About this Abstract |
| Meeting |
2027 TMS Annual Meeting & Exhibition
|
| Symposium
|
Algorithms Development in Materials Science and Engineering
|
| Presentation Title |
BlobNet: A New Lattice-Agnostic Tool for Angstrom Resolution Atom Finding |
| Author(s) |
Darel A. Pates, Austin Houston, Elizabeth Heon, Kai Xiao, Gerd Duscher |
| On-Site Speaker (Planned) |
Darel A. Pates |
| Abstract Scope |
Atomically resolved microscopy can map material function to local ordering of atoms, but this relies on knowledge of precise atomic coordinates. Existing atom-finding workflows often require hand-tuned image-processing parameters or neural networks trained on well-known crystallographic models. These models tend to fail when the structure of interest is irregular or non-periodic such as at interfaces, defects and quasi periodic structures. Here, we introduce BlobNet, a U-Net-based atomic-column localization model trained on synthetic HAADF-STEM-like images containing randomly positioned atomic features instead of a periodic lattice. In controlled tests, BlobNet outperforms geometry-specific models at inhomogeneities while approaching equivalent accuracy on periodic lattices. BlobNet notably avoids many prior-driven false positives that appear when lattice-trained models extrapolate periodic order across edges and voids. We will demonstrate the efficacy BlobNet on experimental micrographs with aperiodic and interfacial behavior with a trained BlobNet fast enough for real time data analysis. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Characterization, Machine Learning, Computational Materials Science & Engineering |