Research map / Video / Video generation
Broadband Ground Motion Synthesis by Diffusion Model with Minimal Condition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HEGGS, a conditional latent diffusion model with pair-exploiting training and an end-to-end amplitude-corrected waveform decoder that synthesizes broadband 3-component earthquake ground motion from minimal station/source metadata.
Problem
Generating seismologically realistic broadband earthquake ground motion waveforms, which GAN-based generators often fail to get right in phase arrival times and amplitude, given only minimal conditioning metadata.
Approach
Exploit the pairing of multiple station recordings of the same earthquake: a diffusion model conditioned on a source-station waveform's latent and a target station's metadata predicts the target spectrogram latent, trained end-to-end with a VAE decoder plus an amplitude correction module (ACM).
Key results
- On SCEDC with source waveform input, HEGGS reaches P MAE 0.4760 s and S MAE 0.5476 s versus 0.5633 s and 0.7808 s for the best LDM baseline, with envelope correlation 0.8187 vs 0.7726.
- On KMA, HEGGS (w/ source waveform) achieves envelope correlation 0.8785 and P MAE 0.2763 s versus LDM's 0.8076 and 1.3521 s.
- Baselines SeismoGen, ConSeisGen and BBGAN produce P/S MAE of several seconds on SCEDC; BBGAN diverged when given source waveforms.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HEGGS instead of LDM | SCEDC · SCEDC test | P_MAE (s): 0.5633 -> 0.4760 | Helped |
| Use HEGGS instead of LDM | SCEDC · SCEDC test | S_MAE (s): 0.7808 -> 0.5476 | Helped |
| Use HEGGS instead of LDM | SCEDC · SCEDC test | env.corr: 0.7726 -> 0.8187 | Helped |
| Use HEGGS instead of LDM | KMA · KMA test | P_MAE (s): 1.3521 -> 0.2763 | Helped |
| Use HEGGS instead of LDM | KMA · KMA test | S_MAE (s): 1.6845 -> 0.4644 | Helped |
| Use HEGGS instead of LDM | KMA · KMA test | env.corr: 0.8076 -> 0.8785 | Helped |
| Use HEGGS instead of LDM | INSTANCE · INSTANCE test | P_MAE (s): 0.8187 -> 0.5085 | Helped |
| Use HEGGS instead of LDM | INSTANCE · INSTANCE test | S_MAE (s): 0.7875 -> 0.6378 | Helped |
| Use HEGGS instead of LDM | INSTANCE · INSTANCE test | env.corr: 0.7898 -> 0.8301 | Helped |
| Add paired-data training (ablation) | LDM on SCEDC · SCEDC test | P_MAE (s): 1.1142 -> 0.5633 | Helped |
| Add paired-data training (ablation) | LDM on SCEDC · SCEDC test | env.corr: 0.6932 -> 0.7726 | Helped |
| Add ACM to end-to-end training (ablation) | HEGGS on SCEDC · SCEDC test | env.corr: 0.6239 -> 0.8187 | Helped |
Surprising findings
- Conventional LDM training on the seismic spectrogram data diverged, forcing normalized-waveform relaxations in the ablation.
- HEGGS without the source waveform still outperforms the benchmark models on most metrics, despite being trained with paired data.
Limitations
- Generation of fictitious earthquakes is limited because only minimal location and magnitude conditions are used; extrapolation to large magnitudes is unverified.
- Models are regional and trained per dataset; a global model is not addressed.
- Synthesized spectrograms show slightly lower resolution than real spectrograms.
Experimental setup
Datasets: SCEDC KMA INSTANCE
Benchmarks: SCEDC KMA INSTANCE
Metrics: P MAE S MAE P/S F1 envelope correlation SNR PSNR spectrogram MSE PGA-distance GMPE analysis corner frequency and M0 analysis
Baselines: SeismoGen ConSeisGen BBGAN LDM
Builds on: latent diffusion model Make-An-Audio / TANGO-style conditional spectrogram generation MP-SeNet EQTransformer SeisBench
Closest papers by meaning
- A Novel Diffusion Model for Pairwise Geoscience Data Generation with Unbalanced Training Dataset (AAAI 2025 · similarity 0.85)
- Synergizing Multigrid Algorithms with Vision Transformer: A Novel Approach to Enhance the Seismic Foundation Model (AAAI 2026 · similarity 0.85)
- A Unified Framework for Forward and Inverse Problems in Subsurface Imaging using Latent Space Translations (ICLR 2025 · similarity 0.84)
- Generating Full-field Evolution of Physical Dynamics from Irregular Sparse Observations (NeurIPS 2025 · similarity 0.81)
- CALM-PDE: Continuous and Adaptive Convolutions for Latent Space Modeling of Time-dependent PDEs (NeurIPS 2025 · similarity 0.80)
- Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs (CVPR 2025 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.