Deep neural network–powered ground‐motion models for acceleration spectrum intensity, Housner’s spectrum intensity, and displacement spectrum intensity
Mao‐Xin Wang, Gang Wang
Abstract
This study develops new ground‐motion models (GMMs) for acceleration spectrum intensity ( ASI ), Housner’s spectrum intensity ( SI ), and displacement spectrum intensity ( DSI ) using the NGA‐West2 database. A deep neural network (DNN)‐based mixed‐effects regression approach is presented for the model development accounting for the between‐event and within‐event variability. The performance evaluation results show the generally unbiased and reliable predictions with R‐squared values exceeding 0.9 on both training and testing data. A hybrid‐scenario method is subsequently employed to conduct the semi‐blind test on large datasets that are not directly used in model development, illustrating favorable model generalization capability. Meanwhile, the DNN models exhibit physically reasonable magnitude and distance scaling characteristics, producing comparable median trends with the existing NGA‐West2 GMM‐embedded indirect approach. Yet, the major advantage of the new models is their ability to achieve lower statistical dispersion using fewer input parameters (i.e. (moment magnitude M w , rupture distance R rup , upper 30‐m shear‐wave velocity V S 30 , and focal mechanism indicator) or ( M w , R rup , V S 30 , focal mechanism indicator, depth to top of rupture Z tor , and depth to 1 km/s shear‐wave velocity, Z 1 )). Incorporating Z tor and Z 1 into the DNN models yields 10%–20% reduction in mean squared error, while the benefit of adding predictor variables is less pronounced for the functional form‐based classical regression models. The new models are applicable to shallow crustal earthquakes with 3 ≤ M w ≤ 7.9, 0 ≤ R rup ≤ 400 km, and 150 ≤ V S 30 ≤ 1500 m/s.