Abstract
In this work, we propose a novel approach that uses an ensemble of deep neural networks (DNN) and convolutional neural networks (CNN) to predict the drag coefficient of small, irregular-shaped particles in low-Reynolds-number flows. In addition, a dataset containing 710 simulations of randomly shaped particle clusters are generated, on which the proposed model is trained upon. Existing literature within the field have mainly focused on symmetrical or regularly shaped objects. The DNN model is trained on tabular data, capturing flow characteristics and particle properties, whereas the CNN processes images of particles taken from three different angles. This combination allows the model to account for the different orientations and shapes of particles. By integrating visual information, the model provides a better understanding of particle geometry, which improves its predictive capabilities compared with methods that rely solely on tabular data. To provide a baseline, we implemented support vector regression (SVR) as a simpler approach, allowing us to assess the added value of our ensemble model. The comparison shows that our ensemble method achieves a notable improvement over SVR and DNN, effectively capturing the influence of the particle structure on drag behavior. The result is validated using k-folds cross-validation with five folds resulting in only a small deviation from the model presented, which is expected due to each fold having less data to train on. This work demonstrates the potential of ensemble learning as surrogate modeling to address complex fluid dynamics problems with irregular-shaped particles, which are common in environmental and industrial contexts.