Deep Learning Prediction of Long-Period Ground Motion and Building Shaking in Nankai Trough Earthquake

In this study, we develop a deep-learning-based real-time prediction framework for ground motions and building responses from large earthquakes along the Nankai Trough, with focus on long-period (LP) ground motions (periods of 2–10 s). LP ground motions pose a serious hazard to modern society because they can shake strong high-rise buildings and the resulting damage. This study aim to rapidly predict ground motions and shaking in high-rise buildings at distant sites using near-source waveform observations.

Most machine-learning approaches to predicting LP ground motions have targeted intensity measures such as peak ground acceleration and response spectra. In contrast, Furumura and Oishi (2023) demonstrated that a time-series forecasting approach using a Temporal Convolutional Network (TCN) can predict LP ground-motion waveforms in distant plains in real time from near-source observations for shallow earthquakes off Tohoku. Building on this concept, this study aim to extend waveform-based prediction to Nankai Trough earthquakes and to predict shaking of high-rise buildings in distant plains.

The novelty of this study lies in two aspects. First, to accommodate the diversity of earthquakes along the Nankai Trough, including both shallow and deep events, we construct a TCN-Transformer model which improve arrival-time and waveform prediction across events with different apparent velocities. Second, for floor-by-floor shaking prediction from ground motions, we newly develop a TCN-PINN (Physics-Informed Neural Network) model that incorporates physics-based constraints derived from the equation of motions for governing building oscillation into loss function. This enabling physically plausible response predictions even with limited training data. Ground motions are first predicted using the TCN-Transformer model, and the results are then used as inputs to TCN-PINN models for each floor of the target building.

For the TCN-Transformer model, we trained and validated on 20 earthquakes of M5.0 or larger that occurred off the Kii Peninsula between 2008 and 2025. Waveforms recorded at the near-source F-net Fujigawa station (FUJF) were used to predict LP ground-motion waveforms at the MeSO-net Ginza station (GNZM) in the Kanto Plain, approximately 130 km away.

Next, we constructed a TCN-PINN model to predict building responses at three locations: B4F, the 13th floor, and the 21st floor, of the Central Government Building No. 2 (CG2), located 1.6 km from GNZM. This model was trained and validated using 25 earthquakes of M6 or larger that occurred off Tohoku between 2010 and 2014.

As a result, for the M6.5 earthquake on 1 April 2016 in the Nankai Trough region, the predictions successfully reproduced the arrival time, duration, and waveform envelope well. However, the overall waveform energy tended to be underestimated, indicating that future improvements is required for damage-assessment applications. For the CG2 responses, the model generally reproducing spectral-peak locations, such as predominant periods, whereas the response amplitudes were underestimated. To improve the prediction accuracy, future work will focus on refining the loss-function design to mitigate underestimation of the LP components, and will strengthen training by combining observed with synthetic scenario waveforms generated from seismic wave-propagation simulations.