Integrated Machine Learning and Process Optimization for Fluidized Bed Separation of Metamorphosed BIF Ore

Abstract

The declining availability of high-grade iron ore has intensified the need to utilize low-grade resources such as banded magnetite quartzite (BMQ) ore and industrial wastes. This study explores the use of a liquid–solid fluidized bed separator to improve the beneficiation of lean-grade BMQ ore through experimental investigation and modeling. Hydrodynamic behavior was evaluated by measuring pressure drop and bed expansion over a range of superficial velocities, demonstrating that particle size, bed height, and superficial velocity significantly influence fluidization characteristics, with coarser particles exhibiting lower flow resistance and more stable bed behavior. Mineralogical studies of the sample confirmed the presence of magnetite, hematite, quartz, and minor silicates. Experimental results aligned well with theoretical predictions, including minimum fluidization velocity and pressure drop. Optimal beneficiation was achieved at a feed size of −850 +500 μm and an overflow tap height of 12 cm, producing 44.26% Fe grade, 83.31% recovery, and 24.54% separation efficiency. A Levenberg– Marquardt algorithm-based artificial neural network (ANN) was employed to iteratively optimize network weights and biases, enabling rapid convergence and highly accurate prediction, with high predictive accuracy (R2 up to 0.9896). Sensitivity analysis identified feed size and bed height as key parameters, with bed height being the most significant. ANN effectively modeled nonlinear relationships and serves as a robust prediction tool. The process offers a sustainable, reagentfree alternative, reducing environmental impact and supporting efficient utilization of low-grade iron resources.