Workflow-Induced Uncertainty in Data-Driven Mineral Prospectivity Mapping

The paper titled delves into the challenges and implications associated with employing data-driven methodologies in mineral prospectivity mapping (MPM). The overarching objective of MPM is to facilitate the identification and delineation of potential areas rich in natural resources. However, the application of data-driven approaches introduces variability in the generated maps based on the specific workflow utilized.

The abstract elucidates that despite the prevalent use of data science frameworks to standardize and streamline MPM workflows, the resulting maps may not necessarily optimize spatial selectivity. The study aims to investigate the interplay between various components within a geodata science-based MPM workflow and their impact on the spatial accuracy and selectivity of the produced maps. Specifically, the study focuses on three modulated factors within the modeling stage: (1) feature space dimensionality, (2) choice of machine learning algorithms, and (3) performance metrics guiding hyperparameter tuning.

The findings highlight the presence of significant local minima within typical geodata science-based MPM workflows. This indicates that different combinations of workflow choices can yield highly discriminatory models, leading to variability in the resultant maps. Moreover, the study identifies inconsistencies in the relationships between variable domain metrics and spatial selectivity, thereby introducing what the authors term "workflow-induced uncertainty."

To address this uncertainty, the paper advocates for the application of the canonical concept of scientific consensus from the broader experimental science framework. It suggests that the reliability of MPM findings is contingent upon the degree of consensus among experimental outcomes. Consequently, the authors propose deliberate modulations of workflow components as a means to understand and quantify workflow-induced uncertainty. By expanding the exploration and experimentation with workflow design, the study contends that more meaningful reductions in the physical search space for natural resources can be achieved.