Data Misfit Function Research Articles

Deep-unrolling architecture for image-domain least-squares migration

Deep-image prior (DIP) is a novel approach to solving ill-posed inverse problems whose solution is parameterized with an untrained deep neural network and cascaded with the forward modeling operator. A key component to the success of such a method is represented by the choice of the network architecture, which must act as a natural prior to the inverse problem at hand and provide a strong inductive bias toward the desired solution. Inspired by the close link between neural networks and iterative algorithms in classical optimization, we apply an unrolled version of the gradient descent (GD) algorithm as our DIP network architecture, denoted as the deep-unrolling (DU) architecture. Each layer of the unrolled network is comprised of two parts: the first part corresponds to the GD step of the data-fidelity term, whereas the second part, formed by a six-layer convolutional neural network (CNN), plays the role of a regularizer in the associated objective function. The developed DU architecture is applied to the problem of image-domain least-squares migration (IDLSM) to invert migrated seismic images for their underlying reflectivity and is denoted as DU-IDLSM. As such, the DU architecture parameterizes the reflectivity, and the input of each layer of the unrolled network is the reflectivity at the previous layer. Similar to the classical DIP approach, the parameters of the DU architecture are optimized in an unsupervised fashion by minimizing the data misfit function itself. Through experiments with a part of the Sigsbee2A model and a marine field data set, we test the effectiveness of the DU-IDLSM approach and highlight two key benefits. First, the DU architecture can effectively regularize the inversion process, resulting in reflectivity estimates with fewer artifacts and higher image resolution than those produced by conventional IDLSM approaches. Second, we indicate that by including dropout layers in the CNN architecture, DU-IDLSM can produce a qualitative measure of the uncertainty associated with the least-squares migration process.

Constrained nonlinear amplitude-variation-with-offset inversion for reservoir dynamic changes estimation from time-lapse seismic data

We develop a novel elastic amplitude-variation-with-offset (AVO) inversion process to estimate the fluid saturation and effective pressure or variations in these properties from time-lapse seismic data sets. These changes occur in oil and gas reservoirs caused by fluid injection or hydrocarbon production that leads to changes in the elastic wave properties, reflectivity, and seismic response. Our method is based on a seismic forward model that consists of a linearized AVO equation and a rock-physics model. The AVO equation links the elastic wave properties to seismic reflection amplitudes, whereas the rock-physics model maps the saturation and pressure into seismic velocities and density. The inversion approach relies on the gradient descent technique to estimate the unknown variables by searching for the minimum of the least-squares misfit between the observed and modeled data. The first-order gradient equations of the least-squares data misfit function with respect to effective pressure and water saturation are derived by using the adjoint-state method and the chain rule. The optimization method used to minimize the misfit function and obtain the best optimal solution is a limited-memory quasi-Newton algorithm. This inversion process allows us to incorporate prior constraints by using the logistic function to map the model variables to a bounded range. To achieve a stable solution to ill-posed inversion problems, optimal regularization weights are applied. The application of the developed workflow on 1D synthetic and real well-log data from the Edvard Grieg oil field simulating various saturation-pressure conditions during production demonstrates the validity of the approach with different noise levels. The inversion is then applied to a 2D synthetic data set modeled from the reservoir model of the Smeaheia field, a potential site for a large-scale offshore CO2 storage field located in the North Sea. Our results illustrate that our inversion method efficiently and accurately estimates reservoir saturation and pressure variations.

Spherical Planting Inversion of GRAIL Data

In large-scale potential field data inversion, constructing the kernel matrix is a time-consuming problem with large memory requirements. Therefore, a spherical planting inversion of Gravity Recovery and Interior Laboratory (GRAIL) data is proposed using the L1-norm in conjunction with tesseroids. Spherical planting inversion, however, is strongly dependent on the correct seeds’ density contrast, location, and number; otherwise, it can cause mutual intrusion of anomalous sources produced by different seeds. Hence, a weighting function was introduced to limit the influence area of the seeds for yielding robust solutions; moreover, it is challenging to set customized parameters for each seed, especially for the large number of seeds used or complex gravity anomalies data. Hence, we employed the “shape-of-anomaly” data-misfit function in conjunction with a new seed weighting function to improve the spherical planting inversion. The proposed seed weighting function is constructed based on the covariance matrix for given gravity data and can avoid manually setting customized parameters for each seed. The results of synthetic tests and field data show that spherical planting inversion requires less computer memory than traditional inversion. Furthermore, the proposed seed weighting function can effectively limit the seed influence area. The result of spherical planting inversion indicates that the crustal thickness of Mare Crisium is about 0 km because the Crisium impact may have removed all crust from parts of the basin.

Determining kernels in linear viscoelasticity

In this work, we investigate the inverse problem of determining the kernel functions that best describe the mechanical behavior of a complex medium modeled by a general nonlocal viscoelastic wave equation. To this end, we minimize a tracking-type data misfit function under this PDE constraint. We perform the well-posedness analysis of the state and adjoint problems and, using these results, rigorously derive the first-order sensitivities. Numerical experiments in a three-dimensional setting illustrate the method.

Lp-Norm Inversion of Gravity Data Using Adaptive Differential Evolution

As a popular population based heuristic evolutionary algorithm, differential evolution (DE) has been widely applied in various science and engineering problems. Similar to other global nonlinear algorithms, such as genetic algorithm, simulated annealing, particle swarm optimization, etc., the DE algorithm is mostly applied to resolve the parametric inverse problem, but has few applications in physical property inversion. According to our knowledge, this is the first time DE has been applied in obtaining the physical property distribution of gravity data due to causative sources embedded in the subsurface. In this work, the search direction of DE is guided by better vectors, enhancing the exploration efficiency of the mutation strategy. Besides, to reduce the over-stochastic of the DE algorithm, the perturbation directions in mutation operations are smoothed by using a weighted moving average smoothing technique, and the Lp-norm regularization term is implemented to sharpen the boundary of density distribution. Meanwhile, in the search process of DE, the effect of Lp-norm regularization term is controlled in an adaptive manner, which can always have an impact on the data misfit function. In the synthetic anomaly case, both noise-free and noisy data sets are considered. For the field case, gravity anomalies originating from the Shihe iron ore deposit in China were inverted and interpreted. The reconstructed density distribution is in good agreement with the one obtained by drill-hole information. Based on the tests in the present study, one can conclude that the Lp-norm inversion using DE is a useful tool for physical property distribution using gravity anomalies.

Marchenko multiple elimination and full-wavefield migration in a resonant pinch-out model

Marchenko multiple elimination schemes are able to attenuate all internal multiple reflections in acoustic reflection data. These can be implemented with and without compensation for two-way transmission effects in the resulting primary reflection data set. The methods are fully automated and run without human intervention, but they require the data to be properly sampled and preprocessed. Even when several primary reflections are invisible in the data because they are masked by overlapping multiples, such as in the resonant wedge model, all missing primary reflections are restored and recovered with the proper amplitudes. Investigating the amplitudes in the primary reflections after multiple elimination with and without compensation for transmission effects shows that transmission effects are properly accounted for in a constant-velocity model. When the layer thickness is one quarter of the wavelength at the dominant frequency of the source wavelet, the methods cease to work properly. Full-wavefield migration relies on a velocity model and runs a nonlinear inversion to obtain a reflectivity model, which results in the migration image. The primary reflections that are masked by interference with multiples in the resonant wedge model are not recovered. In this case, minimizing the data misfit function leads to the incorrect reflector model even though the data fit is optimal. This method has much lower demands on data sampling than the multiple elimination schemes, but it is prone to getting stuck in a local minimum even when the correct velocity model is available. A hybrid method that exploits the strengths of each of these methods could be worth investigating.

Source characterization of airborne pollutant emissions by hybrid metaheuristic/gradient-based optimization techniques

We propose a methodology to estimate single and multiple emission sources of atmospheric contaminants. It combines hybrid metaheuristic/gradient-descent optimization techniques and Tikhonov-type regularization. The dispersion problem is solved by the Galerkin/Least-squares finite element formulation, which allows more realistic modeling. The accuracy of the proposed inversion model is tested under different contexts with experimental data. To identify single and multiple emissions, we use experimental field data. We consider different configurations for both the Tikhonov-type functional and optimization techniques. Several single and composite data misfit functions are tested. We also use a discrepancy-based choice rule for the regularization parameter. The resulting inversion tool is highly versatile and presents accurate results under different contexts with a competitive computational cost.

An accurate strategy to retrieve multiple source emissions in the atmosphere

Abstract We propose a new methodology to retrieve multiple point emission sources in the atmosphere. To allow more realistic modelling, the dispersion problem is solved by a stabilized finite element formulation. The simultaneous estimation of the source strengths, their locations and the total number of sources is made employing Tikhonov-type regularization, with a composite data misfit function. The corresponding minimization of the Tikhonov-type functional is carried out by a combination of a splitting strategy algorithm with a gradient-descent method. We illustrate the accuracy of the proposed technique by using Fusion Field Trials 2007 (FFT07) data.

Gravity Inversion of Blocky Basement Relief Using L0 Norm Constraint with Exponential Density Contrast Variation

Depicting the basement relief of a sedimentary basin with gravity data is vital to predicting the hydrocarbon potential of a sedimentary basin and guiding exploration work. We have developed a gravity inversion method to estimate the depth to a blocky basement of a sedimentary basin. The basement rocks are assumed to be homogeneous and have uniform density, while the density of the sediment over the basement increases exponentially with depth. The density contrast between the sediment and the basement at the surface varies horizontally. The decay factor of density contrast is also nonuniform. The sediment above the basement is divided into vertically juxtaposed prisms, and the depth of the bottom of each prism represents the depth to the basement and is the parameter to be estimated. The L0 norm is introduced to limit the gradient of the parameter vector to obtain the model constraint function. We then establish the objective function for inversion by combining the gravity data misfit function, the known depth constraint function, and the model constraint function. The inversion is performed by minimizing the objective function using the nonlinear conjugate gradient algorithm. The inversion method is evaluated using a 2D and a 3D sedimentary basin model. The results show that our proposed method is capable of delineating the blocky basement relief of a sedimentary basin, and the result is sharper than that obtained using the L1 norm constraint. The method is applied to real data from the western part of the Zhu 1 depression in the Pearl River Mouth Basin, northern South China Sea. The solution reveals a strongly faulted basement, which is in accordance with the known tectonic information indicating the basin is a fully developed graben.

Topography Least-Squares Reverse-Time Migration Based on Adaptive Unstructured Mesh

Least-squares reverse-time migration (LSRTM) attempts to invert the broadband-wavenumber reflectivity image by minimizing the residual between observed and predicted seismograms via linearized inversion. However, rugged topography poses a challenge in front of LSRTM. To tackle this issue, we present an unstructured mesh-based solution to topography LSRTM. As to the forward/adjoint modeling operators in LSRTM, we take a so-called unstructured mesh-based “Grid Method.” Before solving the two-way wave equation with the Grid Method, we prepare for it a velocity-adaptive unstructured mesh using a Delaunay Triangulation plus Centroidal Voronoi Tessellation algorithm. The rugged topography acts as constraint boundaries during mesh generation. Then, by using the adjoint method, we put the observed seismograms to the receivers on the topography for backward propagation to produce the gradient through the cross-correlation imaging condition. We seek the inverted image using the conjugate gradient method during linearized inversion to linearly reduce the data misfit function. Through the 2D SEG Foothill synthetic dataset, we see that our method can handle the LSRTM from rugged topography.

Improving the least-squares image by using angle information to avoid cycle skipping

In the past few decades, the least-squares reverse time migration (LSRTM) algorithm has been widely used to enhance images of complex subsurface structures by minimizing the data misfit function between the predicted and observed seismic data. However, this algorithm is sensitive to the accuracy of the migration velocity model, which, in the case of real data applications (generally obtained via tomography), always deviates from the true velocity model. Therefore, conventional LSRTM faces a cycle-skipping problem caused by a smeared image when using an inaccurate migration velocity model. To address the cycle-skipping problem, we have introduced an angle-domain LSRTM algorithm. Unlike the conventional LSRTM algorithm, our method updates the common source-propagation angle image gathers rather than the stacked image. An extended Born modeling operator in the common source-propagation angle domain is was derived, which reproduced kinematically accurate data in the presence of velocity errors. Our method can provide more focused images with high resolution as well as angle-domain common-image gathers (ADCIGs) with enhanced resolution and balanced amplitudes. However, because the velocity model is not updated, the provided image can have errors in depth. Synthetic and field examples are used to verify that our method can robustly improve the quality of the ADCIGs and the finally stacked images with affordable computational costs in the presence of velocity errors.

Joint inversion methods with relative density offset correction for muon tomography and gravity data, with application to volcano imaging

Muography is a relatively new geophysical imaging method that uses muons to provide estimates of average densities along particular lines of sight. Muography can only see above the horizontal elevation of the detector and it is therefore attractive to attempt a joint inversion of muography data with gravity data, which is also responsive to density but generally requires combination with another geophysical data set to overcome issues related to non-uniqueness and poor depth resolution. Some previous work has investigated this joint inverse problem and demonstrated the potential improvements to be gained by jointly inverting muography and gravity data. However, there has yet to be a thorough investigation of different numerical approaches for formulating the joint inverse problem. Particularly important is how to account for the fact that even though the two data types are sensitive to the same physical quantity, density, they respond through different response functions. Moreover, the two measurements are affected by different systematic uncertainties that are difficult to model. In this work, we considered an approximation where the two density quantities, inferred from the two data types, can be related by an unknown scalar offset. We considered various existing and new joint inversion methods that might solve this problem and we applied them to a synthetic volcano imaging scenario based on the Puy de Dome volcano in the Central Massif region of France. We used unstructured meshes in our modelling to adequately honour the significant topography in that scenario. Our experiments indicated that the most successful joint inversion method for this type of geological scenario was one in which the data misfit function is reformulated to automatically determine the best-fitting offset following a least-squares minimization argument. However, other approaches showed merit and we suggest several of the investigated methods be applied and compared for any specific joint inversion scenario.

Appraising structural interpretations using seismic data — Theoretical elements

Structural interpretation of seismic images can be highly subjective, especially in complex geologic settings. A single seismic image will often support multiple geologically valid interpretations. However, it is usually difficult to determine which of those interpretations are more likely than others. We have referred to this problem as structural model appraisal. We have developed the use of misfit functions to rank and appraise multiple interpretations of a given seismic image. Given a set of possible interpretations, we compute synthetic data for each structural interpretation, and then we compare these synthetic data against observed seismic data; this allows us to assign a data-misfit value to each structural interpretation. Our aim is to find data-misfit functions that enable a ranking of interpretations. To do so, we formalize the problem of appraising structural interpretations using seismic data and we derive a set of conditions to be satisfied by the data-misfit function for a successful appraisal. We investigate vertical seismic profiling (VSP) and surface seismic configurations. An application of the proposed method to a realistic synthetic model shows promising results for appraising structural interpretations using VSP data, provided that the target region is well-illuminated. However, we find appraising structural interpretations using surface seismic data to be more challenging, mainly due to the difficulty of computing phase-shift data misfits.

Full-waveform inversion for microseismic events using sparsity constraints

Full-waveform inversion (FWI) is a powerful tool that can be used to invert for microseismic event locations and the source signature because it can exploit the complete waveform information. We have developed an algorithm to invert for a spatio-temporal source function that encapsulates microseismic events with spatially localized or distributed locations and source signatures. The algorithm does not require assumptions to be made about the number or type of sources; however, it does require that the velocity model is close to the true subsurface velocity. We reformulate the conventional FWI algorithm based on the [Formula: see text]-norm data-misfit function by adding sparsity constraints using a sparsity promoting [Formula: see text]-norm as an additional regularization term to get more focused and less noise-sensitive event locations. The Orthant-Wise Limited-memory quasi-Newton algorithm is used to solve the optimization problem. It inherits the advantageous (fast convergence) properties of the limited memory Broyden-Fletcher-Goldfarb-Shanno method and can easily overcome the nondifferentiability of [Formula: see text]-norm at null positions. We determine the performance of the algorithm on noise-free and noisy synthetic data from the SEG/EAGE overthrust model.

3D gravity inversion of basement relief for a rift basin based on combined multinorm and normalized vertical derivative of the total horizontal derivative techniques

The basement of a rift sedimentary basin, often possessing smooth and nonsmooth shapes, is not easily recovered from gravity data by current inversion methods. We have developed a new 3D gravity inversion method to estimate the basement relief of a rift basin. In the inversion process, we have established the objective function by combining the gravity data misfit function, the known depth constraint function, and the model constraint function composed of the [Formula: see text]-norm and [Formula: see text]-norm, respectively. An edge recognition technology based on the normalized vertical derivative of the total horizontal derivative for gravity data is adopted to recognize the discontinuous and continuous parts of the basin and combine the two inputs to form the final model constraint function. The inversion is conducted by minimizing the objective function by the nonlinear conjugate gradient algorithm. We have developed two applications using synthetic gravity anomalies produced from two synthetic rift basins, one with a single graben and one with six differently sized grabens. The test results indicate that the inversion method is a feasible technique to delineate the basement relief of a rift basin. The inversion method is also tested on field data from the Xi’an depression in the middle of the Weihe Basin, Shaanxi Province, China, and the result illustrates its effectiveness.

Inverse acoustic scattering using high-order small-inclusion expansion of misfit function

This article concerns an extension of the topological derivative concept for 3D inverse acoustic scattering problems involving the identification of penetrable obstacles, whereby the featured data-misfit cost function \begin{document} $\mathbb{J}$ \end{document} is expanded in powers of the characteristic radius \begin{document} $a$ \end{document} of a single small inhomogeneity. The \begin{document} $O(a^6)$ \end{document} approximation \begin{document} $\mathbb{J}_6$ \end{document} of \begin{document} $\mathbb{J}$ \end{document} is derived and justified for a single obstacle of given location, shape and material properties embedded in a 3D acoustic medium of arbitrary shape. The generalization of \begin{document} $\mathbb{J}_6$ \end{document} to multiple small obstacles is outlined. Simpler and more explicit expressions of \begin{document} $\mathbb{J}_6$ \end{document} are obtained when the scatterer is centrally-symmetric or spherical. An approximate and computationally light global search procedure, where the location and size of the unknown object are estimated by minimizing \begin{document} $\mathbb{J}_6$ \end{document} over a search grid, is proposed and demonstrated on numerical experiments, where the identification from known acoustic pressure on the surface of a penetrable scatterer embedded in a acoustic semi-infinite medium, and whose shape may differ from that of the trial obstacle assumed in the expansion of \begin{document} $\mathbb{J}$ \end{document} , is considered.

A Novel Tomographic Reconstruction Method Based on the Robust Student's t Function For Suppressing Data Outliers

Regularized iterative reconstruction methods in computed tomography can be effective when reconstructing from mildly inaccurate undersampled measurements. These approaches will fail, however, when more prominent data errors, or outliers, are present. These outliers are associated with various inaccuracies of the acquisition process: defective pixels or miscalibrated camera sensors, scattering, missing angles, etc. To account for such large outliers, robust data misfit functions, such as the generalized Huber function, have been applied successfully in the past. In conjunction with regularization techniques, these methods can overcome problems with both limited data and outliers. This paper proposes a novel reconstruction approach using a robust data fitting term which is based on the Student's t distribution. This misfit promises to be even more robust than the Huber misfit as it assigns a smaller penalty to large outliers. We include the total variation regularization term and automatic estimation of a scaling parameter that appears in the Student's t function. We demonstrate the effectiveness of the technique by using a realistic synthetic phantom and also apply it to a real neutron dataset.

Fast Nonlinear Generalized Inversion of Gravity Data with Application to the Three-Dimensional Crustal Density Structure of Sichuan Basin, Southwest China

Generalized inversion is one of the important steps in the quantitative interpretation of gravity data. With appropriate algorithm and parameters, it gives a view of the subsurface which characterizes different geological bodies. However, generalized inversion of gravity data is time consuming due to the large amount of data points and model cells adopted. Incorporating of various prior information as constraints deteriorates the above situation. In the work discussed in this paper, a method for fast nonlinear generalized inversion of gravity data is proposed. The fast multipole method is employed for forward modelling. The inversion objective function is established with weighted data misfit function along with model objective function. The total objective function is solved by a dataspace algorithm. Moreover, depth weighing factor is used to improve depth resolution of the result, and bound constraint is incorporated by a transfer function to limit the model parameters in a reliable range. The matrix inversion is accomplished by a preconditioned conjugate gradient method. With the above algorithm, equivalent density vectors can be obtained, and interpolation is performed to get the finally density model on the fine mesh in the model domain. Testing on synthetic gravity data demonstrated that the proposed method is faster than conventional generalized inversion algorithm to produce an acceptable solution for gravity inversion problem. The new developed inversion method was also applied for inversion of the gravity data collected over Sichuan basin, southwest China. The established density structure in this study helps understanding the crustal structure of Sichuan basin and provides reference for further oil and gas exploration in this area.

First-order indicators for the estimation of discrete fractures in porous media

Faults and geological barriers can drastically affect the flow patterns in porous media. Such fractures can be modelled as interfaces that interact with the surrounding matrix. We propose a new technique for the estimation of the location and hydrogeological properties of a small number of large fractures in a porous medium from given distributed pressure or flow data. At each iteration, the algorithm builds a short list of candidates by comparing fracture indicators. These indicators quantify at the first order the decrease of a data misfit function; they are cheap to compute. Then, the best candidate is picked up by minimization of the objective function for each candidate. Optimally driven by the fit to the data, the approach has the great advantage of not requiring remeshing, nor shape derivation. The stability of the algorithm is shown on a series of numerical examples representative of typical situations.

Three-dimensional gravity inversion using graph theory to delineate the skeleton of homogeneous sources

We developed a new strategy, based on graph theory concepts, to invert gravity data using an ensemble of simple point masses. Our method consisted of a genetic algorithm with elitism to generate a set of possible solutions. Each estimate was associated to a graph to solve the minimum spanning tree (MST) problem. To produce unique and stable estimates, we restricted the position of the point masses by minimizing the statistical variance of the distances of an MST jointly with the data-misfit function during the iterations of the genetic algorithm. Hence, the 3D spatial distribution of the point masses identified the skeleton of homogeneous gravity sources. In addition, our method also gave an estimation of the anomalous mass of the source. So, together with the anomalous mass, the skeleton could aid other 3D methods with promising geometric a priori parameters. Several tests with different values of regularizing parameter were made to bespeak this new regularizing strategy. The inversion results applied to noise-corrupted synthetic gravity data revealed that, regardless of promising starting models, the estimated distribution of point masses and the anomalous mass offered valuable information about the homogeneous sources in the subsurface. Tests on real data from a portion of Quadrilátero Ferrífero, Minas Gerais state, Brazil, were performed for complementary analysis of the proposed inversion method.