Estimation of iron ore grade distribution has been attempted using geostatistics and artificial neural network (ANN) based models for a banded iron formation (BIF) type iron ore deposit of East India. Geologically, it is a supergene iron ore formation through an enrichment process of parent material (BIFs) by gradual leaching of impurities, namely SiO2 and Al2O3. As the degree of leaching differs spatially, distribution of Fe varies depending on local controls of ore formation. While geostatistical methods work on models based on spatial correlations of mineralisation parameters, ANN models a non-linear distribution and is widely used as a global approximator. Spatial distribution modelling of Fe values has been carried out employing ordinary kriging (OK) and sequential Gaussian simulation (SGS). Non-linear modelling has been performed using hybrid neural networks (HNN) that provided an alternative option for grade estimation. Classical statistical modelling revealed a negatively skewed 3-parameter lognormal Fe population having a mean of 60.49 per cent with a skewness of -0.08 and a kurtosis of 2.60. Semi-variography and geostatistical estimation employing OK procedure led to block-by-block kriged estimates with associated kriging variance, which provided an overall mean kriged estimate and kriging variance of Fe as 60.29 per cent and 7.76 (per cent)2 respectively. SGS procedure revealed a mean of 60.71 per cent and a standard deviation of 5.89 per cent with minimum and maximum values as 45.12 per cent and 68.70 per cent respectively in respect of Fe. HNN model provided values of mean squared error (MSE) and correlation coefficient (R2) as 12.33 and 0.61 respectively, thereby exhibiting the robustness of the model to sustain maximum heterogeneity within the target domain. A comparison of Fe grade estimates obtained by kriging, simulation and HNN models more or less maintain similar distribution patterns that are close to the sample data. It is thus opined to carry out the grade prediction exercise employing all these three modelling approaches to obtain alternative scenarios for improved mine planning and design exercises. The techniques provide a foundation for cogent grade modelling leading to improved iron ore reserve estimation procedures.
Sarkar, B C, Singh, R K, Ray, D, Kumar, A, Sinha, P K and Sarkar, V, 2017. Iron ore grade modelling using geostatistics and artificial neural networks, in Proceedings Iron Ore 2017, pp 431–438 (The Australasian Institute of Mining and Metallurgy: Melbourne).