172 across contrasting geological settings. The three cases examined are: an orogenic lode gold deposit in Western Australia; a manto-type IOCG copper deposit in the Candelaria–Punta del Cobre district of northern Chile; and an orogenic lode gold deposit in western Kazakhstan. These deposits differ in commodity, geological setting, operational maturity, and data quality, providing a basis for evaluating the robustness and generalizability of the methodology. The methodological emphasis throughout is on techniques that use data already available at operating mine sites and that produce outcomes directly connected to drilling decisions and mine plan economics. 3. METHODOLOGY 3.1. Data Characteristics The data environment is the appropriate starting point for this methodological description, as it is the principal determinant of ML model performance in mining applications. At all three study sites, geological logging was extensive. Every meter of core had been logged for lithology, alteration, and associated visual parameters. However, logging had accumulated over extended operational periods, frequently recorded by different geologists with evolving interpretive conventions. Some lithological codes were applied broadly across large volumes of core, carrying limited discriminating value for grade estimation. Others were infrequent, potentially informative, but insufficiently represented in the database for reliable standalone model training. Assay coverage was spatially variable. At the Chilean IOCG site, approximately half the drillhole database was unassayed core visually classified as barren and excluded from standard assay campaigns. At the Australian and Kazakh gold sites, data gaps were less extensive but remained present, concentrated in zones deemed uneconomic during initial exploration programs. The critical unknown in this data environment was signal-to-noise ratio. For any given logging code, the existence and spatial stability of a correlation with grade was not determinable without systematic screening. Without such a screening step, the investigator faces two suboptimal alternatives: include all available inputs, risking noise contamination and overfitting; or rely on geological judgment alone, risking confirmation bias and inadvertent exclusion of informative signals. 3.2. Pathfinder Screening Algorithm To address the input selection problem, a pathfinder screening algorithm (PSA) was developed. The PSA evaluates each candidate geological log code against assayed grade data using non-linear correlation metrics, ranks codes by their predictive signal-to-noise ratio, and assigns inclusion or exclusion determinations based on a statistical significance threshold. This approach replaces full ablation analysis—the approach of individually training a complete model for every possible input combination to evaluate marginal contribution (Meyes et al., 2019). For deposits with dozens to hundreds of unique logging codes, exhaustive ablation is computationally intractable. The PSA reduces input screening time from weeks to hours while preserving the capacity to identify codes with genuine predictive signal, including those that are infrequent in the database. Notably, the PSA does not impose a fixed input set across deposits. At the Chilean site, it identified a different subset of informative codes than at the Australian or Kazakh sites. At the Kazakh deposit, where logging conventions had evolved considerably over a longer operational history, fewer codes met the significance threshold. This behavior is a design feature rather than a limitation: the algorithm adapts to local data quality rather than enforcing a uniform input template irrespective of signal content.
RkJQdWJsaXNoZXIy MTM0Mzk2