17:15 〜 19:15
[SCG60-P09] Site-specific seismic waveform generation based on the integration of physics theory and deep learning
キーワード：地震波形生成、深層学習、物理損失、位置エンコーディング
Traditional methods for generating seismic data based on physical principles face significant limitations. In recent years, the use of AI-driven techniques for seismic data generation has emerged as a key area of research.
This study focuses on a specific seismic recording site. First, a multilayer neural network is employed to mimic the Ground Motion Prediction Equation (GMPE) and learn the acceleration distribution characteristics at the site. Subsequently, a Generative Adversarial Network (GAN) is utilized to learn the time-series characteristics of seismic data.
The generation process is conditioned on earthquake magnitude, the maximum acceleration at the recording site, and segmented phase features of seismic waveforms. To enhance the model’s ability to capture high-frequency features inherent in the original data, positional encoding(PE) is incorporated into the generator. A combination of time-series processing modules and traditional CNN layers allows the network to effectively capture both long-term variations and localized features of seismic data. Additionally, a physical loss function is introduced to ensure the generated data adheres to physical principles.
The model was trained for 2000 epochs using one NVIDIA A100-SXM4-80GB GPU, achieving strong performance on the training dataset. The model’s tests under given conditions shows that it can generate seismic waveforms similar to real samples. The physical loss remains consistent and PSD analysis also confirms the accurary of the output results. However, its generalization capability under entirely new conditions remains a challenge. This limitation may stem from inaccuracies in control condition labels or the sparsity of the training dataset.
This study focuses on a specific seismic recording site. First, a multilayer neural network is employed to mimic the Ground Motion Prediction Equation (GMPE) and learn the acceleration distribution characteristics at the site. Subsequently, a Generative Adversarial Network (GAN) is utilized to learn the time-series characteristics of seismic data.
The generation process is conditioned on earthquake magnitude, the maximum acceleration at the recording site, and segmented phase features of seismic waveforms. To enhance the model’s ability to capture high-frequency features inherent in the original data, positional encoding(PE) is incorporated into the generator. A combination of time-series processing modules and traditional CNN layers allows the network to effectively capture both long-term variations and localized features of seismic data. Additionally, a physical loss function is introduced to ensure the generated data adheres to physical principles.
The model was trained for 2000 epochs using one NVIDIA A100-SXM4-80GB GPU, achieving strong performance on the training dataset. The model’s tests under given conditions shows that it can generate seismic waveforms similar to real samples. The physical loss remains consistent and PSD analysis also confirms the accurary of the output results. However, its generalization capability under entirely new conditions remains a challenge. This limitation may stem from inaccuracies in control condition labels or the sparsity of the training dataset.