Deciphering the dynamics of distorted turbulent flows: Lagrangian particle tracking and chaos prediction through transformer-based deep learning models

Abstract

Turbulent flow is a complex and vital phenomenon in fluid dynamics, as it is the most common type of flow in both natural and artificial systems. Traditional methods of studying turbulent flow, such as computational fluid dynamics and experiments, have limitations such as high computational costs, experiment costs, and restricted problem scales and sizes. Recently, artificial intelligence has provided a new avenue for examining turbulent flow, which can help improve our understanding of its flow features and physics in various applications. Strained turbulent flow, which occurs in the presence of gravity in situations such as combustion chambers and shear flow, is one such case. This study proposes a novel data-driven transformer model to predict the velocity field of turbulent flow, building on the success of this deep sequential learning technique in areas such as language translation and music. The present study applied this model to experimental work by Hassanian et al., who studied distorted turbulent flow with a specific range of Taylor microscale Reynolds numbers 100<Reλ<120. The flow underwent a vertical mean strain rate of 8 s−1 in the presence of gravity. The Lagrangian particle tracking technique recorded every tracer particle's velocity field and displacement. Using this dataset, the transformer model was trained with different ratios of data and used to predict the velocity of the following period. The model's predictions significantly matched the experimental test data, with a mean absolute error of 0.002–0.003 and an R2 score of 0.98. Furthermore, the model demonstrated its ability to maintain high predictive performance with less training data, showcasing its potential to predict future turbulent flow velocity with fewer computational resources. To assess the model, it has been compared to the long short-term memory and gated recurrent units model. High-performance computing machines, such as JUWELS-DevelBOOSTER at the Juelich Supercomputing Center, were used to train and run the model for inference.