M.S. Candidate: Bevan Deniz Çılğın
Program: Data Informatics
Date: 03.11.2025 / 10:00
Place: A-212

Abstract: Modeling the strong motion data is crucial for seismic hazard assessment. In our work, we aim to unveil the nature of strong motion data and soil characteristics using generative AI. We created a generative model that efficiently extracts underlying physical patterns from historical events with site-specific properties. To achieve this, we trained a Conditional Convolutional Variational Autoencoder model on amplitude and phase spectrograms of earthquake waveforms to generate strong-motion data that accurately represents the soil characteristics. We performed training in two steps: unsupervised learning on 31,249 records and fine-tuning (after minor architectural changes to create a conditional model) on 348 records for five stations. Ultimately, we compared the relevance of the generated signals by calculating the fundamental site frequency distributions of both the generated and original samples, linking the simulation results to a physical property of their sites.

The proposed method employs a data-driven approach using recorded time waveforms and relies solely on station identifiers for conditional generation, differing from existing methods in the literature. Moreover, it uses latent space conditioning techniques to make the model class-aware without requiring a large amount of station-specific data. Unlike traditional methods, it doesn't need theoretical assumptions, extensive parameter tuning, or computational power. We also introduce an evaluation framework to numerically analyze the correctness of the generated waveforms from simulations by comparing fundamental site frequencies calculated from HVSR curves, which are calculated using earthquake recordings. Our model achieves an alignment score of 0.84 with the ideal situation. Generated waveforms accurately represent their fundamental site frequency characteristics and differentiate from the frequency characteristics of other sites.