Flotation is the dominating process in the global copper, lead, and zinc mining industries to separate valuable minerals from waste material. In the upstream process steps, the ore is ground to liberate all mineral grains, and mixed with water to form a slurry. In flotation, chemical reagents are added to improve the hydrophobic properties of selected minerals. When air is added, these minerals follow the air bubbles to the surface and can be extracted in the resulting froth, forming a concentrate. This process is implemented in flotation tanks interconnected in a complex circuit that often includes re-grinding and recirculation. Floatation is a pivotal process step, as it defines the recovery (yield), which has a proportional impact on both environmental aspects and the financial result of the company.
Today, the flotation process is typically controlled semi-manually, where simple control loops stabilize tank levels and flow rates, while operators adjust parameters like airflow, reagent- and lime- addition based on the available measurements and experience. Model predictive control solutions have been attempted, with some success. However, performance is severely limited by poor model accuracy and the inability to adapt to changes in ore properties as new areas of the mine are excavated. To increase efficiency and autonomy of mineral processing, these challenges must be addressed.
Therefore, this PhD project addresses modeling of the flotation process for control purposes. Data-driven modeling through machine learning (ML) techniques holds great potential, but several aspects must be addressed before it can be applied in an industrial setting. In our setting, observation of a process is limited by physical restrictions and the available measurement technology. Furthermore, the effect of the measured properties on the system state are often both complicated and only known conceptually, making the interpretation of the measurements challenging. To use the limited data that is available efficiently, we will combine machine learning with physics-based modeling, avoiding wasting scarce data on learning known laws of physics.
In this project, we will therefore push the state-of-the-art within mining process control by complementing machine learning with physics-based models based on e.g., conservation laws.
We firmly believe that the future of floatation control, and indeed many associated process steps, lies in incorporating informal operator know-how into dynamical models, to enable model-based control solutions, where traditional data-driven paradigms suffer from lack of informative data. This physics-informed machine learning approach to increased autonomy is becoming increasingly feasible thanks to advances within scientific machine learning methodology. Within mining and the process industries, embracing it will provide the opportunity to increase efficiency of resource utilization, thereby enabling the transition to a sustainable technological future.