Using Neural
Networks to Predict Thermal Conductivity from Well
Logs
Goutorbe, Bruno, Francis Lucazeau, Alain Bonneville, Institut de Physique du Globe de Paris, Paris, France
Terrestrial heat-flow is important to predict present-day basin
temperatures and hydrocarbon maturation. Reliable heat-flow estimates can be
obtained from oil exploration well
data, providing that thermal conductivity is
obtained with a sufficient accuracy. Previous methods involve (1) correlations
between
well
logs and calibrated thermal conductivity, (2) empirical models
based on mineral or lithological content. None of these methods gives objective
or universal outcomes.
In order to obtain such a general process, we have developed a
new method based on the neural network technique, which relates directly a set
of geophysical well
logs to thermal conductivity. The method has been
calibrated on ODP data, which accounts for several thousands of conductivity
measurements and five
types
of geophysical
well
logs (Sonic, Density, Neutron,
Resistivity and Gamma-ray). This set has been used to train multi-layer
perceptrons (MLP) and find an empirical relationship between
well
logs (MLP
inputs) and thermal conductivity (MLP output). MLP are a class of neural
networks that can perform efficient function approximation without any a priori
knowledge, providing that a sufficiently large number of data exists.
Validation tests suggest that thermal conductivity can be obtained with a
10-15% level of confidence.
The method is
applied in the ongoing GATOR project (Global Analysis of Temperature from Oil
Exploration) to characterize the worldwide thermal regime of continental
margins. Examples on the South African and Australian margins will be
presented.