Recent advances in artificial intelligence are paving the way for the automation of data acquisition across the scientific landscape. Ongoing research in transmission electron microscopy (TEM) is primarily focused on image processing for super resolution and feature extraction using Deep Convolutional Neural Networks (CNN). Similar methods could use image information and a microscopist-defined goal (such as drift correction or automatic alignment) to automatically adjust the TEM configuration. To assess the feasibility of such an endeavour, we first tested a model's ability to predict the resulting image solely based on the microscope configuration on a modified HF3300 and an in-house paraxial simulation. Then we did the opposite: we predicted a configuration from an image that would result in a microscope image similar to the given image. The ultimate goal is to be able to predict a configuration in which the resulting image meets user-defined properties such as being in-focus, centred, and properly illuminated. As a result, a dataset containing configurations and their associated images was required. We chose to align the microscope first, then slightly modify some parameters while preserving a visible beam because completely randomizing those parameters would not produce a visible beam in most images. We generated 9 datasets with 450-900 configuration/image tuples each with the TEM. The deep learning models were then trained to predict one using the other. This validated the models' ability to simulate the entire microscope and reverse engineer its configuration from an image. Because direct prediction is impossible for the latter since multiple configurations produce similar images, we trained it to produce a configuration whose resulting image is similar to the given one. The simulation was not accurate enough to use the microscope's initial configuration. To address the issue of automation using user-defined goals, we are developing a workflow based on Reinforcement Learning (RL) that will allow the models to train in real time by modifying the microscope configuration directly.