Abstract
Accurately forecasting ground motion during earthquakes requires detailed knowledge of subsurface geological structures, which are often poorly characterized. This paper presents a framework for "online" 3D site characterization that integrates field seismic testing with high-performance inverse modeling. By employing a multiscale optimization strategy and wavelength-adaptive meshing, the researchers demonstrate a computationally feasible method for reconstructing complex underground soil and rock profiles. This approach allows for "steerable" field experiments, where real-time data analysis informs the placement of sensors to resolve areas of highest uncertainty.
Introduction: The Challenge of Subsurface Imaging
Predicting earthquake-induced ground shaking is a significant challenge due to the extreme variability of soil and rock properties, such as shear wave velocity and density, across geological basins. Traditional site characterization methods often rely on sparse laboratory specimens or small-scale field tests, leading to substantial uncertainty in seismic hazard maps. Imaging the Earth's crust is an "ill-posed" problem: deep soil is difficult to access, and the mathematical models used to infer properties from surface vibrations are prone to local minima—getting stuck on incorrect solutions.
Methodology
The researchers addressed these challenges through two primary computational pillars:
Forward Modeling
To simulate wave propagation efficiently, the authors utilized wavelength-adaptive mesh algorithms based on octree-based hexahedral elements. Unlike uniform grids, this simulation uses high resolution only where waves are short and low resolution where they are long. This technique achieved a 2000x reduction in the number of grid points required for a model of the Los Angeles Basin.
Inverse Modeling and Multiscale Strategy
To reconstruct material properties from observed ground motion, the framework employs a multiscale continuation strategy. The optimization begins with a simplified, "smooth" version of the problem and incrementally increases resolution (e.g., from 1x1 to 257x257 grids). This progression prevents the algorithm from becoming trapped in local minima and ensures convergence toward the global solution.
The optimization also incorporates Total Variation (TV) regularization to maintain sharp boundaries between geological layers and a matrix-free Gauss-Newton approach with L-BFGS preconditioning for computational efficiency.
Key Insights: The "Online" Loop
The most innovative aspect of this framework is the concept of the "steerable" experiment. By exciting the soil and transmitting data via satellite to a supercomputer for overnight inversion, researchers can identify areas of high uncertainty. This allows field crews to adjust sensor placement the following day, effectively turning the ground into a steerable laboratory. The study emphasizes that a multiscale approach is mandatory; without it, the inversion process fails to find the correct geological profile.
Results and Validation
The framework was validated using the SCEC LA Basin Model, data from the 1994 Northridge earthquake, and synthetic data for the San Fernando Valley.
High-Fidelity Reconstruction
In 3D synthetic tests, the algorithm successfully reconstructed complex shear-wave velocity profiles, even with the addition of 5% random noise. As shown in the 3D simulations, the inverted waveforms closely matched the target waveforms, even at locations without sensors.
Robustness to Sensor Density
While higher sensor density (e.g., 64 vs. 16 receivers) improves resolution, the multiscale approach proved robust enough to provide a close approximation of the target model even with a sparse sensor network.
Limitations and Future Directions
Despite its success, several challenges remain for real-time application:
- Turnaround Time: At the time of the study, a full 3D inversion required approximately 24 hours on 512 processors. Further speed improvements are necessary for a truly "online" experience.
- Model Physics: Current models assume "lossless" materials. Future iterations must incorporate attenuation and nonlinear soil behavior, which are critical during intense shaking.
- Initial Guesses: Incorporating existing geological knowledge as "priors" could significantly accelerate convergence compared to starting with a homogeneous guess.
Summary updated