KNOWLEDGE-DRIVEN PROSPECTIVITY MODELING OF CLASTIC-DOMINATED LEAD-ZINC MINERAL SYSTEMS FOR AUSTRALIA, CANADA, AND THE UNITED STATES (Invited Presentation)

The Critical Minerals Mapping Initiative (CMMI) is a Tri-National collaboration among Geoscience Australia (GA), the Geological Survey of Canada (GSC), and the US Geological Survey (USGS); the collaboration is designed to facilitate knowledge-sharing about critical mineral resources.

Basin-hosted Zn-Pb deposits, present in all three countries, have been an initial focus of CMMI due to the potential to contain Ga, Ge and In in addition to Zn. A data- driven machine learning approach applied Weights of Evidence and Gradient Boosting to produce prospectivity maps for Mississippi Valley-type and clastic-dominated (also known as SEDEX) lead-zinc deposits (Lawley et al., 2022). The prospectivity maps used numerous evidence layers (e.g., proximity to deep and shallow magnetic source boundaries, satellite gravity, LAB depth, permissive geology etc.).

Our knowledge-driven modeling utilizes the same datasets used in the data-driven approach. However, the knowledge-driven approach allows subject matter experts to 1) establish the permissive values for each evidence layer, 2) determine the mappable location and extent for each evidence layer, thereby “shining light in the black box” often associated with machine learning, and 3) by consensus, apply weights to each evidence layer based on their expert knowledge of the mineralizing system. Expert solicitation establishes 1) the importance of the theoretical criterion to the mineral system, 2) applicability that the mappable proxy reflects the result of the desired process given by the theoretical criterion, and 3) confidence in the data source both in terms of spatial accuracy and overall data quality (Meyer and Brooker, 1991).

In our modeling, known CD deposits/occurrences were randomly divided into “training” and testing data. Training data were used to sample the value of each evidence layer and statistical analyses determined the permissive range within each evidence layer. Evidence layers were then transformed into appropriate fuzzy membership based on the permissive values. Evidence layers with multimodal distributions were assigned fuzzy memberships individually and then combined using fuzzy overlay. The final prospectivity model was combined using a weighted sum.