same range of source, site, and distance characteristics to that of the simulated data set. Our results show that the simulated ground motions fit recorded data similarly to that of empirical models built from a global database of ground motions from various earthquakes. We apply the machine learned model to the recent 2019 Mw7.1 Ridgecrest event in Southern California (about which neither the synthetic nor empirical models have information) and find that it performs similarly to empirical models across the bandwidth range studied here. We highlight our method as one way to utilize the large amount of data that synthetic simulations can produce and propose that a machine learning approach using simulated data should be considered to aid in better constraining seismic hazard in regions where recorded data are lacking.