Aberration Corrector Tuning with Machine-Learning-Based Emittance Measurements and Bayesian Optimization

Aberration-corrected scanning transmission electron microscopes (STEM) use a series of multipole magnets to generate a sub-Ångstrom-sized electron beam for atomic resolution imaging and chemical composition mapping 1 . A new scheme for aberration corrector alignment is proposed, including both new ways of beam quality measurement and aberration corrector control. The new scheme targets fully automated corrector alignment to achieve microscope tuning with greater speed and less human bias. For beam quality measurements, we trained a convolutional neural network (CNN) to determine the beam emittance 2 ⁠ from a single Ronchigram, as shown in Figure 1(a). Beam emittance is a single variable that characterizes the volume occupied by the beam in the phase space. Emittance is convex against the aberration coefficients, and proportional to the root mean square of the phase error, as shown in Figure 1(b). Both indicate that beam emittance is a single robust objective for aberration correction, and avoids the cusp-like instabilities of the individual aberration coefficients 3 . The emittance would be a conserved quantity without the introduction of aberrations, and aberration correction can be guided by minimizing the emittance growth. We trained a CNN with a VGG16 architecture using simulated Ronchigram-emittance pairs to predict emittance from a single Ronchigram. The Ronchigrams are simulated using a random phase plate following ref 4 , with