Learning earthquake ground motions via conditional generative modeling

Pu Ren, Rie Nakata, Maxime Lacour, Ilan Naiman, Nori Nakata, Jialin Song, Zhengfa Bi, Osman Asif Malik, Dmitriy Morozov, Omri Azencot, N. Benjamin Erichson, Michael W. MahoneyNature Communications 2026 · DOI

TL;DR

This paper introduces CGM-GM, a novel AI model that learns to generate realistic earthquake ground motion waveforms by conditioning on earthquake magnitude and location. It overcomes limitations of traditional methods by capturing complex spatial variations and physical characteristics without explicit physics constraints, showing promise for improved seismic hazard assessment.

What Problem Does This Paper Address?

Accurately predicting earthquake ground motions is vital for seismic hazard assessment and infrastructure resilience. Current methods face significant challenges:

  • Empirical simulations: Suffer from sparse sensor networks and limited data for specific earthquake locations.
  • Physics-based simulations: Are computationally very expensive, especially for high-frequency motions, and rely on accurate, often uncertain, models of Earth structures and earthquake sources.

These limitations hinder our ability to generate comprehensive ground motion data across wide geographic areas, especially for scenarios with limited historical records. The paper aims to bridge this gap by developing a data-driven approach that can generate realistic, spatially varying ground motions for various earthquake scenarios.

Methods

The core of the proposed method is a Conditional Generative Modeling for Ground Motion (CGM-GM) framework, built upon a dynamic Variational Autoencoder (VAE). Here's a breakdown of the key techniques:

  • Time-Frequency Decomposition: The ground motion time series are first transformed into the time-frequency domain using the Short-Time Fourier Transform (STFT) to extract amplitude and phase information.
  • Dynamic VAE for Amplitude: A dynamic VAE is employed to learn the amplitude spectrograms. This architecture is designed for sequential data, using recurrent neural networks (RNNs) to capture temporal dynamics within the time-frequency representation.
  • Conditional Embedding: Earthquake magnitudes, depths, and crucially, the geospatial coordinates of earthquake sources and sensors are used as conditional variables. These are embedded into the VAE model using Multilayer Perceptron (MLP) layers. This allows the model to learn spatial heterogeneity and physical interactions (like path and site effects) implicitly from the data.
  • Phase Information: During training, true phase information is used for waveform reconstruction. During generation, phase is estimated using phase retrieval methods (specifically, the Griffin-Lim algorithm).
  • Waveform Reconstruction: The generated amplitude spectrograms and estimated phase information are combined using inverse STFT to reconstruct the final ground motion waveforms.
  • Model Prior: A sequential Gaussian distribution is used as a dynamic model prior, incorporating an auto-regressive recurrence for improved temporal modeling.

The authors highlight that this approach avoids explicit physics constraints, instead learning the underlying physics and spatial heterogeneity directly from observed data.

Datasets

The study primarily utilizes the Northern California Earthquake Data Center (NCEDC) database for small-magnitude earthquakes (M < 4) in the San Francisco Bay Area (SFBA) from 1990-2022.

  • Data Selection: Recordings were filtered for acceptable signal-to-noise ratios (S/N > 3) across the frequency range of [2, 15] Hz. Approximately 15,000 recordings per component were retained, with final datasets of 5,108 for H1 and 5,301 for H2 components.
  • Benchmark Data: For comparison, the study uses existing empirical ground motion models (GMMs), including a non-ergodic GMM developed for the SFBA and the non-ergodic LA21 model for California.

Key Insights and Ideas

  • Implicit Physics Learning: The most significant insight is that a generative AI model, by conditioning on geospatial coordinates, can implicitly learn complex spatial variations in ground motions, including site effects (like amplification by soft bay mud), without being explicitly programmed with physical laws or Earth models.
  • Spatial Continuity in FAS: The CGM-GM model successfully generates Fourier Amplitude Spectra (FAS) maps that exhibit spatial continuity, capturing variations across geological features like faults, which is a key improvement over simpler models.
  • Complementary to Existing Methods: The model shows potential to complement both computationally intensive physics-based simulations and simpler empirical models, offering a more flexible and data-driven alternative.
  • Handling Sparse Data: By learning from available data, the model can generate plausible ground motions for scenarios where direct observations are scarce, a common issue in seismic hazard assessment.
  • Dynamic VAEs for Time Series: The use of dynamic VAEs with sequential modeling (RNNs) is crucial for capturing the temporal evolution of ground motion waveforms, a step beyond static VAEs.

Key Results

  • Waveform Fidelity: CGM-GM generates waveforms that effectively capture shape, frequency content, peak values, and arrival times, showing good agreement with true recordings.
  • FAS Map Accuracy: Generated FAS maps demonstrate spatial continuity and capture local features, aligning well with non-ergodic GMMs and showing improved representation of site effects compared to simpler baseline models.
  • Residual Analysis: The model shows comparable or slightly better residuals (differences between generated and true FAS) than the non-ergodic GMM, especially at lower frequencies, with reduced misfits and uncertainty compared to a baseline model lacking spatial conditioning.
  • Arrival Times and Durations: The model accurately captures P and S wave arrival times, demonstrating a strong correlation with rupture distance. While duration generation shows a good general correlation, it exhibits more scatter, particularly for smaller magnitudes due to lower signal-to-noise ratios in the training data.
  • Magnitude Scaling: When fine-tuned, the model shows a reasonable increase in FAS with magnitude, though it deviates from models trained on large-magnitude events, highlighting the challenge of extrapolating beyond the training data's magnitude range.

Limitations and Future Work

  • Extrapolation to Large Magnitudes: The model, trained primarily on small-magnitude earthquakes, struggles to perfectly replicate the behavior of large-magnitude events, particularly regarding nonlinear effects and saturation of FAS. Future work includes exploring fine-tuning strategies and out-of-distribution generalization methods.
  • Phase Retrieval Artifacts: The reliance on phase retrieval methods for waveform reconstruction can introduce inaccuracies, especially for amplitude-smoothed VAE outputs. The authors suggest exploring time-domain generative models to bypass this issue.
  • Explicit Path Effects: While path effects are implicitly learned, the current framework doesn't explicitly separate source, site, and path effects. Future research aims to incorporate spatial correlation functions (e.g., Matérn covariance) to better model path effects.
  • Focal Mechanisms: Focal mechanisms were not included as conditional variables due to data limitations and the focus on specific fault types in the training data. Future work plans to incorporate them, especially for regions with diverse focal mechanisms.
  • Duration Accuracy: Generating accurate seismic durations, particularly for low S/N data, remains a challenge.

The authors express optimism that their work will spur further research in AI-driven seismic modeling, addressing these limitations to enhance generation quality.

Summary updated

Abstract

Predicting high-fidelity ground motions for future earthquakes is crucial for seismic hazard assessment and infrastructure resilience. Conventional empirical simulations suffer from sparse sensor distribution and geographically localized earthquake locations, while physics-based methods are computationally intensive and require accurate representations of Earth structures and earthquake sources. We propose an artificial intelligence (AI) spectrogram generator, Conditional Generative Modeling for Ground Motion (CGM-GM). CGM-GM leverages earthquake magnitudes and geographic coordinates of earthquakes and sensors as inputs, when postprocessed with phase information, capturing spatially continuous Fourier amplitude spectra (FAS) as well as properties such as P and S arrivals, and waveform durations, without explicit physics constraints. This is achieved through a probabilistic autoencoder that extracts latent distributions in the time-frequency domain and variational sequential models for prior and posterior distributions. We evaluate the performance of CGM-GM using small-magnitude earthquake records from the San Francisco Bay Area, a region with high seismic risks. Here, we report that CGM-GM demonstrates potential for complementing physics-based simulations and non-ergodic empirical ground motion models, as well as shows promise in seismology and beyond.
BibTeX
@inproceedings{ren2026learning,
  author = {Pu Ren and Rie Nakata and Maxime Lacour and Ilan Naiman and Nori Nakata and Jialin Song and Zhengfa Bi and Osman Asif Malik and Dmitriy Morozov and Omri Azencot and N. Benjamin Erichson and Michael W. Mahoney},
  title = {{Learning earthquake ground motions via conditional generative modeling}},
  booktitle = {Nature Communications},
  year = {2026},
  doi = {10.1038/s41467-026-70719-2},
}
Figure