Predicting iron contents in the Tamra-Douahria mining site using a deep neural network approach

The well-known Douahria-Tamra mining site is characterized by the presence of deposits with high variability in composition, colour, and structural-textural peculiarities, especially in the exploitable layers. Thus, understanding the underlying reasons for this heterogeneity is crucial to optimize the extraction processes, ensuring consistent product quality, and maximizing resource utilization. This was the motivation beyond the attempt allocated to shed light on the behaviour of iron and other related ores in this mining district. Iron content was estimated from the measured lead, zinc, manganese, silica and arsenic using unsupervised machine learning tools (HCA and PCA) and deep neural network. For this purpose, 357 iron-rich samples collected in the Tamra-Douahria mining sub-district were used to train, test and validate the obtained models. Out of 357, 285 data sets were selected for training the algorithm while 72 data points were used for model testing and validation. Input variables included lead (Pb), zinc (Zn), manganese (Mn), arsenic (As) and silica (SiO2) contents, while iron content (Fe in %) was considered as output. Our results indicated a mean value of iron content (26.19%) was perfectly predicted to 26.09% by the DNN model. A cross-validation step was necessary to confirm the robustness of the proposed models by using the well-known coefficient of determination (R2). Our results indicated high coefficient of determination (R2 = 0.9978) and Pearson correlation coefficient (0.999) with low RMSE (0.975) which confirm the accurate predictions of the actual values. Therefore, the proposed DNN model was robust in predicting iron contents in the studied Douahria-Tamra mining site.