Low-dimensional representation and prediction of supersonic gas-solid two-phase flow field based on CAE-BiLSTM

Abstract

Dynamic monitoring and rapid prediction of supersonic gas-solid two-phase flow fields are critical for the optimal design and flow control of solid rocket scramjets. This study proposes a hybrid reduced-order model (ROM) based on convolutional autoencoder (CAE) and bidirectional long short-term memory (BiLSTM) network (CAE-BiLSTM) to achieve efficient prediction of solid particle clusters’ distribution in such flow fields. High-dimensional flow field data (102,400 dimensions) were compressed into a low-dimensional latent space (128 dimensions, compression ratio 0.125 %) using CAE, which outperformed the traditional proper orthogonal decomposition (POD) in low-dimensional representation and exhibited a 3 % higher mean structural similarity index (SSIM) in flow field reconstruction. POD analysis revealed that the flow field is dominated by low-order modes, with the first mode contributing 21.1 % of the total energy, while the cumulative energy of the first 50 modes accounted for 61 %. The CAE-BiLSTM model effectively predicted the macroscopic distribution of solid particle clusters within the engine combustor flow field during short-term evolution, achieving a mean SSIM of 0.8814 for single-step prediction and 0.8545 for five-step recursive prediction, capturing the high-dimensional spatiotemporal dynamics of the flow field. However, long-term prediction accuracy deteriorated due to error accumulation. This study provides a novel approach for rapid flow field prediction, demonstrating potential applications in engine design and flow control.