Developing Resource-Efficient Transmission Electron Microscopy for Democratizing Problem Solving in Materials Science

The proposed project, easyTEM, sets out to improve the utilization of complex major research instrumentation by applying modern AI technologies and utilizing otherwise idle time to enable the instrument alignment, operation and data acquisition to work autonomously. This will democratize access and open its usage for less trained and novice researchers, both locally and remotely and will ultimately make the use of these instruments more resource-efficient. While the fundamental concepts and software components will be developed with generic applicability in mind, specific implementations, use cases and tests will be in the area of transmission electron microscopy for materials science. EasyTEM will establish new levels of interaction between the instrument and the user in two ways: (1) non-skilled users will be empowered to steer a TEM using natural language combined with a generic and dynamic graphical user interface (GUI) and be trained along the way, and (2) at the level of the instrument GUI or application programming interface (API), parameters defining the instrument configuration data acquisition processes will automatically be suggested, based on a learned and continually refined model of these operations and adapting to new functionality of the instrument’s API / GUI. These models are trained during idle times using a reinforcement-learning approach that automatically explores parameter ranges and process chains to autonomously find the best possible set-up given a sample and the desired types of measurements. By using self- and pre-training, the software developed in this project will be applicable to a wide range of instruments; its usage will be evaluated on TEMs by four different manufacturers available at the labs run by the group of PIs of easyTEM. All models and software will be open-sourced under a permissive license.