Aftershock Ground Motion Attenuation Relationships Based On Deep Learning

Ground motion attenuation relationship is an important research topic in seismic engineering,which provides theoretical support for earthquake zoning,seismic design of engineering structures,and seismic hazard analysis.The influence of aftershocks on the damage of structures cannot be ignored.A large amount of seismic damage data show that strong aftershocks can even cause the collapse of mainshock-damaged structures.However,there are few researches that focus on the attenuation relationship of aftershocks.In addition,the attenuation relationship obtained by the statistical regression analysis contains some limitations,such as the prediction values of several attenuation models in the same area are too different,and the attenuation model in one area cannot be applied to other regions.This paper hopes to give an aftershock ground motion attenuation relationship with universality and good fitting effect,which can reflect the objective law between ground motion parameters.Considering that deep learning can obtain any non-linear relationship of parameters through training a large amount of sample data,this paper first establishes a global aftershock database,and on this basis,a research on aftershock ground motion attenuation relationships based on deep neural networks is carried out.The main research contents of this article are as follows:(1)This paper obtain seismic and fault information from existing seismic database and finite fault model database.The CRJB-based criteria is used to discriminate the main shocks and aftershocks.According to the source type,earthquakes are divided into shallow crustal earthquake,subduction interface earthquake and subduction intraslab earthquake.Finally,303,019 aftershock ground motion records are selected from all over the world,and the global aftershock database is established,which is the basis of the research on aftershock ground motion attenuation relationship in this paper.(2)Based on Tensorflow framework,the deep neural networks are established by setting hyper-parameters such as initial value of weights and using methods such as Dropout.The prediction results of deep neural networks show that the goodness of fitting is up to 0.986 and the mean square error is less than 0.15.Within-event residual and between-event residual are evenly distributed,and the standard deviation of the residual is as low as 0.63,which reflects that the model proposed in this paper has a good prediction ability.(3)This paper compares the attenuation relationship of shallow crustal aftershocks proposed in this paper with the BSSA14 attenuation model given by Boore.Besides,the paper summarizes the attenuation relationships of subduction interface earthquake and subduction intraslab earthquake.The results show that the attenuation relationships of aftershocks based on the deep neural networks built in this paper fits well for differentmagnitudes and site conditions,and the prediction results are reasonable and reliable.