A Nondestructive Alternative for Kerogen Type Determination in Potential Hydrocarbon Source Rocks Using Hyperspectral Data and Machine Learning

Abstract

In petroleum geology, to assess the hydrocarbon generation potential in source rocks involves the determination of the kerogen type by some destructive method. The usage of such methods is a bottleneck in the process because it is time-consuming, requires specialized tools and personnel, and ends up destroying the rock sample, so it is not possible to do any posterior analysis. This study presents an alternative method for determination of the kerogen type that is fast and nondestructive using hyperspectral data and machine learning techniques. The method is validated using five distinct supervised learning algorithms that were applied to spectral data collected in rock samples from Taubaté Basin, Brazil, of an outcrop whose rocks have a wide range of hydrocarbon generation potential. Cores and samples were collected from the outcrop and had their kerogen type determined by geochemical analyses performed in the laboratory. The robustness of the method is evaluated in two distinct experiments. In the first one, the hyperspectral dataset was collected using a nonimaging spectroradiometer; in the second one, the method uses nonimaging hyperspectral data as training and is tested in hyperspectral images collected. In both experiments, the method was able to establish a relationship between selected spectral features and the kerogen type of the source rocks sampled. The results obtained in this article are prospective for nondestructive classification of kerogen type (and, consequently, the hydrocarbon generation potential) since most of the models generated achieved accuracy above 0.8 in the validation step and 0.75 in the test step.

Keywords

• Classification
• hydrocarbon source rock
• hyperspectral
• kerogen type
• machine learning (ML)