Reproducible geotechnical metrics from AI-assisted core logging

Summary

Abstract

Geotechnical core logging provides crucial input for rock engineering, including mine design, stability assessment, and rock-mass classification. Yet the process remains prone to human error and lacks reproducibility. Core box photographs, widely collected in the industry, represent a largely underutilised primary data source for recomputing geotechnical metrics. This study investigates the practical benefits of applying artificial intelligence (AI) to such images, focusing on reproducibility and enabling multiple rock-mass characterisation workflows from metric-agnostic outputs. An AI-based instance segmentation method (Segment Anything Model; SAM), combined with context-specific filtering, was applied to an image dataset from the Boliden Zinkgruvan mine, Sweden. The workflow automatically extracts the intact core pieces in each image, digitally measures their lengths, and produces a fracture-position database. Natural and mechanical fractures are then differentiated using markings produced by a geologist, enabling computation of rock quality designation (RQD). The results are compared with manual measurements. The workflow achieved a mean absolute difference between AI-computed and manually measured RQD of 7.8 percentage points. Outlier analysis suggested that large discrepancies between manual and computed RQD may be related to human error, indicating that AI-assisted methods can support quality assurance/quality control (QA/QC) of manual logs. The results demonstrate that recording fracture positions rather than interval-based metrics reported over fixed lengths (e.g., RQD every 1 m) enables transparent recalculation of RQD and supports the derivation of additional metrics such as fracture frequency (P10). This highlights how AI-assisted fracture extraction can increase reproducibility in core logging and produce design-independent data that can be transformed into multiple geotechnical metrics and used across different analysis frameworks (empirical methods, numerical simulations, DFNs). The findings confirm that AI-assisted core logging enables reproducible downstream metric computation and design-independent rock-mass metrics from standard core photographs, aiding QA/QC of manual logs and use across different analysis frameworks.