Automations in Chemostratigraphy: Toward Robust Chemical Data Analysis and Interpretation

Abstract

Elemental chemostratigraphy has become an established stratigraphic correlation technique over the last 15 years. Geochemical data are generated from rock samples (e.g., ditch cuttings, cores or hand specimens) for up to c. 50 elements in the range Na-U in the periodic table using various analytical techniques. The data are commonly displayed and interpreted as ratios, indices and proxy values in profile form against depth. The large number of possible combinations between the determined elements (more than a thousand combinations), makes it a time-consuming effort to identify meaningful variations that resulted in correlative chemostratigraphic boundaries and zones between wells. The large number of combination means that 30-40% of the information is not used for the correlations that maybe crucial to understand the geological processes. Automation and artificial intelligence (AI) are envisaged as likely solutions to this challenge. Statistical and machine learning techniques are tested as a first step to automate and establish a workflow to define (chemo-) stratigraphic boundaries, and to identify geological formations. The workflow commences with a quality check of the input data and then with principle component analysis (PCA) as a multivariate statistical method. PCA is used to minimize the number of elements/ratios plotted in profile form, whilst simultaneously identifying multidimensional relationships between them. A statistical boundary picking method is then applied define chemostratigraphic zones, for which reliability is determined utilizing quartile analysis, which tests the overlap of chemical signals across these statistical boundaries. Machine learning via discriminant function analysis (DFA) has been developed to predict the placement of correlative boundaries between adjacent sections/wells. The proposed workflow has been tested on various geological formations and areas in Saudi Arabia. The chemostratigraphic correlations proposed using this workflow broadly correspond to those defined in the standard workflow by experienced chemostratigraphers, while interpretation times and subjectivity are reduced. While machine learning via DFA is currently further researched, early results of the workflow are very encouraging. A user-friendly software application with workflows and algorithms ultimately leading to automation of the processes is under development.