Angular Sparsity Invariant Tilt Series Generation in Scanning/Transmission Electron Microscopy

Scanning/Transmission electron microscopy (S/TEM) is a key technique for probing materials across atomic to micrometer scales. Advances in instrumentation now allow unprecedented spatial and temporal resolution, generating large datasets that are difficult to analyze manually. Automated methods are therefore essential, with deep convolutional neural networks (DCNNs) emerging as promising tools, though they require extensive training data and standardized evaluation. In this work, we introduce a large-scale simulated S/TEM dataset and propose an efficient encoder–decoder DCNN named TSGNet that predicts intermediate tilt frames to enhance sparse tilt series. Our approach improves 3D reconstruction resolution and scalability, achieving an average performance of 5.77e-9 MSE, 31.63 PSNR, and 0.93 SSIM on the held-out aberrated test set across multiple sparsity levels.