Abstract

Geotechnical data for greenfield and brownfield sites is often limited, and the data that is available is often inconsistent and difficult to interpret within domains. Geotechnical characterization is contingent on reliable data to reduce uncertainty in conditions of slope walls or underground workings.

Machine learning (ML) workflows were developed by SRK for use in the geotechnical classification of core box photographs. Deep convolutional neural networks are either trained for each project or as a general model applicable to most projects, with the intention of providing a reliable dataset for rock mass characterization and structural modelling.

Two geotechnical classification systems have been developed using core photos. One methodology involves an assessment of the quantitative degree of brokenness of the drill core using the ratio of clast size to core diameter classified into categories. The trained algorithm, or Core Damage Index (CDI) model identifies and maps similar conditions to these defined categories in available core box photographs for the entire project. The ML workflow produces an interval table (hole ID, from depth, to depth, category) with these categories.

The second methodology calculates a pseudo-RQD and a break frequency based on the breaks observed on drill core images. The algorithm is trained using manually drawn labels of observed breaks and zones of rubbles on training core box photographs. The trained algorithm identifies and maps these features in available core box photographs and the pseudo-RQD and break frequency are calculated from this data. The ML workflow produces an interval table (hole ID, from depth, to depth, pseudo-RQD, break frequency) with these calculations.

Although automated logging presents an advantage in terms of data acquisition speed and reproducibility, inherent limitations must be understood and addressed by the end-user of the data. The model accuracy is dependent on the quality of the labeled training data. Biases in the training data will be propagated to the models. Variations in image quality and parameters or rock type can reduce model accuracy unless additional training data is used. Constant interval lengths used for RQD, pseudo-RQD and break frequency estimations do not capture well fault zones and similar structures. Machine learning mapping does not have yet correlations to conventional rock mass characterization systems.

Both approaches to automated characterization of core images provide the potential to efficiently collect a very large and reliable dataset at early project stages using existing resource, exploration, metallurgical drillholes. This ultimately may produce more confident designs or highlight areas of geotechnical concern sooner than would be possible using limited or unreliable geotechnical data, while significantly reducing costs typically associated with manual core photo review.