Abstract: Deep learning offers new ways to uncover hidden structure and emergent phenomena in physical systems directly from data. In this talk, I will present our three recent studies on using deep learning to identify conserved quantities, detect critical transitions, and discover potential superconducting materials. First, I will revisit neural-network approaches for discovering constants of motion and discuss a more efficient strategy for extracting conserved quantities from dynamical data. I will then introduce a noise-robust contrastive learning framework for detecting critical transitions in stochastic dynamical systems. Finally, I will show how deep learning can extract physically meaningful features from electronic band structures to facilitate the search for superconductors. Together, these works demonstrate how deep learning can move beyond prediction toward discovering physical laws, identifying phase transitions, and accelerating materials discovery.

Zoom link: 
https://zoom.us/j/97920043879
Meeting ID: 979 2004 3879