A physics-guided multi-task deep learning framework for flow field reconstruction and its application in turbomachinery

Accurate and efficient reconstruction of three-dimensional multi-physics fields remains challenging in many engineering applications due to the high-dimensional outputs, strong nonlinearity and complex physical correlations, which restricts the performance improvement of traditional deep learning models. This study proposes a modified Physics-Guided Multi-Task Deep Learning (PG-MTDL) framework to address these issues. The framework first integrates Proper Orthogonal Decomposition (POD) for efficient dimensionality reduction of original flow fields. A multi-task neural network with a self-attention module is employed to capture shared features and learn the intrinsic correlations across multi-physics fields. A hybrid loss function is designed, which combines an uncertainty-weighted data loss to balance task-specific losses adaptively with a physical constraint loss based on gas state equation to enforce thermodynamic consistency. The proposed method is validated employing an industry-standard flow field dataset from a high-pressure turbine stage, containing over 1.4 million grid nodes. Results demonstrate obvious improvements in reconstruction accuracy, physical consistency and computational efficiency compared to conventional single-task models. Specifically, the experiments represent an average reduction in reconstruction error of 18.1%, a remarkable decrease in physical consistency loss by nearly 90%, and a 37.5% reduction in model parameter size. Furthermore, the proposed framework is more accurate in capturing the key aerodynamic features of turbines and maintains relatively better extrapolation performance. These advancements highlight the robustness and efficiency of the PG-MTDL framework, indicating its promising potential for industrial-scale flow field performance analysis, real-time prediction, and design optimization.