Mineral processing in mining focuses on preparing ore for final extraction by concentrating valuable minerals. Froth flotation, the most important and complex concentration technique, is used to separate valuable minerals from gangue. Data-driven machine learning and first-principles methods have been used to model the flotation process. While purely data-driven models are efficient at approximating nonlinear functions, they fail in dynamic industrial environments and are unable to produce physically consistent results, leading to poor generalization. On the other hand, first-principles models are robust but cannot perfectly capture the flotation process complexity due to the interaction of many variables, causing control issues. Integrating mathematical knowledge with machine learning can address these limitations.
This thesis studies the potential and effectiveness of physics-informed machine learning in mineral processing. Using simulated data from two froth flotation cells provided by Metso Oy, the aim was to predict the concentrate gold grade of these cells by incorporating conventional deep learning models and classical mathematical models of the process. To achieve this, first the data were preprocessed to prepare them for model training. Then, three mathematical models, representing the complex behavior of flotation processes, formulated as ordinary differential equations, were employed to construct physics-informed neural network models. Finally, the predictive performance of these models was compared with their purely data-driven counterparts in terms of mean squared error.
The results demonstrate that the physics-informed neural network models yield better predictive accuracy and generalization capabilities compared to the purely data-driven models in both froth flotation cells. The physics-informed neural networks' predictions closely match the actual concentrate gold grades, showing lower mean square errors in both validation and test sets, thereby providing strong generalization capabilities and effectively avoiding overfitting to the training set. In contrast, while data-driven models capture some underlying patterns, their predictions significantly deviate from the actual values, particularly in the test sets, leading to poor generalization, suffering from overfitting and capturing noise.
