GRSG 34th Conference 2023

Title: Can we classify smarter? Adding geothermal and spectral expert knowledge into the mineral classification process

Author: Chris Hecker

Mapping alteration minerals from visible and infrared spectral data is routinely done in the exploration for critical raw materials and geothermal resources. The resulting mineral distribution patterns are then used to home in on the most prospective areas for detailed exploration. This approach depends on a proper assignment of the spectral information to mineral classes. In case of a small number of spectra, this task can be done by expert spectral geologists to a high level of accuracy and . If the dataset becomes large (e.g. millions of spectra in drill cutting imaging) this process needs to be automated which often reduces the accuracy or at least the geologic relevance of the classification results to answer exploration questions.

What if we could make our classifications smarter and include expert knowledge in the classification step to make our results not only understandable but also reproducible and geologically relevant? Opposite to the movement towards Machine Learning and AI, we experiment with low-tech but smart combinations of spectral expert knowledge with a geologic understanding of the system we are investigating. We developed a decision tree classifier that automates the work of a spectral expert by looking at the presence and absence of diagnostic spectral features instead of statistical matches with library spectra. We then structure the tree in such a way that essential minerals for a proper geologic characterization are less likely missed in the classification approach. We demonstrate and assess the approach on three geothermal wells in a system in Sumatra and discuss the applicability of the approach to geothermal and mineralized systems elsewhere.

Mapping alteration minerals from visible and infrared spectral data is routinely done in the exploration for critical raw materials and geothermal resources. The resulting mineral distribution patterns are then used to home in on the most prospective areas for detailed exploration. This approach depends on a proper assignment of the spectral information to mineral classes. In case of a small number of spectra, this task can be done by expert spectral geologists to a high level of accuracy and . If the dataset becomes large (e.g. millions of spectra in drill cutting imaging) this process needs to be automated which often reduces the accuracy or at least the geologic relevance of the classification results to answer exploration questions.

What if we could make our classifications smarter and include expert knowledge in the classification step to make our results not only understandable but also reproducible and geologically relevant? Opposite to the movement towards Machine Learning and AI, we experiment with low-tech but smart combinations of spectral expert knowledge with a geologic understanding of the system we are investigating. We developed a decision tree classifier that automates the work of a spectral expert by looking at the presence and absence of diagnostic spectral features instead of statistical matches with library spectra. We then structure the tree in such a way that essential minerals for a proper geologic characterization are less likely missed in the classification approach. We demonstrate and assess the approach on three geothermal wells in a system in Sumatra and discuss the applicability of the approach to geothermal and mineralized systems elsewhere.