TimesNet-Gen: Deep Learning for Site-Specific Ground Motion Generation
A novel deep generative model, named TimesNet-Gen, has been developed for generating strong ground motion tailored to specific sites. This framework employs a station-restricted, Dirichlet-based latent space resampling method, enabling direct site-specific generation without the need for explicit conditioning inputs or dimensionality reduction. Trained on the AFAD dataset through self-supervised learning, TimesNet-Gen exhibits strong cross-regional generalization by producing station-specific NGA-West2 records without requiring fine-tuning. Its effectiveness is assessed by analyzing distributions in log-HVSR space and conducting a joint evaluation of peak ground acceleration.
Key facts
- TimesNet-Gen is a deep generative framework for strong ground motion generation.
- It uses a station-restricted, Dirichlet-based latent space resampling strategy.
- No explicit conditioning inputs or dimensionality reduction are required.
- Pre-trained on the AFAD dataset via self-supervised learning.
- Generates station-specific NGA-West2 records without fine-tuning.
- Evaluated using log-HVSR space distributions and peak ground acceleration analysis.
- Demonstrates robust cross-regional generalization.
- Addresses earthquake risk reduction through site-specific evaluations.
Entities
—