Broadband Ground Motion Synthesis via Generative Adversarial Neural Operators: Development and Validation (2309.03447v3)

Abstract: We present a data-driven framework for ground-motion synthesis that generates three-component acceleration time histories conditioned on moment magnitude, rupture distance , time-average shear-wave velocity at the top $30m$ ($V_{S30}$), and style of faulting. We use a Generative Adversarial Neural Operator (GANO), a resolution invariant architecture that guarantees model training independent of the data sampling frequency. We first present the conditional ground-motion synthesis algorithm (cGM-GANO) and discuss its advantages compared to previous work. We next train cGM-GANO on simulated ground motions generated by the Southern California Earthquake Center Broadband Platform (BBP) and on recorded KiK-net data and show that the model can learn the overall magnitude, distance, and $V_{S30}$ scaling of effective amplitude spectra (EAS) ordinates and pseudo-spectral accelerations (PSA). Results specifically show that cGM-GANO produces consistent median scaling with the training data for the corresponding tectonic environments over a wide range of frequencies for scenarios with sufficient data coverage. For the BBP dataset, cGM-GANO cannot learn the ground motion scaling of the stochastic frequency components; for the KiK-net dataset, the largest misfit is observed at short distances and for soft soil conditions due to the scarcity of such data. Except for these conditions, the aleatory variability of EAS and PSA are captured reasonably well. Lastly, cGM-GANO produces similar median scaling to traditional GMMs for frequencies greater than 1Hz for both PSA and EAS but underestimates the aleatory variability of EAS. Discrepancies in the comparisons between the synthetic ground motions and GMMs are attributed to inconsistencies between the training dataset and the datasets used in GMM development. Our pilot study demonstrates GANO's potential for efficient synthesis of broad-band ground motions

• The paper introduces cGM-GANO, a data-driven framework that employs generative adversarial neural operators to synthesize realistic three-component acceleration time histories.
• The method uses a two-phase training process with synthetic and real datasets to capture seismic characteristics across a broad frequency range while noting specific biases.
• The study demonstrates the framework's resolution-invariance and scalability, highlighting its potential impact on seismic hazard analysis and real-time data augmentation.

Overview of Broadband Ground Motion Synthesis via Generative Adversarial Neural Operators

The paper introduces a novel data-driven framework for synthesizing broadband ground motions via Generative Adversarial Neural Operators (GANO), named cGM-GANO. This framework is tailored for applications in seismic engineering, focusing on generating realistic three-component acceleration time histories. These ground motions are parameterized by moment magnitude ($\mathbf{M}$), rupture distance ($R_{\text{rup}}$), $V_{S30}$ (time-averaged shear wave velocity at the top 30 m), and the style of faulting or tectonic environment.

Methodology

The proposed method employs a Generative Adversarial Neural Operator, a significant departure from traditional generative models that operate in finite-dimensional spaces. The GANO framework utilized in this paper features a generator and discriminator trained through an adversarial game, allowing it to learn mappings between function spaces with continuous input representations. This approach is contrasted with conventional neural networks, which operate over discrete spaces. GANOs, and specifically the U-shaped Neural Operator (U-NO) architecture, enable resolution invariance, offering computational efficiency and the ability to synthesize data at arbitrary resolutions.

The architecture's training is bifurcated into two phases: verification using synthetic data generated by the Southern California Earthquake Center's Broadband Platform (SCEC BBP) and validation using real recorded ground motions from Japan's KiK-net. This strategic training approach assists in understanding both the synthetic reproduction of ground motions and the applicability of the model to empirical datasets.

Results and Findings

• Verification with SCEC BBP Simulations: The cGM-GANO successfully reproduces low-frequency characteristics ($f < 1Hz$) in comparison with BBP synthetic data. However, it faces challenges at higher frequencies, likely due to the distinct stochastic processes utilized in BBP for synthesizing those components. The discrepancy suggests potential limitations of neural operators when distinguishing between deterministic and stochastic components of synthetic ground motions.
• Validation with KiK-net Observed Data: The cGM-GANO model performs well across a broad range of frequencies (up to 30 Hz) on empirical datasets. However, it tends to exhibit positive biases for short rupture distances and certain soft soil site conditions, emphasizing data scarcity in these scenarios as a probable cause. Notably, the model consistently matches traditional Ground Motion Models (GMMs) in terms of median ground motion scaling across various tectonic conditions but occasionally underestimates aleatory variability.

Implications

The success of cGM-GANO illustrates the viability of using neural operators for broadband ground motion synthesis in engineering applications. The framework's ability to generate realistic seismic time series efficiently with trained models offers promising prospects for seismic hazard analysis where high-frequency and realistic broadband ground motions are critical. Furthermore, the framework’s resolution-invariance and scalability indicate potential applications beyond the current setting, such as real-time seismic data augmentation for early warning systems.

Future Directions

The paper recognizes the necessity for future work to address the observed biases, particularly focusing on expanding the training datasets with high-quality recorded data from diverse tectonic environments. In addition, exploring the integration of non-linear site response effects into the model's training regime is crucial, as this would enhance the applicability and fidelity of synthesized ground motions for engineering purposes. More sophisticated augmentation strategies, potentially combining physics-informed data generation and advanced data-driven approaches, might further bolster the framework's performance.

This research contributes significantly to the growing field of machine learning applications in seismology, presenting a clear pathway toward improved data-driven earthquake ground motion modeling capabilities. The GANO approach, offering both theoretical and practical advantages, may be instrumental in advancing robust seismic hazard assessments and risk mitigation strategies.