A machine-learning based approach to predict facies associations and improve local and regional stratigraphic correlations

The geological record has challenged stratigraphers through time. Many depositional, tectonic and paleobiological events require stratigraphic positioning to determine temporal relationships among such events. This task is complicated and challenging, especially in sedimentary succession with scarcity or lack of paleontological content bearing biostratigraphic value and radiometric ages. Therefore, subjective personal criteria adopted during correlations and field mapping activities make stratigraphic correlations more complex and confusing. New methodological approaches are necessary to test human expertise in recognizing stratigraphic units with environmental significance and to contribute to stratigraphic correlations based on quantitative data. We used cores obtained by coal drilling campaigns during the 1970s and 1980s on the southern border of the Parana' Basin, southern Brazil, to generate a quantitative database. Data obtained from gamma spectrometry in the cores of Carboniferous to Permian age measured total count (cps), potassium (K), uranium (U) and thorium (Th). We used machine learning (ML) to predict facies associations and to confront the quantitative database with the three facies associations mapped through time based on qualitative geological mapping criteria. The k-nearest neighbors (KNN) algorithm reached maximum accuracy values of 86.0%% and f1-score of 90.0%, 73.0% and 91.0% for facies associations 1, 2 and 3, respectively, using exponential moving average and normalized data. This KNN-based method using gamma spectrometric data opened new possibilities to perform local and regional stratigraphic correlations using quantified data. New tests must be performed to improve this promising correlation method in other regional settings and related basins.