Machine learning prediction of screening efficiency of new drum screen based on Harris Eagle algorithm optimization

Abstract

As a core screening device in mineral processing and construction industries, the efficiency of a drum screen is governed by multiple nonlinearly coupled parameters, rendering traditional empirical models inadequate for high-precision prediction. This study presents an integrated machine learning framework combining the Discrete Element Method (DEM) with the Harris Hawks Optimization (HHO) algorithm to dynamically model the trommel sieving rate of a novel screen. Initially, a high-dimensional dataset comprising 810 samples—including particle size distribution, rotational speed, and drum inclination—was generated via DEM simulations. Subsequently, the HHO algorithm was employed to optimize the hyperparameters of a deep multilayer perceptron (MLP), achieving coefficients of determination (R²) of 0.9864 and 0.9760 on the training and test sets, respectively. To generalize the model across varying drum lengths, a variable-screen-length iterative prediction algorithm was developed; results demonstrate that for drum lengths exceeding 300 mm, the prediction error remains below 3 % and decays exponentially with increasing length. Global sensitivity analysis identified drum inclination (26.01 %) and fine-particle feed rate (24.73 %) as the predominant factors influencing screening efficiency. Moreover, interpretability analysis indicated that fine-particle feed rates (“d < 3 mm”, “3 mm < d < 4.5 mm”) and drum inclination consistently serve as the primary drivers of model predictions, although the model exhibits context-dependence with notable variations across output regimes. The proposed approach offers robust theoretical and technical support for the intelligent design and parameter optimization of drum screening equipment.