Joint interpretation of geophysical data: Applying machine learning to the modeling of an evaporitic sequence in Villar de Canas (Spain)

An optimal strategy for building realistic geological models must combine different geophysical techniques, each with its advantages and limitations. However, dealing with multiple geophysical datasets to derive a geological interpretation is not straightforward since geophysical parameters are not always functionally related. In this work, we propose an innovative approach consisting of using machine learning techniques to jointly interpret three geophysical datasets (a pseudo-3D resistivity model, a 3D velocity model, and 4 well-logs). These datasets, among others, were acquired to characterize the suitability of an evaporitic sequence for hosting a temporary storage facility of hazardous radioactive waste, which was planned in Villar de Canas (Spain). Our strategy consisted of integrating both models in a single 3D bi-parametric grid that nested the velocity and resistivity values in each node. We then used a supervised learning algorithm to lithologically classify each node according to a training set based on the borehole data, which acts as ground truth. The training set is composed of classifiers that lithologically label resistivity-velocity pairs. However, the very shallow nodes lack classifiers due to the poor well-log coverage at the top part of the evaporitic sequence. To fill this gap, we computed an unsupervised cluster analysis that provided new classes to complete the training set. Finally, the supervised classification was applied, providing a new 3D lithology model that is far more consistent with the geology than the models derived from each parameter independently. The 3D model also revealed geological features previously unknown, notably the existence of an inactive fault. The proposed method can be applied to integrate and jointly interpret any kind of multidisciplinary datasets in a wide range of geoscientific problems, including natural resource exploration, geological storage, environmental monitoring, civil engineering practice, and hazard assessment.

Automated summary

The study proposes an innovative approach using machine learning techniques to jointly interpret multiple geophysical datasets for creating more realistic geological models by integrating velocity and resistivity values into a 3D bi-parametric grid and lithologically classifying nodes with a supervised learning algorithm. This method fills the missing functional relationships between different datasets and provides a more consistent 3D lithology model that uncovers previously unknown geological features such as an inactive fault.