Macassa Mine—the last producing mine in the Kirkland Lake camp—faces a pressing resource‑replacement challenge with less than a decade remaining in its declared mine life. Historically, geological interpretation relied largely on gold assays, supported by limited multi‑element geochemistry data, constraining or limiting a broader geological insight. To overcome these limitations, a comprehensive core‑scanning program was launched in 2022 to collect XRF data, hyperspectral mineralogy, RQD, fracture data, and high‑resolution imagery. This initiative aimed to reduce inconsistencies inherent in traditional logging, particularly those affecting the recognition of alteration intensity and subtle lithological variability.
These datasets provide objective, high‑density measurements: hyperspectral signatures delineate alteration halos, while XRF‑derived sulfur and immobile‑element chemistry improve mineralization indicators and constrain lithologic boundaries. To effectively utilize this data‑rich environment, the geology team developed a modified workflow incorporating machine‑learning methods. A validated training dataset was constructed from an updated lithologic column informed by XRF and hyperspectral inputs, and an additional lithologic class was introduced to represent the Main Break alteration corridor—a unit that shares immobile‑element characteristics with its unaltered equivalent but exhibits distinct hyperspectral alteration patterns. These differences were confirmed through feature‑ranking and feature‑engineering analyses.
Multiple algorithms (Random Forest, Neural Networks) were evaluated, with preprocessing and feature engineering improving classification accuracy toward a target performance of over 80 percent. The top‑performing model was applied to a blind dataset to classify lithologies, emphasizing detection of the newly defined alteration corridor unit. Core‑scan intervals collected at 1‑ft spacing were composited to 3‑ft, 5‑ft, and 10‑ft lengths using Python-based scripting to optimize inputs for interpretation and modeling. Model outputs were benchmarked against hyperspectral alteration indicators, gold assay trends, and the current mineral resource wireframe. Results demonstrate that AI‑assisted geological workflows can enhance classification consistency and reduce observer‑driven variability, offering tangible benefits for advanced interpretation and model refinement in mature underground settings.
In the current Macassa resource model, evolving logging standards, legacy data from multiple project owners, and the inherent subjectivity involved in lithology and alteration logging create challenges in producing a geologically consistent and well-constrained low-grade halo. Core‑scan alteration signatures highlight a laterally continuous and data-supported feature that can be used to define this low‑grade halo. Identifying this footprint also provides an opportunity to resample missed intervals and reassess domain boundaries, improving variographic characterization and allowing a more geologically defensible transition between hard and soft domains. Incorporating this newly defined zone into the modeling workflow could refine estimation parameters, reduce artificial grade discontinuities, and enhance our understanding of the broader mineralized system.