Predicting ore content throughout a machine learning procedure – An Sn-W enrichment case study

Abstract

The distribution patterns of trace elements are very useful for predicting mineral deposits occurrence. Machine learning techniques were used for the computation of adequate models in trace elements' prediction. The main subject of this research is the definition of an adequate model to predict the amounts of Sn and W in the abandoned mine area of Lardosa (Central Portugal). Stream sediment samples (333) were collected within the study area and their geochemical composition-As, B, Be, Cd, Co, Cr, Cu, Fe, Ni, P, Sn, U, V, W, Y, and Znused as input attributes. Different machine learning techniques were tested: Decision Trees (CART), Multilayer Perceptron (MLP) and Support Vector Machines (SVM). For regression and clustering, CART, MLP approaches were tested and for the classification, problem SVM was used. These algorithms used six different inputs-N1 to N6-aiming to pick out the best-performing model. The results show that CART is the optimized predictor for Sn and W. Concerning the regression approach, correlation coefficients of 0.67 for Sn (with Input N1) and 0.70 for W (with Input N3) were obtained. Regarding the classification problem, an error rate of 0.10 was reached for both Sn (Input N1) and W (Input N2). The classification process is the best methodology to predict Sn and W, using as input the trace element concentrations in the collected stream sediment samples, Lardosa area, Portugal.

Estimation of the dolomite content of carbonate rock outcrops based on spectral knowledge and machine learning

“…SVR is a regression analysis method derived from the principles of SVM. The foundational concept of SVR is to seek an optimal balance between model accuracy and generalizability (Smola and Schölkopf, 2004;Iglesias, 2020). The primary aim of SVR is to identify a smooth model function that minimizes the discrepancy between the estimated and actual values.…”

“…Due to their consideration of training samples and feature interdependencies, embedded approaches typically yield superior model performance compared with other feature selection methods (Kumar, 2014;Thomas and Gupta, 2020). Algorithms frequently used in the modeling stage include partial least squares regression (PLSR) (Knox et al, 2015), ensemble learning (Carranza, 2015;Tan et al, 2020;Lin et al, 2022), SVM (Iglesias, 2020;Chatterjee et al, 2022), artificial neural networks (ANN) (Saikia et al, 2020), and deep learning (Zhang T. et al, 2023). Nonlinear models are widely regarded to outperform linear models in a variety of instances (Zhou et al, 2021;Sim et al, 2023).…”