Spatio-temporal Modeling of High-fidelity Turbulence with Convolutional Long Short-Term Memory Neural Networks

A major challenge in machine learning for turbulence is the chaotic, high dimensional and spatio-temporal nature of the data; which can make the learning process ineffective and/or expensive. Previous work [1] had demonstrated the capability of Long Short-Term Memory (LSTM) neural networks to capture temporal dynamics of turbulence. In this work, we extend this capability to modeling spatio-temporal dynamics using the Convolutional LSTM (ConvLSTM) neural network. ConvLSTM augments the traditional architecture of the LSTM cell with a convolutional layer to capture spatial correlations in multidimensional data. We demonstrate the potential of ConvLSTM in learning and predicting the dynamics of a DNS homogeneous isotropic turbulence dataset. We perform statistical tests on the predicted turbulence to quantify the quality of the“learned” physics and develop physics-inspired neural network constraints for improved predictions. Finally, we study the feasibility of this approach for large datasets and explore strategies to increase computational efficiency.

[1] https://arxiv.org/abs/1804.09269