Machine learning, and specifically deep learning techniques applied to geophysical inverse problems, is an attractive subject, which has promising potential and, at the same time, presents some challenges in practical implementation. Some of the obstacles relate to scarce knowledge of the searched geological structures, a problem that can limit both the interpretability and the generalizability of the trained deep learning networks when applied to independent scenarios in real applications. Commonly used (physics-driven) least squares optimization methods, are very efficient local optimization techniques but require good starting models close to the correct solution to avoid local minima. We develop a hybrid workflow, which combines both approaches in a coupled physics-driven/deep learning inversion scheme. We exploit the benefits and characteristics of both inversion techniques to converge to solutions that typically outperform individual inversions results and bring the solution closer to the global minimum of a non-convex inverse problem. The completely data-driven and self-feeding procedure relies on a coupling mechanism between the two inversion schemes taking the form of penalty functions applied to the model term. Predictions from the deep learning network are used to constrain the least-square inversion while the feedback loop from inversion to the deep learning scheme consists of the network re-training with partial results obtained from inversion. The self-feeding process tends to converge to a common agreeable solution, which is the result of two independent schemes with different mathematical formalisms and different objective functions on data and model misfit. We demonstrate that the hybrid procedure is converging to robust and high-resolution resistivity models when applied to the inversion of both synthetic and field transient electromagnetic (TEM) data. We finally speculate that the developed procedure may be adopted to recast the way we solve inverse problems for several different disciplines.

Reservoir characterization and monitoring represent some of the most ambitious goals for geophysical methods. Several challenges are involved, including sensitivity to the parameter changes and resolution of the obtained results. Electromagnetic (EM) methods are attractive for reservoir applications due to the high sensitivity of the resistivity parameter to oil/water saturations. Crosswell EM and surface-to-borehole EM provide opportunities for reservoir monitoring. The EM inverse problem, however, is highly nonconvex and ill-posed so as to necessitate significant preconditioning in the form of a priori information and regularization that impact resolution. We explore the use of machine-learning (ML) techniques in the form of deep-learning neural networks for implementing EM-based reservoir monitoring coupled with a dynamic fluid flow simulator. A crosswell acquisition setup is modeled in the framework of a realistic water-alternating-gas reservoir simulation scenario for enhanced oil recovery. Several reservoir saturation instances are generated and converted into resistivity, and corresponding crosswell EM data are generated using an electric source and a multicomponent (electric-magnetic) receiver assemblage. The U-Net deep-learning network is modified for the purpose of training and validation in which saturation models and the corresponding EM data are used. We also test the sensitivity of the deep-learning inversion to multiple EM components, noise in the data, generalization problems, and 3D reconstruction ability in which we use 3D convolutional neural network layers. In all cases, ML inversion proves to be robust with good resilience to increased noise levels. Prediction results indicate excellent reconstruction capabilities with resolution comparable to the reservoir models used by the simulator. Our results suggest that ML inversion through deep learning can become an efficient approach to data-driven and physics-constrained reservoir monitoring in which the sensitivity of EM-based techniques to fluid saturations can be fully exploited without compromising the resolution and accuracy of the results.

We have developed a new framework for performing surface-consistent amplitude balancing and deconvolution of the near-surface attenuation response. Both approaches rely on the early arrival waveform of a seismic recording, which corresponds to the refracted or, more generally speaking, to the transmitted energy from a seismic source. The method adapts standard surface-consistent amplitude compensation and deconvolution to the domain of refracted/transmitted waves. A sorting domain specific for refracted energy is extended to the analysis of amplitude ratios of each trace versus a reference average trace to identify amplitude residuals that are inverted for surface consistency. The residual values are either calculated as a single scalar value for each trace or as a function of frequency to build a surface-consistent deconvolution operator. The derived operators are then applied to the data to obtain scalar amplitude balancing or amplitude balancing with spectral shaping. The derivation of the operators around the transmitted early arrival waveforms allows for deterministically decoupling the near-surface attenuation response from the remaining seismic data. The developed method is fully automatic and does not require preprocessing of the data. As such, it qualifies as a standard preprocessing tool to be applied at the early stages of seismic processing. Applications of the developed method are provided for a case in a complex, structure-controlled wadi, for a seismic time-lapse [Formula: see text] land monitoring case, and for an exploration area with high dunes and sabkhas producing large frequency-dependent anomalous amplitude responses. The new development provides an effective tool to enable better reservoir characterization and monitoring with land seismic data.

Land seismic velocity modeling is a difficult task largely related to the description of the near surface complexities. Full waveform inversion is the method of choice for achieving high-resolution velocity mapping but its application to land seismic data faces difficulties related to complex physics, unknown and spatially varying source signatures, and low signal-to-noise ratio in the data. Large parameter variations occur in the near surface at various scales causing severe kinematic and dynamic distortions of the recorded wavefield. Some of the parameters can be incorporated in the inversion model while others, due to sub-resolution dimensions or unmodeled physics, need to be corrected through data preconditioning; a topic not well described for land data full waveform inversion applications. We have developed novel algorithms and workflows for surface-consistent data preconditioning utilizing the transmitted portion of the wavefield, signal-to-noise enhancement by generation of CMP-based virtual super shot gathers, and robust 1.5D Laplace-Fourier full waveform inversion. Our surface-consistent scheme solves residual kinematic corrections and amplitude anomalies via scalar compensation or deconvolution of the near surface response. Signal-to-noise enhancement is obtained through the statistical evaluation of volumetric prestack responses at the CMP position, or virtual super (shot) gathers. These are inverted via a novel 1.5D acoustic Laplace-Fourier full waveform inversion scheme using the Helmholtz wave equation and Hankel domain forward modeling. Inversion is performed with nonlinear conjugate gradients. The method is applied to a complex structure-controlled wadi area exhibiting faults, dissolution, collapse, and subsidence where the high resolution FWI velocity modeling helps clarifying the geological interpretation. The developed algorithms and automated workflows provide an effective solution for massive full waveform inversion of land seismic data that can be embedded in typical near surface velocity analysis procedures.

No abstract