Introduction
In seismic hazard analysis, Ground-Motion Models (GMMs) are essential for predicting the intensity of shaking at specific locations during an earthquake. Traditionally, these models have relied on "ergodic" assumptions, treating the variability in ground motion at a specific site as a random deviation from a global average. However, geological structures are not random; specific underground paths consistently amplify or dampen seismic waves.
To capture these repeatable patterns, seismologists are moving toward "non-ergodic" models. While Gaussian Processes (GPs) are currently the gold standard for this task, they face significant hurdles in computational scaling and flexibility. This paper introduces CGM-FAS (Conditional Generative Modeling for Fourier Amplitude Spectra), a deep-learning framework designed to overcome these limitations using Conditional Variational Autoencoders (CVAEs).
The Limitations of Gaussian Processes
Gaussian Processes have historically been favored for non-ergodic modeling because they provide a robust statistical framework for spatial interpolation. However, they suffer from three primary drawbacks:
- Computational Complexity: GPs require the inversion of large covariance matrices, a process with $O(n^3)$ complexity. As datasets grow, the time required for simulations becomes prohibitive.
- Rigid Functional Forms: GPs typically rely on pre-defined correlation functions (kernels). These mathematical assumptions may fail to capture the true complexity of geological structures, often resulting in "azimuthal artifacts" or unrealistic streaks in predicted shaking maps.
- Frequency Isolation: Most GP-based models treat different seismic frequencies independently. In reality, different parts of the seismic spectrum are physically coupled, and ignoring these correlations can lead to physically inconsistent results.
The CGM-FAS Framework
CGM-FAS reframes the problem of modeling path effects as a generative task rather than a traditional regression problem. The core of the framework is a Conditional Variational Autoencoder (CVAE).
Architecture and Conditioning
The model consists of an encoder that compresses seismic residuals into a low-dimensional latent space and a decoder that reconstructs the full Fourier Amplitude Spectra (FAS). Crucially, the model is "conditioned" on the geographic coordinates of both the earthquake source and the recording station. This allows the neural network to learn spatial continuity and complex geological patterns directly from the data without being constrained by a fixed kernel.
Balancing Accuracy and Variability
A key innovation in this approach is the introduction of a hyperparameter, $\alpha$, which balances reconstruction accuracy against the regularization of the latent space. By carefully tuning $\alpha$, the authors ensure that the model's "aleatory variability" (the unpredictable component of ground motion) aligns with established empirical values (specifically $\phi_{SP,NE} = 0.40$). This ensures the model remains statistically consistent with observed seismic behavior.
Experimental Validation
The authors trained and evaluated CGM-FAS using a dataset of small-magnitude earthquakes ($M_w$ 1 to 4) from the San Francisco Bay Area, sourced from the NCEDC database. The dataset included 5,108 recordings across 911 frequencies (2–15 Hz). The model's performance was benchmarked against LANN25, a state-of-the-art GP-based non-ergodic model.
Key Results
- Computational Efficiency: The most striking result is the speedup. CGM-FAS generated path effect maps for 10,000 sites across 1,000 frequencies in less than 10 seconds. A standard GP approach would require approximately 10 hours to complete the same task. This linear scaling makes dense, high-resolution hazard maps practical for the first time.
- Spatial and Spectral Realism: The model successfully identified spatial correlation lengths of approximately 16 km, consistent with known Bay Area geology. It accurately captured higher path effects in the South Bay’s shallow basement and lower effects in the attenuative sediments east of the Hayward Fault. Furthermore, because the CVAE processes the entire spectrum simultaneously, it naturally preserved interfrequency correlations.
- Statistical Integrity: Analysis of within-path residuals showed a Gaussian distribution centered at zero, confirming that the generative approach does not introduce significant bias compared to traditional methods.
Discussion and Future Directions
While CGM-FAS represents a major leap in efficiency, it introduces new challenges. Unlike GPs, CVAEs do not inherently provide "epistemic uncertainty"—a measure of model confidence based on data density. The authors suggest that ensemble methods or k-means clustering could be used to estimate this uncertainty in future iterations.
Additionally, while this study focused specifically on path effects, a truly non-ergodic model must also account for spatially varying source and site effects. Future work will focus on integrating these components into the generative framework to provide a comprehensive tool for seismic risk assessment.
Summary updated