Grounding a RHEED AI Analytical Reward Function in Post-Growth XPS Film Characterization

• Niyati Gupta, Physics, University of Illinois Urbana-Champaign

• Ryan Comes, Materials Science and Engineering, University of Delaware

Molecular beam epitaxy (MBE) is used to grow films for a diverse range of applications. The application of artificial intelligence (AI) can be used to increase efficiency during the growth of MBE films, and speed up the path to material characterisation. X-ray photoelectron spectroscopy (XPS) is a spectroscopic technique that can analyze the topmost layer of films to identify the elemental composition of samples. Therefore, automating the analysis of XPS spectra is essential, as it provides a post-growth metric to train a real-time, during-growth AI system that uses reflection high-energy electron diffraction (RHEED) to guide the growth of MBE films. In order to achieve this goal, a convolutional neural network (CNN) was trained on a dataset generated from XPS survey spectra of FeSe, SrIrO3, and LaCoO3 films. The design of the CNN was based on an existing model used to classify transition metal oxides, but modified to quantify elemental composition from more complex chalcogenide and oxide films. The reference spectra were labelled by their relative intensity of the selected elements by expert human quantification, and were then randomly linearly combined and transformed to generate the training dataset. The metrics collected about the performance of the different models trained, including mean absolute error and mean squared error, show that the final model outperformed other models that used smaller binding energy ranges and different methods of dataset generation. In conclusion, the model showed success at classifying unseen spectra according to their elemental composition, as compared to expert quantification of these spectra. Future work will focus on incorporating additional spectra from new oxide and chalcogenide films grown, and those grown in the recently installed Theia XPS system, increasing the model’s applicability to different compounds and systems.