Unsupervised Construction of Sequence Stratigraphie Framework

Abstract Sequence stratigraphy contributes to the improved understanding of the origin and architecture of sedimentary basins, Earth's geological record of local to global changes, and serves as crucial data source for the prediction of exploration for natural resources. The interpretation process involves analysis of heterogeneous data by highly qualified experts, and takes months of work. We propose a workflow aimed to assist in creating the basic stratigraphic framework that can essentially speed up the interpretation. Interpretation workflow for a single well is based on the application of changepoint detection technique that came from time series data analysis. In this work, we used gamma ray well log as a form of time series, and Bayesian online changepoint detection algorithm as a segmentation technique. The Bayesian method is based on the computation of the probability for a point in a time series to be a point in a sequence of length n. The resulting probability curve is then used to determine sequence boundaries as rapid drops in the probability values. The technique has been tested on four wells from the F3 block of the Dutch sector of North Sea. For each of the four wells potential sequence stratigraphic boundaries were selected with pre-defined minimum thickness limit of 100 m corresponding to the third order sequences. The only one controlling hyperparameter of the workflow (maximum thickness limit) allow to work in low scales as well as in large scales. Sequences derived from the segmentation are consistent with characteristic patterns in grain size changes related to sea level changes (aggrading, prograding, retrograding). In wells located in zones of layered continuous underground structures defined segments are stable and can be easily correlated both in automatic (using clustering e.g.) and manual manner. For the wells located in complex geological settings additional expertise and sources of data are required to justify segments connections. The approach for sequence stratigraphic boundaries identification helps optimize the interpretation process. Using logging data as the main input allows boundaries to be predicted more accurately due to the higher resolution compared to seismic data. The derived zonation can be used as a primary knowledge on geological structure in absence of 3D geological model or 3D seismic. Next steps will include development of the automatic inter-well correlation workflow in cross-sections with complex geological structures.