An Approach Based on Hierarchical Bayesian Graphical Models (HBGM) for Measurement Interpretation and Parameter Estimation Under Uncertainty

Monday, February 22, 2016 - 10:00am - 10:25am
Lind 305
Maja Skataric (Schlumberger-Doll Research)
In the oilfield, multiple measurements encompassing a variety of modalities are often available for analysis and interpretation for determining underlying states of nature of interest. Despite and sometimes due to the richness of data, significant challenges arise in interpretation manifested as ambiguities and inconsistencies due to various uncertain factors in the physical properties (inputs), environments, tool properties, human error, and measurements (outputs). Most of these uncertainties cannot be described by any rigorous mathematical means, and modeling of all possibilities is usually infeasible for many real time applications.
In this talk, I will discuss an approach based on HBGM for the improved interpretation of complex (multi-dimensional) problems with parametric uncertainties that lack usable physical models. In this setting, the input space of the physical properties is specified through prior distributions based on domain knowledge and expertise, and forward models are used offline to generate the expected distribution of the proposed measurements. The uncertainties in input and output spaces are modeled as variance imposed in each individual space. In the Bayesian framework, all model parameters are treated as random variables, and inference of the parameters is made on the basis of posterior distribution given observed data. Parameters of the posterior distribution can then be used to build an efficient classifier for differentiating new observed data in real time on the basis of pre-trained models.