Stimulation Initiative for European Neural Applications
Esprit Project 9811
Recognition of Exploitable Oil and Gas Wells
Company background
Koninklijke/Shell is a multinational
Oil company, based in the Netherlands and the U.K. . Number of
personnel: > 100,000.
Shell has its own research departments.
The problem
Whether economically recoverable oil or gas is present in subsurface
reservoirs is highly dependent on physical properties of the reservoir rock
such as porosity and permeability. These properties can be determined
directly from actual rock samples (cores) taken from wells.
For economical reasons, however, cores are only available from a limited
number of wells. In most instances, formation analysts have to rely on
measurements from wireline tools. These tools are lowered into a well to measure
a suite of physical parameters, such as sound velocity and gamma radiation
as a function of depth. The resulting data are interpreted by formation analysts
to identify the rock types, which in turn, are indicative for
the reservoir properties. The problem is the inherent variability
of the wireline data, due to gradations in rock characteristics, effects
of data acquisition and statistical fluctuations in radiation measurements.
This makes the identification of rock types very difficult, in particular
with complex and heterogeneous reservoirs.
Neural network application
Shell Research trained neural networks to classify rock types
as function of the wireline data. For different geological environments
different neural networks were trained. The desired outputs were
provided by expert geologists. The degree of consensus between
the neural network and its trainer were roughly equivalent to the
degree of consensus among different geologists.
The neural network has been incorporated in Shell's geological
computing environment, in which the formation analyst can train and
apply the neural network via a user-friendly graphical interface.
Benefits
Previously, the identification of rock types was done manually by the
formation analysts. Now the neural network enables the formation analysts to
classify rock types from wireline data at reduced effort.
For this problem, the neural network
performs significantly better than standard statistical techniques.
Generalization
The identification of rock types from wireline data is a typical problem
for which no good numerical model exists.
The problem is very noisy, and many variables are involved. This
makes the problem very time-consuming for the specialists. However,
the fact that human specialists are able to perform reasonable on this
task, purely on the basis of the wireline data, was indicative that the
task could be learned by a neural network. This types of identification
tasks, or pattern recognition tasks, often occurs in complex
engineering.
Contact person
M. Kraaijveld, Shell Research, Rijswijk, the Netherlands
References for further reading
W. Epping et al.: Lithofacies identification from wireline logs
in: ICANN'93: Proceedings of the International Conference on
Artificial Neural Networks (Kappen, B. and Gielen, C., ed.), pp 876-881,
Springer-Verlag, London, UK (1993).