Geophysical inversion and machine learning both provide solutions for inverse problems in which we estimate model parameters from observations. Geophysical inversions such as impedance inversion, amplitude-variation-with-offset inversion, and traveltime tomography are commonly used in the industry to yield physical properties from measured seismic data. Machine learning, a data-driven approach, has become popular during the past decades and is useful for solving such inverse problems. An advantage of machine learning methods is that they can be implemented without knowledge of physical equations or theories. The challenges of machine learning lie in acquiring enough training data and selecting relevant parameters, which are essential in obtaining a good quality model. In this study, we compare geophysical inversion and machine learning approaches in solving inverse problems and show similarities and differences of these approaches in a mathematical form and numerical tests. Both methods aid in solving ill-posed and nonlinear problems and use similar optimization techniques. We take reflectivity inversion as an example of the inverse problem. We apply geophysical inversion based on the least-squares method and artificial neural networks as a machine learning approach to solve reflectivity inversion using 2D synthetic data sets and 3D field data sets. A neural network with multiple hidden layers successfully generates the nonlinear mapping function to predict reflectivity. For this inverse problem, we test different L1 regularizers for both approaches. L1 regularization alleviates effects of noise in seismic traces and enhances sparsity, especially in the least-squares method. The 2D synthetic wedge model and field data examples show that neural networks yield high spatial resolution.
30 November 2022 in Journal of Geophysical Research: Earth Surface
Small‐scale polar ice‐sheet composite structures can influence coarse‐grained climate models. For example, surface melting can hydro‐fracture the ice and may eventually lubricate the base. The drainage process will change the petrophysical properties of the subsurface and is predictable from high‐resolution seismic imaging. Ray‐based imaging methods, though limited by the approximations made, have mapped seismic velocities in 1D, but have yet to offer definitive information about the lateral heterogeneity. Here we use the wave‐equation‐based seismic imaging method to estimate shear‐wave velocities and to image an englacial reflector in 2D. We use 7‐day ambient‐noise data recorded by a linear array near the West Antarctic Ice Sheet (WAIS) Divide drilling site to perform the study. We obtain 2D vertical and horizontal shear‐wave velocity models and observe a lateral variation of the radial anisotropy along the acquisition line. The inverted velocity model is further used to estimate the thicknesses of firn‐air and firn layers. We also observe evident SH reflections and calculate the reflector image using robust zero‐offset imaging and a more precise reverse time migration imaging method. The imaged reflector is at about 1,700 m depth and is dipping to the southwest. We anticipate that a porosity change at that depth may cause this englacial seismic discontinuity near WAIS Divide. Our results demonstrate the feasibility of estimating the petrophysical properties of polar ice using seismic methods. We anticipate our work to be further used for understanding ice dynamics and thermodynamics, such as the stability of ice shelves, the retreat of the Antarctic ice sheet, and restoring the paleo‐sedimentary environment.
Subsurface structures play important roles in seismic ground motion, crustal hydrology, stability of the built environment, and more. Constraining temporal changes of subsurface shear wave velocity ( VS ) can provide useful information to all these topics and the growing field of hydrological monitoring with seismic velocity. Using borehole records at Garner Valley, CA, we estimate seasonal subsurface VS variations from impulse response functions (IRFs) of earthquake data (2005–2018) along with IRFs and cross‐correlation of cross‐hole experiment data (2015–2018). The inferred VS variations are up to ∼25% in the top 6 m and ∼10% at 2–5 m in depth. The VS variations correlate strongly with the water table depth changes, suggesting that the changes are mostly due to fluctuations of pore pressure in the shallow material. The shallow velocity changes alter the near‐surface conditions, can affect seismic hazard estimation, and may be improperly attributed to deeper processes without careful analysis.
27 September 2018 in Journal of Geophysical Research: Solid Earth
We find that higher‐order ambient seismic field correlations, specifically C3 (correlation of the coda of correlations) of long‐duration data sets from Southern California are symmetric despite the fact that the background wavefield is strongly directional. We demonstrate that this symmetry arises primarily from the structure of the C3 time series analysis rather than the diffuse character of the coda. For that reason, the resulting ambient‐field Green's function symmetry may not be diagnostic of correlation quality. We propose an alternative prestack processing scheme that constructs C3 correlations from each C1 (cross correlation) coda time window, but stacks the results only at the final step of processing. Prestack C3 leads to faster convergence and extracts arrivals at higher frequencies than standard poststack processing.
The waveforms generated by the 2014 Long Valley Caldera earthquake swarm recorded at station MLH show clear reflected waves that are often stronger than direct P and S waves. With waveform analyses, we discover that these waves are reflected at the top of a low‐velocity body, which may be residual magma from the ∼767 ka caldera‐forming eruption. The polarity of the reflection compared to direct P and S waves suggests that the reflection is SP waves (S from hypocenters to reflector and then convert to P waves to the surface). Because the wavefields are coherent among different earthquakes and hold high signal‐to‐noise ratios, we apply them to a wavefield migration method for imaging reflectors. The depth of the imaged magmatic system roof is around 8.2 km below the surface. This is consistent with previous studies. Even though we use only one station and waveforms from one earthquake swarm, the dense cluster of accurately located earthquakes provides a high‐resolution image of the roof.
Distributed acoustic sensing (DAS) has great potential for monitoring natural-resource reservoirs and borehole conditions. However, the large volume of data and complicated wavefield add challenges to processing and interpretation. In this study, we demonstrate that seismic interferometry based on deconvolution is a convenient tool for analyzing this complicated wavefield. We also show the limitation of this technique, in that it still requires good coupling to extract the signal of interest. We extract coherent waves from the observation of a borehole DAS system at the Brady geothermal field in Nevada. The extracted waves are cable or casing ringing that reverberate within a depth interval. These ringing phenomena are frequently observed in the vertical borehole DAS data. The deconvolution method allows us to examine the wavefield at different boundary conditions and separate the direct waves and the multiples. With these benefits, we can interpret the wavefields using a simple 1D string model and monitor its temporal changes. The velocity of this wave varies with depth, observation time, temperature, and pressure. We find the velocity is sensitive to disturbances in the borehole related to increasing operation intensity. The velocity decreases with rising temperature. The reverberation can be decomposed into distinct vibration modes in the spectrum. We find that the wave is dispersive and the fundamental mode propagates with a large velocity. This interferometry method can be useful for monitoring borehole conditions or reservoir property changes using densely-sampled DAS data.