Optimizing Parameters for High-resolution and Low-dose Ptychography

Ptychography is an increasingly popular technique in scanning transmission electron microscopy and has demonstrated sub-angstrom spatial resolution and high dose-efficiency, with applications ranging from single dopant detection to imaging of magnetic skyrmions. In practice, obtaining accurate reconstructions requires optimization of numerous types of experimental and algorithm parameters that are highly data-dependent and often selected by trial and error. To reduce human involvement and biases, we leverage physics-informed Bayesian optimization with Gaussian processes (BO-GP) to develop automatic data analysis and simulation workflows in ptychography. Our method assumes minimal prior knowledge about experimental datasets and can consistently produce high-resolution reconstructions similar or even superior to those processed by human experts. We also used BO-GP to explore the ideal experimental conditions for low-dose ptychography – a challenging task for humans due to the complex trade-offs between various physical requirements. The workflow was recently utilized to optimize high-order aberrations when designing probe structures that further push the limit of electron ptychography.