Predicting Mixing:\nA Strategy for Integrating Machine\nLearning and Discrete Element Method

Segregation, the opposite of mixing, poses a common challenge\nin\ngranular systems. Using a rotating drum as the basic mixing equipment,\nthe fundamental focus of this study is to quantify undesirable segregation.\nThe impact of particle level parameters (size, density, their combination,\nmass fraction) and system parameters (filling %, rotational speed,\nand baffle) on the segregation index within the rotating drum is first\nassessed using the discrete element method (DEM). Later, the machine\nlearning (ML) model is applied in conjunction with DEM to expand and\nfill in the parameter space for particle-level parameters in a computationally\nefficient way, providing accurate predictions of segregation in less\ntime. The DEM results are validated by comparing them with experimental\ndata, ensuring their accuracy and reliability. The results show that\noptimal mixing is achieved when the total filling percent in a system\nis 36.3% while maintaining an equal proportion of particles. The highest\nlevel of mixing occurs at 60 rotations per minute, with fine particles\nconcentrating near the drum’s core and coarser particles distributed\naround the periphery. The presence of 3–4 baffles optimally\nenhances mixing performance. Four ML modelslinear regression,\npolynomial regression, support vector regression, and random forest\n(RF) regressionare trained using data from DEM simulations\nto predict the segregation index (SI). An error analysis is performed\nto pick the best model out of the four ML models. The analysis reveals\nthat the RF model accurately predicts the SI. Using the RF model,\nthe SI can be reliably predicted for any value of the seven features\nstudied using DEM. An example 3D surface plot is generated by considering\njust two (out of 7) of the most important particle level parameters:\nsize and density. The result shows that while both particle size and\ndensity contribute to segregation, variations in particle size appear\nto have a more pronounced effect on the SI compared to particle density.