Downhole density estimation using multielement geochemistry and machine learning

Abstract

Listen

Translate article

Machine learning (ML) was used to estimate bulk density from multielement geochemistry. At the Wheeler River site, host to the Phoenix and Gryphon uranium deposits, multielement geochemistry data were acquired for 829 exploration holes. Of those holes, 41 were logged with a downhole dual-spaced density probe during several mobilizations between 2009 and 2019. Density measurements were collected to provide constraints for inversions of airborne gravity data. To improve the density model's spatial resolution, ML models were trained to estimate bulk density from collocated multielement geochemistry data. Two geochemical laboratory methods were used (251 holes for the old method and 578 holes for the new method); therefore, two separate models were trained. Leave-one-hole-out cross-validation mean absolute error (MAE) results from the old and new geochemistry models showed similar scores of 0.027 g/cm3 and 0.025 g/cm3, respectively. Eight test holes were removed from the training data and used for final evaluation once the model was trained. Test hole results showed MAE scores of 0.026 g/cm3 for the old geochemistry model and 0.043 g/cm3 for the new geochemistry model. A unique aspect of this data set was the presence of repeat logs for multiple boreholes over a decade-long logging campaign. This provided the opportunity to assess the measurement uncertainty across time, density probes, operators, and boreholes conditions. The process of estimating downhole density from multielement geochemistry data could be used for many exploration projects to help generate better starting density models for use in geophysical inversions and other applications.