OreGenes: An optimized neural network tool for ore deposits classification based on gold grain geochemistry

Assessment of gold grains and their characterization in abundance and morphology can be used as a discriminatory tool for mineral deposits. However, the processes that lead to the formation of an ore deposit are multiscale and nonlinear. As a result, the elemental compositions recorded in gold grains during mineralization seem mostly irregular and unpredictable. Here, we took advantage of the capabilities of neural networks, computational models composed of layers of interconnected nodes. After training, it processed data through nonlinear transformations that approximate a complex natural function. It enabled the recognition of complex patterns and relationships in chemical variability observed in gold and relating it to the mineral system that formed the mineral, allowing for predictive modeling. This was achieved by using published trace element data in gold of four types of gold deposits (i.e., Orogenic, VMS, porphyry, and epithermal) from 47 different localities, obtained by laser ablation-inductively coupled plasma-mass spectrometry (LA- ICP-MS). The model was optimized by training five different architectures on the most influential elements, determined by principal component analysis (PCA). As a result, two hidden layers with ten neurons (series of nodes that process and transmit information) each were found to be the best architecture for using trace elements as predictors of the type of deposit that formed a natural gold grain. In order to encourage the use of the findings made in this paper, we introduce OreGenes, an app developed in Matlab2023b based on the best-obtained model. It allows any user who possesses compatible data the ability to import and process it to obtain a prediction with an average accuracy level of 88.9 % confidence to assess the mineral deposit type that an unknown gold grain came from, which grants the user the ability to have a powerful tool for exploration or research.