171 higher-confidence targets elsewhere. The resource estimation methodologies that underpin the majority of targeting decisions have remained fundamentally unchanged for several decades. Ordinary Kriging and its variants constitute the industry standard (Glacken and Blackney, 2022). These approaches perform reliably under conditions of dense drill spacing and spatially continuous mineralization, but exhibit well-documented limitations in deposit geometries that are increasingly characteristic of contemporary mining operations: structurally controlled systems, nuggety grade distributions, and ore bodies in which economic value is concentrated in discrete high-grade domains surrounded by marginal or barren rock. As a linear spatial interpolator, Kriging is inherently smoothing in its estimation behavior. In deposits where grade continuity is poor and economic value is concentrated rather than distributed, this averaging characteristic constitutes a limitation with material consequences for mine planning (Glacken and Blackney, 2022). A secondary, underappreciated problem relates to data utilization. Operating mines routinely collect extensive geological logging data—lithology, alteration, mineralization codes, geotechnical parameters—for every meter of drill core recovered. The preponderance of this information is used, at most, for structural domain definition before being archived. Additionally, at many underground operations, substantial portions of drillhole databases remain unassayed as a cost-saving activity: core visually classified as barren during logging is frequently excluded from resource model construction entirely, despite its spatial information content regarding deposit geometry and host rock architecture. ML and DL architectures are technically capable of working with heterogeneous data of this kind. Neural networks accommodate non-linear relationships, process mixed data types, and identify spatial patterns that linear estimation methods cannot capture (Goodfellow, Bengio and Courville, 2016; DumakorDupey and Arya, 2021). However, direct application of these techniques to geological modelling is not trivial. Finite computational resources, overfitting risks, inconsistencies in geological core logs, assay bias, and the inherent subjectivity of visual core logging each introduce substantive complications. The predictive utility of individual geological log codes for grade estimation is not a priori knowable; among economically informative codes, the correlation with grade is not guaranteed to be spatially uniform across the deposit. Economic mineralization tends to occur in geologically anomalous zones, which by definition are underrepresented in the bulk of any logging database. This paper addresses two interconnected questions: (i) how to systematically determine which geological data streams merit inclusion in a DL resource estimation model; and (ii) how to quantify whether the resulting drill targets are sufficiently reliable to support capital expenditure decisions. 2. OBJECTIVES AND SCOPE The first objective was to develop and empirically test a framework for quantifying prediction confidence across DL resource models. The intent is to move beyond single-model point estimates toward a multi-model framework capable of identifying which geological data channels function as effective pathfinders and of assigning calibrated confidence levels to block-level predictions. The second objective was to compare alternative strategies for integrating geological logging data into DL models. This comparison includes evaluation of how data are encoded—as input features, as target labels, or as both—along with quantitative measurement of relative accuracy relative to site Kriging models, and assessment of model performance through prospective blind reconciliation rather than retrospective insample metrics. The third objective was to validate the methodology across three geologically distinct deposits. The intent was not merely to demonstrate applicability at a single site, but to assess methodological transferability
RkJQdWJsaXNoZXIy MTM0Mzk2