[S-CG60] Toward Discoveries in Solid Earth Sciences by Machine Learning

[SCG60-04] Neural Network-Based Ground Motion Model Learning Site Property from Data

*Tomohisa Okazaki¹, Tomoharu Iwata^1,2, Asako Iwaki³, Hiroyuki Fujiwara³, Naonori Ueda¹ (1.Center for Advanced Intelligence Project, RIKEN, 2.NTT Communication Science Laboratories, 3.National Research Institute for Earth Science and Disaster Resilience)

Keywords:ground motion model, neural network, site property

We constructed a neural network-based ground motion prediction model estimating spectral accelerations. Instead of specifying physical quantities as site properties, we input site ID and let the network learn the site property from strong motion data. We demonstrated that this model improves the prediction performance by applying it to KiK-net data in Tohoku region of Japan. Moreover, the obtained site property indicates that, for intermediate depth earthquakes, the short-period (< ~1 sec) spectrum is mainly governed by the distance from the volcanic front. This suggests that the proposed model successfully learned not only the ground condition under the individual sites but also propagation path effects known as anomalous seismic intensity in this region.