Integrated ore classification using stand-alone and hybridised machine learning algorithms

A state-of-the-art, robust classification model using artificial intelligence (AI) algorithms was developed in this study to estimate the gold ore classes based on trace elements measured by inductively coupled plasma (ICP). This approach involves the utilisation of stand-alone machine learning (ML) algorithms integrated within a committee machine (CM) framework. To this end, 19 trace elements including arsenic, bismuth, cadmium, cobalt, chromium, copper, iron, mercury, manganese, molybdenum, nickel, lead, antimony, selenium, tin, strontium, thorium, uranium and zinc acquired from eight drill holes in the Sari-Gunay gold-polymetallic mine (SGGM), Iran, were used as the inputs and three gold grade classes (ore, low-grade ore, and waste) were specified as outputs classes. The classes were defined based on their gold grades and mineable zones, using threshold values of > 1 g/t, 0.5-1 g/t, and < 0.5 g/t for ore, low-grade ore and waste, respectively. Eight stand-alone ML-based classifiers, including the back-propagation neural network (BP-NN), Takagi and Sugeno fuzzy inference system (TS-FIS), adaptive boosting (AdaBoost), GentleBoost, LogitBoost, random under-sampling boosting (RUSBoost), extreme gradient boosting (XGBoost) and adaptive neuro-fuzzy inference system (ANFIS) were used to perform the initial classification. The accuracy, precision and recall metrics were employed to evaluate the performance of the algorithms by comparing predicted results with reference data derived from ICP analysis and geological studies, providing a consistent reference benchmark. The parameters within each algorithm were tuned using the parameter-tuning method, aiming to identify the optimal model. The AdaBoost algorithm outperformed the other stand-alone algorithms. To further improve the stand-alone-derived classification outcomes and integrate the algorithms into a unified algorithm, genetic (GA) and simulated annealing (SA) optimisation algorithms were configured in the framework of the CM. Parameter-tuning of both optimisation algorithms produced multiple distinct models to determine the optimal weights and the best-performing model. The CM with SA (CMSA) outperformed the GA optimiser. Employing CMSA resulted in a 7.28% improvement in overall accuracy compared with the average performance of the stand-alone algorithms.