PETROPHYSICAL DECISION SUPPORT
Looking for oil and gas using Bayesian networks.
The exploration for oil and gas requires real-time petrophysical expertise to interpret measurement data acquired in boreholes in terms of the mineral composition of the potential well and to recommend further steps. High time pressure and the far reaching nature of these decisions, as well as the limited opportunity to gain in depth petrophysical experience suggests that a decision support system that can aid the petrophysicist will be very useful.
Shell Exploration & Production
We have developed a decision support system for borehole appraisal in the search of oil and gas. The system helps petrophysicists in the interpretation of measurement data acquired from boreholes by providing a probability distribution of the mineral composition of the borehole content given the data. Furthermore the system helps the user in the decision for subsequent measurements by computing the expected information of each of the possible measurements.
ARTIFICIAL REASONING
The system provides decision support by answering questions like: what could be the causes of my observational data? Which additional measurements should I do to make this answer more precise? This is a typical application for a reasoning model, such as a Bayesian network.
DOMAIN KNOWLEDGE
Based on domain expertise provided by Shell, we have developed a Bayesian network that encodes the relation between mineral content of a borehole and the possible outcomes of borehole measurements.
ADVANCED COMPUTATION
The model differs from standard Bayesian networks by the continuous-valued variables and the nonlinear relations between variables. Discretization of the variables turned out not to be an option, so a standard Bayesian network approach was infeasible. We therefore had to resort to approximate inference methods. Pilot studies indicated that sampling methods gave the best performance. In the application, hybrid Monte Carlo methods are used for inference, i.e., to compute the probability distributions of the possible mineral compositions of the borehole content given the measurement data.
PHASED APPROACH
The development of the system consisted of an initial research phase to investigate models and methods in Mat lab. After establishing these, we developed the application in C/C++ including an optimized inference engine and a graphical user interface.
TAILOR MADE SOLUTION
This application is an example of a tailor-made product. This means that (1) the development of the product, including the research, started from scratch and (2) the product is not sold to other parties. However, it can be expected that customers that look for solutions in similar problems can benefit from similar methods. Ask us for more information.
Shell Exploration & Production
We have developed a decision support system for borehole appraisal in the search of oil and gas. The system helps petrophysicists in the interpretation of measurement data acquired from boreholes by providing a probability distribution of the mineral composition of the borehole content given the data. Furthermore the system helps the user in the decision for subsequent measurements by computing the expected information of each of the possible measurements.
ARTIFICIAL REASONING
The system provides decision support by answering questions like: what could be the causes of my observational data? Which additional measurements should I do to make this answer more precise? This is a typical application for a reasoning model, such as a Bayesian network.
DOMAIN KNOWLEDGE
Based on domain expertise provided by Shell, we have developed a Bayesian network that encodes the relation between mineral content of a borehole and the possible outcomes of borehole measurements.
ADVANCED COMPUTATION
The model differs from standard Bayesian networks by the continuous-valued variables and the nonlinear relations between variables. Discretization of the variables turned out not to be an option, so a standard Bayesian network approach was infeasible. We therefore had to resort to approximate inference methods. Pilot studies indicated that sampling methods gave the best performance. In the application, hybrid Monte Carlo methods are used for inference, i.e., to compute the probability distributions of the possible mineral compositions of the borehole content given the measurement data.
PHASED APPROACH
The development of the system consisted of an initial research phase to investigate models and methods in Mat lab. After establishing these, we developed the application in C/C++ including an optimized inference engine and a graphical user interface.
TAILOR MADE SOLUTION
This application is an example of a tailor-made product. This means that (1) the development of the product, including the research, started from scratch and (2) the product is not sold to other parties. However, it can be expected that customers that look for solutions in similar problems can benefit from similar methods. Ask us for more information.
our customer
Publication
W. Burgers e.a. A Bayesian Petrophysical Decision Support System for Estimation of Reservoir Compositions. In: Expert System With Applications, Elsevier, 2010. In press.
W. Wiegerinck e.a. Bayesian Networks for Expert Systems, Theory and Practical Applications. In: Interactive Collaborative Information Systems, series Studies in Computational Intelligence, R. Babuska, F.C.A. Groen, Editors, Springer, 2010. In press.
Demo
The petrophysical decision support system cannot be publicly disclosed. However, to give an impression of the system, we have developed a toy system that uses similar methods. The toymodel is a decision support system to reconstruct the chemical composition of a wine given sensory information (measurements) such as smell, taste and color.
In the underlying model, relations are modeled from chemical composition to sensory information. By Bayesian inference, using a so-called Monte Carlo method, the relations are "inverted", and a probability distribution over possible chemical compositions is returned.
Note that the system is only for demonstration purposes. It does not at all pretend to be realistic. All relations in the model are completely arbitrary modeled, and the outcome of the model has therefore no meaning.
Download demo [for windows only]
Publication
W. Burgers e.a. A Bayesian Petrophysical Decision Support System for Estimation of Reservoir Compositions. In: Expert System With Applications, Elsevier, 2010. In press.
W. Wiegerinck e.a. Bayesian Networks for Expert Systems, Theory and Practical Applications. In: Interactive Collaborative Information Systems, series Studies in Computational Intelligence, R. Babuska, F.C.A. Groen, Editors, Springer, 2010. In press.
Demo
The petrophysical decision support system cannot be publicly disclosed. However, to give an impression of the system, we have developed a toy system that uses similar methods. The toymodel is a decision support system to reconstruct the chemical composition of a wine given sensory information (measurements) such as smell, taste and color.
In the underlying model, relations are modeled from chemical composition to sensory information. By Bayesian inference, using a so-called Monte Carlo method, the relations are "inverted", and a probability distribution over possible chemical compositions is returned.
Note that the system is only for demonstration purposes. It does not at all pretend to be realistic. All relations in the model are completely arbitrary modeled, and the outcome of the model has therefore no meaning.
Download demo [for windows only]