TimesNet-Gen uses deep learning for site-specific earthquake motion generation

Researchers have developed TimesNet-Gen, a deep learning framework for generating site-specific strong ground motion from earthquake accelerometer records. This model uses a novel station-restricted latent space resampling strategy, eliminating the need for explicit conditioning inputs. Pre-trained on the AFAD dataset using self-supervised learning, TimesNet-Gen demonstrates strong cross-regional generalization capabilities without fine-tuning, as validated by comparisons in log-HVSR space and joint analysis of peak ground acceleration and fundamental site frequency. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Enhances seismic risk assessment accuracy by improving simulation of local site effects on ground motion.

RANK_REASON This is a research paper detailing a new deep learning framework for earthquake simulation.

Read on arXiv cs.LG →

• TimesNet-Gen
• AFAD
• NGA-West2
• arXiv

• paper
• other