Efficient regional seismic risk assessment via deep generative learning of surrogate models

Efficient regional seismic risk assessment including ground motion prediction and damage risk estimation is needed for emergency response planning. However, a conventional regional assessment suffers from low data- and time- efficiency as it generally involves a large number of locations and infrastructure systems that have specific soil conditions, and geometric, material, and structural properties, requiring access to large data and massive individual calculations with complicated procedures. To achieve efficient regional seismic risk assessment, this work presents a deep generative learning framework to construct input-output surrogate models of regional seismic risk by learning the underlying complex relation between earthquake source parameters and regional seismic risk involving many locations and structures from data. The learned deep surrogate models directly output the ground motion intensity map and the risk map of a region given earthquake source parameters, circumventing massive individual calculations and data access to individual locations and structures. The presented framework is validated on the bridge network risk assessment using simulated scenario earthquakes of the San Francisco Bay Area. We observe that the obtained deep surrogate models perform well without the need of data access to locations and structures and are time-efficient. We also discuss the applicability and limitations of the presented framework.