Knowledge Bank

The challenges in predicting the highly nonlinear and chaotic nature of turbulent flow have been of particular interest in the realm of machine learning (ML) in recent years. Turbulent flow is a common phenomenon in fields such as acoustics, aeronautics, meteorology, and automobile engineering. Traditional simulation methodologies like Large-eddy simulation (LES), Reynolds-averaged Navier-Stokes (RANS), and Direct Numerical Simulation (DNS) have limitations in terms of computational intensity and the inability to predict intermittent events in turbulent flow. ML, particularly Long Short-Term Memory Neural Networks (LSTMs) and its variants such as Bidirectional LSTM (BiLSTM) and Gated Recurrent Units (GRUs), offer promising ways for forecasting such complex phenomena in a less computationally intensive manner. This study focuses on using these ML models to predict intermittent spikes in turbulent flow. Data from a Computational Fluid Dynamics (CFD) analysis of an impinging jet is decomposed to isolate the acoustic fluctuating stagnation enthalpy for ML training. A grid search methodology is employed to optimize the ML models to find the models with the highest accuracy by changing the number of layers, cells, and type of activation functions for each model. The grid search demonstrated that most of the models found predicting intermittencies incredibly difficult, with average test accuracies being around 6%. However, for each of the LSTM, BiLSTM, and GRU models, test accuracies rose to be as high as 53.5% in some models. This shows promise towards being able to predict intermittencies, but is still too low to make any significant predictions for practical usage. In future work, physics can be intertwined with the loss functions during the learning process to improve accuracy at the expense of increased computational costs.