Recent Advances in the Application of Neural Network for Flotation Prediction and Optimization

Flotation is the most widely used mineral processing method, and its separation efficiency and resource utilization depend largely on the perception and control of flotation conditions. Owing to the multivariable, nonlinear, strongly coupled, and highly disturbed nature of the flotation process, traditional control strategies—which rely on human expertise and rule−based approaches—are insufficient to meet the dynamic regulation requirements of complex flotation conditions. In recent years, neural networks, with their capabilities in deep feature extraction and nonlinear modeling, have demonstrated significant advantages and have been widely applied in flotation prediction and optimization. This paper systematically reviews research progress in three core aspects of flotation: froth feature recognition, reagent dosage optimization, and prediction of concentrate grade and recovery. Specifically, cross−scale convolutional modeling, Vision Transformers (ViT) , and three−dimensional vision techniques have enabled high−accuracy recognition of both static and dynamic froth characteristics; multi−information fusion modeling, visual memory networks, and Visual Mixture of Experts (V−MoE) have been employed for reagent dosage optimization; and multimodal flotation variables combined with temporal modeling methods have been used to achieve real−time prediction of concentrate grade and recovery. Existing studies reveal an evolution from single−modality modeling to cross−scale multimodal fusion, from static modeling to dynamic perception, and from experience−driven approaches to the integration of data−driven methods with process mineralogy. Future research should focus on improving the efficient fusion of multi−source heterogeneous flotation data and reducing the computational complexity (lightweighting) of models, with the goal of constructing a closed−loop control architecture that integrates cross−scale perception, lightweight models, and digital twins, thus providing a reference for the application of neural networks in flotation prediction and optimization.