A flow field reconstruction based on cylindrical wake flow data using the PSO-CNN-LSTM algorithm.

Area of Science:

• Fluid Dynamics
• Computational Science

Background:

• Accurate reconstruction of unsteady wake flow fields is crucial for understanding vortex shedding dynamics.
• Data-efficient analysis in engineering applications requires reliable flow field reconstruction.

Purpose of the Study:

• To propose a spatiotemporal reconstruction framework for 2D incompressible cylindrical wake flows.
• To evaluate the framework's generalization capabilities for missing-frame reconstruction and analyze error accumulation.
• To introduce uncertainty quantification for trustworthy flow-field reconstruction.

Main Methods:

• Utilized a Particle Swarm Optimization-Convolutional Neural Network-Long Short-Term Memory (PSO-CNN-LSTM) architecture.
• Employed a convolutional encoder for spatial features and LSTM for temporal dynamics and vortex shedding phase.
• Incorporated Particle Swarm Optimization for automatic hyperparameter tuning and Monte Carlo dropout for uncertainty quantification.

Main Results:

• The PSO-CNN-LSTM framework achieved high-fidelity reconstruction of dominant wake structures.
• Streamwise velocity error was of the order 10^-3, while cross-stream velocity error was of the order 10^-2.
• The method demonstrated lower errors and improved stability compared to LSTM, RNN, and Transformer baselines in rolling predictions.
• Error-uncertainty coupling analysis showed positive correlations between absolute error and predicted uncertainty.

Conclusions:

• The proposed framework enhances reconstruction accuracy for unsteady wake flows.
• It provides reliable spatial indicators of potential mismatch regions, supporting trustworthy flow-field reconstruction.
• The uncertainty quantification strategy aids in assessing the reliability of reconstructed flow fields.