Machine learning framework for the integration of drill-core hyperspectral and geochemical point data to enhance geological modeling

In mineral exploration, on-site analytical techniques provide tools for real-time data acquisition, supporting informed decision-making. Portable instruments such as handheld X-ray fluorescence (pXRF) and short-wave infrared (SWIR) hyperspectral spectrometers enable rapid, non-destructive collection of geochemical and mineralogical information directly from drill cores. When effectively integrated and interpreted, these datasets offer powerful tools for advancing geological understanding and refining 3D models, ultimately improving vectoring toward mineralization and supporting more efficient, sustainable exploration

Where traditional interpretation methods are often subjective and time-consuming, data-driven approaches, particularly machine learning, can identify patterns and correlations within large datasets, accelerating analysis. In this study, we propose a machine learning framework for fusing drill-core hyperspectral and geochemical point data to enhance geological modeling.

Methodologies were applied and tested in two gold target sites hosted by an Archean Ilomantsi Greenstone Belt in eastern Finland. The geology at the selected sites is dominated by visually homogeneous schistose metasediments exhibiting intense sericite–chlorite alteration. Hence, these target areas provide an ideal natural environment for evaluating machine-learning approaches aimed at refining lithological and lithogeochemical discrimination and alteration mineralogy interpretations. The data-fusion and predictive modeling approach has the potential to significantly extend the data-driven geological models in 3D to enhance geological understanding and controls of the Au mineralization.

Lithogeochemical data were first partitioned into distinct compositional groups using the K-means clustering algorithm. The resulting cluster assignments served as training labels for a supervised learning framework aimed at linking geochemical classes to hyperspectral signatures. Selected SWIR spectral parameters corresponding to geochemical sampling points, together with their assigned labels, were used to train a Random Forest (RT) classifier. The trained model was applied to unclassified spectral data to infer lithogeochemical classes to produce a predictive model.

Despite the generally noisy nature of both pXRF and spectral point data and overall, rather poor probability measures of the RT model (< 50% for most classes), in 3D, a clear and spatially reasonable model is produced. Along-strike continuation of lithogeochemical stratigraphy provides a validation argument supporting the success of the predictive model beyond areas with both lithogeochemical and hyperspectral data.

This approach leverages existing drill holes in a fast and cost-efficient manner by utilizing portable data-acquisition technologies. Machine-learning-based integration of multi-sourced datasets is demonstrated to improve lithological/lithogeochemical discrimination and predict subsurface geological features. This aids in the delineation of drilling targets more accurately, supporting dynamic, data-driven decision-making in mineral exploration.