10 November 2022 in Geophysics
Seismic velocity inversion plays a vital role in various applied seismology processes. A series of deep learning methods have been developed that rely purely on manually provided labels for supervision; however, their performances depend heavily on using large training data sets with corresponding velocity models. Because no physical laws are used in the training phase, it is usually challenging to generalize trained neural networks to a new data domain. To mitigate these issues, we have embedded a seismic forward modeling step at the end of a network to remap the inversion result back to seismic data and thus train the neural network through self-supervised loss, i.e., the misfit between the network input and output. As a result, we eliminate the need for many labeled velocity models, and physical laws are introduced when back-propagating gradients through the seismic forward modeling step. We verify the effectiveness of our approach through comprehensive experiments on synthetic data sets, where self-supervised learning outperforms the fully supervised approach, which accesses much more labeled data. The superior performance is even more significant when compared with a new data domain that has velocity models with faults and more geologic layers. Finally, in case of unknown and more complex data types, we develop a network-constrained full-waveform inversion (FWI) method. This method refines the initial prediction of the network by iteratively optimizing network parameters other than the velocity model, as found with the conventional FWI method, and demonstrates clear advantages in terms of interface and velocity accuracy. With these measures (self-supervised learning and network-constrained FWI), our physics-driven self-supervised learning system successfully mitigates issues such as the dependence on large labeled data sets, the absence of physical laws, and the difficulty in adapting to new data domains.
31 October 2023 in Geophysics
Tunnel and underground engineering construction often encounter unfavorable geology, leading to disasters such as water and mud inrushes and landslides. To prevent geologic hazards, it is important to look ahead and predict the location and distribution of adverse geology ahead of the tunnel face. This process is known as seismic forward prospecting in tunnels, and it typically requires an accurate calculation of velocity. Seismic waveform inversion methods based on deep learning have demonstrated potential in estimating velocity from synthetic seismic data. However, the superiority of these methods over traditional ones on field data is still an area of active research. Here, the Pearl River Delta Water Resources Allocation Project in China is used as an example to develop a self-supervised learning waveform inversion method for building a reliable velocity distribution in front of the tunnel. By introducing the background velocity as large-scale information and implementing multiscale loss functions, the previous self-supervised learning inversion method on synthetic data is improved. In addition, the corresponding network-based workflow for field data is developed. To demonstrate the effectiveness of our method, a comparison is conducted with practical tunneling exposure, for which the low-velocity zone corresponds with the fault-fractured zones and the water-flowing zones. This indicates that the results obtained from our method can be used as geologic guidance for safe tunneling practices. In the end, the applicability and disadvantages of our deep-learning inversion method for seismic forward prospecting in tunnels are discussed.
14 October 2026 in SSRN Electronic Journal
Multichannel analysis of surface waves (MASW) is a widely used approach for characterizing near-surface shear-wave velocity (Vs) structure. Conventional MASW generally estimates one-dimensional Vs profiles from surface-wave dispersion information, including both fundamental and higher-mode dispersion, making direct reconstruction of laterally varying two-dimensional structures less straightforward. This study presents Modal Dispersion Network (ModalDispNet), an end-to-end deep learning framework that reconstructs a two-dimensional Vs section directly from a single-shot record. The shot gather is first reorganized into a sliding-window dispersion stack of 13 images indexed by source-receiver offset, which encodes both the lateral velocity variation and the physical transition from fundamental-mode-dominated near-offset dispersion to higher-mode-rich far-offset dispersion. A window projection then constrains each window to reconstruct only the subsurface region it represents. Analytically computed Rayleigh-wave dispersion curves are used to build energy-modulated, per-mode masks that supervise four serial-parallel modal attention branches, enabling the network to separate and weight fundamental and higher-mode energy according to surface-wave physics. A bidirectional semi-supervised mechanism transfers the mask knowledge learned on horizontally layered models to laterally varying models. Numerical experiments on synthetic models, together with comparisons against CNN and U-Net, demonstrate that ModalDispNet can recover subsurface interfaces and lateral velocity variations, with lower velocity errors than the CNN and U-Net baselines, showing promise for near-surface 2D Vs imaging.