A data adaptation method to classify inter-borehole lithology proportion from geophysical logs

Geophysical logs provide a basis for estimating lithology proportions in boreholes where direct lithological observations are limited, a common scenario in mining exploration. The same lithology can produce different log responses in different boreholes, and this data mismatch limits the transfer of automated lithology prediction between boreholes. This study evaluates whether quantile-based data adaptation improves the inter-borehole prediction of lithology proportions and how its effect depends on the reference borehole(s). The adaptation is a label-free affine transformation that aligns the quantiles of each log with those of a single reference borehole or of a divergence-weighted combination of several. It was tested with three machine learning models on 11 boreholes in the Bowen Basin, Australia. Relative to scaling each borehole on its own range, the adaptation raised the major-lithology accuracy of the unlabeled boreholes from about 0.5 to 0.56–0.57 with a single reference and 0.63–0.68 with a group of references (Macro-F1), with a similar gain in litho-proportion similarity (cosine similarity). With each borehole used in turn as the reference, the adaptation improved prediction for 9–10 of the 11 references, depending on the model; only the northernmost borehole failed for all three models. The most effective references lay in the structurally quiet part of the area, but position alone did not determine suitability. More broadly, the results show how a limited number of labeled boreholes can be used more effectively for lithology estimation in subsurface characterization, which underpins engineering and mining applications.