Machine Learning for Automated Dispersion Curve Picking: A Survey of Recent Progress

Dispersion curve picking serves as a fundamental step in the interpretation of surface wave data, specifically for Multi-channel Analysis of Surface Waves (MASW) and Passive Surface Wave methods. Between 2018 and 2024, the geophysical community experienced a significant transition from manual or semi-automated picking methods toward fully automated systems driven by deep learning. This shift was motivated by the need to process increasingly large seismic datasets generated by dense sensor arrays and fiber-optic distributed acoustic sensing (DAS) systems.

The traditional workflow involves transforming time-space (t-x) domain seismic records into spectral images, such as frequency-phase velocity (f-c) or frequency-wavenumber (f-k) plots. In these images, surface wave energy appears as high-amplitude ridges. Historically, analysts manually selected the centers of these ridges—a process prone to subjectivity and significant time requirements. Modern Convolutional Neural Networks (CNNs) have redefined this task as an image segmentation or classification problem, allowing for the rapid extraction of dispersion curves with high consistency and minimal human intervention.

At a glance

• Temporal Shift:The move from conventional signal processing to deep learning dominance occurred largely between 2018 and 2024.
• Architecture Focus:Convolutional Neural Networks (CNNs), particularly U-Net and ResNet variants, represent the standard for ridge detection in spectral images.
• Data Sources:The Stanford Exploration Project (SEP) and various synthetic datasets provide the primary training and benchmarking grounds for new algorithms.
• Efficiency Gains:AI-driven workflows can reduce interpretation time from hours to seconds per seismic record while maintaining 90-95% agreement with expert picks.
• Physical Constraints:Current research focuses on integrating physical laws into neural network loss functions to prevent the picking of non-physical velocity jumps.

Background

Surface wave dispersion occurs because different frequencies of seismic waves penetrate to different depths. In a stratified medium where velocity varies with depth, these frequencies travel at different phase velocities. By measuring these velocities—expressed as a dispersion curve—geophysicists can invert the data to determine the shear-wave velocity (Vs) profile of the subsurface. This profile is critical for site classification, earthquake hazard analysis, and infrastructure assessment.

For decades, the standard approach involved the f-k transform or the phase-shift method to create a power spectrum. The analyst would identify the fundamental mode (and occasionally higher modes) by tracking the maxima in the spectrum. This manual step became a bottleneck as seismic surveys evolved from 12-channel geophone spreads to thousands of channels in high-resolution urban imaging. Furthermore, noise from traffic or industrial activity often obscures the dispersion energy, requiring experts to make subjective judgments that are difficult to replicate.

The Transition to Convolutional Neural Networks (2018–2024)

The application of computer vision techniques to geophysical spectral images gained momentum in 2018. Early researchers demonstrated that CNNs could recognize the patterns of energy ridges even in the presence of significant background noise. Unlike traditional threshold-based automation, which often fails when the signal-to-noise ratio (SNR) is low, CNNs are capable of learning the contextual shape of a dispersion curve, allowing them to interpolate through areas of low energy.

U-Net Architectures and Image Segmentation

A prominent development in this period was the adaptation of the U-Net architecture. Originally designed for medical image segmentation, the U-Net consists of an encoder path to capture context and a symmetric decoder path to enable precise localization. In the context of surface waves, the spectral image is treated as a single-channel input, and the network outputs a binary mask or a probability map highlighting the dispersion ridge. This approach allows the simultaneous identification of multiple modes (fundamental and higher order) without requiring the user to specify the number of modes beforehand.

Benchmark Performance with SEP Archives

The Stanford Exploration Project (SEP) archives have served as a vital resource for validating these automated workflows. By comparing AI-picked curves against legacy picks within the SEP datasets, researchers established that deep learning models could achieve a mean absolute error (MAE) comparable to the variance between two human experts. In studies conducted between 2021 and 2023, automated systems outperformed traditional f-k analysis in robustness, particularly when dealing with aliased data or records with missing traces.

Verification Protocols and Physical Constraints

As the adoption of AI grew, concerns regarding the "black box" nature of neural networks led to the development of rigorous verification protocols. Because a CNN might pick a curve that is mathematically plausible in an image but physically impossible for a geological medium, developers began implementing Physics-Informed Neural Networks (PINNs).

Consistency with Elastic Theory

Verification involves checking that the picked phase velocities do not violate the fundamental limits of elastic wave propagation. For example, in a normally dispersive medium, phase velocity typically decreases with increasing frequency. Algorithms are now often layered with post-processing filters or constrained loss functions that penalize non-monotonic behavior or sudden velocity spikes that cannot be explained by the underlying lithology. This ensures that the output is suitable for the subsequent inversion process.

Standardized Quality Control (QC)

Modern workflows include a QC stage where the AI-predicted curve is used to generate a synthetic seismogram, which is then cross-correlated with the original field data. A high degree of correlation provides empirical evidence that the picked curve accurately represents the physical wavefield. This feedback loop is essential for building trust in automated systems within the engineering and regulatory communities.

What researchers emphasize

While the technical feasibility of automated picking is well-established, there is an ongoing discussion regarding the generalization of models. A model trained on high-frequency MASW data for shallow foundations may perform poorly on low-frequency microtremor data for deep crustal studies. Researchers emphasize that the "data gap" between synthetic training sets and heterogeneous field conditions remains the primary challenge. While synthetic data can be generated in infinite quantities, it often lacks the complex noise characteristics found in real-world urban environments.

Another point of emphasis is the role of the analyst. Rather than replacing the geophysicist, AI is increasingly viewed as a pre-processing tool that handles the bulk of the repetitive labor, allowing the human expert to focus on interpreting anomalous zones or complex multi-mode interference patterns that the AI may flag as uncertain.

Technological Impact on Non-Destructive Testing

The automation of dispersion curve extraction has had a profound impact on the non-destructive testing (NDT) of engineered structures. For bridges and tunnels, where rapid assessment is required to minimize service disruptions, automated surface wave analysis allows for near-real-time monitoring of structural integrity. By analyzing the dispersion curves of waves induced by controlled sources or ambient vibration, systems can now detect delamination or void formation with significantly higher throughput than was possible in the pre-2018 era.

Integration with Distributed Acoustic Sensing

The emergence of DAS technology, which utilizes fiber-optic cables as continuous sensors, has produced datasets of unprecedented scale. Traditional manual picking is impossible for the millions of traces generated by DAS. The survey of recent progress confirms that without the parallel development of CNN-based automated picking, the utility of DAS for surface wave characterization would be severely limited. The cooperation between high-density sensing and AI-driven analysis represents the current frontier in empirical study and practical application within the field.