Hybrid Surrogate Approach for Modelling a Flotation Process - Combining Machine Learning and Physics

The large computational resources required and the slowness to model and simulate complex real-world processes have led to the need for faster and lighter surrogate models. These accelerated computational representations of complex models, for example, in the field of mineral processing, can help predict and optimise the productivity of industrial systems. A hybrid surrogate modelling methodology integrates numerous surrogate modelling methodologies in order to combine their benefits while minimising their flaws. This thesis investigates different approaches to develop machine learning and physics-based hybrid surrogate models. The aim is to create a hybrid surrogate model for the flotation cell in a gold mine, combining historical process data and process dynamics-related information. The goal is to enhance the machine learning model's capability to describe process dynamics, thereby improving its performance in different process scenarios.

The main approaches to combine machine learning and physics information are physics-informed features and labels, physics-guided loss functions, physics-guided initialization, physics-guided architectures, and residual modelling. The physics-informed features and physics-guided loss function approaches were selected to be investigated further in the experimental part, where future concentrate grades of the flotation cell are predicted. The physics-informed features were generated from a data set collected from the real-world process and using a physics-based simulator, used as a digital twin. The experiments with features included feature ranking and selection based on the model’s performance. Two different physics-guided loss functions using the information on how the mass of gold flows in the flotation cell were implemented, and their performance was compared to a purely data-driven approach.

The best-observed model employs a physics-guided loss function, targeting actual gold flow balance values to align the predicted concentrate grade's gold flow balance with the process physical realities. As an input, the model used a combination of physics-informed features and features taken directly from the data set. When compared to the best-performing data-driven model, the physics-guided model performed slightly better, showing to be a potential approach to this surrogate modelling problem.