[S-SS05] Natural hazards and uncertainty: Informing societal decisions

[SSS05-P05] Ground motion model using Deep Neural Network in Taiwan

*Jia-Cian Gao¹, Chung-Han Chan^1,2, Kuo-Fong Ma^1,3 (1.Earthquake-Disaster and Risk E valuation and Management (E-DREaM) Center, National Central University, Taiwan, 2.Department of Earth Sciences, National Central University, Taiwan, 3.Institute of Earth Sciences, Academia Sinica, Taiwan)

Keywords:Ground Motion Model, Deep Neural Network

The traditional ground-motion model (GMM) select a functional form and determine it based on nonlinear regression. The forms from this method, however, sometimes cannot fully explain all natural effects due to human limitations. To better approximate complex nonlinear ground motion behaviors, this study aims to establish a GMM using a deep neural network (DNN). In order to train our DNN model, we implemented parameters including moment magnitude, rupture distance, and S-wave velocity down to the depth of 30 meters (V_s³⁰) obtained by the Taiwan Strong Motion Instrument Program database from 1991 to 2014. Using magnitude-distance cutoffs, all analyses are ensured to be well within the reliable range. Our approach first adjusts the model’s hyperparameters to determine the DNN's setpoints. After testing, we set learning rate of 0.001, epoch of 6000 times, batch size to a size that covers all data sets, and the network with 5 layers and 90 neurons for each layer. Comparing to multiple neural networks, a single neural network is more stable and has a period dependence in predicting multiple ground motion parameters. Oversampling is implemented to solve the flaws caused by imbalanced data in the model. Lastly, ensemble learning is performed under multiple independent trainings to improve the final model's stability and performance. As a result of the training, A DNN model fits nonlinear ground motion models well, and the model converges effectively, that is, a coefficient of determination between 0.87 and 0.92, and a standard deviation between 0.62 and 0.70. In general, Our GMM is stable and reasonable, with the exception of the interval with less data, and it is able to predict the ground motions in the forms of PGA, PGV, and full response spectrum (0.01 to 10 seconds), as well as to measure the credibility of the validation indexes.