Electron Microscope. Properly setting up electron microscopes is a labour intensive and repetitive job. It is also a continuous process, depending on a whole host of factors, from the microscope itself, to its operating environment, or simply the time it has been in operation. Short of human intuition there isn’t a fixed recipe to tune microscopes. Operators do commonly look at the acquired images to infer clues on the current state of the microscope, one of these being the so called Ronchigrams. In short there are three components in play. Firstly, there must be a form of feature extraction that takes microscope images and maps or transforms them into parameters. Secondly, these parameters can then be used as input for a reinforcement learning algorithm. Lastly, this means that a digital twin also needs to be able to generate synthetic images that adhere to the same (unknown) distribution as the real images. Currently the only digital twin (model) we have deployed is a surrogate simulator that generates synthetic Ronchigram images according to pre-set aberration parameters. We also have a trained regressor that can blindly ascertain the original aberration parameters from a single synthetic image. Given that there is no complete digital twin as of yet, and the fact that the simulator is governed by known equations, we currently see a number of ways forward.
