Chirality Detection in Scanning Tunneling Microscopy Data Using Artificial Intelligence
Author Information
Author(s): Tim J. Seifert, Mandy Stritzke, Peer Kasten, Björn Möller, Tim Fingscheidt, Markus Etzkorn, Timo de Wolff, Uta Schlickum
Primary Institution: Institute of Applied Physics TU Braunschweig
Hypothesis
Can synthetic data improve the training of AI models for detecting chirality in scanning tunneling microscopy images?
Conclusion
The study demonstrates that training AI models on synthetic data significantly enhances their ability to detect chirality in real scanning tunneling microscopy images.
Supporting Evidence
- Training on synthetic data improved model performance significantly compared to models trained on augmented datasets.
- A Faster R-CNN model trained solely on synthetic data achieved a mean average precision of 99% on real data.
- The approach demonstrated high robustness against experimental noise and varying zoom levels.
- Models trained on synthetic data showed better generalizability to different structures.
Takeaway
This study shows that using fake images to teach computers can help them better understand real images of tiny structures, making it easier to find specific shapes.
Methodology
The study used synthetic data to train AI models for detecting chiral unit-cells in scanning tunneling microscopy images, comparing different model architectures and training datasets.
Potential Biases
Models trained on augmented data showed overfitting, leading to poor performance on real images that differed from the training data.
Limitations
The performance of the models is limited by the resolution of the images and the number of labeled real images available for training.
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website