An Early Forecast of Long‐Period Ground Motions of Large Earthquakes Based on Deep Learning

Abstract

Abstract Long‐period (LP; approximately 2–10 s) ground motions generated by large earthquakes are amplified in large basins and threaten high‐rise buildings in modern cities. In this study, we accomplished an early forecast of LP ground motions in distant basins based on deep learning technology using waveforms observed near the epicenter. A Temporal Convolutional Network was first trained using waveform data from past large earthquakes in the Japan Trench. LP ground motions of recent large earthquakes, including the 2011 Off the Pacific coast of Tohoku earthquake (Mw 9.0), were forecasted in the Kanto (Tokyo) and Osaka basins. This study effectively forecasted LP ground motions of large earthquakes regarding amplitude, waveform envelope shape, spectral characteristics, and duration. Faster forecasts (in 0.05 s or less) allow for updating forecasts as data is acquired, improving forecast accuracies and ensuring 1–2 min of lead time before large and prolonged shakes occur.

Keywords

• deep learning
• temporal convolutional network
• long‐period ground motion
• early forecast
• earthquake