The reliability of rock-type prediction using down-hole density, gamma ray response, and magnetic susceptibility measurements was evaluated at the Victoria property, Sudbury, ON. A supervised neural network, trained using lithological information from drill hole FNX1168, yields a predictive accuracy of 83% for the training data. Applying the trained network on drill hole FNX1182 resulted in 64% of the rock types being correctly classified when compared with the classification produced by geologists during logging of the core. The homogenous rock types, like quartz diorite, had a high accuracy of classification; while the heterogeneous rock types such as diabase were poorly classified. Overlap between physical properties of rock types caused by heterogeneity or inherent similarity in physical properties of rock types, which were verified by observing the cores, accounts for most of the misclassification. To reduce the misclassification, the network was trained from physical log units in FNX1168 derived from clustering of physical properties measurements. Four physical log units mainly represented four groups of rocks: i) quartz diorite; ii) metabasalt and metagabbro; iii) metasediment and quartzite; and iv) sulfide and diabase. The predictive accuracy in the training process rose to 95%. The trained network then was applied to predicting the physical log units in FNX1182. Given the relationships between physical log units and rock types from FNX1168, the results of physical-log-unit classification in FNX1182 were interpreted from a geological point of view. Although in ideal cases we would like to be able to extract the same classification that a geologist provides, the extraction of physical log units is a more realistic goal. The interpretation of the lithological units from the physical log units can be compared with the geologist's, classification and discrepancies or anomalies analyzed in greater detail. (C) 2015 Elsevier B.V. All rights reserved.
Dc1cx<br/>Times Cited:1<br/>Cited References Count:24
