The time taken to generate a super-resolution image and the quality of the final synthetic image depends on the performance of the localization algorithm which is used in the localization microscopy pipeline. The most precise and accurate algorithms are mostly iterative and they take a long time to generate the localization list while the faster ‘one-shot algorithms’ are not very accurate and precise. A deep learning method smNet (single-molecule Net) was developed by Zhang et al which was claimed to perform one-shot localization with precision close to the theoretical limit and very accurately, along with performing aberration estimation and dipole-emitter orientation angle estimation. The deep learning model smNet was trained either by augmenting experimental data or using simulated data generated with an erroneously simplified simulation model and a phase retrieval method. The purpose of this work was to characterize the performance of smNet when it was trained with simulated images generated using an accurate vector model for a range of physical conditions. Along with the characterization of smNet’s performance in doing 3D localization and aberration estimation with the accurate vector model, a pipeline was also designed which made the training process of smNet more efficient and computationally cheaper while performing accurate and precise 3D localization and aberration estimation. The pipeline was designed to implement the concept of simulator learning where a smNet model could be trained on simulated data and used to perform 3D localization and aberration estimation directly on experimental data without any retraining or domain adaptation techniques.