Deep Learning Predicts Drag Reduction in Pulsating Turbulent Pipe Flow

This paper demonstrates the generalization capability of deep learning models (CNN and LSTM) in predicting drag reduction in complex fluid dynamics scenarios. The key innovation lies in the model's ability to predict unseen, non-sinusoidal pulsating flows after being trained on a limited set of sinusoidal data. This highlights the importance of local temporal prediction and the role of training data in covering the relevant flow-state space for accurate generalization. The study's focus on understanding the model's behavior and the impact of training data selection is particularly valuable.

• •Deep learning models (CNN and LSTM) can predict drag reduction in pulsating turbulent pipe flow.
• •The models generalize well to unseen, non-sinusoidal flow conditions after training on sinusoidal data.
• •Local temporal prediction is crucial for generalization.
• •Training data selection is critical; covering the local flow-state space is key for accurate prediction.
• •Incorporating intermittent laminar-turbulent transition regimes in training data improves prediction accuracy.