We propose to use simulations of exoplanet transits to improve training outcomes for Machine Learning models. Machine learning has huge potential in exoplanet detection but faces challenges due to data imbalance and lack of ground truth in observational data. Most stars do not show transits, leading to datasets being skewed towards non-transit light curves, which can result in over-fitting and poor recall. Furthermore, the absence of ground truth complicates understanding the effects of noise and errors on detection outcomes. To address these issues, we simulate exoplanet transits using key astrophysical parameters and diverse noise profiles to create balanced training datasets. This simulation-based approach will improve machine learning models, enhancing their outcomes in detecting exoplanets in real-world data.