Abstract
dc:descriptionIn the present Doctoral thesis, we develop and present the Cosimulated Iterative Spatial Resampling (C-ISR) method for stochastic solution of the inverse problem of hydrofacies characterization in a groundwater flow system and we establish the theoretical and mathematical background of the method. The development of the method stems from the need to characterize the geological formations before any other geostatistical estimation of hydrological parameters, due to the complexity of soil types as natural entities. The spatial distribution of hydrofacies could be estimated first by classical geostatistical simulations, such as Boolean methods, Sequential indicator simulation (SIS) and truncated Gaussian simulation (TGS) or Plurigaussian simulation (PGS), then they are populated with heterogeneous hydraulic and transport parameters. However, the solution of the inverse problem is considered the most efficient practice to model the structure of hydrofacies while the physical law governing the groundwater flow system is taken into account. The parameters of the physical law, such as hydrofacies distribution, are defined by optimizing the response of the system while solving the physical law (forward problem), compared to the observations of a physical variable such as the hydraulic head. In most cases, the inverse problems are ill-posed, which means they have not a unique solution, the solutions do not depend continuously on data, or a solution does not exist. Moreover, the solutions of the problem may be affected significantly by small changes in the observations and cause large variations in the parameters estimation, making the system ill-conditioned. So, the limited number of observations and the nature of system may impose difficulties in solving an inverse problem. Thus, the inverse problem is a classical optimization problem when the system of equations is linear, where the best-unbiased estimator can be found by the minimization of a least squares error criterion. In ill-conditioned or ill-posed problems, the objective is not only to find an unbiased estimator, but also one with a stable behavior. In weakly linear or non-linear systems, the objective function is linearized and iterative gradient methods are applied to find the estimator. However, the iterative process of optimization in those cases is often a demanding task with high computational cost, while gradient methods can get stuck at local optima. Therefore, the use of stochastic methods allowing the sampling of model space, such as the Markov chain Monte Carlo (McMC) methods, is often preferable. McMC methods are often used under the Bayesian perspective in solving the inverse problem. In Bayesian inference, the prior information is updated iteratively as new members are added to the chain. The members of the chain constitute the final a-posteriori distribution of the parameters. The McMC instead of Monte Carlo methods preferentially visit the model space where the posterior density is high, even if the dimension of the model space is large. The implementation of McMC requires the definition of a transition kernel from an accepted model to a new model, a criterion to accept a member in the chain and a criterion to interrupt the chain. In this work, we adopt the iterative spatial resampling (ISR) technique as the transition kernel, either when the objective is to sample the posterior distribution of parameters or to reach an optimal solution. In sampling the posterior, the candidate models must be independent of the last accepted member of the chain, while the criterion of Metropolis-Hastings or Mosegaard and Tarantola (1995) could be used to accept a model to the chain. In the case of optimization, we adopt the implementation of independent Markov chains to reach an optimal solution. A candidate model is accepted in a chain when its likelihood is better than the last member of the chain, while the interruption of the chain is stochastic with the probability to interrupt becoming higher when the likelihood of a candidate model is high. This way, bias with the opposite direction is introduced to the samples of the posterior. Thus, an approach to posterior distribution is achieved. This mechanism reaches an optimal solution avoiding a large number of geostatistical simulations and forward problem solutions. The novelty of the C-ISR algorithm is the iterative use of cosimulation between hydrofacies and the reference data as an auxiliary variable, in order to gradually improve the path to the optimal solution within a constantly improving Markov chain. Cosimulation for modeling a discrete variable such as the hydrofacies distribution has not been applied in inversion yet, due to the nonlinearity relation between the response variable and unknown parameters. In most cases, the response variable is used as indirect data to evaluate the prior models and drive the search path. The C-ISR method exploits the available information on the relationship between hydrofacies and the physical variable, to produce valid realizations by using cosimulation. More specifically, our method is based on the approach that, instantaneously, the Normal scores transformed hydrological measurements can be correlated with the Gaussian variable of hydrofacies, through a linear coregionalization model, although the underground flow problem is not linear. The approach is used repeatedly within a Markov chain, while the members of the chain result from an iterative spatial resampling transition kernel. In this case, apart from their indirect use for inversion, groundwater pressure measurements are used directly, in order to evaluate the prior geological model of the subsurface. This, results in a narrower and more informed prior distribution, due to the support of the reference variable. The effectiveness of our method is demonstrated by an example application on a synthetic underdetermined inverse problem in aquifer characterization. The results show that the C-ISR method is faster and more accurate as compared to plain ISR.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2018
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Valakas, Georgios
- Βαλάκας, Γεώργιος
Subjects
dc:subject × 11Rights
dc:rights- Statement dc:rights
-
- BY_NC_SA
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/45707