Application of Autoencoders Neural Network and K-Means Clustering for the Definition of Geostatistical Estimation Domains

Abstract

The objective of this study was the definition of estimation domains through the application of an artificial neural network Autoencoders and K-Means clustering. The study was based on the analysis of 5,654 composites obtained from an exploratory drilling campaign in a copper deposit. The specific architecture of the autoencoder included an encoder and a decoder, each composed of multiple layers and ReLU activation functions. The encoder, with four hidden layers of 600, 600, 800 and 10 neurons, respectively, was complemented by a decoder that replicated this structure. Application of the K-Means algorithm, with 30 initializations on these encoded representations, culminated in a silhouette score of 0.261 and an inertia of 17,447.44, revealing the optimal formation of two distinct estimation domains: domain 1, with 4,204 samples and an average copper grade of 0.44%, and domain 2 with 1450 samples and an average grade of 0.41% copper. Compared to the geochemical modeling approach in definition of estimation domains, a significant reduction in the mean error (0.29 vs. 0.05) and in the error variance (0.04 vs. 17.36) was observed. In conclusion, this approach not only complements geostatistical estimation techniques, but also improves accuracy and reliability in geological resource estimation.