Skip to main content

Predicting the quality of lithostratigraphic data from borehole records using machine learning

Often, geological models are generated based on information obtained from exploration data, e.g. borehole records. These borehole records contain descriptions and interpretations about petrography (lithology) and stratigraphy, respectively. The information is crucial for modeling the spatial distribution of lithostratigraphic layers. However, the interpretation of the drilling profiles is error-prone as it depends on several factors, including date of recording, exploration target, quality of digitalization of the borehole record and the human interpretation bias of the responsible expert, among others. Due to the large number of boreholes drilled during explorations, separating adequate from insufficient drilling profiles is of great importance, yet rather difficult. While visual inspection of the inferred geological model is a viable approach it results in numerous iterations to filter inadequate drilling profiles which is time-consuming and expensive.

               We present a Python-based software package that predicts the quality of lithostratrigraphic data from borehole records based on several criteria. Using pre-checked reference drilling profiles, we train a random forest model to predict the quality of non-checked drilling profiles for geologically comparable regions. The aforementioned selection criteria as well as the predictors are individually definable by the software user.

               As a study area, we selected a former lignite mining area of Lusatia bordering the Federal State of Brandenburg and the Free State of Sachsen (Saxony). Here, 71 pre-checked and classified drilling profiles exist which are combined with erroneous synthetic profiles and used to train the random forest model to predict the quality of the >3000 remaining profiles.

Details

Author
Elisabeth Schönfeldt1, Thomas Hiller1, Marcus Fahle1, Mathias Hübschmann2, Friedemann Grafe2
Institutionen
1Bundesanstalt für Geowissenschaften und Rohstoffe, Forschungs- und Entwicklungszentrum Bergbaufolgen (FEZB), Germany; 2Sächsisches Landesamt für Umwelt, Landwirtschaft und Geologie, Germany
Veranstaltung
GeoSaxonia 2024
Datum
2024
DOI
10.48380/79jg-g702