Lithological interpretation of core samples is a decisive early stage of many geoscience workflows. However it is often time-consuming, although significant parts of the samples may be covered by very similar facies alternations.
The target of this project is to accelerate core interpretation in assisting expert geologists with artificial intelligence. Doing so, they can focus their time and energy where they are the most valuable, and not on core segments where the interpretation can be statistically deduced from earlier work.
Our approach uses a supervised classification method based on a Deep Learning model. A convolutional neural network is trained with the interpretation of a geologically representative part of the data set, then used to infer facies identification on the remaining samples.
We carried out a proof of concept on a real-life data set from the Gulf of Corinth, with 3 drilling sites and 17 facies associations. We performed a thorough parametric study, assessing 6 different methods to generate the training data from the humanly interpreted segments and comparing 3 reference neural network architectures.
The early results are promising and pave the way for further adaptation of Deep Learning technologies to the core interpretation context. Besides, core images and wireline logs could be combined in the future, to build a global model for the assisted interpretation of well data.