This paper proposes an unsupervised neural network model for facies pattern recognition and formation characterization using borehole images. The goal is to create an automated workflow for rock fabric identification using high resolution acoustic or electrical borehole images with the aim of supporting 3D geological modeling. The results are compared and validated with geological and petrophysical interpretation.
Image-based facies recognition is challenging when applying Deep Learning techniques: 1/ the volume of released labeled data constrains the abilities to build a robust neural network model 2/ data classification itself is subject to geologist interpretation. Additionally, indirect measurements can bias data, hindering the correlation between log response and any particular classification.
We propose, therefore, an application of a fully convolutional autoencoder for borehole image data clustering to extract the most representative information displayed by the images without relying on labeled data.
The data set corresponds to electrical borehole images with high-resolution at 0.2in and 80% borehole coverage. First, we apply an autoencoder reconstruction loss for network pre-training, then a joint training using cluster assignment hardening. After training and applying the model, patterns represented by each cluster of geological facies or geomechanical features constitute a library that can be assigned by the user to specific facies or can be automatically correlated to the core description. The method provides pattern recognition and facies prediction with higher resolution and accuracy than conventional Machine Learning methods based on the clustering of petrophysical properties.