Reduced order model based on autoencoder and long short-term memory network

Autoencoder is an effective dimensionality reduction method that can learn the hidden information and features implicated in the data, and reconstruct the original input data. We propose a model reduction method with improved accuracy based on a multi-layer autoencoder and a long short-term memory network. The method is analyzed and verified through a two-dimensional flow past a cylinder. Firstly, the multi-layer autoencoder is used to reduce the order and extract features of the original data. Then, a prediction model based on a long short-term memory network is established. At last, the autoencoder and the prediction model are spliced into a single network to obtain a fine-tuned reduced order model. This model is further compared with another one based on the principal component analysis. Results show that the multi-layer autoencoder can improve the data compression ratio while ensuring the accuracy. The proposed reduced order method can effectively improve the model accuracy since the root mean square error between the predicted and the original velocity fields is reduced to within 3×10^-3.