[3G01-03] Applied Numerical Simulations

[3G03] Prediction of Flow Behavior with Machine Learning Using Convolutional LSTM

Application to Unsteady Backward-facing Step Flow

*Aya Kitoh¹, Kei Maeshima¹, Kenya Takiwaki¹, Hideki Horie¹ (1. TOSHIBA ESS)

Keywords:machine learning, deep learining, ConvLSTM, CFD

Recently, research on the application of machine learning and deep learning to numerical analysis has made rapid progress. In this study, we investigate the applicability of machine learning to fluid analysis for the purpose of supporting the design and development of fluid equipment. So far, we have evaluated the flow behavior of a steady backward-facing step flow using an auto-encoder, which is widely used in image recognition and other applications, and the obtained results are successful. In this presentation, we report the results of the training and prediction of unsteady backward-facing step flow at high Reynolds number by Convolutional LSTM. Convolutional LSTM is a method that adds a convolutional layer to LSTM (long short-term memory) to retain spatial information.