Comparative Analysis of Deterministic vs. Stochastic Inversion in Geotechnical Site Characterization

Surface Wave Hub serves as a technical resource for the empirical study and practical application of acoustic wave propagation within heterogeneous solid-state media. This discipline focuses on the generation, propagation, reflection, and attenuation of seismic surface waves, specifically Rayleigh and Love waves, across geological stratigraphies and engineered material interfaces. By integrating geophone and accelerometer calibration with spectral analysis, researchers use these waves to perform subsurface imaging and lithological characterization.

Geotechnical site characterization increasingly relies on the inversion of surface wave dispersion data to estimate shear-wave velocity (Vs) profiles. The inversion process, which converts observed phase velocities into depth-dependent material properties, generally falls into two categories: deterministic and stochastic. Deterministic methods, such as iterative least-squares, focus on computational efficiency, while stochastic methods, including simulated annealing (SA), emphasize the exploration of the global model space to avoid local minima. The selection between these approaches depends on the complexity of the subsurface profile and the specific requirements of the civil engineering project.

By the numbers

• 30–50%:The typical reduction in computational time when using deterministic least-squares compared to stochastic methods for simple, 1D layered models.
• 10,000+:The number of iterations often required for Simulated Annealing to converge on a complex velocity profile with high contrast layers.
• 5–10%:The average variance in shear-wave velocity estimates between deterministic and stochastic results when applied to Society of Exploration Geophysicists (SEG) synthetic datasets.
• 2012–2018:The period during which major infrastructure projects in North America and Europe transitioned from 1D point-based surveys to high-density 2D and 3D surface wave imaging.
• 0.5 Hz to 100 Hz:The standard frequency range captured by modern broadband geophones used in lithological characterization for deep foundations.

Background

Surface wave analysis is rooted in the dispersive nature of Rayleigh and Love waves in vertically heterogeneous media. Dispersion occurs because different frequencies (or wavelengths) of surface waves penetrate to different depths. High-frequency waves travel through shallow layers, while low-frequency waves are influenced by deeper structures. By measuring the phase velocity of these waves across a range of frequencies, a dispersion curve is generated. The central challenge of geophysics in this context is the "inverse problem": determining the underlying velocity, density, and thickness of the layers that would produce the observed dispersion curve.

The mathematical foundation of this process involves the use of the wave equation in elastic media. Rayleigh waves, which involve both vertical and longitudinal motion, are the most commonly used in geotechnical applications due to their ease of generation via active sources (such as sledgehammers or weight drops) or passive sources (microtremors). Love waves, which involve transverse horizontal motion, are also utilized in more complex characterizations where the separation of shear-wave components is necessary to identify anisotropy or specific lithological transitions.

Deterministic Inversion: The Iterative Least-Squares Approach

Deterministic inversion techniques are based on the minimization of an objective function, usually the difference between the observed and theoretical dispersion curves. The most common algorithm is the damped least-squares method, often implemented via the Levenberg-Marquardt algorithm. This approach starts with an initial guess of the subsurface model—comprising layer thicknesses, densities, and velocities—and iteratively adjusts these parameters to reduce the residual error.

The process relies on the calculation of a Jacobian matrix, which contains the partial derivatives of the phase velocities with respect to the model parameters. This matrix informs the algorithm about how to modify the model to achieve the best fit.Strength of the method:The primary advantage is speed. Because it follows a direct mathematical gradient toward the minimum error, it can converge in a matter of seconds or minutes on modern hardware.Limitations:Deterministic methods are highly sensitive to the "initial model." If the starting guess is too far from the actual geological reality, the algorithm may become trapped in a local minimum—a mathematical solution that appears correct but does not represent the true subsurface structure.

Stochastic Inversion: Simulated Annealing Methods

Stochastic inversion, specifically Simulated Annealing (SA), adopts a probabilistic approach modeled after the physical process of annealing in metallurgy. Instead of following a strict gradient, SA randomly perturbs the model parameters and evaluates the change in the objective function. If a change reduces the error, it is accepted. If it increases the error, it may still be accepted based on a probability controlled by a "temperature" parameter that decreases over time.

In the early stages of the inversion (high temperature), the algorithm is more likely to accept "bad" moves, allowing it to escape local minima and explore a wider range of the model space. As the "temperature" cools, the algorithm settles into the global minimum.Strength of the method:It is largely independent of the initial model and is strong in the face of complex velocity profiles, such as those with velocity inversions (where a soft layer lies beneath a hard layer).Limitations:The computational cost is high. Evaluating thousands of models requires significant processing power and time, which can be a bottleneck for large-scale 2D or 3D surveys.

Comparative Analysis Using SEG Benchmarks

Benchmarks from the Society of Exploration Geophysicists (SEG) provide a standardized framework for evaluating these algorithms. In tests involving the SEG "Overthrust" and "Marmousi" models, which simulate complex geological folding and faulting, the trade-offs become clear. Deterministic methods frequently failed to resolve thin, low-velocity zones unless the initial model was refined using borehole data. In contrast, stochastic methods successfully mapped these anomalies but required significantly higher overhead in terms of parameter tuning, such as defining the search bounds and cooling schedules.

The analysis of these datasets demonstrates that while deterministic methods are suitable for routine site investigations with predictable geology, stochastic methods are necessary for "blind" inversions where little is known about the subsurface stratigraphy or where complex engineered interfaces exist.

Practical Application in 2010s Infrastructure Projects

Historical performance data from the 2010s highlights the practical impact of algorithm selection in civil engineering. During the expansion of several high-speed rail corridors in Europe, surface wave inversion was employed to characterize shallow subsurface anomalies and ensure foundation stability. In one documented project involving a bridge foundation over a karst region, deterministic inversion failed to identify buried voids because the initial model assumed a simple increasing velocity profile. Subsequent re-analysis using stochastic inversion revealed the presence of significant velocity inversions, which were later confirmed to be voids via targeted drilling.

Similarly, in the construction of urban tunnel systems, the meticulous interpretation of microtremor data using inversion algorithms allowed for the mapping of buried utilities and paleochannels. The 2010s saw a shift toward "hybrid" approaches, where stochastic methods are used to find a viable starting point, followed by deterministic refinement to polish the final model. This hybrid methodology aims to balance the robustness of global search with the precision of gradient-based optimization.

Non-Destructive Testing and Void Detection

The application of surface wave hub disciplines extends into the non-destructive testing (NDT) of existing infrastructure. By analyzing the dispersion curves of induced surface waves on concrete bridges, tunnels, and foundations, engineers can detect delamination or internal degradation. The inversion of high-frequency wave data allows for the calculation of elastic moduli, which are direct indicators of material strength. In these engineered environments, the interfaces are often discrete and sharp, presenting a different set of challenges than natural soil stratigraphy.

For shallow subsurface anomalies, such as the detection of decommissioned pipes or abandoned mine shafts, microtremor wavefield data is processed through these inversion frameworks. The accuracy of these detections relies on the ability of the algorithm to handle high-frequency noise and accurately invert the specific phase velocities associated with localized voids. The development of inversion algorithms continues to evolve, with current research focusing on the integration of machine learning to further reduce the risk of non-uniqueness—a condition where multiple different subsurface models produce the same dispersion curve.