This study presents a machine learning (ML) approach using a random forest classifier to predict blockages in multiphase flows with cohesive particles. The machine learning model aims to predict blockage under varying parameters, using binary classification as the result. We chose the random forest classifier because it can handle complex datasets with many variables, which is suitable for our research with multiple parameters and features. We applied a classifier using the Scikit-learn library in Python. The model is trained using a combination of datasets from flow loop experiments with ice slurry in decane and Computational Fluid Dynamics with the Discrete Element Method (CFD-DEM) simulations. Experiments and simulations provided a complete view of the dynamics of the considered multiphase system, including the effects of parameters. The collected dataset contains parameters like the Reynolds number, concentration, capillary number, and an indication of blockage.
The model is evaluated using a cross-validation method with a parameter k=5, repeating this process five times with different fold combinations to verify its performance. Each model was trained on 80% of the data and tested on the remaining 20% in each iteration. Some parameters were adjusted, including a fixed random seed, the number of estimators, and the maximum depth of decision trees, while all other parameters were at their default values. After finalising the hyperparameters, the final random forest classifier was trained on the complete dataset to construct the blockage boundaries. The model's performance was evaluated using precision, recall, and F1-score metrics, demonstrating high values. Notably, the model achieved a maximum precision score of 1 and a recall of 0.8 for blockage cases.
The key result of our study is the flow map, which compares machine learning boundaries with experimental and CFD-DEM simulation results. The flow map demonstrates that the results of the applied model closely align with the upper boundaries observed in simulations and experiments, especially at high Reynolds numbers. Furthermore, we considered how changes in cohesion affect blockage boundaries, observing an increase in cohesion.
In conclusion, our method combines experiments and simulations to create an accurate predictive ML model. The successful application shows how machine learning can benefit fluid dynamics and blockage prevention, indicating possibilities for advanced models.