Data driven analysis of particulate systems for development of reliable model to determine drag coefficient of non-spherical particles

Abstract

Non-spherical particles are extensively encountered in the process industry such as feedstock or catalysts e.g., energy, food, pharmaceuticals, and chemicals. The design of equipment used to process these particles is highly dependent upon the accurate and reliable modeling of hydrodynamics of particulate media involved. Drag coefficient of these particles is the most significant of all parameters. A universal model to predict the drag coefficient of such particles has not yet been developed due to the diversity and complexity of particle shapes and sizes. Taking this into consideration, we propose a unique approach to model the drag coefficient of non-spherical particles using machine learning (ML) to move towards generalization. A comprehensive database of approximately five thousand data points from reliable experiments and high-resolution simulations was compiled, covering a wide range of conditions. The drag coefficient was modeled as a function of Reynolds number, sphericity, Corey Shape Factor, aspect ratio, volume fraction, and angle of incidence. Three ML techniques—Artificial Neural Networks, Random Forest, and AdaBoost—were used to train the models. All models demonstrated strong generalization when tested on unseen data. However, AdaBoost outperformed the others with the lowest MAPE (20.1%) and MRD (0.069). Additional analysis on excluded data confirmed the robust predictive abilities and generalization of the proposed model. The models were also evaluated across three flow regimes—Stokes, transitional, and turbulent—to further assess their generalization. A comparative analysis with well-known empirical correlations, such as Haider and Levenspiel and Chien, showed that all ML models outperformed traditional approaches, with AdaBoost achieving the best results. The current work demonstrates that new generated ML techniques can be reliably used to predict drag coefficient of non-spherical particles paving way towards generalization of ML approach.