Deep Learning vs. Traditional Inversion: Seismic Algorithm Frontiers

Seismic inversion is a fundamental computational process used to derive a physical model of the earth's subsurface from observed geophysical data. This discipline converts seismic measurements, such as those captured by geophones and accelerometers, into quantitative descriptions of material properties including elastic moduli, density, and porosity. Traditionally, the field has relied upon iterative least-squares algorithms to minimize the residuals between observed wavefields and predicted synthetic data. However, the emergence of deep learning has introduced alternative methodologies that use neural networks to map seismic observations directly to subsurface parameters.

The study of surface wave propagation, specifically Rayleigh and Love waves, provides the necessary data for these inversion techniques within heterogeneous solid-state media. By analyzing the dispersion characteristics of these waves—where different frequencies travel at different velocities depending on the depth and stiffness of the material—researchers can characterize geological stratigraphies and engineered material interfaces. As computational capacity increases, the debate between the rigorous but slow iterative methods of the past and the high-speed, data-driven approaches of machine learning has become a central focus of modern geophysical research.

Timeline

Early 1990s:Initial experiments with Multi-Layer Perceptrons (MLPs) were conducted to perform simple seismic pattern recognition and first-break picking. These early models were limited by hardware constraints and the lack of large-scale digital datasets.
2000–2010:Iterative least-squares inversion remained the industry standard. Refinements in Full Waveform Inversion (FWI) improved the resolution of subsurface imaging, though the process remained computationally expensive and prone to local minima issues.
2012–2015:The resurgence of deep learning, particularly the success of Convolutional Neural Networks (CNNs) in computer vision, prompted a re-evaluation of neural networks for geophysical applications.
2016–2019:Researchers began using the Open Seismic Repository and synthetic datasets to train Encoder-Decoder architectures for seismic denoising and velocity model estimation.
2020–Present:Modern CNNs and Transformers are utilized to perform end-to-end seismic inversion. These models are now routinely benchmarked against traditional iterative solvers using regional survey data and verified against borehole logs.

Background

Surface Wave Hub operates at the intersection of empirical geophysics and practical structural engineering. The study of seismic surface waves involves the precise measurement of ground motion signatures to infer the internal characteristics of solid media. In a heterogeneous environment, seismic waves do not travel in a straight line or at a constant speed; instead, they reflect, refract, and attenuate based on the interfaces they encounter. Rayleigh waves, which move in an elliptical motion, and Love waves, which exhibit horizontal shear, are particularly sensitive to the shallow subsurface (the first 100 meters).

Capturing these signals requires the deployment of high-sensitivity sensors. Geophones convert ground motion into voltage, while accelerometers measure the rate of change of velocity. Once collected, this raw data is subjected to spectral analysis. The goal is to produce dispersion curves, which plot velocity against frequency. Inverting these curves—solving for the physical properties that would produce such a curve—is the core challenge. Whether through traditional mathematics or neural networks, the objective remains the same: to produce an accurate 2D or 3D image of the subsurface for applications ranging from earthquake hazard assessment to the non-destructive testing of civil infrastructure like bridges and foundations.

Traditional Least-Squares Inversion: Mechanics and Limitations

Traditional seismic inversion typically employs a least-squares optimization framework. This is an iterative process where an initial geological model is proposed, and synthetic seismic data is generated from that model using the wave equation. The difference between the synthetic data and the actual field observations is calculated as a residual or "misfit." The algorithm then calculates the gradient of this misfit to determine how the model parameters should be adjusted to reduce the error in the next iteration.

This method is mathematically rigorous and honors the underlying physics of wave propagation. However, it suffers from several practical drawbacks. First, the computational cost is immense. Each iteration requires solving the forward wave equation, which involves complex partial differential equations. For large-scale 3D surveys, this can take weeks of supercomputer time. Second, the method is highly sensitive to the "starting model." If the initial guess is too far from reality, the algorithm may converge on a local minimum—a mathematically plausible but geologically incorrect solution. This phenomenon, often called "cycle skipping," is a primary hurdle in low-frequency seismic data processing.

Deep Learning and Convolutional Neural Networks

Deep learning approaches seismic inversion as a regression problem. Instead of iteratively refining a single model, a neural network is trained on thousands or millions of examples consisting of seismic wavefields paired with their corresponding velocity models. Many of these training sets are drawn from the Open Seismic Repository or generated through high-fidelity synthetic simulations. Once the network—typically a Convolutional Neural Network (CNN) or a U-Net architecture—has learned the complex, non-linear mapping between the data and the model, it can perform inversion on new data in a single "forward pass."

CNNs are particularly effective because they are designed to recognize spatial patterns. In seismic data, these patterns represent the reflections and diffractions caused by geological layers. By using multiple layers of filters, a CNN can identify both large-scale structures and subtle lithological variations. Unlike least-squares methods, deep learning does not require a starting model and is significantly less susceptible to the local minima problem once the training phase is complete.

Comparative Study: Latency and Accuracy

A central research focus involves comparing the performance of these two paradigms. Studies using datasets from the Open Seismic Repository have revealed a stark contrast in computational latency. While a traditional iterative inversion may require several hours to process a 2D seismic profile, a pre-trained deep learning model can achieve a similar result in fractions of a second. This near-instantaneous inference makes deep learning attractive for real-time applications, such as monitoring the integrity of a bridge during a seismic event or detecting utility voids during active excavation.

However, accuracy remains a point of contention. Least-squares methods generally provide higher fidelity for fine-scale geological features, provided the physics of the wave propagation is perfectly understood and modeled. Deep learning models can sometimes produce "hallucinations"—features in the subsurface image that are not actually present but resemble patterns the network saw during its training phase. To mitigate this, researchers are increasingly looking toward "physics-informed neural networks" (PINNs), which incorporate the wave equation into the loss function of the machine learning model, ensuring that the predictions remain physically plausible.

Verification via Borehole Log Data

The ultimate test of any inversion algorithm is its performance against "ground truth." In geophysics, this ground truth is provided by borehole logs. A borehole log is a direct measurement of subsurface properties obtained by drilling a hole and lowering sensors to record density, gamma radiation, and acoustic velocity at various depths. By comparing the predictions of a deep learning model with the data from a nearby borehole, researchers can quantify the model's reliability.

Regional surveys documented in the Surface Wave Hub archives show that while deep learning can accurately predict the general trend of lithological boundaries, it may struggle with abrupt, high-contrast interfaces where traditional least-squares methods excel. Conversely, in areas with significant noise or missing data, deep learning models often show greater robustness, as they can rely on learned statistical priors to fill in the gaps that would cause a traditional iterative solver to fail.

Applications in Non-Destructive Testing and Infrastructure

The practical application of these algorithms extends beyond pure geology into the area of civil engineering. In non-destructive testing (NDT), the dispersion curves of induced surface waves are analyzed to detect internal flaws in infrastructure. For instance, in a concrete bridge foundation, an anomaly in the Rayleigh wave velocity might indicate a void or a zone of delamination. Traditional least-squares inversion has been the standard for these assessments, but the need for rapid, on-site diagnostics is driving the adoption of machine learning.

For shallow subsurface characterization—such as detecting buried utilities or abandoned tunnels—the interpretation of microtremor data is critical. Microtremors are low-amplitude ambient seismic vibrations caused by human activity or atmospheric conditions. Because these signals are weak and often incoherent, traditional inversion methods struggle to extract meaningful information. Deep learning models, trained to distinguish signal from noise in highly chaotic environments, offer a promising solution for mapping urban underground spaces without the need for destructive excavation.

Conclusion of Current Findings

The integration of deep learning into seismic inversion does not represent the end of traditional least-squares methods but rather an expansion of the geophysical toolkit. Current research suggests a hybrid approach may be the most effective: using deep learning to provide a highly accurate and rapid "initial guess" that is then refined using a few iterations of a physics-based least-squares solver. This combination leverages the speed and pattern recognition of neural networks while maintaining the mathematical rigor and physical consistency of traditional wavefield modeling. As Surface Wave Hub continues to document these advancements, the focus remains on the meticulous calibration of sensors and the development of inversion algorithms that can operate across the increasingly complex geological and engineered interfaces of the modern world.