Error reduction in long-term mine planning estimates using deep learning models

Abstract

The long-term mine planning model (LTMP) and short-term mine planning model (STMP) are two approaches that describe the ore content in a mine; they are essential intangible resources that determine a mining operation and its feasibility. These models are obtained with geostatistical methods and, given their nature, are prone to discrepancies with one another. To reduce these differences, we studied the performance of deep learning (DL)-based models in ore grade estimation for a copper mine in Chile. Specifically, feedforward neural network (FNN), one-dimensional (1D) convolutional neural network (CNN), and long short-term memory (LSTM) models were analyzed. The experiment consisted of a dataset with 732,870 samples, obtained after data cleaning and selection. The use of spatial information in the samples was also studied, adding contextual information for estimation. This led to a dataset of 545,768 samples that were used to evaluate 1D CNN and LSTM models. Architecture tuning was performed by the k-fold cross-validation (CV) method, and hyperparameters such as the number of layers, number of neurons, activation function, and kernel size were optimized. The resulting hyperparameters were used to perform a final evaluation. Performance was measured based on the mean absolute error (MAE), mean squared error (MSE), and coefficient of determination (R2) metrics. A baseline is created using LTMP and STMP estimates to quantify the improvement in the performance of DL-based models. The experimental results revealed the ability of DL-based models to significantly improve copper grade estimates provided by standard mining industry methods in the context under study. For MSE in the final tests, FNN improved by 21%, CNN by 37%, and LSTM by 39% over baselines.