Physics-informed Bayesian Optimization of an Electron Microscope

Precise alignment of the electron beam is critical for successful application of scanning transmission electron microscopes (STEM) to understanding materials at atomic level. However, the nature of magnetic lenses introduces various orders of aberrations and makes aberration corrector tuning a complex and time-consuming procedure. Here we approach the problem from the perspective of accelerator physics and demonstrate the equivalence between aberration correction and beam emittance minimization in phase space. We show a deep neural network can accurately capture phase space variations from electron Ronchigrams, enabling fast beam quality measurements. A Bayesian optimization framework is developed to minimize emittance growth, while providing the full posterior distribution of the responses across control parameters to account for uncertainties in each query. Furthermore, a deep kernel is implemented and shown to effectively learn the correlations between input dimensions, which can generalize to other accelerator tuning tasks as well. Both simulation and experimental results validate that the proposed method outperforms existing alignment approaches. This new scheme enables fully automated aberration corrector tuning, achieving greater speed and less human bias.