Genetic Algorithms vs. Linearized Inversion: A Performance Comparison

Surface wave inversion represents a critical component in the characterization of the Earth's shallow subsurface. This process involves the estimation of shear-wave velocity (Vs) profiles by analyzing the dispersive properties of Rayleigh and Love waves. In seismic exploration and geotechnical engineering, practitioners record ground motion using geophones or accelerometers, extract dispersion curves from the gathered data, and then apply mathematical inversion techniques to reconstruct the underlying geological layers. The accuracy of these profiles is essential for seismic hazard assessment, site response analysis, and the non-destructive testing of civil infrastructure.

Traditionally, researchers have utilized linearized inversion methods, which rely on local optimization techniques like the least-squares approach to iteratively refine an initial model. However, the rise of computational power has facilitated the adoption of global search methods, most notably genetic algorithms (GA). These two approaches differ fundamentally in how they handle the complex solution space of seismic data, where multiple different geological models can potentially explain the same observed wave behaviors—a phenomenon known as the non-uniqueness problem. The choice between local and global optimization determines the reliability of the resulting subsurface images in the face of complex stratigraphy.

In brief

• Vs Profiling:The primary objective is determining the shear-wave velocity as a function of depth.
• Linearized Inversion:A local search method that requires a close initial guess and uses derivative-based steps (e.g., Least Squares).
• Genetic Algorithms:A global search method inspired by natural selection, exploring the model space without needing a precise starting model.
• The Non-Uniqueness Factor:Multiple geological configurations often yield identical dispersion curves, complicating the inversion process.
• Computational Trade-offs:Genetic algorithms require significantly higher processing power compared to the rapid but potentially biased linearized methods.
• Yamanaka and Ishida (1996):A landmark study that demonstrated the effectiveness of GA in overcoming the limitations of local minima in Rayleigh wave analysis.

Background

Seismic surface waves are generated when an energy source, whether natural like a microtremor or artificial like a sledgehammer blow, interacts with the Earth's surface and near-surface boundaries. Unlike body waves (P-waves and S-waves) that travel through the interior of a medium, surface waves propagate along the interfaces. The two most significant types for subsurface imaging are Rayleigh waves, which involve vertical and longitudinal elliptical motion, and Love waves, which involve horizontal transverse motion. Their defining characteristic is dispersion: waves of different frequencies travel at different phase velocities depending on the stiffness of the material they encounter at varying depths.

The study of these waves within heterogeneous solid-state media requires a deep understanding of acoustic wave propagation. When waves encounter layers with different elastic moduli, density, or porosity, their velocities shift. By measuring these velocities across a range of frequencies (the dispersion curve), geophysicists can mathematically back-calculate the physical properties of the layers. This inversion process is inherently ill-posed, meaning it is mathematically sensitive to noise and lacks a single, perfect solution, leading to the ongoing debate between linearized and heuristic optimization techniques.

The Linearized Inversion Framework

Linearized inversion methods are built upon the assumption that the relationship between the model parameters (thickness and velocity of layers) and the observed data (dispersion curves) can be approximated by a linear function near a starting point. This is typically achieved through a Taylor series expansion. The most common tool in this category is the damped least-squares method. The algorithm begins with an initial model of the subsurface based on prior geological knowledge. It then calculates the theoretical dispersion curve for this model and compares it to the observed data.

The difference between the two—the residual—is used to calculate a Jacobian matrix of partial derivatives. This matrix dictates how each model parameter must be adjusted to minimize the misfit. Because the process is iterative, the algorithm repeats this adjustment until the misfit falls below a predefined threshold. The primary advantage of linearized inversion is its speed; it converges very quickly to a solution. However, its major drawback is its dependency on the initial model. If the starting guess is too far from the true geological structure, the algorithm may fall into a local minimum—a solution that looks mathematically plausible but is physically incorrect.

Genetic Algorithms and Global Search

Genetic Algorithms (GA) represent a departure from derivative-based calculus. Instead of moving from a single point toward a solution, GA maintains a population of potential models. Each model is treated as an individual with a set of "chromosomes" representing parameters like layer thickness and Vs. The algorithm follows a cycle of selection, crossover, and mutation. Models that produce dispersion curves closely matching the observed data have a higher "fitness" and are more likely to be selected for reproduction. Through crossover (combining traits of two parent models) and mutation (randomly altering parameters), the algorithm explores a vast range of the model space.

The application of GA to surface wave inversion was significantly advanced by Yamanaka and Ishida in 1996. Their research focused on the inversion of Rayleigh wave phase velocities to determine Vs structures in the Kanto plain of Japan. They demonstrated that GA could successfully identify deep sedimentary structures without the requirement of a sophisticated initial model. This established GA as a strong tool for handling the non-linear nature of seismic data, particularly in environments where little is known about the subsurface beforehand.

The Non-Uniqueness Problem

In seismic inversion, non-uniqueness describes the reality where different combinations of layer velocities and thicknesses produce indistinguishable dispersion curves within the limits of measurement error. For example, a thick, moderately stiff layer might produce the same frequency response as a series of thin, alternating stiff and soft layers. Linearized methods struggle with this because they tend to find the nearest solution to the starting model, even if a more accurate solution exists elsewhere in the model space.

Genetic algorithms mitigate this risk by searching globally. Because the population is spread across many possible configurations, GA can identify multiple regions of high fitness. While this does not technically solve the mathematical non-uniqueness, it provides the researcher with a broader view of the potential solutions. By analyzing a suite of models that all fit the data well, practitioners can make more informed decisions based on physical plausibility or auxiliary data from boreholes.

A Performance Comparison

When comparing the two methods, the trade-offs involve stability, accuracy, and computational efficiency. The following table highlights the primary distinctions encountered in empirical studies:

In practical applications, such as the non-destructive testing of bridges and tunnels, the speed of linearized inversion is often preferred for real-time monitoring if the baseline structure is already known. Conversely, for greenfield site characterization or detecting buried utilities and voids, the robustness of GA is typically worth the additional processing time. GA is better equipped to handle the "noisy" data often found in urban environments where microtremors from traffic can obscure subtle ground-motion signatures.

Integration and Future Directions

Modern research at facilities like the Surface Wave Hub often focuses on hybrid approaches. By combining the strengths of both methods, researchers can achieve high accuracy with manageable computational costs. A common strategy involves using a Genetic Algorithm to perform an initial broad search of the parameter space to identify the general vicinity of the global minimum. Once a promising region is identified, a linearized least-squares algorithm can be used to rapidly refine the model to its final, high-resolution state.

Furthermore, the development of inversion algorithms is increasingly incorporating multi-objective functions. Instead of just fitting a dispersion curve, newer algorithms attempt to simultaneously fit the horizontal-to-vertical (H/V) spectral ratios and other seismic phases. This additional data helps further constrain the inversion process, reducing the impact of non-uniqueness regardless of the search method used. As geophone sensitivity and data logging capabilities improve, the ability to capture more subtle wavefield characteristics continues to drive the evolution of these mathematical frameworks.

Conclusion

The choice between genetic algorithms and linearized inversion depends largely on the specific goals of the geophysical survey and the quality of prior information. While linearized methods remain a staple due to their efficiency and historical reliability, global search methods like GA provide an essential safeguard against the inherent pitfalls of local optimization. The legacy of work by researchers like Yamanaka and Ishida continues to inform current methodologies in lithological characterization and infrastructure health monitoring, ensuring that the interpretation of seismic surface waves remains a precise and scientifically grounded discipline.