APPLICATION OF SEQUENTIAL INDICATOR SIMULATION TO MODEL NON-STATIONARY GEOLOGICAL DOMAINS COMBINING WITH A MACHINE LEARNING ALGORITHM

Abstract

Resource estimation is an essential aspect of the development process for any mining project. The geological domains are defined based on data obtained from boreholes, with the goal being to determine the mineral grades in the geological domains. Geostatistics assumes that the joint distribution of geological attribute values is consistent across homogeneous domains and is defined by a stationary covariance function. However, the nature of geological systems often contains uncertainties and variations in structure and behaviour. Sequential Gaussian and Sequential Indicator Simulation are one of several methods used for simulating continuous and categorical variables in 3D geological modelling. Despite its advantages, this method and other conventional techniques have been criticized for not effectively capturing local mean values, variance, and spatial continuity changes. The traditional algorithms used in the industry are not suitable for non-stationary geological domains, as they are designed for stationary target simulation variables. This thesis proposes using Multinomial Logistic Regression as an alternative method for simulating the spatial properties of non-stationary geological domains. The technique will be applied to a copper-porphyry deposit that shows clear signs of non-stationarity. The mineral resource model will be created by weighting the copper grade estimates based on the probability of occurrence of different rock types in various geo-domains. The generated probability maps will be evaluated using various criteria, including visual inspection of realizations, probability maps, replicas of each geo-domain fraction, connectedness metrics, and trend analysis.