4:15 PM - 4:30 PM
[SCG55-13] A forecast of long-period ground motions by using a convolutional neural network
Keywords:long-period ground motion, machine learning, real-time prediction
1. Introduction
We applied a convolutional neural network (CNN) for forecasting long-period (LP; 2-10 s) ground motions that occur in large basins during large earthquakes. LP ground motions with long wavelengths travel hundreds of kilometers and resonate with high-rise buildings, so their early forecast is important for disaster prevention. During the 2004 Niigata Chuetsu Earthquake (Mw6.5), LP ground motion occurred in the Kanto basin, 200 km away, damaging elevator cable in a high-rise building. The Nankai Trough earthquake of M8-9 range is feared to cause even greater damage. To address these issues, we forecast the velocity response spectrum of LP ground motions at a point in the Kanto basin using a CNN with observed waveform records near the epicenter as input.
2. Data and Method
In this study, we predicted the velocity response spectrum of the Hi-net YFTH station (Yokohama) in the Kanto basin for large earthquakes occurred in central Japan from the velocity waveform data recorded at five stations 150 km far from YFTH.
We used Hi-net two horizontal component seismic velocity waveform with a data length of 5 minutes for 56 earthquakes of Mw>4.5 that occurred in central Japan (Figure 1). The velocity response spectrum with period 0.1 to 20 seconds at YFTH were obtained and paired with the input waveform data to prepare training datasets. In addition, we prepared training data for 45 hypothetical earthquakes (Figure 1, marked with x) using 3-D FDM simulations with OpenSWPV (Maeda et al., 2017). The JIVSM (Koketsu et al., 2012) was used as the subsurface model. In order to learn efficiently, the input waveform data were normalized by their maximum value (PGV) and multiplied by the logarithm of PGV to retain their information of amplitude. The base-10 logarithm was applied to the output response spectrum and then divided by 10 to keep their value between 0 and 1.
The CNN model consists of seven convolutional layers with ReLU activation functions, which are flattened and finally fed to a fully connected layer with tanh activation functions. The loss was evaluated by the squared error between the observed and predicted response spectrum. Using the training datasets, the CNN was trained for 50 epochs.
3. Result
Using this trained CNN model, we checked the forecast performance of LP seismic motions for 11 earthquakes of Mw>4.5 occurred after 2011. The performance of the model was evaluated using the average of the forecast / observed ratio of the velocity response spectrum for periods 1-10 s (PF value). First, in the study using natural earthquakes, good prediction results (PF : 1.0) were obtained for the Noto Peninsula earthquake in 2021 (Mw5.1), however, for the 2019 Yamagata-oki earthquake (Mw 6.7), the PF value was poor at 0.4, especially underestimated on the LP side (>2 s). In contrast, the CNN model, which was additionally trained with hypothetical earthquakes using FDM simulations, showed improved prediction performanc, especially for the Yamagata-oki earthquake at LP (>2 s). Examining the quartile distribution of PF values for the 11 earthquakes that were predicted, a large effect (PF : 0.5–5.5) was observed from the additional learning of synthetic seismograms in the FDM simulations (Figure 3), compared to the learning results for natural earthquakes alone (PF : 0.3–7)
4. Conclusion
The prediction performance of machine learning models generally depends on the quality of the training data sets. However, when training data sets are prepared using natural earthquakes, it is difficult to construct a versatile prediction model for a variety of future earthquakes due to the bias in seismic activity and scale. It is effective to add synthetic waveforms for diverse earthquakes from seismic wave propagation simulations to the training. However, there may be a limitation of the uncertainty of the subsurface model used for seismic wave propagation calculations, and a combination of natural earthquakes and simulations may be effective.
We applied a convolutional neural network (CNN) for forecasting long-period (LP; 2-10 s) ground motions that occur in large basins during large earthquakes. LP ground motions with long wavelengths travel hundreds of kilometers and resonate with high-rise buildings, so their early forecast is important for disaster prevention. During the 2004 Niigata Chuetsu Earthquake (Mw6.5), LP ground motion occurred in the Kanto basin, 200 km away, damaging elevator cable in a high-rise building. The Nankai Trough earthquake of M8-9 range is feared to cause even greater damage. To address these issues, we forecast the velocity response spectrum of LP ground motions at a point in the Kanto basin using a CNN with observed waveform records near the epicenter as input.
2. Data and Method
In this study, we predicted the velocity response spectrum of the Hi-net YFTH station (Yokohama) in the Kanto basin for large earthquakes occurred in central Japan from the velocity waveform data recorded at five stations 150 km far from YFTH.
We used Hi-net two horizontal component seismic velocity waveform with a data length of 5 minutes for 56 earthquakes of Mw>4.5 that occurred in central Japan (Figure 1). The velocity response spectrum with period 0.1 to 20 seconds at YFTH were obtained and paired with the input waveform data to prepare training datasets. In addition, we prepared training data for 45 hypothetical earthquakes (Figure 1, marked with x) using 3-D FDM simulations with OpenSWPV (Maeda et al., 2017). The JIVSM (Koketsu et al., 2012) was used as the subsurface model. In order to learn efficiently, the input waveform data were normalized by their maximum value (PGV) and multiplied by the logarithm of PGV to retain their information of amplitude. The base-10 logarithm was applied to the output response spectrum and then divided by 10 to keep their value between 0 and 1.
The CNN model consists of seven convolutional layers with ReLU activation functions, which are flattened and finally fed to a fully connected layer with tanh activation functions. The loss was evaluated by the squared error between the observed and predicted response spectrum. Using the training datasets, the CNN was trained for 50 epochs.
3. Result
Using this trained CNN model, we checked the forecast performance of LP seismic motions for 11 earthquakes of Mw>4.5 occurred after 2011. The performance of the model was evaluated using the average of the forecast / observed ratio of the velocity response spectrum for periods 1-10 s (PF value). First, in the study using natural earthquakes, good prediction results (PF : 1.0) were obtained for the Noto Peninsula earthquake in 2021 (Mw5.1), however, for the 2019 Yamagata-oki earthquake (Mw 6.7), the PF value was poor at 0.4, especially underestimated on the LP side (>2 s). In contrast, the CNN model, which was additionally trained with hypothetical earthquakes using FDM simulations, showed improved prediction performanc, especially for the Yamagata-oki earthquake at LP (>2 s). Examining the quartile distribution of PF values for the 11 earthquakes that were predicted, a large effect (PF : 0.5–5.5) was observed from the additional learning of synthetic seismograms in the FDM simulations (Figure 3), compared to the learning results for natural earthquakes alone (PF : 0.3–7)
4. Conclusion
The prediction performance of machine learning models generally depends on the quality of the training data sets. However, when training data sets are prepared using natural earthquakes, it is difficult to construct a versatile prediction model for a variety of future earthquakes due to the bias in seismic activity and scale. It is effective to add synthetic waveforms for diverse earthquakes from seismic wave propagation simulations to the training. However, there may be a limitation of the uncertainty of the subsurface model used for seismic wave propagation calculations, and a combination of natural earthquakes and simulations may be effective.