Regensburg 2025 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

DY: Fachverband Dynamik und Statistische Physik

DY 28: Poster: Machine Learning, Data Science

DY 28.2: Poster

Mittwoch, 19. März 2025, 15:00–18:00, P4

Neural Networks for Phase Recognition on Lattice Systems — •Shashank Kallappara and Martin Weigel — Institut für Physik, Technische Universität Chemnitz, Chemnitz, Germany

The Ising model undergoes a phase transition in dimensions d > 1, with its magnetisation as the order parameter. Fully connected neural networks have been shown to learn the translational invariance of the Ising model when learning its phases. This requires only a single hidden layer; analytic solutions for the same exist for highly compact networks that are constructed to obey the translational invariance automatically. Here, we show this learning of the invariance in single-layer networks of different widths and compare the networks performance in classifying the phases. We also consider a highly compact network to study the gradient descent dynamics during training over its loss landscape; we suggest a few changes to this that greatly improve its performance while preserving interpretability. Another problem we consider is percolation on a square two-dimensional lattice. Here, we leverage convolutional neural networks to detect the phase transition by training on different properties of the systems and evaluate its effectiveness.