Ore Body Characterisation using Machine Learning and Measure-While-Drilling Data

The Challenge

Bench-scale orebody characterization is fundamentally constrained by sparse and expensive exploration drilling, leading to large spatial gaps and high geological uncertainty. Conventional wireline logging introduces operational risks and physical limitations, while MWD data, despite being abundant, remains underutilized due to high variability, alignment challenges with other datasets, and the inability of traditional statistical methods to capture non-linear relationships between drilling responses and rock properties. This results in limited, inconsistent subsurface understanding and suboptimal decision-making across mining operations.

Proposed Solution

Machine learning models (notably Random Forests, Decision Trees and Gaussian Processes) can reliably predict geotechnical, geophysical and geochemical properties from MWD data, achieving high accuracy (up to R² ≈ 0.98) when supported by appropriate feature engineering. Feature importance analysis reveals that pressure-related variables (e.g., bit air pressure) and derived ratio features outperform traditional indicators like penetration rate and torque, challenging established assumptions. The research establishes a unified, scalable framework capable of integrating multi-domain data to deliver sub-meter resolution orebody characterization.

Proposed Benefits to WA

For WA’s iron ore operations, particularly in the Pilbara, this approach enables real-time, high-resolution orebody knowledge using low-cost production drilling data, reducing reliance on expensive exploration programs. Improved prediction of rock strength, geotechnical conditions and ore characteristics support optimized drill-and-blast design, enhanced grade control and more efficient extraction strategies. This translates to reduced waste processing, improved resource recovery, lower operational costs and safer mining practices, directly enhancing productivity and economic returns in WA’s primary export industry.

Thesis DOI

https://doi.org/10.25917/curtin.31769020