High-Confidence and Efficient Drill Programs for Expansion and Brownfields Exploration Using Assay Data and Lithological Logging Using Machine Learning; Data Evaluation Techniques, Practical Input Data Construction, and Metrics for Precision Targeting * Farzi Yusufali¹, Julia Oliveira² ¹ Stratum AI, CEO, Toronto, Canada (farzi@stratum.ai) ² Stratum AI (julia@stratum.ai) ABSTRACT Due to the increased demand for critical minerals, expanding and exploiting existing resources is as important as finding new deposits. Given rising exploration costs in both greenfield and brownfield settings, efficient targeting methods are an active research area. Machine learning (ML) models—algorithms that learn patterns in data without explicit instructions—have been applied mainly in greenfield exploration, typically learning from soil samples, limited drilling and hyperspectral sensor data. Similarly, these models can be used in existing operations to create value, directly impacting net present value (NPV), increasing resources and reserves, and providing a cost-effective method for adding de-risked economic blocks to a site's mine plan. While being a powerful set of algorithms, ML and deep learning (DL) in particular cannot be applied without significant data analysis and tailored evaluation criteria. Applying these systems to geological modelling requires quantitative analysis of which data should be incorporated. Adding all available data is unfeasible: finite computational resources, overfitting, signal-to-noise ratio management, false trends from assay bias, and the subjective nature of geological logging are all significant detriments to target generation. The authors present and test a method for measuring confidence across multiple resource models to identify which data can serve as pathfinders for target generation. The technology has been applied and validated through successful drill programs on three deposit types: (1) gold at a Western Australian orogenic lode deposit, (2) copper expansion at a Chilean Manto-type ironoxide copper-gold deposit, and (3) gold modelling at an orogenic lode deposit in western Kazakhstan. The authors also test different methods of integrating data, quantifying relative accuracy, and empirically measuring prediction confidence on a block-by-block basis within resource models. KEYWORDS Deep learning, resource estimation, target confidence, geological logging, pathfinder analysis, drill program optimization, IOCG, orogenic gold
RkJQdWJsaXNoZXIy MTM0Mzk2