Broadband Ground Motion Synthesis by Diffusion Model with Minimal Condition

TL;DR

HEGGS reframes broadband seismic waveform synthesis as a conditional diffusion problem operating on spectrograms, trained on paired earthquake observations to maximize seismological realism with minimal conditioning. It introduces a Pair-Exploiting Diffusion Model that transfers morphology from $W^{src}$ to $W^{tgt}$ via a latent transform $\eta$ and a denoiser $\mathbf{x}_{\theta}$, coupled with an end-to-end autoencoder and Amplitude Correction Module. Across Europe, East Asia, and North America, HEGGS outperforms GAN- and LDM-based baselines in phase-arrival fidelity, envelope correlation, SNR/PSNR, and GMPE-consistent PGA distributions, and can synthesize 3-component ($E$-$N$-$Z$) waveforms at arbitrary stations or even fictitious events. By leveraging minimal metadata (e.g., $s_{lat},s_{lon},e_{lat},e_{lon},M_L$) and the intrinsic pair-structure of seismic data, the method offers practical utility for earthquake modeling, early warning, and hazard assessment while advancing AI-based seismology.

Abstract

Shock waves caused by earthquakes can be devastating. Generating realistic earthquake-caused ground motion waveforms help reducing losses in lives and properties, yet generative models for the task tend to generate subpar waveforms. We present High-fidelity Earthquake Groundmotion Generation System (HEGGS) and demonstrate its superior performance using earthquakes from North American, East Asian, and European regions. HEGGS exploits the intrinsic characteristics of earthquake dataset and learns the waveforms using an end-to-end differentiable generator containing conditional latent diffusion model and hi-fidelity waveform construction model. We show the learning efficiency of HEGGS by training it on a single GPU machine and validate its performance using earthquake databases from North America, East Asia, and Europe, using diverse criteria from waveform generation tasks and seismology. Once trained, HEGGS can generate three dimensional E-N-Z seismic waveforms with accurate P/S phase arrivals, envelope correlation, signal-to-noise ratio, GMPE analysis, frequency content analysis, and section plot analysis.

Figures (25)

• Figure 1: Left : Visualization of SCEDC data using paired waveforms. It shows that earthquake events can be detected at greater distances depending on magnitude. Right : Diagram of the waveform generative model architecture of HEGGS and its training loss.
• Figure 2: Result of GMPE analysis in PGA values with respect to the distance. The points in the figures represent the PGA values calculated from randomly selected waveforms from the test set containing earthquakes filtered by the magnitude range indicated by the black solid lines, and synthetic waveforms using the corresponding metadata. The subfigures correspond to the earthquake source: SCEDC (North America), KMA (East Asia), and INSTANCE (Europe).
• Figure 3: Comparison of 3-component real and synthetic waveforms from earthquake datasets SCEDC (North America), KMA (East Asia), and INSTANCE (Europe). For each panel, top shows a real waveform and the bottom shows a synthetic waveform generated with the same metadata. The phase arrivals marked as red (P) and blue (S) lines are detected by EQT.
• Figure 4: Comparison of real and synthetic spectrograms.
• Figure 5: (a)-(c): Frequency contents of synthetic waveform compared to the real waveform

Theorems & Definitions (1)

• Remark 10.1