Rock Mass Joint Sets Identification Through Stereographic Projection and Unsupervised Learning: A Comparative Study

Geological and geotechnical studies have been usually based on experimentation, analytical analysis, graphical representations, and numerical modeling. However, with the development of the artificial intelligence field and the sharp increase of its applications, using machine learning and deep learning methods has gained a significant importance in these studies. Within this context, this paper presents, first, an overview about recent research articles that studied the application of machine learning approaches to rock mass studies. Second, a case study of rock mass joint sets identification at Draa Sfar mine in Morocco is presented. Based on a database of 400 measurements of dips and dips directions for different rock mass discontinuities such diaclases and schistosity, a comparative study between stereographic projection as a classical method and K-means clustering as an unsupervised machine learning method is established, to identify the mine’s main joint sets. The results show that the clustering of the joints sets through both two methods is concordant. However, the use of stereographic projection for joints sets identification does not need prerequisites while K-means clustering presents the limitation of requiring the initial number of clusters as an input. In this case, the use of pre-analysis methods, such as the elbow method or the silhouette score, is mandatory. In addition, both of the methods need sufficient number of inputs in order to give accurate values of joint sets’ spatial orientation characteristics. To sum-up, our comparative study shows the advantages of classical modeling for rock mass joint sets identification, but also introduces the potential and the limitations that have one of the clustering algorithms within this topic. The challenge now is to train and test more sophisticated artificial intelligence methods that will allow advanced analysis of geotechnical and geological data, to confirm if these tools can partially or entirely substitute classical modeling.