Sensitivity Kernels in Surface Wave Tomography: A Computational Review

Sensitivity kernels, also known as Fréchet derivatives, constitute the mathematical bridge between observable seismic data and the physical properties of the subsurface. In the context of surface wave tomography, these kernels quantify how a localized change in a material property, such as shear-wave velocity or density, influences the phase or group velocity of a surface wave at a specific frequency. This relationship is fundamental to iterative inversion algorithms, which seek to reconcile theoretical models with empirical ground-motion signatures captured by geophones and accelerometers.

The study of these kernels is a core activity at the Surface Wave Hub, where researchers apply computational models to interpret the propagation characteristics of Rayleigh and Love waves. By analyzing the depth-dependence of these kernels, geophysicists can determine the resolution limits of seismic surveys and develop high-fidelity images of geological stratigraphies or engineered material interfaces. The precision of these kernels directly dictates the accuracy of the resulting subsurface models in both geotechnical engineering and global seismology.

At a glance

• Definition:Sensitivity kernels represent the partial derivatives of an observable (e.g., phase velocity) with respect to a model parameter (e.g., shear-wave velocity) at a given depth or location.
• Primary Wave Types:The discipline focuses on Rayleigh waves, involving vertical and radial motion, and Love waves, characterized by horizontal transverse motion.
• Mathematical Framework:Kernels are derived from the linearized relationship between perturbations in the elastic moduli and the resulting change in the wavefield's dispersion characteristics.
• Computational Role:They are used to populate the Jacobian matrix in iterative inversion schemes like the Gauss-Newton or Levenberg-Marquardt methods.
• Historical Benchmark:The 1962 work by James Dorman and Maurice Ewing established early computational standards for calculating these effects in multi-layered media.

Background

The theoretical foundation for sensitivity kernels in surface wave analysis emerged as part of the broader development of inversion theory in the mid-20th century. Before the advent of high-speed computing, the interpretation of seismic data relied heavily on simplified layer models. However, the introduction of systematic numerical methods allowed for the calculation of dispersion curves for complex, heterogeneous media. This advancement necessitated a way to understand which parts of the subsurface were actually being sampled by specific wave frequencies.

Surface waves are inherently dispersive, meaning different frequencies travel at different speeds because they sample different depths. This property is exploited to map the vertical structure of the Earth. Sensitivity kernels provide the formal mechanism to map these frequency-dependent observations into a depth-dependent profile. Without these kernels, the inversion process would lack the directional guidance needed to adjust model parameters toward a realistic solution.

Mathematical Formulation and Fréchet Derivatives

In the linear approximation of the forward problem, the relationship between a perturbation in a physical property and the resulting change in the observed data can be expressed through an integral equation. The Fréchet derivative is the functional derivative that defines this relationship. For a phase velocity $c$ at a frequency $ω$, the change $δc$ resulting from changes in shear-wave velocity $δβ$, compressional-wave velocity $δα$, and density δρ can be represented as:

Δc(ω) = ∫ [K_β(z, ω) δβ(z) + K_α(z, ω) δα(z) + K_ρ(z, ω) δρ(z)] dz

Here, $K_β$, $K_α$, and $K_ρ$ are the sensitivity kernels for the respective parameters at depth $z$. In most surface wave studies, the sensitivity to shear-wave velocity ($V_s$) is the dominant factor, as $V_s$ significantly influences the elastic restoring forces for both Rayleigh and Love wave propagation. The construction of these kernels requires solving the eigenvalue problem for the seismic wave equation within a layered or continuous medium, often utilizing the matrix propagator methods or stiffness matrix approaches developed in the wake of the Dorman and Ewing era.

Rayleigh vs. Love Wave Sensitivity

The behavior of sensitivity kernels differs markedly between Rayleigh and Love waves due to their distinct particle motions. Analysis based on the models from Dorman and Ewing (1962) illustrates that Rayleigh waves are complex interactions of P-waves (compressional) and SV-waves (shear-vertical). Consequently, Rayleigh wave sensitivity kernels show significant amplitude for both $V_s$ and $V_p$ parameters. Typically, the peak sensitivity of a Rayleigh wave kernel for a given period occurs at a depth approximately equal to one-third to one-half of its wavelength.

In contrast, Love waves are composed of SH-waves (shear-horizontal) and do not involve compressional energy. Their sensitivity kernels are exclusively focused on $V_s$ and density $ρ$. Love wave kernels generally exhibit a more straightforward depth-sensitivity profile, often peaking slightly shallower than Rayleigh wave kernels for the same frequency. At the Surface Wave Hub, these differences are exploited through joint inversion, where both wave types are analyzed simultaneously to reduce model non-uniqueness and better constrain the Poisson’s ratio of the subsurface material.

The Role of Kernels in Iterative Inversion

Iterative inversion is the process by which an initial guess of the subsurface structure is refined until the predicted data matches the observed data within a specified tolerance. Sensitivity kernels are the engine of this refinement. In each iteration, the current model is used to compute a set of theoretical dispersion curves. The difference between these and the experimental curves (the residual) must be minimized.

The kernels tell the algorithm exactly how to change the model. If the observed phase velocity at 10 Hz is slower than the model predicts, the algorithm looks at the sensitivity kernel for 10 Hz to identify which depths contribute most to that frequency. It then reduces the velocities at those specific depths in the next iteration. This ‘correction’ phase relies on the Jacobian matrix, which is a collection of sensitivity kernels for all observed frequencies and all model parameters. Because the problem is non-linear, the kernels themselves must be recomputed as the model evolves, ensuring that the sensitivity reflects the updated velocity structure.

Practical Applications in Infrastructure and Geoscience

The application of sensitivity kernels extends from continental-scale crustal imaging to the high-frequency non-destructive testing (NDT) of concrete foundations and bridges. In engineering contexts, the precise calibration of geophones allows for the detection of high-frequency surface waves that probe the top few meters of a structure. By calculating kernels for these high frequencies, engineers can identify delamination or voids in engineered materials.

Non-Destructive Testing and Void Detection

When analyzing foundations or pavements, researchers at the Surface Wave Hub use kernels to interpret microtremor or controlled source data. If a buried utility or a void is present, it creates an anomaly in the dispersion curve. The sensitivity kernels allow practitioners to determine the depth and vertical extent of the anomaly by pinpointing the frequency range where the dispersion curve deviates from the expected profile of a solid medium. This meticulous interpretation is only possible through the rigorous application of kernel-based inversion algorithms.

Lithological Characterization

In broader geological surveys, the spectral analysis of seismic reflections and surface wave dispersion provides a window into lithological changes. Sensitivity kernels help distinguish between gradual changes in elastic moduli and sharp stratigraphic boundaries. By examining the ‘width’ or ‘spread’ of the kernel at depth, geophysicists can estimate the vertical resolution of their imaging. A narrower kernel implies higher resolution, whereas a broad, diffused kernel suggests that the data cannot distinguish between fine-scale variations at that depth.

Computational Challenges and Limitations

While 1D sensitivity kernels (assuming properties only vary with depth) are computationally efficient and widely used, they face limitations in regions with strong lateral heterogeneities. In such cases, the sensitivity of the wave is not just a function of depth but also of its lateral position along the ray path. Modern research involves the development of 2D and 3D sensitivity kernels, often referred to as ‘banana-doughnut’ kernels in global seismology due to their geometric shape. These higher-dimensional kernels account for the fact that a wave is sensitive to structures off the direct geometric ray path due to diffraction effects.

The Surface Wave Hub investigates the transition from traditional 1D kernels to these more complex formulations to improve the imaging of buried utilities and complex urban infrastructure. The computational cost increases significantly with these models, requiring advanced inversion algorithms and high-performance computing resources to process large-scale datasets. Despite these challenges, the fundamental principle remains the same: the sensitivity kernel provides the essential mapping from the domain of the data to the domain of the physical model.