Machine learning for classification of stratified geology from MWD data
Katherine L. Silversides, Arman Melkumyan
Abstract
Measure while drilling (MWD) data can be collected during routine drilling at a mine site. This produces large datasets that are not easily processed to provide geological information. MWD data measures the performance of the drill using multiple variables such as penetration rate, pulldown pressure, torque and rotational frequency. These can be related to mechanical properties of the rock such as hardness, however, it is frequently difficult to relate them to geological properties that can be used in modelling and directing mine planning. Stratigraphic deposits such as BIF-hosted iron ore, coal, and oil and gas have layers of distinctly different minerals and are ideal candidates for MWD classification. This paper reviews recent research into applying machine learning techniques to prevent drilling wells in unproductive rock, prevent coal seam penetration, detect weak strata and improve geological models in iron ore mines. Common machine learning methods such as neural networks, boosting and Gaussian Processes are compared in different situations. The results and implications for implementation are discussed.