Resource estimation of a placer gold deposit with uncertainty: applying the Bayesian neural network model

Transactions of the Society for Mining, Metallurgy, and Exploration , 2011, Vol. 330, No. 1, pp. 606-617

Chatterjee, S.; Bandopadhyay, S.

A spatial modeling technique based on the Bayesian neural network (BNN) is proposed. Incorporation of the Bayesian method for posterior probability of the output parameter helps calculate the uncertainty associated with the estimate. Parameters of the BNN model were selected using a grid search algorithm. The expected value and the variance of block support were calculated by Markov chain Monte Carlo sampling of the posterior distribution at discretized points within the block. The BNN model was validated by applying it to the Walker Lake data set and comparing the results with ordinary kriging results. The comparison revealed that the proposed BNN method performs marginally better than ordinary kriging. The variance map is less smooth than with ordinary kriging, and the proportional effect is less in the BNN-based model than in the ordinary kriging model. The developed BNN model was applied to the Coho block of a placer gold deposit in Nome, AK. Results reveal that the BNN model is globally biased for the Coho block and underestimates the sample mean. The grade volume curve was developed by calculating the block mean and block variance after discretizing the block into a number of points and applying the BNN model to each discretized point.