Abstract

Hydrometallurgy is the discipline that deals with the selective extraction of metals of interest, with leaching as one of its fundamental stages. This process operates across multiple scales, with complex chemical reactions between the mineral and the leaching agent occurring at the particle level scale. Given the complexity of modeling its kinetics, machine learning techniques have emerged as valuable tools to surrogate these unknown mechanisms. In this context, a hybrid model was developed in Python for the heap leaching process of copper oxide ores as a proof of concept. The model combines a phenomenological description of fluid flow in porous media with neural networks that surrogate the dissolution kinetics in acidic media to predict the copper recovery and the acid consumption. The model was trained using batch-scale experimental data and subsequently validated with column leaching scale tests. Training results showed a consistent and progressive reduction of the value for the loss function, reaching a value of 3.2⋅10^(-02). The column scale validation showed an RMSE between 1.68% to 5.96%. These discrepancies are likely attributable to mineralogical heterogeneity and variations in environmental temperature, which were not explicitly accounted for during training of the neural network. Results show that integrating neural networks with first principles models is a useful tool to represent microscale complex systems. This modelling approach combines the robustness of the phenomenological aspects at a large scale with the capabilities of the neural networks to capture unknown phenomena at a microscale, allowing broader mathematical representation for engineering applications.