ELLIS Delft Talk by Guillaume Rongier

Abstract

While machine learning has a long history in geology, empirical relationships remain widely used. Through the example of total organic carbon (TOC), this talk will illustrate the close links between empirical relationships and machine learning, and the benefits of turning to machine learning. TOC is a measure of the proportion of organic carbon in rock samples typically gathered from boreholes. It can be used to assess the potential for hydrocarbons, understand rock mechanics, or assess reducing conditions for basin-hosted mineral systems, and is paramount when seeking to understand variations in paleo-environmental conditions. Since gathering and analyzing rock samples is expensive, empirical relationships have been developed to predict TOC from well logs, which are based on more widely available geophysical measurements into boreholes. Those empirical relationships come from geological and petrophysical principles implemented in mathematical models manually fitted to the data. This leads to several limitations, mainly poor generalization, inability to quantify uncertainties, time-consuming and subjective calibration that leads to reproducibility issues. But those empirical relationships can be rewritten as linear regressions, a simple change that solves many of the previous limitations. Turning to more advanced machine learning methods improves predictions by taking into account the non-linearity and variability in the data. Using the expert knowledge behind empirical relationships as input besides well logs improves the predictions as well: this shows that leveraging geological and petrophysical concepts through feature selection and engineering boosts machine learning performances.

To join this event, please contact Frans Oliehoek.