5. DISCUSSION 5.1 Principal Findings The most significant finding from this study is that the DL architecture in isolation does not account for the observed performance improvements. At the Chilean site, Method 1—DL trained on assay data only—performed essentially equivalently to Kriging, confirming that the representational capacity of the network does not generate value in the absence of appropriately structured input data. The performance gains were attributable to the manner in which geological logging was integrated: specifically, the dual-channel Method 3 approach, in which visual barren classification is encoded as both a spatial input signal and a ground-truth training target. The implication is significant: the relevant question in applying DL to resource estimation is not only to consider an alternate DL-based algorithm to evaluate a resource, but rather whether the available data environment contains information that linear spatial interpolation structurally cannot leverage. The PSA screening step proved more consequential than initially anticipated. Preliminary investigation included a temptation to incorporate the maximum number of available geological log codes, on the reasoning that additional input features would expand the information available to the model. Systematic evaluation demonstrated the contrary: unscreened inputs degraded model performance through noise introduction. At the Kazakh site, where the logging database had the lowest signal-to-noise ratio of the three sites, the PSA was the critical determinant of model performance. This finding reinforces a broader principle applicable to ML in mining contexts: the performance bottleneck is rarely the model architecture. It is, in the large majority of cases, data curation and preparation. 5.2 Block-Level Confidence in Operational Practice The block-level confidence scoring framework materially influenced drill program design at the Chilean site. Target selection was not based solely on predicted grade, but on the intersection of predicted grade and the achieved confidence for de-risking each zone to Measured. Zone 65 was classified as both high-grade and high-confidence despite being located beyond a structural feature that the existing geological model treated as a mineralogical boundary. Considered on the basis of grade prediction alone, this target would have been assigned priority eventually; the confidence score advanced its prioritization. The subsequent 100% drill hole intercept rate across all 11 holes in high-confidence zones—while recognized as a small sample from a statistical standpoint—provides empirical validation that the framework successfully separated reliable from uncertain predictions under field conditions. The exclusion of low-confidence zones from the drill program is equally important: capital was not deployed in areas of low return on investment. 5.3 Structural Reinterpretation at the Chilean IOCG site The geological reinterpretation prompted by Zone 65 at the IOCG site merits specific attention. The DL model identified a spatial grade pattern inconsistent with the prevailing structural interpretation—it did not furnish an explanation for the inconsistency. The site
RkJQdWJsaXNoZXIy MTM0Mzk2