Data-driven semi-supervised clustering for oil prediction

We present a new graph-Laplacian based semi-supervised clustering method. This new approach can be viewed as an extension/improvement of previously published work, both in terms of areas of applicability and computational speed. Our clustering method is capable of handling very large datasets with millions of data points using very limited amounts of labelled data. In this work, we apply our clustering method to 3D oil prospectivity, based on amplitude-versus-angle inversion parameters and borehole information. We cluster the synthetic Life of Field dataset, which has a fault-block constrained central oil reservoir, where we also perform a cross-validation check of the predictive power of our method. Furthermore, we cluster a field dataset, which is characterized by a stratigraphic trapped channelling system. In both cases we find appealing results.

Automated summary

A new graph-Laplacian based semi-supervised clustering method is proposed in this research, which is capable of handling large datasets and yielding satisfactory results in oil prospectivity analysis.