5:15 PM - 7:15 PM
[SCG60-P10] Preliminary study on earthquake ground motion time-history generation models using generative AI
Keywords:Machine Learning, Generative AI, Ground Motion, Time-history Waveform
1. Introduction
Conventional earthquake ground motion evaluation methods fall into two approaches: a simplified method using ground motion prediction equations and a detailed method employing fault models. The simplified method represents specific ground motion indices through empirical regression formulas based on representative parameters. While limited in scope, it offers advantages in simplicity and low computational cost. On the other hand, the detailed method allows for the evaluation of time-histories, preserving the inherent information of ground motion. However, it requires high expertise and significant computational resources, as it involves setting fault parameters, modeling subsurface structures, and selecting appropriate waveform synthesis techniques.
Recently, pioneering studies have explored applying neural networks, which have been successfully used in speech synthesis, to earthquake ground motion (e.g., Matsumoto et al., 2023; Yamaguchi et al., 2024; Shi et al., 2024). In these studies, similar parameters to those used in the simplified method, or uniform random numbers, are used as inputs, enabling the direct and efficient generation of artificial ground motion time-histories with similar characteristics to observed ground motions used as training data. By conditioning each ground motion on source and site-specific parameters, it is possible to construct a non-ergodic ground motion time-history generation model, which could lead to further advancements in strong ground motion prediction and seismic hazard assessment. However, limited discussion exists on the differences among time-history waveforms generated by various neural networks and the influence of feature parameters conditioned on the observed ground motions used as training data.
In this study, as a preliminary study on the application of generative AI to earthquake ground motion, we employ two existing neural networks, WaveGAN and DiffWave, to train models using the same dataset of observed ground motions as training data. We then compare the reproducibility of ground motions generated by each model. Additionally, we examine the impact of feature parameters used to train the observed ground motions on the generated ground motions.
2. Model and Training Dataset
We adopted WaveGAN (Donahue et al., 2019), originally proposed in the field of speech synthesis, and DiffWave (Kong et al., 2021), a model based on diffusion processes, as the foundational neural networks. Using these, we developed two types of earthquake ground motion time-history generation models.
The training ground motion data consists of 14,313 acceleration time-histories observed at the K-NET and KiK-net stations. These were recorded during 80 earthquakes with magnitudes ranging from Mw 5.0 to 7.8, which occurred in and around Japan. The acceleration time-histories were represented by a single component, obtained by rotating the two horizontal components 45 degrees. A band-pass filter with a 0.1–10 s period range was applied to all time-histories. For training, the time-histories were normalized so that the absolute maximum amplitude value was scaled to 1, and the length was set to 160 seconds.
The feature parameters used for conditioning each time-history included the following variables: moment magnitude, hypocentral distance, hypocentral depth, average S-wave velocities in the upper 10 m and 30 m layers, top depth of the 1,400 m/s S-wave velocity layer, and top depth of the seismic bedrock.
3. Evaluation for Generated Ground Motions
The reproducibility of the training ground motions was evaluated using the acceleration response spectrum, the response duration spectrum (Ishii, 2012), and the mean group delay time. The constructed models for the acceleration response spectrum and mean group delay time performed better in the short-period range but were less accurate in the long-period range. In contrast, the constructed models for the response duration spectrum showed consistently good performance in both the short- and long-period ranges.
For data with relatively short hypocentral distances, some ground motions generated by the DiffWave model exhibited a delayed arrival of wave packets compared to the observed time-history waveforms, while others showed an excessively amplification of long-period components. This suggests that the influence of feature parameters in the DiffWave model differs from that in the WaveGAN model, making it more susceptible to the effects of training ground motions of distant large earthquakes.
In this presentation, we will discuss the differences in the ground motions generated by each model and their implications.
Conventional earthquake ground motion evaluation methods fall into two approaches: a simplified method using ground motion prediction equations and a detailed method employing fault models. The simplified method represents specific ground motion indices through empirical regression formulas based on representative parameters. While limited in scope, it offers advantages in simplicity and low computational cost. On the other hand, the detailed method allows for the evaluation of time-histories, preserving the inherent information of ground motion. However, it requires high expertise and significant computational resources, as it involves setting fault parameters, modeling subsurface structures, and selecting appropriate waveform synthesis techniques.
Recently, pioneering studies have explored applying neural networks, which have been successfully used in speech synthesis, to earthquake ground motion (e.g., Matsumoto et al., 2023; Yamaguchi et al., 2024; Shi et al., 2024). In these studies, similar parameters to those used in the simplified method, or uniform random numbers, are used as inputs, enabling the direct and efficient generation of artificial ground motion time-histories with similar characteristics to observed ground motions used as training data. By conditioning each ground motion on source and site-specific parameters, it is possible to construct a non-ergodic ground motion time-history generation model, which could lead to further advancements in strong ground motion prediction and seismic hazard assessment. However, limited discussion exists on the differences among time-history waveforms generated by various neural networks and the influence of feature parameters conditioned on the observed ground motions used as training data.
In this study, as a preliminary study on the application of generative AI to earthquake ground motion, we employ two existing neural networks, WaveGAN and DiffWave, to train models using the same dataset of observed ground motions as training data. We then compare the reproducibility of ground motions generated by each model. Additionally, we examine the impact of feature parameters used to train the observed ground motions on the generated ground motions.
2. Model and Training Dataset
We adopted WaveGAN (Donahue et al., 2019), originally proposed in the field of speech synthesis, and DiffWave (Kong et al., 2021), a model based on diffusion processes, as the foundational neural networks. Using these, we developed two types of earthquake ground motion time-history generation models.
The training ground motion data consists of 14,313 acceleration time-histories observed at the K-NET and KiK-net stations. These were recorded during 80 earthquakes with magnitudes ranging from Mw 5.0 to 7.8, which occurred in and around Japan. The acceleration time-histories were represented by a single component, obtained by rotating the two horizontal components 45 degrees. A band-pass filter with a 0.1–10 s period range was applied to all time-histories. For training, the time-histories were normalized so that the absolute maximum amplitude value was scaled to 1, and the length was set to 160 seconds.
The feature parameters used for conditioning each time-history included the following variables: moment magnitude, hypocentral distance, hypocentral depth, average S-wave velocities in the upper 10 m and 30 m layers, top depth of the 1,400 m/s S-wave velocity layer, and top depth of the seismic bedrock.
3. Evaluation for Generated Ground Motions
The reproducibility of the training ground motions was evaluated using the acceleration response spectrum, the response duration spectrum (Ishii, 2012), and the mean group delay time. The constructed models for the acceleration response spectrum and mean group delay time performed better in the short-period range but were less accurate in the long-period range. In contrast, the constructed models for the response duration spectrum showed consistently good performance in both the short- and long-period ranges.
For data with relatively short hypocentral distances, some ground motions generated by the DiffWave model exhibited a delayed arrival of wave packets compared to the observed time-history waveforms, while others showed an excessively amplification of long-period components. This suggests that the influence of feature parameters in the DiffWave model differs from that in the WaveGAN model, making it more susceptible to the effects of training ground motions of distant large earthquakes.
In this presentation, we will discuss the differences in the ground motions generated by each model and their implications.