The development of advanced imaging and modeling techniques has given rise to powerful predictive tools to diagnose heart defects or assess risks of postoperative complications, allowing for the surgical planning prior to a cardiac surgery is done. For those surgical planning methodologies based on numerical simulations such as computational fluid dynamic (CFD) or finite element analysis (FEA), boundary conditions (BCs) are applied for patient-specific geometries to simulate postoperative hemodynamics. While postoperative geometries can be created from medical-image-based preoperative anatomies by virtual surgery tools, the acquisition of postoperative BCs, i.e., realistic characterizations of postoperative flows into the geometries, remains unsolved due to their unclear relationship to preoperative BCs. This results in the direct use of preoperative BCs towards cardiac surgical planning by current clinicians, where the accuracy of predicted postoperative conditions is questioned by the dissimilarity between preoperative and postoperative BCs usually found in patients. Therefore, this work proposes a machine learning framework to find the relationship between preoperative and postoperative BCs, particularly incorporating generative active machine learning to address model generalizability on small data sets - a limitation to a great deal of cardiovascular flow studies, such that the prediction of postoperative BCs can be made from preoperative ones, and permanently substitute them in the surgical planning pipeline for all future cardiac surgical planning endeavors to achieve the optimal surgical outcomes. The diversity-aware active learning framework, embodied with informative data query strategies, unique query metrics, and data set augmented by generative adversarial networks (GAN), is firstly developed on a synthetic flow database mimicking congenital single-ventricle patients done with Fontan procedures. The framework is then transferred to a real clinical cohort of TAVR patients for method validation on real-world implementation and for extension to other clinical scenarios. The study provides a new machine learning rationale to cardiovascular flow prediction/regression with reduced labeling requirements and augmented learning, bringing the opportunity to shifting the paradigms of a wide variety of pre-surgical planning of structural heart diseases.