Real-Time Prediction of Material-Holding Capacity in Disc-Type Air Jet Mills via Hybrid CFD-DEM and PSO-Optimized LSTMModeling

Introduction: Disc-type air jet mills are critical for ultrafine pulverization of neodymium iron boron (NdFeB) due to their contamination-free operation and high precision. However, real-time monitoring of material-holding capacity (holdup mass) within the grinding chamber remains challenging, as it directly affects grinding efficiency, throughput, and energy consumption. This study aims to establish an intelligent real-time prediction model for material-holding capacity in disc-type air jet mills by integrating CFD-DEM simulations and deep learning, thereby supporting operational optimization and intelligent control. Methods: A hybrid methodology was developed: (1) CFD-DEM simulations (EDEMFluent coupling) quantified the effects of key parameters (feed pressure, grinding pressure, and classifier speed) on holdup mass; (2) A Particle Swarm Optimization-optimized Long Short-Term Memory (PSO-LSTM) model was constructed using simulation-derived datasets, with hyperparameters autonomously tuned via PSO to enhance prediction accuracy. Results: The PSO-LSTM model achieved superior performance with MAE = 16.55, RMSE = 20.82, and R² = 0.94, outperforming both BP neural networks (MAE = 191.89, RMSE = 231.67, R² = 0.78) and unoptimized LSTM (MAE = 29.29, RMSE = 37.78, R² = 0.87). CFD-DEM analysis further revealed that increasing feed pressure raised holdup mass by 6.25%, while higher grinding pressure reduced it due to enhanced particle discharge. Discussion: This study employed CFD-DEM coupled simulations and simulation data to construct a Particle Swarm Optimization-Long Short-Term Memory (PSO-LSTM) prediction model, integrating it for real-time material holdup monitoring. This approach overcomes the limitations of empirical models in capturing nonlinear dynamics and explores new strategies for predicting material holdup in grinding processes. Furthermore, it provides a theoretical basis and technical support for developing intelligent control systems and optimizing the operation of disc-type jet mills. Conclusion: The PSO-LSTM framework provides a robust solution for real-time prediction of material-holding capacity, reducing prediction errors by over 83% compared with conventional methods. This approach enables precise control of comminution processes and lays a foundation for intelligent optimization of disc-type air jet mill systems.