Varying dimensional Bayesian acoustic waveform inversion for 1D semi-infinite heterogeneous media
Introduction
A subsurface earth model is composed of complex geophysical formations, which embodies a wide range of physical and mechanical heterogeneities. The aim of probabilistic inverse modeling is to reconstruct the random field structure of these subsurface properties, while accounting for various sources of uncertainty stemming from ground seismic geologic measurement errors, aleatory formations, and limited theoretical understanding about underground wave propagation.
In practice, one of the main goals of geophysical investigations is to identify the main geomorphological features of an unknown medium, meaning the spatial location and concentration of geological features such as the transition between materials, discontinuities and material concentrations [34]. In the case of a vertical 1D profile, this requires the definition of the location of the sharp transitions between layer interfaces (i.e. material properties), and the characterization of its corresponding mechanical properties.
In a horizontally stratified earth model, prior to making an inference about the likely variation of the elastic parameters within the geological layers, an assumption must be made concerning the number of layers in a certain depth range of interest. This assumption defines the dimensionality (i.e., the number of unknowns) of the inverse problem. In reality, however, such information is rarely available for the dimension and definition of the parameter space to be fixed. Consequently, fixing the number of layers based on an incorrect assumption results in an erroneous subsurface characterization.
To relax the hypothesis about the subsurface structure or spatial layering of the media's mechanical parameters, before the forward model is calibrated, it is proposed to define the number of layers, their locations, and their corresponding mechanical parameters, as random variables. From a Bayesian perspective, this set up is closely associated with probabilistic model selection, where a collection of models with varying number of parameters are presented for inversion, and the task is to select the models that most likely describe the experimental observations.
To illustrate the applicability of the proposed probabilistic calibration method, a one dimensional horizontally stratified medium is presented in terms of a Bayesian partition model [8]. The partition model divides the unknown material field into a number of non-overlapping regions, where each region represents a soil layer. Formulating the inverse medium problem in terms of a partition model may help reduce the dimensionality of the parameter space. Hence, regularizing the solution through specific prior distributions, which bears smoothness constraints (in a Bayesian inversion framework [45], [12], [20]), or regularization terms (in deterministic optimization problems [44], [36], [15]), is precluded.
A generalization of the simulation-based Markov Chain Monte Carlo methods, the so-called reversible jump [18], is used to sample the posterior distribution of varying dimensionality. In this setting, the Markov Chain is capable of undergoing dimension changes while moving among a number of candidate models. The key aspect of the reversible jump algorithm is the introduction of some auxiliary random variables to equalize the dimensionality of the parameter space across models. A series of one-to-one deterministic functions are defined to perform dimension matching such that the detailed balance condition is satisfied. Balance condition is the necessary condition for a Markov Chain to converge to the target density [7].
Since the introduction of Bayesian inference methods to the geophysical community, this has received a great deal of attention in a variety of applications [14], [16], [17], [45], [41], [43]. However, a limited number of studies have addressed the subsurface parameter estimation as a model selection problem, many of which resort to approximate methods to fulfill the model determination [10], [13]. The varying dimensional formulation was first introduced to the geophysics literature by Malinverno [31] in a 1D-DC resistivity sounding inversion, and later implemented in a number of geophysical probing inverse problems [39], [11], [1], [35].
The major impact of utilizing probabilistic calibration via a Bayesian approach is the systematic exploration of all combinations of the model parameters through a transparent definition of the impact of the participating uncertainty sources. During such exhaustive parameter exploration, a probability metric is defined to assess the likelihood of selecting sets of parameters that serve to approximate the experimental observations (likelihood); but also a probability density is defined to reflect the degree of knowledge on the model parameters (prior) before the model inversion. The combination of these two states of knowledge about the model of interest yields the following benefits: a transition from deterministic to probabilistic model parameters, assessment of the type and degree of correlation between the model parameters (e.g. linear or non-linear), measurement of the impact of the varying experimental observations (e.g. the effect of the number of observations on the prediction of confidence levels), assessment of the model performance, and most importantly, that among a number of competing models to choose from, selecting the best model which can describe the process that generated the observations. Key applications of the probabilistic subsurface imaging include integrated site investigation, since the recovery of geophysical mechanical parameters allows enhanced geomechanical characterization. [33], [34]. The varying parameter dimensionality is formulated through a Bayesian inversion, to populate likely configurations of a heterogeneous elastic medium occupying a semi-infinite domain.
A key defining characteristic of full waveform inversion is the numerical solution of the equations of motion. The governing forward physics consist of a 1D transient scalar acoustic wave propagation, where in order to model the semi-infinite extent of the physical domain, a perfectly matched layer (PML) is introduced at the truncation boundary to emulate the infiniteness of the earth structure [22]. A displacement–stress mixed finite element scheme is used for the numerical solution of the PML-augmented wave PDE.
Section snippets
Bayesian approach to inverse problems
An inverse problem is described as the process of estimating some characteristics or parameters of a physical system from a set of directly measurable responses of the system (observations). The vector of model parameters , and the vector of observable quantities are mapped through a forward model. The forward model operator G relies on a physical theory to predict the outcome of a possible experiment, or in other words to approximate the reality: , orwhere is the
Forward model
This section introduces the forward model used in the model inversion. The forward physics describes seismic vertical propagation of compressional waves within a horizontally stratified semi-infinite elastic earth, when the media is subjected to a uniform excitation over the surface. This problem can be treated as a one dimensional problem along the depth direction. In a computational setting, a major issue associated with this geo-acoustic inverse problem is to model the semi-infinite
Bayesian partition models
As described in the preceding section, our geo-acoustic inverse problem identifies the spatially dependent coefficient of a PML augmented wave equation , given the probed medium's response to a known excitation. This describes a functional inverse problem where the unknown quantity is a function of the spatial coordinate. Hence, in our Bayesian probabilistic setup, the inverse problem parameter comprises a real-valued random field (of infinite dimensionality), which assigns a
Bayesian model selection
The proposed Bayesian partition model is categorized within a special class of models namely variable dimension models. A variable dimension model is defined as a model whose number of unknowns is an unknown itself [9]. A Bayesian variable dimension model is composed of a set of plausible models , each reflecting a hypothesis about the data. Having K such competing models, it is desired to find the model that best describes the observations. Here, represents a k-layer subsurface
Prior elicitation
The first step in a Bayesian inference analysis setting is to specify prior densities to the model parameters (given the model representation is chosen). The prior distribution is basically a tool to summarize the initially available information on the process, and to quantify the uncertainty associated with this information. Selecting standard vague or non-informative priors is favored in this case in order to base the inference only on the experimental observations.
A number of
Posterior computation
A customary burden of using Bayes factors (Eq. (9)) is the computation of, oftentimes, high dimensional marginal likelihood integrals (Eq. (10)). To circumvent this difficulty, one may resort to alternative solutions such as Monte Carlo simulation based methods (e.g., pseudo-priors [4], [5]), or asymptotic approximation to Bayes factors (e.g., Schwartz's criteria also known as BIC) [42]. The latter is widely used in a variety of applications including geophysical modeling (e.g., see [13], [10],
Application to a synthetic case
The inversion scheme outlined in the preceding sections is applied to a synthetic data set to deduce the subsurface elastic properties of a soil model. We consider the horizontally stratified semi-infinite soil medium depicted in Fig. 4. The medium is modeled as a one-dimensional PML-truncated domain, with the regular domain extending to z=100 m, and the PML buffer zone thickness being 10 m. Fig. 4 illustrates the target wave velocity profile, which reflects sharp transitions between different
Results
In this section, we illustrate the applicability of the Bayesian varying dimensional inversion, and model determination using the methodology introduced in the preceding sections. The inversion is allowed for a maximum of 40 soil layers (up to the truncation interface), which indicates maximum number of 83 model unknowns. This maximum resolution is attributed to the frequency of the exerted load (maximum frequency 40 Hz). The simplest earth model is k=1, which corresponds to the state of a
Conclusions
This paper introduces a probabilistic calibration approach via a Bayesian formulation for the solution of inverse problems, defined by the random field characterization of heterogeneous media, for an acoustic one-dimensional velocity field with horizontally layered structure. A self-regularized varying structure model is formulated based on the notion of Bayesian partition models in order to parameterize the acoustic wave velocity random field. The method offers a reduced dimensional inversion
Acknowledgments
The authors would like to thank Simula School of Research and Innovation (Computational Geoscience) and Statoil for sponsoring the development of this work. Also, the constructive comments of our reviewers are gratefully acknowledged.
References (46)
- et al.
Bayesian partition modelling
Comput. Stat. Data Anal.
(2002) A multi-resolution, non-parametric, Bayesian framework for identification of spatially-varying model parameters
J. Comput. Phys.
(2009)- et al.
Receiver function inversion by transdimensional Monte Carlo sampling
Geophys. J. Int.
(2010) A new look at the statistical model identification
IEEE Trans. Autom. Control
(1974)- et al.
Accurate and stable Bayesian model selectionthe median intrinsic Bayes factor
Sankhya Ser B
(1998) - et al.
Bayesian model choice via Markov chain Monte Carlo methods
J. R. Stat. Soc. B Methodol.
(1995) - et al.
Marginal likelihood from the Metropolis–Hastings output
J. Am. Stat. Assoc.
(2001) - et al.
Automatic Bayesian curve fitting
J. R. Stat. Soc. B Methodol.
(1998) - Y. Fan, S. Sisson, Reversible jump MCMC, in: Handbook of Markov Chain Monte Carlo, Chapman and Hall/CRC, Boca Raton,...
- D.G.T. Denison, C.C. Holmes, B.K. Mallick, A.F.M. Smith, Bayesian methods for nonlinear classification and regression,...
Model selection and Bayesian inference for high-resolution seabed reflection inversion
J. Acoust. Soc. Am.
Trans-dimensional geoacoustic inversion
Geophys. J. Int.
Quantifying uncertainty in geoacoustic inversion. I. A fast Gibbs sampler approach
J. Acoust. Soc. Am.
Bayesian matched-field geo-acoustic inversion
Inverse Probl.
Bayesian estimation in seismic inversion Part Iprinciples
Geophys. Prospect.
A Newton-CG method for large-scale three-dimensional elastic full-waveform seismic inversion
Inverse Probl.
Bayesian estimation in seismic inversion Part IIuncertainty analysis
Geophys. Prospect.
Bayesian seismic waveform inversionparameter estimation and uncertainty analysis
J. Geophys. Res. Solid Earth
Reversible jump Markov Chain Monte Carlo Computation and Bayesian model determination
Biometrika
Bayesian prediction via partitioning
J. Comput. Graph. Stat.
Uncertainty analysis in matched-field geoacoustic inversions
J. Acoust. Soc. Am.
Ockham's Razor and Bayesian analysis
Am. Sci.
Mixed unsplit field perfectly matched layers for transient simulations of scalar waves in heterogeneous domains
Comput. Geosci.
Cited by (10)
An efficient PDE-constrained stochastic inverse algorithm for probabilistic geotechnical site characterization using geophysical measurements
2018, Soil Dynamics and Earthquake EngineeringCitation Excerpt :It had its birth with the development of Kalman filter [27]. Though computationally expensive, with the availability of faster computers, it is increasingly being used for stochastic inverse analysis in many fields of science and engineering, including geophysics [23,56,9,14,52,53,12,18,37,57,62,25,51,20,57,42,15,46,19]. However, the main novelty of this paper lies, in addition to first time application of such technique in full 3D geotechnical site characterization, in efficient probabilistic solution of the uncertain forward problem.
Seismic Attributes and Acoustic Inversion for Shallow Marine Slope Stratigraphy Analysis
2021, Proceedings of the Annual Offshore Technology ConferenceCharacterization of railway subgrade using an elastic wave full-waveform inversion method
2020, Journal of the Korean Society for RailwayProbabilistic seismic inversion for shallow site characterization on thehydrate ridge area
2020, Proceedings of the Annual Offshore Technology ConferenceBayesian stratigraphy integration of geophysical, geological, and geotechnical surveys data
2019, Proceedings of the Annual Offshore Technology Conference