ST-FlowNet: A lightweight framework for long-term spatio-temporal flow field prediction.

Area of Science:

• Fluid Dynamics
• Computational Science
• Artificial Intelligence

Background:

• Unsteady flow prediction is crucial for aerospace and energy but faces challenges due to high dimensionality and complexity.
• Traditional numerical methods are computationally expensive, while intelligent methods struggle with long-term prediction accuracy due to error accumulation.

Purpose of the Study:

• To develop a lightweight framework (ST-FlowNet) for accurate and efficient long-term spatio-temporal flow field prediction.
• To address the limitations of existing methods in terms of computational cost and prediction accuracy over time.

Main Methods:

• Utilized proper orthogonal decomposition (POD) to simplify flow characteristics and reduce computational costs.
• Developed an attention-enhanced model integrated with a physics-informed loss function to mitigate error accumulation in long-term temporal predictions.

Main Results:

• ST-FlowNet achieved accurate long-term simulation of unsteady flows.
• Demonstrated a two-order-of-magnitude speedup compared to traditional numerical simulations.
• Attained state-of-the-art prediction accuracy and generalization capability with a minimal parameter size.

Conclusions:

• ST-FlowNet offers an efficient and accurate solution for long-term spatio-temporal flow field prediction.
• The framework effectively balances computational cost, prediction accuracy, and model complexity.
• Presents a significant advancement over existing intelligent and traditional approaches for unsteady flow analysis.