Machine learning-based concentrate grade prediction in froth flotation: a review of data representation, model development, and evaluation challenges

Online prediction of concentrate grade is important for flotation monitoring, optimization, and control. However, robust industrial prediction remains difficult because grade-related information is usually inferred from indirect and heterogeneous data. This review examines machine-learning-based concentrate grade prediction in froth flotation from Jan 2015 to Aug 2026. It focuses on four main data routes: froth vision, radiometric/XRF, laser spectroscopic, and process-variable data. The review discusses data representation, feature engineering, label alignment, model development, data quality, and explainability. Recent literature shows an increasing use of deep and structured learning architectures alongside conventional machine-learning methods. These approaches include CNNs, attention mechanisms, Transformer-based temporal or spatiotemporal models, semi-supervised learning, transfer learning, and hybrid models. They have been used to represent froth appearance, temporal dynamics, multimodal information, and limited-label data. Their reported advantages, however, depend strongly on data quality, label alignment, acquisition conditions, data partitioning, and validation design. Supervised learning remains the main route for direct continuous-grade prediction. Semi-supervised and unsupervised methods are mainly used to reduce label dependence or improve representation learning. Transfer learning is useful when target-domain labels are limited, but its reliability depends on source–target similarity. Future studies need clearer label-alignment rules, time or condition-based validation, modality-aware fusion, robustness tests under operating-condition shifts, and explanation-guided model diagnosis. These improvements are needed to develop grade prediction models that are accurate, interpretable, and transferable in industrial flotation.