Environment
Marco Antonio Cotrina Teatino; Jairo Jhonatan Marquina-Araujo; Jose Nestor Mamani-Quispe; Juan Antonio Vega-Gonzalez; Moises Bartolome Guia-Pianto
Abstract
The Quiulacocha tailings deposit in central Peru, containing 70 Mt of historical mine waste, presents both environmental risks and opportunities for secondary metal recovery. This study applies data-driven machine learning techniques to estimate the remaining silver resources using 927 one-meter composites ...
Read More
The Quiulacocha tailings deposit in central Peru, containing 70 Mt of historical mine waste, presents both environmental risks and opportunities for secondary metal recovery. This study applies data-driven machine learning techniques to estimate the remaining silver resources using 927 one-meter composites from 40 vertical drillholes. Three supervised learning models—Random Forest (RF), k-Nearest Neighbors (KNN), and Extreme Gradient Boosting (XGBoost)—were trained using spatial coordinates (X, Y, Z) as the sole input features. Model validation was performed using leave-one-out cross-validation (LOOCV), and results were benchmarked against ordinary kriging (OK). Among the models, RF delivered the highest predictive performance (mean error = 0.53 g/t, RMSE = 7.21 g/t, R = 0.82), outperforming OK (R = 0.63, RMSE = 10.47 g/t). Block model predictions indicated higher silver content from machine learning models: 1,532.86 t (RF), 1,542.16 t (XGBoost), and 1,492.09 t (KNN), compared to 1,463.73 t from OK. Additionally, XGBoost maintained superior grade-tonnage relationships under elevated cutoff thresholds, highlighting its potential to delineate high-grade subdomains within the deposit. These findings confirm the value of machine learning in resource estimation under conditions of low spatial continuity, such as tailings, where material mixing and irregular deposition patterns limit correlation across space.
Exploration
Jairo Jhonatan Marquina Araujo; Marco Antonio Cotrina Teatino; José Nestor Mamani Quispe; Eduardo Manuel Noriega Vidal; Juan Antonio Vega Gonzalez; Juan Vega-Gonzalez; Juan Cruz-Galvez
Abstract
The objective of this research work to employ machine learning techniques including Multilayer Perceptron Artificial Neural Networks (ANN-MLP), Random Forests (RFs), Extreme Gradient Boosting (XGBoost), and Support Vector Regression (SVR) to predict copper ore grades in a copper deposit located in Peru. ...
Read More
The objective of this research work to employ machine learning techniques including Multilayer Perceptron Artificial Neural Networks (ANN-MLP), Random Forests (RFs), Extreme Gradient Boosting (XGBoost), and Support Vector Regression (SVR) to predict copper ore grades in a copper deposit located in Peru. The models were developed using 5654 composites containing available geological information (rock type), as well as the locations of the samples (east, north, and altitude) and secondary ore grade (Mo) obtained from drilling wells. The data was divided into 10% (565 composites) for testing, 10% (565 composites) for validation, and 80% (4523 composites) for training. The evaluation metrics included SSE (Sum of Squared Errors), RMSE (Root Mean Squared Error), NMSE (Normalized Mean Squared Error), and R² (Coefficient of Determination). The XGBoost model could predict the ore grade with an SSE of 15.67, RMSE = 0.17, NMSE = 0.34, and R² = 0.66, the RFs model with an SSE of 16.40, RMSE = 0.17, NMSE = 0.36, and R² = 0.65, the SVR model with an SSE of 19.94, RMSE = 0.19, NMSE = 0.43, and R² = 0.57, and the ANN-MLP model with an SSE = 21.00, RMSE = 0.19, NMSE = 0.46, and R² = 0.55. In conclusion, the XGBoost model was the most effective in predicting copper ore grades.
