Feedforward neural network (FFNN) optimization and modelling approach for the upgrading of South African coal fines via flotation process

Coal is South Africa's main energy source, with increasing demand for high-quality products requiring upgrading of low-grade coal fines through flotation. This study developed a feedforward neural network (FFNN) using MATLAB's Fitnet to model and optimize the flotation process for South African coal fines. Characterization by X-ray fluorescence and diffraction revealed quartz and kaolinite as dominant minerals. The FFNN showed strong prediction accuracy with R² > 0.9, and statistical tests confirmed time, solids concentration, impeller speed, collector, and frother dosages as significant factors (p < 0.05). Correlation analysis indicated that coal yield and calorific value increased with these variables, while ash content decreased. Flotation kinetics fitted well with the Kelsall model (R² > 0.9); however, kinetic constants K_f and K_s were lower than values reported previously, likely due to kaolinite coating on coal particles. Optimal flotation conditions were identified as 7 min flotation time, 20 % solids, 1600 rpm impeller speed, 2000 g/t collector, and 150 g/t frother dosages. Under these conditions, the flotation process achieved a coal yield of 34 %, ash content of 16 %, and calorific value of 24.92 MJ/kg. These results demonstrate that FFNN modeling combined with kinetic analysis effectively optimizes flotation, enhancing coal upgrading for South African fines.