The investigation of separated flows over air foils is notoriously difficult due to three dimensional and unsteady effects. These flows require extensive experimental or computational data that can be analysed using a variety of tools. In this work, various data driven methods have been used to examine flow over stalled wings to understand the flow physics and develop reduced-order models for predictions. It is shown that sparsely distributed sensors in the flow field can also predict the state of the flow. Performance of multiple data-driven reduced-order models (linear and non-linear) together with pseudo probes in the flow are used to reconstruct the separated flow. A non-linear neural network-based approach is found to perform better in reconstructions across different cases. To enhance physical interpretation of non-linear reduced-order modelling (such as autoencoders), a hierarchical approach is examined. Subnetworks are trained to rank the non-linear modes according to their contributions to the reconstruction. By forcing the latent space distributions towards a unit normal distribution, with a variational autoencoder, it becomes possible to disentangle the separate modes. It has been shown that with the proper regularisation the non-linear modes become nearly orthogonal and offer a better reconstruction than the truncated proper orthogonal decomposition. A large computational data set of flow over a NACA 0012 wing has been created with different types of flow ranging from attached to fully separated flow. The importance of the flow characteristics near the surface of the wing has been indicated. It is shown that surface pressure can be used to predict these flow characteristics in liaison with a data-driven stall detection model. Finally, leveraging flow visualisation using tufts, a data driven model that estimates the unsteady lift fluctuations based on tuft motions is developed. A proof of concept is examined with computational data and subsequent wind tunnel experiments together with a neural network reduced order model to provide accurate estimates of lift and pitching moment fluctuations.