Deep Learning with Reflection High-Energy Electron Diffraction Images to Predict Cation Ratio in Sr₂Ti_2(1-O₃ Thin...

Area of Science:

• Materials Science and Engineering
• Thin Film Deposition
• Artificial Intelligence in Materials Science

Background:

• Traditional thin film synthesis relies on indirect property measurements, limiting real-time control and acceleration.
• In-situ diagnostics like reflection high-energy electron diffraction (RHEED) are typically used for qualitative analysis.
• Machine learning (ML) offers potential for quantitative analysis of complex materials data.

Purpose of the Study:

• To apply deep learning to predict the stoichiometry of strontium titanate (Sr₂Ti_2(1-O₃) thin films.
• To demonstrate the use of RHEED images for quantitative, ML-driven property prediction during pulsed laser deposition.
• To explore explainable AI (XAI) for uncovering structure-property relationships in thin films.

Main Methods:

• A gated convolutional neural network was developed for regression analysis.
• The model was trained on a dataset of 31 pulsed laser deposition samples with RHEED data.
• Explainable AI techniques were employed to interpret model predictions and identify key features.

Main Results:

• The deep learning model accurately predicted the strontium (Sr) atomic fraction in Sr₂Ti_2(1-O₃ thin films.
• A novel correlation between RHEED diffraction streak features and cation stoichiometry was discovered.
• The study successfully transformed a qualitative RHEED diagnostic into a quantitative measurement tool.

Conclusions:

• ML, combined with in-situ RHEED, enables quantitative prediction of thin film properties.
• This approach accelerates materials synthesis by providing real-time feedback and enabling autonomous workflows.
• The findings highlight the potential of AI to revolutionize materials discovery and manufacturing.