The time it takes to bring a deposit from initial discovery to operation is often a decade, or longer. During this time, much effort is spent on data collection to improve the understanding and characterization of the resource, with limited focus on geotechnical data collection until the project is at a more advanced stage. Early-stage mine designs may therefor be significantly conservative due to a lower density and reliability of geotechnical characterization data. However, with the use of deep learning and computer vision techniques, the existing core photographs from earlier exploration drilling may be geotechnically characterized to produce a far more robust dataset from which initial mine design inputs can be derived, or areas of design risk can be identified.
This discussion summarizes a machine learning (ML) workflow developed by SRK for use in the geotechnical photologging of exploration core photographs. The methodology involves the qualitative description of rock mass character along the length of each drill run with a range from; Disintegrated Core (ex. Faults, Overburden) to Stick Rock. This workflow has been used at a Copper-Gold porphyry mine in Central Asia, where 44,500 m of exploration drill core photos were automatically characterized using the ML model. The output from the ML directly supplemented the available geotechnical logging data to build a more reliable domain model of the deposit, which is the foundation for the geotechnical analysis. As a proof of concept, approximately 25,000 m of core photos were manually classified and compared with the ML results.
This approach to automated characterization of core images provides the potential to leverage already existing data at earlier stages in a mining project. This ultimately may produce more confident designs, or highlight areas of geotechnical concern sooner than would be possible using conventional geotechnical characterization methods.
For more information about this ML workflow, please contact Andrew LeRiche.
