Neural-Network Modeling Of Placer Ore Grade Spatial Variability

Abstract

Traditional geostatistical methods have been used in ore reserve estimation for decades. Research in the last two decades or so has added a number of other statistical methodologies for ore reserve estimation procedures. Recent advances in neural networks have provided a new approach to solve this problem. This thesis is focused on the Neural-network modeling for the estimation of placer ore reserve. Due to the spatial variability, multiple dimensional inputs and very noisy drill hole sample data from the selected region, it requires that the neural-network be organized in a multiple-layers to handle the non-linearity and hidden slabs for smoothing the predicted results. Various neural-network architectures are investigated and the Back-propagation is selected for modeling the ore reserve estimation problem. Sensitivity analysis is performed for the following parameters: the type of neural-network architecture, number of hidden layers and hidden neurons, type of activation functions, learning rate and momentum factors, input pattern schedule, weight updated, and so on. The influences of these parameters on the predicted output are analyzed in details and the optimal parameters are determined. To investigate the accuracy and promise of neural network modeling as a tool for ore reserve estimation, the ore grade and tonnage of Neural-network output is compared with those estimated by geostatistical methods under various cut-off grades. In addition, the overall performance is also validated by the analysis of R-squared (R2), Root-Mean-Squared (RMS), and the comparison between predicted values and 'actual' values. As the final part of this study, the optimized Neural Network was used to estimate the distribution of placer gold grade and volume of gold resource in offshore Nome. The predicted results for all the mining blocks in the lease area are validated by checking the values of RMS, R2, and Scatter plots. The estimated gold grades are also presented as contour maps for visualization.