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.

Generalizability of transformer-based deep learning for multidimensional turbulent flow data

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

Prediction of Turbulent Boundary Layer Flow Dynamics with Transformers

Super-resolution reconstruction of turbulent flows with a transformer-based deep learning framework

A transformer-based synthetic-inflow generator for spatially developing turbulent boundary layers

A generative adversarial network based on an efficient transformer for high-fidelity flow field reconstruction

Forecasting small scale dynamics of fluid turbulence using deep neural networks

Feel the Force: From Local Surface Pressure Measurement to Flow Reconstruction in Fluid-Structure Interaction

Transformer Neural Autoregressive Flows

Physical consistency and invariance in machine learning of turbulent signals

Stochastic Latent Transformer: Efficient Modelling of Stochastically Forced Zonal Jets

Generalization techniques of neural networks for fluid flow estimation

Feel the force: From local surface pressure measurement to flow reconstruction in fluid–structure interaction

A Swin-transformer-based model for efficient compression of turbulent flow data

Generalizability of Convolutional Encoder–Decoder Networks for Aerodynamic Flow-Field Prediction Across Geometric and Physical-Fluidic Variations

β-Variational autoencoders and transformers for reduced-order modelling of fluid flows

A Fluid Flow Data Set for Machine Learning and its Application to Neural Flow Map Interpolation

A deep learning super-resolution model for turbulent image upscaling and its application to shock wave–boundary layer interaction

Neural Multigrid Memory For Computational Fluid Dynamics

Self-supervised learning based on Transformer for flow reconstruction and prediction