Abstract: Real-time analysis and prediction of flow field characteristics are essential for flight safety and necessitate in-flight data acquisition, an almost impossible task for traditional techniques such as temperature/pressure-sensitive paint. Flexible smart skins demonstrate great promise in acquiring in-flight data. Nevertheless, the multi-physical data they collected are usually sparsely distributed and limited in quantity, making it difficult to accurately identify transition and stall locations. To overcome this limitation, we introduce a novel method for rapidly reconstructing high-resolution flow fields around aircraft wings from sparse data. Leveraging the flow fields computed by CFL3D, a dual-branch attention fusion model is developed based on an encoder-decoder network. This model facilitates rapid prediction of 2D flow fields for 52 types of NACA airfoils at 156 different freestream conditions with an average relative error of 3.68%, providing a substantial amount of high-fidelity training data for the reconstruction process. Furthermore, using flow fields around the M6 airfoil generated by the rapid prediction model, a flow field reconstruction model was developed based on a Shallow Neural Network, which fulfills rapid reconstruction of high-resolution flow fields with a relative error of less than 5%. This work establishes a valuable data foundation for the in-flight application of smart skins, enhancing real-time monitoring capabilities and improving flight safety by enabling accurate flow field reconstructions and insights.