The development of machine learning (ML) algorithms has led to promising advances in lithology
classification from geophysical logs, which is indispensable in various underground engineering
applications. Due to the resolution and availability of logging data, ML models can recognise lithology
change in a definitive form at a fine scale (0.01 m). While the results from previous models indicated
promising performance in the classification, the robust capability of generalising lithology prediction
to unseen data may not be true for those models due to: (1) data bias caused by borehole selections,
and (2) logging data mismatches among boreholes. To this end, this paper aims to investigate the
generalisability of ML models and seek potential improvements. Four ML models were selected to
be trained on single reference boreholes to test the other (as unseen) in the same region, which
includes Support Vector Classifier (SVC), Random Forest (RF), eXtreme Gradient Boosting
(XGBoost) and Residual Neural Network (ResNet10). The data set involves 11 boreholes from a
coalmine in north-western Bowen Basin (Queensland, Australia): density, gamma ray, neutron, and
sonic logs are selected as inputs. Additionally, a data adaptation method is applied for better
generalisability. The results show that there is an accuracy trade-off between the same borehole and
unseen boreholes across the models. It also indicates a 16–43 per cent reduction in model
performance when generalising the predictions (in marco-F1 score), while the adapted data sets can
contribute to a around 12 per cent improvement. This study provides a fundamental understanding
of the model generalisability when using a single borehole, essential for further correlating boreholes
for regional lithology classification; the adaptation method can improve the generalised accuracy,
and reduces the labour to label lithology, which facilitates the identification of gas storage
mechanisms, geological/geophysical modelling, and stratigraphic analysis.