Comparison of Deterministic vs. Stochastic Inversion Algorithms in Lithological Characterization

Surface wave inversion stands as a cornerstone of modern geophysics, enabling the non-invasive mapping of the Earth's shallow subsurface. By analyzing the dispersive nature of Rayleigh and Love waves—where different frequencies propagate at different velocities depending on the stiffness of the material they traverse—researchers can reconstruct shear-wave velocity (Vs) profiles. This process, known as lithological characterization, requires the application of sophisticated mathematical frameworks to transform observed dispersion data into physical models of the earth. At the center of this discipline is the debate between deterministic and stochastic inversion algorithms, each offering distinct advantages in terms of computational efficiency, accuracy, and the ability to resolve complex geological stratigraphies.

The study of acoustic wave propagation within heterogeneous solid-state media has advanced significantly through the standardization of geophysical site characterization protocols. Modern research initiatives, such as the Surface Wave Hub, focus on the empirical calibration of ground-motion signatures. By deploying high-precision geophones and accelerometers, geophysicists capture subtle seismic reflections and microtremors. These data serve as the input for inversion algorithms designed to infer critical material properties, including elastic moduli, density, and porosity. The reliability of these inferences depends heavily on the chosen mathematical approach, particularly when addressing the non-linearity and non-uniqueness inherent in seismic data interpretation.

At a glance

• Primary Methodology:Surface wave analysis utilizes the dispersion of Rayleigh and Love waves to characterize subsurface lithology and engineered material interfaces.
• Algorithm Types:Inversion techniques are categorized into deterministic (e.g., Occam’s inversion) and stochastic (e.g., Monte Carlo, Genetic Algorithms) methods.
• Benchmark Standard:The 2004 InterPacific project remains the definitive reference for comparing seismic site characterization methods across international datasets.
• Key Variables:Critical outputs include shear-wave velocity (Vs), Poisson’s ratio, density, and porosity, with varying degrees of sensitivity to the inversion process.
• Applications:Non-destructive testing (NDT) of infrastructure, including bridges and foundations, and the detection of buried utilities or subsurface voids.

Background

Surface waves are generated when seismic energy interacts with the free surface of the Earth. Unlike body waves (P-waves and S-waves) that travel through the interior, surface waves are confined to the upper layers, making them ideal for shallow geotechnical investigations. Rayleigh waves, characterized by a retrograde elliptical motion, and Love waves, which exhibit horizontal shear motion, are the primary focus of these studies. The depth of penetration for these waves is frequency-dependent: low-frequency waves sample deeper strata, while high-frequency waves are sensitive only to the near-surface.

The fundamental challenge in surface wave analysis is the "inverse problem." Given a measured dispersion curve—a plot of phase velocity versus frequency—how does one find the corresponding layers of soil or rock? This problem is inherently ill-posed because multiple different geological models can theoretically produce the same dispersion curve. To handle this ambiguity, geophysicists employ inversion algorithms that seek a "best fit" between the theoretical model and the observed data. The evolution of these algorithms has been marked by a shift from rigid deterministic frameworks to more flexible, though computationally intensive, stochastic methods.

Deterministic Inversion: The Least-Squares and Occam’s Approach

Deterministic inversion methods are gradient-based techniques that start with an initial guess of the subsurface model and iteratively refine it to minimize the difference between observed and calculated dispersion curves. The most prevalent among these is the Least-Squares approach, often implemented as Occam’s inversion. Named after the principle of parsimony (Occam’s Razor), this algorithm seeks the "smoothest" possible model that fits the data within a specified error tolerance.

Occam’s inversion is highly valued for its speed and stability. By imposing a smoothness constraint, it prevents the algorithm from creating unrealistic, jagged layers that might otherwise result from attempts to fit noise in the data. However, the reliance on an initial model is a significant drawback. If the starting profile is too far from reality, the algorithm may become trapped in a "local minimum," providing a mathematically valid but geologically incorrect solution. This limitation is particularly pronounced in heterogeneous environments where velocity inversions (softer layers beneath harder ones) are present.

Stochastic Inversion: Monte Carlo and Genetic Algorithms

Stochastic methods depart from the gradient-based path by employing random search strategies to explore the entire parameter space. These techniques do not require a detailed initial model and are better suited for finding the "global minimum" in complex geological settings. Two of the most common stochastic approaches are Monte Carlo simulations and Genetic Algorithms (GAs).

• Monte Carlo Inversion:This method involves generating thousands or even millions of random subsurface models. For each model, a theoretical dispersion curve is calculated and compared to the field data. While computationally expensive, it provides a detailed map of the solution space, allowing researchers to quantify uncertainty.
• Genetic Algorithms:Inspired by biological evolution, GAs treat model parameters as "chromosomes." A population of models "evolves" over generations through processes of selection, crossover, and mutation. Models that better fit the data are more likely to pass their characteristics to the next generation. GAs are particularly effective at handling high-dimensional spaces where density and thickness are unknown.

The trade-off for the robustness of stochastic methods is their demand for processing power. Unlike the seconds or minutes required for an Occam’s inversion, a detailed GA or Monte Carlo run can take hours or days, depending on the number of layers and variables involved.

The 2004 InterPacific Benchmark

The 2004 InterPacific project provided a critical turning point in the validation of these algorithms. This international collaborative study was designed to assess the reliability of seismic site characterization by distributing identical datasets to different research teams. The results, published in subsequent geophysical journals, highlighted the inherent variability in results when different inversion strategies were applied to the same experimental data.

The InterPacific benchmarks revealed that while all methods could accurately determine the average shear-wave velocity of the top 30 meters (Vs30), they diverged significantly when defining the exact depths of lithological boundaries. Deterministic methods tended to blur sharp interfaces due to the smoothness constraint, whereas stochastic methods often produced many "equivalent" models, emphasizing the need for independent geological constraints, such as borehole logs, to narrow the results.

Error Margins in Porosity and Density Estimations

While shear-wave velocity is the primary output of surface wave inversion, there is increasing interest in inferring secondary properties such as density and porosity. Data analyzed in theGeophysical Journal InternationalSuggests that these properties are significantly more difficult to resolve than velocity. Because Rayleigh wave phase velocity is less sensitive to density than to shear modulus, the error margins for density estimations can exceed 15–20% in the absence of a priori information.

Porosity estimation introduces further complexity, as it involves the interaction of the solid matrix with pore fluids. Inversion algorithms must integrate Biot-Gassmann theory or similar rock physics models to link seismic velocities to porosity levels. Research indicates that stochastic methods generally outperform deterministic ones in this regard, as they can better account for the non-linear coupling between density and porosity. However, even with advanced algorithms, the uncertainty in porosity often remains high, particularly in the shallow subsurface where degree of saturation varies.

“The accurate characterization of lithology from surface waves requires a transition from finding a single ‘best’ model to understanding the distribution of all probable models.”

Practical Applications and Infrastructure NDT

The practical application of these algorithms extends beyond academic research into the area of civil engineering and infrastructure maintenance. Non-destructive testing (NDT) of bridges, tunnels, and foundations relies on the analysis of dispersion curves to detect degradation or structural anomalies. For instance, an increase in the attenuation of high-frequency surface waves can signal the presence of micro-cracking or moisture ingress in concrete foundations.

In urban environments, characterizing shallow subsurface anomalies is a primary task. The meticulous interpretation of microtremor data—ambient seismic noise—allows for the detection of buried utilities and voids without the need for active seismic sources. By applying inversion algorithms to this passive data, engineers can identify zones of low velocity that may indicate potential sinkholes or poorly compacted fill material. The choice between deterministic and stochastic methods in these cases often depends on the urgency of the assessment; a rapid Occam’s inversion may be used for initial screening, followed by a detailed Monte Carlo analysis for high-risk zones.

What researchers disagree on

Despite the advancements in inversion theory, there remains significant debate regarding the "regularization" of the inverse problem. Some researchers argue that the smoothness constraints in deterministic models are physically unjustified and may mask important geological features like thin clay lenses or sharp bedrock transitions. Conversely, critics of stochastic methods point out that without proper constraints, random searches can produce models that are mathematically plausible but physically impossible, such as layers with unrealistic Poisson’s ratios.

Furthermore, the integration of higher-mode data remains a point of contention. While including the first or second higher modes of Rayleigh waves can theoretically improve the resolution at depth, identifying these modes in noisy field data is notoriously difficult. Misidentifying a higher mode as the fundamental mode can lead to catastrophic errors in the resulting shear-wave velocity profile, a risk that persists regardless of the inversion algorithm used.