181 structurally complex deposits such as IOCG systems. 5.4. Limitations The methodology is dependent on geological logging quality. Deposits in which logging is inconsistent, poorly standardized, or dominated by codes with limited discriminating power will yield smaller improvements relative to conventional estimation—a conclusion directly demonstrated by the Kazakh case. The PSA mitigates this dependency by filtering low-information inputs, but cannot manufacture predictive signals from data that does not contain exist. Prospective validation was conducted at all three sites; however, the depth of validation differs. The Chilean site provides the most comprehensive dataset: eight quarters of blind reconciliation and a completed 2,200 m drill program. The Australian and Kazakh sites have shorter validation windows. Continued operational deployment will generate additional data against which long-term performance stability can be assessed. An organizational dimension not captured by technical performance metrics also warrants acknowledgement. Mine planning workflows are institutionally structured around Kriging-based outputs, and the integration of DL model predictions requires adaptations that extend beyond software installation. Geological and engineering teams must exhibit a dependence on the in model outputs, a process that is built incrementally through demonstrated field performance rather than through hypothetical exercises. The Chilean drill program provides a substantive evidentiary foundation, but each new operational site required its own period of parallel evaluation before model outputs were incorporated into binding mine plan decisions. 5.5. Methodological Transferability The methodology was designed to be deposit-agnostic with respect to commodity, geological setting, and data architecture. The PSA, data integration strategies, confidence quantification framework, and evaluation protocol make no assumptions specific to any particular deposit type. The three case studies— spanning two commodities, three geological environments, and a range of data quality levels—provide empirical support for the claim that the approach is transferable across geological contexts without fundamental methodological modification. Input selection varies between deposits as a function of PSA outputs; the framework itself does not. The methodology is likely to deliver greatest value at sites where three conditions coincide: high within-domain grade variability, structural geological complexity, and substantial volumes of unassayed or under-utilized logging data. Underground IOCG, orogenic gold, and intrusion-related deposits are representative examples. Extension to VMS, porphyry, and sediment-hosted systems is a logical next step for evaluation. 6. CONCLUSIONS This study presents and validates a DL-based resource estimation methodology applied across three geologically distinct deposits, demonstrating consistent improvement over conventional Kriging-based estimation, with the magnitude of improvement scaling with geological logging data quality. At the Chilean IOCG site, the DL model incorporating proxy logging integration (Method 3) reconciled 1.67 times more mineralization than the site Kriging model at an equivalent false positive rate. A targeted 2,200 m underground drill program verified 7.7 kT of in-situ Cu (approximately USD 64M) in volumes previously classified as waste, with all 11 drill holes intersecting high-grade intercepts and drilling costs 25% lower per kiloton verified than the site baseline. Block-level confidence scores correctly distinguished reliable predictions from uncertain ones; high-confidence target zones achieved a 100% drill
RkJQdWJsaXNoZXIy MTM0Mzk2