The computational cost of simulating scanning transmission electron microscopy (STEM) images limits the curation of large enough data sets to train accurate and robust machine learning networks for deep feature extraction from atomically resolved STEM images. For nanoparticle size estimation in particular, a diverse data set is essential due to the large variations in size, shape, crystallinity, orientation, and dynamical diffraction effects in experimental data. To address this, we train a 3D convolutional neural network to predict STEM images from voxelized atomic models, achieving a 100x speed-up compared to traditional multislice simulations while maintaining high image quality. We then generate a data set of 100.000 synthetic multislice images and investigate the performance of different size-estimator architectures as a function of training set size. A ResNet18-based model trained on 4000 real and 100.000 synthetic images is found to perform the best, reducing the median size-estimation error from 9.89% without synthetic data to 5.26%.