Christopher Mark Yeomans , Robin Shail, Stephen Grebby, Vesa Nykänen, Maarit Middleton, Paul Lusty

Novel mineral prospectivity modelling presented here applies knowledge-driven feature extraction to a data-driven machine learning approach for tungsten mineralisation. The method emphasises the importance of appropriate model evaluation and develops a new Confidence Metric to generate spatially refined and robust exploration targets. The data-driven Random Forest™ algorithm is employed to model tungsten mineralisation in SW England using a range of geological, geochemical and geophysical evidence layers which include a depth to granite evidence layer. Two models are presented, one using standardised input variables and a second that implements fuzzy set theory as part of an augmented feature extraction step. The use of fuzzy data transformations mean feature extraction can incorporate some user-knowledge about the mineralisation into the model. The typically subjective approach is guided using the Receiver Operating Characteristics (ROC) curve tool where transformed data are compared to known training samples. The modelling is conducted using 34 known true positive samples with 10 sets of randomly generated true negative samples to test the random effect on the model. The two models have similar accuracy but show different spatial distributions when identifying highly prospective targets. Areal analysis shows that the fuzzy-transformed model is a better discriminator and highlights three areas of high prospectivity that were not previously known. The Confidence Metric, derived from model variance, is employed to further evaluate the models. The new metric is useful for refining exploration targets and highlighting the most robust areas for follow-up investigation. The fuzzy-transformed model is shown to contain larger areas of high model confidence compared to the model using standardised variables. Finally, legacy mining data, from drilling reports and mine descriptions, is used to further validate the fuzzy-transformed model and gauge the depth of potential deposits. Descriptions of mineralisation corroborate that the targets generated in these models could be undercover at depths of less than 300 m. In summary, the modelling workflow presented herein provides a novel integration of knowledge-driven feature extraction with data-driven machine learning modelling, while the newly derived Confidence Metric generates reliable mineral exploration targets.