(ZFCU)—and the right side specifies the ground truth against which the model trains—in this case, Cu assays only. This notational distinction reflects a deliberate design principle: data streams of differing reliability can contribute information to the model in distinct roles without being treated as equivalent in terms of label quality. Implementation was conducted in Python, utilizing PyTorch for neural network construction and CUDA for GPU-accelerated computation. The convolutional neural network (CNN) architecture processes multiple input channels simultaneously, learning non-linear spatial relationships across them (O’Shea and Nash, 2015). 3.4 Data Integration Methods Three data integration strategies were evaluated. The Chilean copper site is used for illustrative purposes here, but the same methodological logic was applied at each of the three deposits with appropriate commodity-specific adaptations. Method 1 — Assay-only baseline. D(Cu) ~ D(Cu). Only geochemically assayed grades are used as both input and training target. All unassayed core is excluded. This configuration tests the performance ceiling achievable from geochemical data alone and serves as the primary benchmark against which the additional value of geological logging is assessed. Method 2 — Geological logging as ancillary input. D(Cu, ZFCU) ~ D(Cu). Unassayed core visually logged as barren is introduced as a separate input channel (ZFCU) but is excluded from the training target. This configuration provides the model with spatial context derived from visual classification without treating the visual assessment as a quantitative grade estimate. Method 3 — Geological logging as input and training target. D(Cu, ZFCU) ~ D(Cu, ZFCU). Identical to Method 2 on the input side, but barren-logged unassayed core is additionally assigned a grade of 0.0% Cu on the training target side. This encoding instructs the model that visual barren classification is associated with negligible copper grade—an imperfect label, given that some proportion of such core will carry weak mineralization, but one that constitutes the best available information. Critically, DL architectures trained on probabilistic or noisy labels can exploit this signal in ways that are structurally unavailable to linear spatial interpolators such as Kriging. The decision to evaluate all three configurations rather than proceeding directly to Method 3 reflects a methodological principle: if Method 3 outperforms Method 1 while Method 2 does not, this outcome provides specific mechanistic information—the performance gain derives not merely from the presence of additional spatial data, but from the manner in which the model is trained to interpret the semantic content of visual barren classification.
RkJQdWJsaXNoZXIy MTM0Mzk2