Please use this identifier to cite or link to this item:
|Title:||Application of physical properties measurements to lithological prediction and constrained inversion of potential field data, Victoria Property, Sudbury, Canada.|
|Keywords:||Down-hole physical properties;Fuzzy k-means clustering;Neural network;Classification;Constrained geophysical inversion;Litho-prediction model;Derivative analysis;Layer boundaries|
|Abstract:||In recent years the number of near-surface deposits has decreased significantly; consequently, exploration companies are transitioning from surface-based exploration to subsurface exploration. Geophysical methods are an important tool to explore below the surface. The physical property data are numerical data derived from geophysical measurements that can be analyzed to extract patterns to illustrate how these measurements vary in different geological units. Having knowledge of links between physical properties and geology is potentially useful to obtain more precise understanding of subsurface geology. Firstly, down-hole density, gamma radioactivity, and magnetic susceptibility measurements in five drillholes at the Victoria property, Sudbury, Ontario were analyzed to identify a meaningful pattern of variations in physical property measurements. The measurements grouped into distinct clusters identified by the fuzzy k-means algorithm, which are termed ‘physical log units’. There was a meaningful spatial and statistical correlation between these physical log units and lithological units (or groups of lithological units), as classified by the geologist. The existence of these relationships suggests that it might be possible to train a classifier to produce an inferred function quantifying this link, which can be used to predict lithological units and physical units based on physical property data. A neural network was trained from the lithological information from one hole, and was applied on a new hole with 64% of the rock types being correctly classified when compared with those logged by geologists. This misclassification can occur as a result of overlap between physical properties of rock types. However, the predictive accuracy in the training process rose to 95% when the network was trained to classify the physical log units (which group together the units with overlapping properties). Secondly, lithological prediction based on down-hole physical property measurements was extended from the borehole to three-dimensional space at the Victoria property. Density and magnetic susceptibility models were produced by geologically constrained inversion of gravity and magnetic field data, and a neural network was trained to predict lithological units from the two physical properties measured in seven holes. Then, the trained network was applied on the 3D distribution of the two physical properties derived from the inversion models to produce a 3D litho-prediction model. The lithologies used were simplified to remove potential ambiguities due to overlap of physical properties. The 3D model obtained was consistent with the geophysical data and resulted in a more holistic understanding of the subsurface lithology. Finally, to extract more information from geophysical logs, the density and gamma-ray response logs were analyzed to detect boundaries between lithological units. A derivative method was successfully applied on the down-hole logs, and picked the boundaries between rock types identified by geologists as well as additional information describing variation of physical properties within and between layers not identified by the geologist.|
|Appears in Collections:||Doctoral Theses|
Files in This Item:
|Omid Mahmoodi_Final_corrected_Thesis.pdf||4.81 MB||Adobe PDF|
Items in LU|ZONE|UL are protected by copyright, with all rights reserved, unless otherwise indicated.