Speaker
Description
Classifying local structure is a central task in condensed-matter physics and materials science. Traditional descriptors such as Steinhardt bond-orientational order parameters are widely used for this purpose, but they often struggle to distinguish structurally similar phases and perform poorly in heterogeneous or partially ordered systems. This project investigates symmetry-aware machine-learning order parameters for local structure classification as a more flexible alternative.
The proposed models operate directly on local particle neighborhoods while explicitly respecting physical symmetries such as rotation and permutation invariance. Using symmetry-aware neural architectures, including invariant encoders and autoencoders, the approach is benchmarked against classical descriptors and existing machine-learning methods on crystal and soft-matter datasets. The results show improved separation of local environments, greater robustness to disorder and interfaces, and strong potential for discovering previously unknown structural motifs. Overall, the project demonstrates that symmetry-aware machine learning provides a promising framework for physically meaningful and transferable local structure analysis.