Loading...

Stochastic Ground Motion Modeling Using Deep Learning

Masoudifar, Mohsen | 2025

0 Viewed
  1. Type of Document: Ph.D. Dissertation
  2. Language: Farsi
  3. Document No: 58126 (09)
  4. University: Sharif University of Technology
  5. Department: Civil Engineering
  6. Advisor(s): Mahsuli, Mojtaba
  7. Abstract:
  8. This dissertation proposes a probabilistic framework for simulating three-dimensional ground motion time series using deep learning based on Generative Adversarial Networks (GANs) and Convolutional Neural Networks (CNNs). The proposed framework for generating synthetic ground motion consists of two distinct deep learning-based modules: signal generation and amplitude prediction. The first module generates normalized three-dimensional ground motion time series given source and site characteristics. For this purpose, a conditional Wasserstein GAN (WGAN) is designed. The second module predicts the peak ground acceleration (PGA) for the three spatial components of the generated normalized ground motion, given the normalized motion and the source and site characteristics. For this purpose, a CNN is designed with “inception” layers. The learning performance of both modules is improved by realistic data augmentation techniques designed explicitly for three-dimensional ground motions. In this study, two training approaches are employed to train the proposed framework based on existing dataset. The first approach, “learning from scratch”, follows the aforementioned methodology and is employed when the module is initially trained on sufficient data. The second approach is used when a pre-trained module undergoes “transfer learning” with limited data to adapt to a new region. In the first training approach, the proposed framework is trained using the dataset of over 200,000 records of the KiK-net database. For the second approach, the framework is fine-tuned using 15,000 records registered in California as part of the NGA-West2 dataset to develop a region-specific model. The site and source characteristics utilized in the application of the study comprise the moment magnitude, distance, fault mechanism, and shear wave velocity. The validation of the proposed framework is assessed using both classical and innovative approaches. The signal generation module is validated by comparing the diversity of the generated signals with the diversity of real ground motions in the time and frequency domains. Classical metrics, such as the correlation coefficient between real and predicted PGA, are used to evaluate the accuracy of the amplitude prediction module. This correlation coefficient is approximately 0.97 for both training approaches for the test data, which demonstrates a satisfactory prediction quality and the absence of overfit. Moreover, the performance of the developed framework is qualitatively assessed by comparing the ground motion responses in the time and frequency domains for real and synthetic data. Finally, this study introduces three novel procedures for quantitatively validating the synthetic ground motion generation model. The first procedure quantitatively compares the distribution of intensity measures between synthetic and real ground motions using Jensen-Shannon divergence (JSD). The average JSD is 0.18 for the first training approach and 0.25 for the second, demonstrating substantial similarity between synthetic and real distributions. In the second procedure, a CNN-based classifier is trained to quantify the quality of the generated signals. This procedure shows that about 96% of the synthetic motions are classified as high quality. In the third procedure, another CNN-based model is trained to estimate the source and site characteristics by receiving normalized three-dimensional ground motions and their PGAs. This procedure estimates the GAN quality index at about 78%, confirming the adequacy of synthetic ground motions. All validation procedures confirm that the developed framework generates synthetic ground motions with sufficient accuracy in both training approaches
  9. Keywords:
  10. Earthquake ; Machine Learning ; Deep Learning ; Generative Adversarial Networks ; Transfer Learning ; Signal Analysis ; Seismic Site Effects ; Stochastic Ground Motion ; Ground Motion Modeling ; Fault Mechanism

 Digital Object List

 Bookmark

No TOC