Predicting the settling velocity of solid particles using machine learning

The prediction of settling velocity (SV) of solid particles is critical for optimizing processes in various engineering fields, including water treatment, sedimentation, mineral processing, and water distribution systems. Traditional theoretical and empirical models, while valuable, often suffer from narrow applicability and limited accuracy across diverse conditions. Recent advances in artificial intelligence (AI) offer a promising alternative to address these limitations. While previous research heavily studied the application of neural networks and linear models, we presented a critical comparison of linear, ensemble, and neural network models. We incorporated Extreme Gradient Boosting, Categorical Boosting (CatBoost), Elastic Net Regression, Bayesian Ridge, RuleFit, Artificial Neural Network, and Long Short-Term Memory. We tuned the model hyperparameters using a Bayesian and randomized search approach. CatBoost outperformed all the models, achieving an R 2 > 0.985 and an RMSE < 0.046. A k-fold analysis was conducted, where the RuleFit model was the most consistent. Although the model’s performance was high on the testing dataset, it drastically plummeted in the k-fold Analysis. Feature analysis identified that particle density and diameter were the most important features across the models. The outcome of this study highlights the significant potential of machine learning in modeling settling velocity, providing a dependable and scalable approach to enhancing efficiency in particle-fluid interaction processes across a wide range of engineering applications.