12:00 〜 12:15
[SSS11-17] Generative adversarial network for ground motion augmentation compatible with given spectral acceleration and durations
キーワード：Generative Adversarial Network, Deep learning, Spectral acceleration, Significant duration, Ground motion augmentation
1. Introduction
The impact of ground motion duration on liquefaction and slope stability is widely recognized. Even under the same intensity conditions, variations in duration can influence structural collapse risk (Raghunandan & Liel, 2013). Modern infrastructure growth intensifies this issue by increasing vulnerabilities, leading to greater variability in earthquake-induced damage.
Generating large volumes of ground motion time history data with adjustable parameters provides a way to overcome the limitations of existing recordings and address associated uncertainties. However, integrating detailed duration segments and ground motion intensities as variables remains an area that requires further study. Accordingly, this study employs a Generative Adversarial Network (GAN)-based approach (Goodfellow et al., 2014) to generate multiple accelerograms compatible with significant durations and spectral acceleration (SA) of interest. This method has gained considerable attention in generative modeling and has recently been applied to earthquake waveform synthesis (e.g., Kim & Kim, 2024; Matinfar et al., 2023). This approach is expected to offer greater flexibility for exploration and support extensive evaluations across various scenarios, thereby advancing earthquake risk assessment.
2. Data and results
The acceleration time histories along the east-west and north-south directions were obtained from the seismograph network of Japan, Kyoshin Network, i.e., K-NET (National Research Institute for Earth Science and Disaster Resilience, 2019). We designed the GAN-based generative model to produce multiple accelerograms from a predefined design spectrum and significant durations corresponding to 5–35%, 35–65%, and 65–95% intervals (i.e., D5–35, D35–65, and D65–95). The ability of the generator to produce accelerograms that align with the specified input conditions was evaluated.
The figure illustrates (a) three generated samples (SampleG), (b) comparisons of SAs, where SamplesG refers to the SAs from 40 generated samples, MeanG represents the mean SA of SamplesG, and BestG denotes the SA with the minimum Root Mean Squared Error (RMSE) compared to the design spectrum. Lastly, (c) presents comparisons of three significant duration components, including the input value, the BestG value, and the value range of SamplesG. A visual inspection shows that the generated samples align closely with the conditional information.
Acknowledgments
This research was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (RS-2024-00412801).
References
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., & Bengio, Y. (2014). Generative adversarial nets. Advances in neural information processing systems, 27, 2672-2680.
Kim, J., & Kim, B. (2024). Generative adversarial network to produce numerous artificial accelerograms with pseudo-spectral acceleration as conditional input. Computers and Geotechnics, 174.
Matinfar, M., Khaji, N., & Ahmadi, G. (2023). Deep convolutional generative adversarial networks for the generation of numerous artificial spectrum-compatible earthquake accelerograms using a limited number of ground motion records. Computer-Aided Civil and Infrastructure Engineering, 38(2), 225-240.
National Research Institute for Earth Science and Disaster Resilience. (2019). NIED K-NET, KiK-net, National Research Institute for Earth Science and Disaster Resilience.
Raghunandan, M., & Liel, A. B. (2013). Effect of ground motion duration on earthquake-induced structural collapse. Structural Safety, 41, 119-133.
The impact of ground motion duration on liquefaction and slope stability is widely recognized. Even under the same intensity conditions, variations in duration can influence structural collapse risk (Raghunandan & Liel, 2013). Modern infrastructure growth intensifies this issue by increasing vulnerabilities, leading to greater variability in earthquake-induced damage.
Generating large volumes of ground motion time history data with adjustable parameters provides a way to overcome the limitations of existing recordings and address associated uncertainties. However, integrating detailed duration segments and ground motion intensities as variables remains an area that requires further study. Accordingly, this study employs a Generative Adversarial Network (GAN)-based approach (Goodfellow et al., 2014) to generate multiple accelerograms compatible with significant durations and spectral acceleration (SA) of interest. This method has gained considerable attention in generative modeling and has recently been applied to earthquake waveform synthesis (e.g., Kim & Kim, 2024; Matinfar et al., 2023). This approach is expected to offer greater flexibility for exploration and support extensive evaluations across various scenarios, thereby advancing earthquake risk assessment.
2. Data and results
The acceleration time histories along the east-west and north-south directions were obtained from the seismograph network of Japan, Kyoshin Network, i.e., K-NET (National Research Institute for Earth Science and Disaster Resilience, 2019). We designed the GAN-based generative model to produce multiple accelerograms from a predefined design spectrum and significant durations corresponding to 5–35%, 35–65%, and 65–95% intervals (i.e., D5–35, D35–65, and D65–95). The ability of the generator to produce accelerograms that align with the specified input conditions was evaluated.
The figure illustrates (a) three generated samples (SampleG), (b) comparisons of SAs, where SamplesG refers to the SAs from 40 generated samples, MeanG represents the mean SA of SamplesG, and BestG denotes the SA with the minimum Root Mean Squared Error (RMSE) compared to the design spectrum. Lastly, (c) presents comparisons of three significant duration components, including the input value, the BestG value, and the value range of SamplesG. A visual inspection shows that the generated samples align closely with the conditional information.
Acknowledgments
This research was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (RS-2024-00412801).
References
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., & Bengio, Y. (2014). Generative adversarial nets. Advances in neural information processing systems, 27, 2672-2680.
Kim, J., & Kim, B. (2024). Generative adversarial network to produce numerous artificial accelerograms with pseudo-spectral acceleration as conditional input. Computers and Geotechnics, 174.
Matinfar, M., Khaji, N., & Ahmadi, G. (2023). Deep convolutional generative adversarial networks for the generation of numerous artificial spectrum-compatible earthquake accelerograms using a limited number of ground motion records. Computer-Aided Civil and Infrastructure Engineering, 38(2), 225-240.
National Research Institute for Earth Science and Disaster Resilience. (2019). NIED K-NET, KiK-net, National Research Institute for Earth Science and Disaster Resilience.
Raghunandan, M., & Liel, A. B. (2013). Effect of ground motion duration on earthquake-induced structural collapse. Structural Safety, 41, 119-133.